Three Essays in Political Methodology

by
Arthur Peter Andrew Spirling
Submitted in Partial Fulfillment
of the
Requirements for the Degree
Doctor of Philosophy
Supervised by
Professor Curtis Signorino
Department of Political Science
The College
Arts & Sciences
University of Rochester
Rochester, New York
2008
ii
For my mother and for my father.
iii
Curriculum Vitae
The author was born in Carshalton in the London Borough of Sutton, United
Kingdom, on August 26, 1979. He attended the London School of Economics
from 1997 to 2001 and graduated with a Bachelor of Science degree in 2000.
He received a Master of Science degree in 2001 from the same institution, and
then attended Nuffield College, Oxford University, from 2001 to 2003. He
came to the University of Rochester in the Fall of 2003 and began graduate
studies in Political Science. He pursued his research in political methodology
under the direction of Professor Signorino and received the Master of Arts
degree from the University of Rochester in 2007.
iv
Acknowledgements
Social reformer Sir Thomas Fowell Buxton once modestly observed that
“with ordinary talents and extraordinary perseverance, all things are at-
tainable.” Those things apparently include my completion of a dissertation
in political science at the University of Rochester, although even Sir Thomas
would be surprised to learn quite how ordinary one’s talents can be in prac-
tice. With this in mind, there are several individuals who I wish to thank
for their contribution to my education, work and emotional well-being over
these last several years.
Chronologically, the story begins at Nuffield College, Oxford, where Iain
McLean was a sage source of counsel and helped me garner admission to the
University of Rochester in 2003. David Firth tried patiently to teach me
some statistics, and has been a constant source of helpful advice since that
time. Working and coauthoring with both these individuals has taught me
much about our related interests, and the travails of getting published.
At Rochester, Curtis Signorino, my advisor, allowed me the flexibility and
support to explore new avenues of methods research without the micro-
management that would have pained us both. Preferring to gently steer
rather than direct, Curt provided much needed pastoral care when I felt
particularly frustrated and anxious—especially in the last year of my disser-
tation writing, which coincided with the stress of the market. Emphasizing
the ‘grand plan’ as well as the details, Curt has taught me a great deal
v
about how to ‘play the game’ of our discipline. Also serving on my commit-
tee, Tasos Kalandrakis is one of the hardest working men in the business,
and his diligence in reading and commenting on my papers is second-to-
none. Our projects together have been the most challenging and rewarding
experiences of my career to date, and I expect our cooperative efforts will
continue for some time yet. Larry Rothenberg stepped forward when I
needed an extra body on my committee for bureaucratic reasons, but he has
become an important member of my team in the meantime. I thank him
for his time and advice. Though not on my committee, I am very grateful
to Michael Peress and Tony Lenzo who both supported my work in quite
different ways. Michael’s deep understanding of econometrics allowed him
to provide valuable technical feedback on my papers. Tony, who managed
the star lab during the writing phase of my dissertation, ran a very tight
ship and ensured that both the hardware and software I had access to was
in the best possible working condition.
In the broader methods community, I am very grateful to Jeff Gill and An-
drew Martin at Wash U, Michael Herron at Dartmouth and Kevin Quinn at
Harvard. All these people kindly spent time helping me improve my papers
with lengthy and valuable comments.
My comrades in the department were constant sources of amusement and
encouragement throughout the writing of this dissertation; in my cohort,
I thank Matthew Platt, David Carter, Michael Epstein, Arnd Plagge and
Subhashish Ray. Peter Loewen spent an all too short visit at Harkness,
vi
during which time we became firm friends. In the early days, Justin Fox,
Antoine Yoshinaka and Kris Ramsay were a strong influence on my outlook
and behavior as a graduate student.
I wish to convey a very special thanks to my family that loved me whether
my research was progressing or stagnating: my mother, my father, my sister
and my brother. They were constant in their support, affection and financial
assistance; I hope I can, one day, repay these debts. Finally, perhaps most
of all, I thank Sonia Tasbas: without her love and kindness, I know only
that my outlook—academic or otherwise—would be the poorer. I remain
truly grateful that I have found such a wonderful woman to spend my life
with.
vii
Abstract
This political methodology dissertation consists of several distinct essays
that apply three new estimation techniques to all three subfields of the
discipline—international relations, American politics and Comparative pol-
itics. The first is an application of reversible jump Markov chain monte
carlo, a more general form of MCMC popular for model search problems
in statistics. We apply RJMCMC to the current Iraq conflict in order to
identify change points in terms of civilian casualty numbers. The second
chapter uses the Bradley-Terry model for pairwise contests to estimate the
‘power’ of actors in legislatures. We apply this generalized linear model
to the United States Senate. We use both an unstructured and covariate
based model to show that, inter alia senators’ party identification, ideo-
logical tendency, leadership rank, service length and gender all affect their
ability to influence others in the chamber. The third chapter shows that
British Members of Parliament’s voting decisions on roll calls can affect
their constituency performance at election time. In particular, ‘rebeling’
on government bills apparently hurts them, while voting independently on
less important matters tends to benefit them. We use a non-parametric,
random forest algorithm to estimate the relationship, since the number of
parameters (the number of roll calls) far exceeds the number of observations
(constituency performances) to be predicted.
viii
Contents
1 Introduction: Purpose and Plan 1
1.1 Plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 ‘Turning Points’ in the Iraq Conflict: Reversible Jump Markov
Chain Monte Carlo in Political Science 6
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Background and Data . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Estimation Problem . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 Measuring Power in Political Science: A New Method with
Application to the Senate 22
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 Problems with Traditional Approaches to Measuring Power . 24
3.3 Statistical Theory and Model . . . . . . . . . . . . . . . . . . 26
3.4 Application: 108th Senate . . . . . . . . . . . . . . . . . . . . 32
3.5 Structured Modeling . . . . . . . . . . . . . . . . . . . . . . . 36
ix
3.5.1 Party and Ideology . . . . . . . . . . . . . . . . . . . . 38
3.5.2 Committee and Agenda Control . . . . . . . . . . . . 42
3.5.3 Geographic Factors . . . . . . . . . . . . . . . . . . . . 46
3.5.4 Career Factors . . . . . . . . . . . . . . . . . . . . . . 50
3.5.5 Summary of Findings . . . . . . . . . . . . . . . . . . 52
3.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Appendix to Chapter 3 61
3.7 The “Busy” Senator Problem . . . . . . . . . . . . . . . . . . 61
3.8 Logit Model with Order Effect . . . . . . . . . . . . . . . . . 62
3.9 Amendment Proposers . . . . . . . . . . . . . . . . . . . . . . 63
3.10 Goodness-of-Fit . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4 Rebels with a Cause? Legislative Activity and the Personal
Vote in Britain, 1997–2005 70
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.2 Members, Constituents and the Personal Vote in Britain . . . 74
4.3 Formalizing Intuitions . . . . . . . . . . . . . . . . . . . . . . 77
4.4 Data and Method . . . . . . . . . . . . . . . . . . . . . . . . 81
4.4.1 Data: 1997–2001, 2001–2005 . . . . . . . . . . . . . . 83
4.4.2 Estimation Problem: p > n . . . . . . . . . . . . . . . 84
4.4.3 Tree-based Regression and Random Forests . . . . . . 85
4.4.4 Estimation Plan . . . . . . . . . . . . . . . . . . . . . 93
4.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.5.1 First term: 1997–2001 . . . . . . . . . . . . . . . . . . 96
4.5.2 Second term: 2001–2005 . . . . . . . . . . . . . . . . . 104
x
4.5.3 Summary of Results . . . . . . . . . . . . . . . . . . . 109
4.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Appendix to Chapter 4 113
4.7 Regression Tree Split Criteria . . . . . . . . . . . . . . . . . . 113
4.8 Model Fit, Variable Importance and Effect . . . . . . . . . . 114
4.9 Imputing Missingness . . . . . . . . . . . . . . . . . . . . . . 116
5 Discussion 117
Bibliography 122
xi
List of Tables
2.1 Change point median dates [90% HPD] in Iraq casualty data, with rates
between breaks. Comments note political and other activities approxi-
mately contemporaneous to break. . . . . . . . . . . . . . . . . . . 21
3.1 Baseline results for model of power for US Senators in the 108th
Senate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.2 Effect of majority party and senior leadership status on power in the
Senate. Asterisked coefficients (***) imply p < 0.01. AIC: 11052. . . . 58
3.3 Effect of majority party membership and any leadership status on power
in the Senate. Asterisked coefficients imply p < 0.01 (***), p < 0.05
(**). AIC: 11152. . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.4 Effect of conservatism, distance from median and majority status on
power in the Senate. Asterisked coefficients imply p < 0.01 (***), p <
0.05 (**). AIC: 11169. . . . . . . . . . . . . . . . . . . . . . . . . 58
3.5 Effect of extremism, distance from median and majority status on power
in the Senate. Asterisked coefficients imply p < 0.01 (***), p < 0.05
(**). AIC: 11136. . . . . . . . . . . . . . . . . . . . . . . . . . . 59
xii
3.6 Effect of committee membership, majority party membership and chair
status on power in the Senate. Asterisked coefficients imply p < 0.01
(***), p < 0.05 (**). AIC: 9434.9. . . . . . . . . . . . . . . . . . . . 59
3.7 Effect of incomes, southern state representation, land area (state size)
and state population on senator’s power. Asterisked coefficients imply
p < 0.01 (***), p < 0.05 (**). AIC: 10954. . . . . . . . . . . . . . . . 60
3.8 Effect of incomes, southern state representation, land area (state size)
and state population on senator’s power. Asterisked coefficients imply
p < 0.01 (***), p < 0.05 (**). AIC: 10205. . . . . . . . . . . . . . . . 60
3.9 Baseline results for model of power for US Senators in the 108th
Senate, logit with ‘order effect’ (party advantage) estimated. . . . 68
3.10 Number of amendments per senator. . . . . . . . . . . . . . . . . . 69
4.1 ‘Most important’ parliamentary votes for predicting electoral perfor-
mance by backbench Labour MPs in 2001 general election. First column
is data code for the division with date after point: so, div323.001107 is
division 323 taking place on November 7, 2000. Name of bill in column
2; random forest ‘coefficient’ in column 3 for being a ‘rebel’(in the mi-
nority and/or frustrating the Government’s ambitions); OLS coefficient
in column 6; p-value for that coefficient in column 7. . . . . . . . . . . 100
xiii
4.2 ‘Most important’ parliamentary votes for predicting electoral perfor-
mance by backbench Labour MPs in 2005 general election. First column
is data code for the division with date after point. Name of bill in column
2; random forest ‘coefficient’ in column 3 for being a ‘rebel’(in the mi-
nority and/or frustrating the Government’s ambitions) ; OLS coefficient
in column 6; p-value for that coefficient in column 7. . . . . . . . . . . 107
xiv
List of Figures
2.1 Iraq Casualty Incidents, May 2003–May 2007. Solid line represents cu-
mulative counts; solid dots are jittered incident occurrences; various
dates of interest demarcated. . . . . . . . . . . . . . . . . . . . . 11
2.2 Posterior of k: number of change points, Iraq casualty data. . . . . . . 16
2.3 Change in rate of casualty incidents over time: rate on y-axis, actual
incidents as jittered points. Dates correspond to breaks found. . . . . . 19
3.1 NOMINATE score versus power (λi); open squares are Republicans,
closed circles are Democrats; solid line is loess. Top figure is for all
senators; bottom figure is for all proposing senators. . . . . . . . . . 43
3.2 Extremism versus power (λi); open squares are Republicans, closed cir-
cles are Democrats; solid line is loess. Top figure is for all senators;
bottom figure is for all proposing senators. . . . . . . . . . . . . . . 44
3.3 Boxplot showing relative power of committee chairs, non-chairing Re-
publicans and Democrats who propose amendments. . . . . . . . . . 47
xv
3.4 Map of contiguous United States with shading proportional to ‘power’
given by coefficients in Table 3.7: lighter states are associated with more
powerful senators. . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5 Boxplot showing relative power of male and female senators. . . . . . 53
3.6 Amendment numbers: distribution. . . . . . . . . . . . . . . . . . 64
4.1 Percentage of respondents naming different qualities as “important for
an MP to have”: solid line is percentage naming ‘independent minded’,
broken line is percentage naming ‘loyal to party’. Multiple choice, more
than one choice permissible; survey dates are 1983, 1994, 1996, 2000,
2005. Source: Johnson and Rosenblatt (2007, 166, Table 1). . . . . . . 75
4.2 Example of a classification tree for American states, predicting (catego-
rizing) the percentage of vote for George Bush in the 2004 Presidential
election. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.3 Example of a tree that uniquely classifies American states, in terms the
percentage of vote for George Bush in the 2004 Presidential election.
Node labels removed for clarity of presentation. . . . . . . . . . . . . 90
4.4 Random forest schematic. Here the data set consists of ten observations
labeled 1 through 10. These are ‘bootstrap sampled’ into bags (1 thor-
ough B). For each bag, some observations are ‘left out.’ A tree is grown
from each bag, and the results aggregated. . . . . . . . . . . . . . . 94
4.5 Variable importance plot for 1997–2001 data. Controls and ‘important’
roll calls used in subsequent analysis shown as subsets of all variables.
x-axis gives the increase in the regression performance from including
each variable. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
xvi
4.6 Variable importance plot for 2001–2005 data. Controls and ‘important’
roll calls used in subsequent analysis shown as subsets of all variables.
x-axis gives the increase in the regression performance from including
each variable. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
1
Chapter 1
Introduction: Purpose and
Plan
As befits its title, the following dissertation is composed of three self-contained
essays within the broad rubric of ‘political methodology.’ For current pur-
poses, this term refers to the use of quantitative techniques that can give
scholars insights into political phenomena. The stress on ‘quantitative’ is
important since the chapters below all make claims that attributes of po-
litical events and actors can be measured and compared. In line with this,
the essays develop tools that other political scientists can use to analyze the
data scenarios they face. Though these themes are consistent threads, the
dissertation is eclectic: in terms of the models of inference, the nature of
the techniques used, the substantive topics of investigation, and the ques-
tions that are actually answered. To this end, one of the papers is avowedly
Bayesian, while two are not. Two papers utilize parametric methods, while
Section 1.0 2
the third uses non-parametric techniques. The papers cover the three broad
areas of substance seen in the discipline: International Relations, American
and Comparative politics respectively.
Pursuing such a diverse agenda in one work is not uncontroversial; King
(1989), for example, argues for a ‘unified’ methodology (in practice, max-
imum likelihood) in political science that would facilitate the application
of consistent standards, communication and the cumulation of knowledge
within the discipline. With good reason, he warns against techniques which
are “imported intact and without adaptation” from other disciplines (King,
1989, 3). The current dissertation is more sanguine about such endeav-
ors: it borrows techniques more commonly seen in bio-statistics, pattern-
recognition and market research literatures. The contention is that, if as-
sumptions (and thus limitations) are clearly described, readers may judge for
themselves whether or not the model is appropriate. In any case, in review-
ing the technical literature in each essay, an argument is made for the absence
of an existing method to solve the specific problem addressed. Moreover,
as will be explored, the particular problems here are neither uniquely rare
nor trivial—that is, the chapters are not ‘solutions in search of applications’.
A second concern for work in political methodology is that of specifying
appropriate ‘theory’. Achen (2002, 424), for example, argues that “we have
yet to give most of our statistical procedures legitimate theoretical micro-
foundations.” For Achen (2002, 437), a micro-foundation is “a formal model
of the behavior of the political actors under study. The model might emerge
Section 1.1 3
from decision theory, game theory, or some other formalism. Then the statis-
tical setup is derived from the model, with no further ad hoc adjustments.”
In the second chapter, this prescription is broadly followed: a model of hu-
man behavior is suggested and an estimator derived directly. In the first and
third chapters it is not. That is in part because these other essays are more
concerned with data exploration and summary, rather than establishing “re-
liable empirical generalizations.” In statistics, such endeavors are common
and their usefulness is in part derived from the fact that they simplify and
describe (for the analyst) an otherwise complex picture of information. As a
first pass on the data they are helpful: previously hidden regularities emerge
and help scholars (including the present author) to think more systemati-
cally about what is driving behavior ‘under the hood.’ Of course, it is a thin
line between ‘data mining’ and ‘data dredging’ but these essays make a good
faith attempt to avoid the latter. In any case, as argued above, introducing
new techniques to the discipline where current ones simply cannot be used
is a fruitful exercise.
1.1 Plan
The first essay (Chapter 2) uses a reversible jump Markov chain monte carlo
approach for examining structural breaks in the current Iraq conflict. The
interest is in identifying patterns of violence, especially as they relate to
polity building. We find four ‘turning points’—all of which correspond to
increased civilian casualties. These are arguably congruent with insurgent
responses to widely publicized events including the arrest of Saddam and the
Section 1.1 4
holding of elections. Thus the chapter makes a contribution to the system-
atic study of international conflict and comparative political development.
Methodologically, the chapter adds to the political science toolkit by sug-
gesting a more general Bayesian approach to change point detection that is
philosophically appealing.
The second essay (Chapter 3) suggests a new way of measuring ‘power’
for actors in structured settings like legislatures and courts. In contrast to
a priori pivotality indices (of which Shapley-Shubik and Banzhaf are most
famous) which have been criticized for their measurement of ‘luck’ rather
than ‘power’, this essay introduces a data-driven Bradley-Terry approach
that proceeds directly from a theoretical model of (pairwise) actor behav-
ior. We apply the model to the 108th Senate. In its unstructured form, we
are able to produce a valid rank ordering of senators, with individuals such
as John McCain, Bill Frist and Robert Byrd appearing in the power top
ten. When a structured, covariate model is used, we show that, empirically,
chamber centrists lack power relative to party medians. Moreover, we are
able to demonstrate that, contrary to some expectations, female senators
tend to be as powerful (on average) as males.
The third essay (Chapter 4) uses a non-parametric regression approach to
examine the relationship between a member of the British parliament’s leg-
islative activity and their constituency electoral performance. In contrast to
assumptions of no effect, we show evidence of a subtle reward/punishment
dynamic, contingent on the nature of the roll calls in question. This pa-
Section 1.1 5
per combines quantitative and qualitative research on roll calls in the years
1997–2001. Methodologically, we describe a new ‘random forests’ approach
for political scientists interested in problems for which the number of pa-
rameters far exceeds the number of observations, and for which the causal
story is potentially complex and non-linear.
We conclude in Chapter 5 with a discussion of future applications and ex-
tensions to the work here, along with some comments on the role and future
of political methodology more generally.
6
Chapter 2
‘Turning Points’ in the Iraq
Conflict: Reversible Jump
Markov Chain Monte Carlo
in Political Science
President Bush believes that the region is at a true turning point.
He believes that the people of the Middle East have a real chance
to build a future of peace and freedom and opportunity.
—Condoleezza Rice, National Security Advisor (2003)
2.1 Introduction
The study of inter– and intra– state conflict is a mainstay of political sci-
ence. As an international conflict that increasingly resembles a civil war,
the current situation in Iraq provides both a testing ground for theories on
the duration and termination of different types of conflicts (e.g. Filson and
Werner, 2004; Stam and Bennett, 2006), as well as a rich source of data
for empirical work. This is quite separate from its obvious importance as a
political, military and economic event in progress. In part due to its contem-
Section 2.1 7
poraneous nature, political scientists have access to carefully, daily recorded,
military and civilian casualty information: an unusual and excitingly fine
level of detail. Of course, the utility of any data is only as good as the way
it is explored and analyzed. Here, we suggest that a fruitful approach for
political scientists lies in examining the time series for (potentially multiple)
structural breaks and their effects. For scholars of American politics and
public policy, the way that these change points correspond with administra-
tion statements on the progress of the war may be particularly intriguing.
This notion extends to Comparative institutions scholars interested in the
potentially pacifying effect of various post-war ‘state-building’ activities. In
keeping with the increasing acceptance and popularity of Bayesian methods
in political science, in undertaking our study we justify and adopt a novel (to
political science) approach that uses a more general form of Markov chain
monte carlo (MCMC) techniques, well-known to statisticians as ‘reversible
jump’ MCMC (Green, 1995). We do so primarily for computational reasons.
Examining civilian casualty data from the official cessation of hostilities
(May 2003) to May 2007, we find evidence of four change points. These
breaks are approximately contemporaneous with (1) the capture of Sad-
dam, and the emergence of the Abu Graib scandal (late 2003 to Spring
2004); (2) the installation of the Iraqi Interim Government, and the sub-
sequent handover of power to the Iraqi Transitional Government (Summer
2004 to early 2005); (3) the legislative elections for, and negotiations to
form, the first full-term Iraqi government (the early months of 2006); (4)
the assumption of security and some military responsibilities by the Iraqi
Section 2.2 8
government (August/September 2006) . In every case, the frequency with
which such incidents occur is increasing after the break.
2.2 Background and Data
The United States and allied forces attacked Iraq with aerial bombardments,
followed by a land invasion, on March 20th, 2003. By mid-April, Iraq’s cap-
ital city, Baghdad and Saddam Hussein’s home region of Tikrit was under
allied control—bringing a de facto end to the war. A fortnight later, on May
3rd, 2003, President Bush declared that allied combat operations would now
officially cease. As with all conflicts, the war has not been costless. What
marks the Second Iraq War though, is the continued loss of life after the Iraqi
army was formally defeated. At the time of writing, some 3,000 coalition
force members had died in addition to at least 57,000 civilian fatalities since
military operations began (sources are http://icasualties.org/oif/ and
http://www.iraqbodycount.org/ respectively. Some studies have placed
the number of civilian fatalities at a much higher number. For example,
Burnham et al. (2006) claim up to 600,000 deaths). Violence has not yet
abated despite the passing of some presumably important landmarks in what
some characterize as the development of Iraq’s polity and stability: for ex-
ample, the capture of Saddam (December, 2003), the placing of the former
dictator on trial from ‘crimes against humanity’ (July 2004) and his execu-
tion (December 2006); the killing of Saddam’s sons, Uday and Qusay (July
2003); National Assembly elections (January 2005); the drafting (December
2003–March 2004) and subsequent referendum approval (October 2005) of a
Section 2.2 9
constitution; the election of a new president (April 2005) and the forming of
a governing coalition (May 2006); the execution of an Al-Qaeda ringleader,
Abu Musab al-Zarqawi, thought responsible for planning many terrorist at-
tacks (June 2006); the assumption of security responsibility by the Iraqi
government (September 2006). We are interested in violence for the post-
(official) war period: although we certainly cannot make firm causal claims,
our study will enable us, for example, to make statements about the plau-
sibility of various events as “turning points” and allows us to pass some
exploratory comments on how new democratic institutions and state appa-
ratus developments are effecting Iraqis. Hence our study focuses on May
3rd 2003 through to the present time of writing (May 2007).
Our data are drawn from iraqbodycount.org a (online) data base that
records civilian deaths in Iraq “that have resulted from the 2003 military
intervention by the USA and its allies. The count includes civilian deaths
caused by coalition military action and by military or paramilitary responses
to the coalition presence (e.g. insurgent and terrorist attacks)” (Dardagan
and Sloboda, 2006). The data in raw form record deaths at the day level,
from January 2003 through to the present and are compiled from (primarily
Western) media reports and other sources. Since uncertainty often exists
on precise numbers, especially when different agencies have conflicting fig-
ures for the same incident, the data base reports a range of possible death
numbers from a ‘minimum’ to a ‘maximum.’ Potential ‘over-counting’ is
a concern, so we use the ‘minimum’ and define a ‘casualty incident’ as in-
volving five deaths or more (our findings below are similar when we define
Section 2.3 10
the incident threshold at ten or twenty deaths). For the purposes of this
paper, we focus on the (changing) frequency of attacks, rather than their
size (above our minimum). In part this is a behavioral assumption: we
would contend that, at least initially, terrorists were able to control how
often they planned to inflict casualties, rather than how many. There were
1682 such incidents in our time series, and we graph their occurrence in
Figure 2.1; there, the solid line is the cumulative incident count, the solid
dots are simply jittered incident occurrences (for which the y-axis is not
the scale). We also report various dates that may of interest and to give
readers a sense of timing perspective. Although univariate time series work
is not regularly encountered in political science, it is valuable in the current
context as a ‘first glance’ exploration before covariate information becomes
available. We think that such work helps to prompt both theorizing and
data gathering for more nuanced and sophisticated analysis.
2.3 Estimation Problem
The single change point problem, estimated using Markov chain monte carlo
techniques, has been discussed for and by political scientists elsewhere (see
Western and Kleykamp, 2004). That treatment is similar to the (hierarchi-
cal) presentation given by Carlin, Gelfand and Smith (1992): suppose y =
(y
1
, . . . , y
T
) is a vector of observations of the random variable Y (casualty
incidents) over time and let f and g be unknown densities in the same para-
metric family with y
i
∼ f(Y |λ
1
), i = 1, . . . , k, y
i
∼ g(Y |λ
2
), i = k + 1, . . . T.
We wish to estimate k the (single) change point which takes (discrete) values
Section 2.3 11
0 500 1000 1500
0
5
0
0
1
0
0
0
1
5
0
0
Time (days after May 03, 2003)
C
u
m
u
l
a
t
i
v
e

C
o
u
n
t
Uday, Qusay killed
Saddam captured
Transitional Admin Law
Elections for transitional govt
Constitution writing finished
First full term govt takes power
Figure 2.1: Iraq Casualty Incidents, May 2003–May 2007. Solid line represents cu-
mulative counts; solid dots are jittered incident occurrences; various dates of interest
demarcated.
Section 2.3 12
in {1, 2, . . . , T}. A frequentist approach proceeds by maximizing
L(y) =
k

i=1
f(y
i

1
)
T

i=k+1
g(y
i

2
) (2.1)
to obtain k and the parameters λ
1
and λ
2
(which for the count case are
arrival rates for a Poisson) if they are of interest. A Bayesian approach
proceeds by placing a prior τ(k) on the change point. There are computa-
tional advantages of a Bayesian MCMC approach here since (a) maximizing
(2.1) requires optimization in a space that is not continuous (recall that k is
discrete) which, say, Gibbs sampling does not; (b) the resultant non-nested
models may be straightforwardly compared using Bayes factors (Chib, 1998);
(c) missingness in y is handled systematically. This is quite apart from the
philosophical appeal of Bayesian approaches of which political scientists are
increasingly aware (see, for example, Gill 2002, 1–6 and Jackman 2004, 486).
Here, we are interested in exploring multiple change point and such work
(Bayesian or otherwise) is much less common in political science. In part this
is because, with respect to the logic above, there are profound computational
difficulties in generating proposals for situations where we suspect there are
more than a couple of change points. One approach, suggested by Chib
(1998) and applied to American politics by Park (2006), treats the change
point model as a type of time series Markov mixture model, where the ob-
servations are (assumed) drawn from latent state variables. Notice that this
approach requires separate Markov chain monte carlo runs for the different
numbers of change points hypothesized (Leonte, Nott and Dunsmuir, 2003).
Section 2.3 13
An alternative solution is to use reversible jump Markov chain monte carlo
which allows us to complete the computational operations in one ‘go’ as well
as allowing us to be a priori agnostic over the number of parameters to be
estimated.
Typically when MCMC is used in political science the parameter vector
θ has a known number of components, denoted n. For the single change
point problem n = 3 (these are k, λ
1
and λ
2
). Now consider a very different
scenario which arises for an unknown number of k change points: for every
possible k, we need to estimate 2k +1 parameters—the change points them-
selves and then parameters of the densities before, between and after them.
That is, we have a set of M
k
= {1, . . . , K} candidate models of our data
generating process, each with a different number of parameters. Otherwise
put, the number of parameters is, of itself, a parameter. More formally,
the k
th
model in M
k
has associated parameter vector θ
k
which contains n
k
parameters such that θ
k
∈ R
n
k
.
Continuing to denote our data vector y, the joint distribution becomes:
p(k, θ
k
, y) = p(y|k, θ
k
)p(k, θ
k
)
= p(y|k, θ
k
)p(θ
k
|k)p(k). (2.2)
Section 2.3 14
Since we have a constant of proportionality we can rearrange and reexpress
(2.2) into the more familiar
p(k, θ
k
|y)
. ¸¸ .
posterior
∝ p(k)p(θ
k
|k)
. ¸¸ .
prior(s)
p(y|k, θ
k
)
. ¸¸ .
likelihood
. (2.3)
Notice that p(y|k, θ
k
) is simply the likelihood, while p(θ
k
|k) is the prior for
the parameter vector, given a particular data generating process and p(k)
is the prior on the model itself. We wish to generate samples from (2.3).
Setting up a Markov chain to do this may be difficult though, because it is
required not simply to move around the parameter space for any particular
θ
k
, but to also ‘jump’ from space to space (from model to model) depending
on the k in question.
This type of problem is given a general formulation by Green (1995), known
as reversible jump MCMC (RJMCMC), of which standard MCMC algo-
rithms are special cases. Green explicitly discusses a Poisson count change
point problem and we followed his approach for our application (though we
varied the priors somewhat to ensure that our results were robust to such
alternative specifications). Although well known to statisticians, the details
somewhat technical, and readers are guided to Brooks (1997) who gives an
accessible overview for political scientists.
The implementation of RJMCMC, in particular the efficiency of propos-
als, can be problematic in practice and Hastie (2005) devotes considerable
Section 2.4 15
attention to designing a technique to do this. We used his Automix sam-
pler (with a maximum of ten possible change points) for our estimation.
Though the full details are somewhat technical, drawing on Hastie (2005,
202–203), it is instructive to summarize the way that the model of the data
generating process is selected. The first two stages of the sampler produce
a Normal mixture distribution for every possible value of k. In the third
stage, assuming the Markov chain is currently in state (k, θ
k
), Automix al-
locates the parameter vector θ
k
to a component l
k
of the mixture and uses
it to standardize θ
k
. Then a new model k

is proposed, along with a com-
mensurate (new) mixture which has component l

k
. To obtain the new state
vector θ

k

, the standardized vector is transformed using the mean and the
covariance matrix of the mixture component l

k

. Automix then accepts the
proposed state (k

, θ

k

) with some specified acceptance probability. A par-
ticularly pleasing feature of this software is that issues such as burn in and
the requisite number of post-burn iterations are handled automatically.
2.4 Results
There are three sets of (posterior) distributions that interest us here:
1. the posterior of k: this enables us to answer the question “how many
change points in the data?” This will have support k = 1, . . . , k
max
where k
max
= 10.
2. the posterior of change point positions conditional on some estimated
k. More intuitively, this enables us to answer the question “given a
particular number of change points, when did they occur in the data?”
Section 2.4 16
1 2 3 4 5 6 7 8 9 10
k
0
.
0
0
.
1
0
.
2
0
.
3
0
.
4
0
.
5
0
.
6
Figure 2.2: Posterior of k: number of change points, Iraq casualty data.
3. the posteriors of the rates for each period, conditional on some esti-
mated k: that is, given the number of change points, and when they
occurred, we can answer “what were the effects of the change points?”
In Figure 2.2 we display the posterior for k, the number of change points.
The strongest evidence (in the sense of Kass and Raftery (1995)) is for k = 4
and we will explore this possibility exclusively. In Table 2.1 we summarize
the results for k = 4 model in a way that answers questions 2 and 3 above.
The first break, in late January 2004 occurs between incidents that may be
of import. The first was the capture and arrest of Saddam at a farmhouse
near Tikrit in December 2003. The subsequent months saw both an insur-
gency uprising lead by rebel Shia cleric Muqtada al-Sadr in Baghdad and the
diffusion of abusive photographs taken at the Abu Graib prison where coali-
Section 2.4 17
tion forces were holding Iraqi detainees. The political fallout of the latter
was profound, and criticism of the Bush administration by allied, Arab and
other politicians was widespread. This event, arguably, rallied and spurred
sectarian hatreds and violence. The break marks a sharp increase in the
casualty rate, doubling from one incident every four days, to one every two
days.
The second break occurs in August of 2004, a little while after the Iraqi
Interim Government assumed power from the Coalition Provisional Author-
ity (in June 2004). This new entity, under the Premiership of Iyad Allawi
was subsequently recognized as the legitimate sovereign government of Iraq
by both the United Nations and the Arab League (an important regional
player). Allawi quickly announced new security measures to tackle insur-
gency forces and was criticized by some for their draconian nature. As part
of this offensive, the Iraqi Interim Government began to censor the critical
reports of media outlet al Jazeera. Included in the highest posterior density
interval for this break is the January 2005 democratic elections for the Iraqi
Transitional Government. This change point saw an increase in violence
from one incident every two days, to four incidents every five days.
The third break itself, in February 2006, occurs not long after the elections
for the first full term Iraqi government (December 2005) and at around the
time of the protracted negotiations to form a new coalition government.
These talks were deadlocked for some time, lasting from December through
to April of the following year. Jawad al-Maliki, leader of the Islamic Dawa
Section 2.5 18
Party would become Prime Minister after the original candidate Ibrahim
al-Jaafari proved unacceptable to the Sunni and Kurdish representatives in
parliament. Once again, violence surged after this point with, on average,
five incidents occurring every three days.
The fourth and final break occurs in September of 2006, a time when the
Iraqi government assumed control of national security for approximately 70
percent of the country. The first specific task of the Iraqi Security Forces
was, and is, to tame insurgency (with coalition logistical and medical sup-
port). By now incident rates were approaching three per day. In Figure
2.3 we summarize our findings in a different way: the open circles represent
the median incident rate between the relevant breaks which are demarcated
by the broken lines. For reference, we again draw the jittered incidents
themselves on the plot.
2.5 Discussion
Our study—to our knowledge the first that uses RJMCMC in a politi-
cal science context—suggests that violence is increasing and that impor-
tant state-building activities, like democratic elections, are contemporaneous
with upticks in casualties. Apart from this rather grim substantive conclu-
sion, we found that investigating time series on violence to be an interesting
and fruitful exercise. If a Bayesian approach is pursued, then reversible
jump techniques seem most helpful. We hope that our discussion here will
encourage others in political science to consider such methods in future.
Section 2.5 19
0 500 1000 1500
Date

λ
0
1
2
3
Jan 2004
Aug 2004
Feb 2006
Sep 2006
Figure 2.3: Change in rate of casualty incidents over time: rate on y-axis, actual inci-
dents as jittered points. Dates correspond to breaks found.
Section 2.5 20
As noted above, we do not establish causation in any sense: the events
we noted were simply occurring at around the same time as the breaks in
the time series and it is speculative that they may be of direct importance.
This suggests some interesting avenues for future research: for example, one
possibility is that increasing violence is a product of an increasingly orga-
nized insurgency. On this point, notice that the solid line in Figure 2.1
resembles a exponential curve of form F(t) = ae
rt
+ (we are indebted to an
anonymous TAS referee for this observation). Modeling—both theoretical
and statistical—of this apparent pattern would allow us to think about the
development of the conflict in Iraq in a more systematic way. We leave this
for future research.
Section 2.5 21
P
e
r
i
o
d
T
i
m
e
R
a
t
e
(
λ
)
C
o
m
m
e
n
t
s
b
e
f
o
r
e
b
r
e
a
k
1
0
.
2
8
3
1
i
n
c
i
d
e
n
t
e
v
e
r
y
4
d
a
y
s
b
r
e
a
k
1
J
a
n
2
6
,
2
0
0
4
C
a
p
t
u
r
e
o
f
S
a
d
d
a
m
(
D
e
c
2
0
0
3
)
[
J
u
l
1
8
,
2
0
0
3

A
p
r
3
0
,
2
0
0
4
]
A
b
u
G
r
a
i
b
s
c
a
n
d
a
l
(
A
p
r
2
0
0
4
)
b
t
w
n
b
r
e
a
k
1
a
n
d
2
0
.
4
7
6
1
i
n
c
i
d
e
n
t
e
v
e
r
y
2
d
a
y
s
b
r
e
a
k
2
A
u
g
7
,
2
0
0
4
I
r
a
q
i
I
n
t
e
r
i
m
G
o
v
e
r
n
m
e
n
t
a
s
s
u
m
e
s
p
o
w
e
r
(
J
u
n
2
0
0
4
)
[
J
u
l
1
0
,
2
0
0
4

M
a
y
4
,
2
0
0
5
]
e
l
e
c
t
i
o
n
s
f
o
r
I
r
a
q
i
T
r
a
n
s
i
t
i
o
n
a
l
G
o
v
t
(
J
a
n
2
0
0
5
)
b
t
w
n
b
r
e
a
k
2
a
n
d
3
0
.
8
1
4
4
i
n
c
i
d
e
n
t
s
e
v
e
r
y
5
d
a
y
s
b
r
e
a
k
3
F
e
b
2
2
,
2
0
0
6
e
l
e
c
t
i
o
n
s
f
o
r

r
s
t
f
u
l
l
t
e
r
m
I
r
a
q
i
g
o
v
e
r
n
m
e
n
t
(
D
e
c
2
0
0
5
)
[
F
e
b
1
6
,
2
0
0
6

M
a
r
2
4
,
2
0
0
6
]
b
t
w
n
b
r
e
a
k
3
a
n
d
4
1
.
6
6
0
5
i
n
c
i
d
e
n
t
s
e
v
e
r
y
3
d
a
y
s
b
r
e
a
k
4
S
e
p
t
4
,
2
0
0
6
I
r
a
q
i
g
o
v
t
a
s
s
u
m
e
s
c
o
n
t
r
o
l
o
f
c
o
u
n
t
e
r
-
i
n
s
u
r
g
e
n
c
y
[
A
u
g
2
3
,
2
0
0
6

S
e
p
t
2
0
,
2
0
0
6
]
o
p
e
r
a
t
i
o
n
s
f
o
r
m
u
c
h
o
f
c
o
u
n
t
r
y
(
S
e
p
t
2
0
0
6
)
a
f
t
e
r
b
r
e
a
k
4
2
.
8
5
0
8
i
n
c
i
d
e
n
t
s
e
v
e
r
y
3
d
a
y
s
T
a
b
l
e
2
.
1
:
C
h
a
n
g
e
p
o
i
n
t
m
e
d
i
a
n
d
a
t
e
s
[
9
0
%
H
P
D
]
i
n
I
r
a
q
c
a
s
u
a
l
t
y
d
a
t
a
,
w
i
t
h
r
a
t
e
s
b
e
t
w
e
e
n
b
r
e
a
k
s
.
C
o
m
m
e
n
t
s
n
o
t
e
p
o
l
i
t
i
c
a
l
a
n
d
o
t
h
e
r
a
c
t
i
v
i
t
i
e
s
a
p
p
r
o
x
i
m
a
t
e
l
y
c
o
n
t
e
m
p
o
r
a
n
e
o
u
s
t
o
b
r
e
a
k
.
22
Chapter 3
Measuring Power in Political
Science: A New Method
with Application to the
Senate
Power is not revealed by striking hard or often, but by striking
true.
—Honor´e de Balzac
3.1 Introduction
Scholars have long conceptualized politics as a process of conflict over re-
sources. To determine “who gets what, when and how” (Lasswell, 1936), po-
litical scientists have invoked notions of ‘power’ (Dowding, 1996, 1–8). Early
positive political theory made the case that median voters—be they in the
electorate or in parliaments—were particularly decisive and hence ‘powerful’
(Black, 1948; Downs, 1987; Riker, 1962). Later theorists—particularly those
associated with the ‘Rochester school’ (Amadae and Bueno de Mesquita,
Section 3.1 23
1999)—have focused on the role of institutional rules and norms in giving
‘players’ advantages over one another in settings like Congress. This litera-
ture is, by now, enormous, and includes foundational work by Riker (1982,
1986) and others (e.g. Shepsle (1979), Shepsle and Weingast (1987), Kre-
hbiel (1991)). Despite this theoretical work, the empirical investigation and
measurement of power—even in structured settings like parliaments, com-
mittees and courts—has proved difficult and contentious, concentrating on
a priori metrics that are removed from the data that we have regarding
agents’ actual actions and decisions (Felsenthal and Machover, 1998, 2004).
This paper is an attempt to address this problem and to bridge a gap be-
tween the theoretical treatments of power familiar to positive political theo-
rists and the empirical work of political methodologists. Following Dowding
(1996), we suggest an actor-based, data driven approach. Thus, we treat
‘power’ as a latent variable or ‘ability’ possessed to varying extents by actors
like Congressmen (Senators in particular). We demonstrate that this trait
is straightforwardly uncovered by studying easily available voting records.
A strength of this approach is that the definition of power is theoretically
removed from factors that influence how powerful individuals actually are.
Methodologically, our contribution to the discipline is the introduction of a
new version of the Bradley-Terry (Bradley and Terry, 1952) model for pair-
wise comparisons, which includes regressors that are thought to covary with
actors’ power. We show the utility of both the conceptual and empirical
suggestions by studying ‘power’ in the 108th United States Senate where
we show the role that ideology, institutional arrangements, geography and
Section 3.2 24
personal factors play in determining the power of its members. To wit, in
Section 3.2 we discuss the problems with traditional voting index measures
of power; in Section 3.3 we suggest, derive and discuss a new data driven
model; in Section 3.4 we apply the model to the 108th Congress and pro-
duce a ‘power list’ of Senators therein; in Section 3.5 we show how the power
can be predicted by its proposed ‘causes,’ and we discuss the impact of, in-
ter alia, party, committee assignment, geographic and personal factors. In
Section 3.6 we conclude and suggest future avenues for research.
3.2 Problems with Traditional Approaches to Mea-
suring Power
Scholars have been interested in measuring the power of actors in structured
settings for over half a century.
1
Beginning with Shapley and Shubik’s sem-
inal 1954 contribution, with extensions most famously by Banzhaf (1965),
this branch of positive theory has failed to find widespread acceptance in
political science for at least three reasons. First, when voters have different
weights, different indices yield different results for the same data and, gener-
ally, the resulting ambiguity cannot be resolved because there is no objective
evidence on the actual distribution of power (Leech, 2002a,b). Second, for
institutions with large numbers of actors, there are computational difficul-
ties with performing the requisite calculations (Leech, 2002a,b). Third, there
1
Social scientists have discussed ‘power’ in various ways for well over a century: at least
since, for example, Marx and Das Kapital. The American Political Science Association felt
it was sufficiently important and unresolved even in 2006 to devote their annual meeting
to “Power Reconsidered.”
Section 3.2 25
is much sceptism on the question of whether the indices actually measure
‘power’ at all (Barry, 1991; Dowding, 1996; Riker, 1964).
The third criticism is the central issue here. Barry (1991) argues that,
since power indices are based on the probability that an actor is pivotal,
they measure something akin to ‘luck’ or ‘decisiveness’ rather than power.
2
The point is that any definition or measurement of power must incorporate
the notion of getting one’s preference in the face of resistance from other
actors. Moving on from this notion, Dowding (1996) considers pivotality as
a resource among many.
3
The implication is that one’s power index score is
simply one of several independent variables explaining or ‘causing’—but the-
oretically distinct from—an individual’s ‘power’ (which might be thought of
as a latent variable). Thinking this way enables political scientists to break
away from some nonsensical statements such that the power of the Chief
Justice in the Supreme Court is (
1
9
), or that the Chairman of the Finance
Committee in the United States Senate is is equal in power terms to a fresh-
man senator who has never held a committee post (
1
100
).
In sum: power indices whatever their hue, are not ideal tools for studying
what political scientists conventionally think of as ‘power.’ A new method
is needed which, inter alia,
2
A natural defence is that power indices are intended for a priori analysis; that is,
power indices tell us about the distribution of power before we consider the preferences
of actors or their ability to set an agenda (see Felsenthal and Machover (1998) for a
comprehensive review. See also: Lane and Berg (1999) and Holler and Widgren (1999) in
their response to Garrett and Tsebelis (1999)).
3
See also Krehbiel (1998) for a development of ‘pivotality’ with respect to US legisla-
tures.
Section 3.3 26
1. defines power as the capacity to get what an actor wants in the face
of resistance. . .
2. treats power as a (latent) capability. . .
3. estimates an actor’s power as a function of independent variables. . .
4. incorporates commensurate statements of uncertainty and. . .
5. allows for explicit comparison of the effects of different predictors.
In the next section, we show one way to achieve these aims.
3.3 Statistical Theory and Model
‘Power,’ as considered here, is a latent variable. Since we cannot measure
it directly, we need a statistical model that takes observable data, estimates
an unobservable trait, and outputs a metric. We made the case above that,
in the broadest sense, power is an actor’s ability to obtain his preferred
outcome. In the case of legislatures, which are our concern here, members
require majorities to pass bills and an actor’s power in these settings will
turn on his capability to form legislative coalitions for the issues he cares
most about. There is resistance, in the form of actors who vote against the
bill, precisely because they stand to lose from its passage. After introduc-
ing a little notation, we now provide a way to systematically measure actors
—in fact, senators’—capabilities to achieve their preferences in such settings
Let i = 1, . . . , I index the senators and j = 1, . . . , J index the bills on
Section 3.3 27
which they vote. For any particular j, there exists a senator who proposes
the bill, denoted i
p
j
. In any session of the Senate, there will be not more
than J such individuals, since all bills are proposed by someone, but not all
senators propose. The proposer of bill j, i
p
j
, is assumed to seek a coalition
to vote for the bill and we denote that coalition c
i
p
j
which has |c
i
p
j
| mem-
bers (one of whom is the proposer). It is natural to assume that i
p
j
prefers
a coalition that is a majority of all senators voting on j, but this is not
strictly required here.
4
For now, we suppose that the benefits of passing
j—be they pork, credit, news headlines—will accrue only to i
p
j
and that
other legislators pay some cost in supporting the bill, perhaps because the
total size of the pork or publicity ‘pie’ available is now decreased for them
and their home districts. Else, there is some simple opportunity cost of the
time they take to vote. In attempting to form the coalition, we say that
i
p
j
‘convinces’ a senator to join c
i
p
j
if she subsequently votes in favor of the
bill (and hence in favor of i
p
j
’s ideal point relative to the status quo). The
nature of the convincing could take many forms: the proposer may attempt
to reason, argue, threaten or bargain with his colleagues explicitly, or else
the inducement may be implicit. In subsequent votes, say j + 1, the situa-
tion may well be reversed with the previous proposer now joining a coalition
c
i+1
p
j+1
backing a senator i +1 who had been a member of c
i
p
j
. We say that
two senators ‘interact’ in a particular session of Congress if one proposes a
bill and the other joins its backing coalition. Notice that for any particular
bill j proposed by i, there are c
i
p
j
−1 interactions.
4
We assume rather that a proposer is coalition size maximizing: his utility is increasing
in |c
i
p
j
|. We could think of this as increasing the probability that j is passed.
Section 3.3 28
For each senator, the statistical model we discuss below will determine a
score λ
i
such that if we know only that Senator A and Senator B are in-
volved in an interaction, the probability that it is A convincing B (i.e. B is
backing A’s proposal) rather than B convincing A, is the difference of the
scores λ
A
−λ
B
:
logodds(A convinces B|A and B interact ) = λ
A
−λ
B
. (3.1)
The interpretation is then straightforward: the greater the value of λ
i
rela-
tive to other senators, the more powerful we hold that senator to be.
The derivation of the estimator begins by assuming that there exists a latent
and hence unobservable power for each senator denoted α
i
. For any given in-
teraction (any particular bill) between a pair of senators, let π
AB
∈ (0, 1) be
the probability that the interaction involves Senator A convincing Senator
B. Since an interaction must involve either A convincing B or B convincing
A it must be true that π
AB
+ π
BA
= 1.
5
Write the odds that A convinces
B as a function of their latent powers, such that
π
AB
π
BA
=
π
AB
1 −π
AB
=
α
A
α
B
. (3.2)
5
An objection may be that party or some other factor compels senators to vote one
way or the other, in which case this equality will not hold. By studying personal amend-
ments in Section 3.4 we ameliorate these concerns somewhat but see also our ‘order effect’
specification below.
Section 3.3 29
This formulation has the obvious consequence that, if α
A
> α
B
, in any given
interaction, π
AB
> π
BA
. That is, for any given interaction between A and
B, A is more likely to be forming the coalition (and gaining his preferred
outcome) than B is.
The problem of estimating α
i
can be approached via a logistic regression.
To see how, first let α
i
= exp(λ
i
). Then, some rearrangement yields
6
π
AB
=
exp(λ
A
)
exp(λ
A
) + exp(λ
B
)
. (3.3)
Suppose that A and B interact a total of N
AB
times. Of these interactions,
let n
AB
—where n
AB
≤ N
AB
— be the number of times that A convinces B.
Then, so long as the N
AB
interactions are independent of one another, and
the same probability π
AB
applies to each interaction, n
AB
has a binomial
(N
AB
, π
AB
) distribution. If we are also willing to assume that all the other
interactions between the other senators are also independent, we have a logit
model that can be estimated via maximum likelihood:
logit[Pr(A convinces B)] = λ
A
−λ
B
. (3.4)
We can then compare the relative size of λ
A
and λ
B
to see which senator is
more powerful.
7
6
From Equation (3.2), it is obvious that π
AB
(1 +
α
A
α
B
) =
α
A
α
B
; substitute
α
B
α
B
for 1 and
note that α
i
= exp(λ
i
) to obtain (3.3).
7
Importantly, the model does not require that the matrix of interactions is ‘complete’
in the sense that every senator interacts at some point with every other. The model
implicitly assumes transitivity: if Senator A is more powerful than B, and B is more
powerful than C, then A is also more powerful than C.
Section 3.3 30
A concern with this formulation might be that, in fact, party pressures
and thus a priori similar preferences between senators, obviates the need to
‘convince’ other legislators in any real sense. Otherwise put, some senators
are disposed to supporting others like them, whatever the proposers power.
We will tackle this issue head-on by estimating a version of our logistic
regression as
logit[Pr(A convinces B)] = λ
A
−λ
B
+ δw (3.5)
where w is an indicator taking the value 1 if the proposer’s party is the same
as that of an individual he convinces and −1 otherwise. Hence, δ will cap-
ture the ‘natural advantage’ of proposing to someone in a senator’s party,
and the λ
i
will reflect the ability absent this advantage.
Notice that our method here is very different to that of previous endeavors
to measure power. Unlike index methods, it is an a posteriori in the sense
that we infer ‘power’ after observing actor’s actual decisions—rather than
before. Hence the terms ‘data driven’ and ‘actor based.’ Second, unlike
subjective rankings it is objective and relies on a clearly defined definition
of power. This means, in contrast to ‘panel of expert’ surveys, that our
findings (below) are exactly replicable by any other political scientist who
chooses so to do. Moreover, it is much cheaper and faster to calculate.
This estimator is not new to this paper; it is the Bradley-Terry (Bradley
Section 3.4 31
and Terry, 1952) pairwise comparison (with equation 3.5 representing an
‘order effect’ specification) method in a novel setting. This approach is well
known and well studied by statisticians working in fields as diverse as sport
team rankings (Agresti, 2002), journal citation patterns (Stigler, 1994) and
competition for mates in the biological sciences (Stuart-Fox et al., 2006).
8
This is not the first paper in the discipline to suggest thinking of some
political phenomena as pairwise interactions: for example, Groseclose and
Stewart (1998) study the value of Congressional committee assignments us-
ing (dyadic) transfers of representatives.
9
The paper is perhaps closest in
spirit (though not in execution) to work by Wawro (2000) who studies leg-
islative entrepreneurship in the House. Designing a novel “entrepreneurship
scales score” based on five observables of behavior, Wawro shows that party-
based career prospects are strongly linked to a representative’s records of
actively introducing legislation. Below, we build on this work by presenting
congruent findings—though with a different causal direction: in particular,
that party and committee advancement aids senators in their quest to pass
(potentially controversial) legislation. Also, like Wawro we provide scholars
with a metric that may be used for further research.
8
Such models can be fitted with many standard statistical packages. For this paper, R
(R Development Core Team, 2006) was the environment of choice in conjunction with the
BradleyTerry library (Firth, 2005).
9
See also King (2001) who discusses some possible extensions of standard international
relations models that assume dyadic interactions, and Fowler (2006) who studies the ‘con-
nectedness’ of legislators in both chambers—though not in an explicitly pairwise way.
Section 3.4 32
3.4 Application: 108th Senate
We use data from the 108th Senate, that met between January 2003 and
December 2004. The universe of cases for the present analysis is the 255
amendments, by 74 different senators, proposed and voted upon during this
time. Amendments are preferred to bills as a whole since they “tend to
reflect more specific changes to a bill that are less susceptible to deviations
from the sponsor’s original intent” (Fowler, 2006, 9). That is, they are
more amenable to the notion of personal coalitions formed to achieve an
individual’s goals. Studying the Senate in particular has the further ben-
efit that, subject to some constraints negotiated via Unanimous Consent
Agreements, its members may propose amendments at essentially any time,
in any order, without permission from a Committee of the Whole. This
means that, unlike the House, the status quo is afforded less protection—an
ongoing concern for those studying power (see Bachrach and Baratz, 1962,
for example). There are 101 actors in the current data set: 74 proposing
senators, 26 non-proposing but voting senators and the President (where
his views on the amendments are known). Of the amendments, 112 were
winning, 139 were losing, and four were tied votes. Recall that the number
of observations is the number of interactions, and is thus well in excess of
the number of amendments: otherwise put, each amendment corresponds to
multiple interactions and hence multiple observations.
One concern might be that senators vote strategically: against their first
preferences for an alternative that is a priori less preferred. They might
Section 3.4 33
also propose strategically, to ensure their preferred alternative is selected by
majority rule (Riker, 1982, 1986). However, it is not immediately apparent
that strategic (let alone ‘killer’) amendments are common. Second, strate-
gic voting on amendments, as opposed to bills, should be relatively rare:
we can generally assume that those voting for an amendment want it to
pass. Third, strategic voting is presumably a product of the organizational
structure—like the timetabling procedures—of the Senate: but this is pre-
cisely the sort of consideration that ought to be included in the calculation
of our metric. Lastly, votes in the Senate are part of the public record, so
we expect senators to treat their vote seriously and think through its con-
sequences.
10
In Table 3.1, we report a selection of the power estimates for this specifica-
tion. Recall that there is no sense in which the power measure is absolute,
and all the coefficients are scaled relative to zero which is assigned to Pres-
ident Bush.
Examining the table, we note that the long serving (since 1979) John Warner
of Virginia (senior senator) is ranked first. Second is Mitch McConnell of
Kentucky (senior senator). In the 108th Senate he was elected as the ma-
jority whip by the Republicans and, at the time of writing, was the leader
of Republicans in the Senate. Judd Gregg, the senior (Republican) sena-
10
Notice that simply proposing (and forming coalitions) for more amendments cannot
in and of itself increase a senator’s power in this model: see Appendix 3.7 and Appendix
3.9.
Section 3.4 34
tor from New Hampshire is at third. Gregg chaired the Health, Education,
Labor and Pensions Committee in the 108th Congress, and became chair
of the Budget Committee in the 109th. Robert Byrd, the longest serving
senator in history is the first Democrat on this power list. From West Vir-
ginia, Byrd is, at the time of writing, the President Pro Tempore—a role he
had previously held in the 1980s and 1990s. John McCain of Arizona, then
chair of the Commerce committee, is followed by Bill Frist of Tennessee, the
Senate Majority leader for the Congress in question. Thad Cochran and
Charles Grassley, chairs of the Appropriations and Finance Committees re-
spectively (in the 109th Congress) hold the 7th and 8th spots. Barbara
Boxer, the junior senator from California is the first woman to feature on
the list. At the bottom of the ranking, we see both senators from Hawaii
(Daniel Akaka and Daniel Inouye) along with Jim Jeffords (Republican to
2001, Independent thereafter).
11
Fitting an ‘order effect’ model such that
the ‘natural advantage’ of proposing to like-minded partisans is controlled
for—as specified in equation 3.5—makes little appreciable difference to this
rank order: see Appendix 3.8 for more details.
Subjective rankings—published in magazines such as Time and by consul-
tancy groups like Knowlegis—reach very similar conclusions to ours. The
Knowlegis power list for 2005 for example, one session after the 108th, places
Thad Cochran at 1 (7 in our ranking), Mitch McConnell at 4 (2), Charles
Grassley at 7 (8), John McCain at 8 (5), Bill Frist at 10 (6), Arlen Spec-
11
Importantly, the rank ordering is very different to that which might be garnered from
looking solely at the number of amendments proposed by each senator. See Appendix 3.9
for this rank ordering.
Section 3.4 35
tor at 9 (15) and Orrin Hatch at 2 (19). This similarity to our ranking is
remarkable given that (a) they are discussing a different Congress and that
(b) our approach is based on a simple, model-based voting metric that is
relatively straightforward to compute. The Knowlegis list, by contrast, is
based upon a vast and expensive survey incorporating what is theoretically
much more information.
12
As noted above, a pleasing feature of the current estimator is that the power
estimates λ
i
have meaning outside of a simple rank ordering. Recall that
λ
A
− λ
B
is the (anti-logged) probability that, conditioned on two senators
interacting, it is A that convinces B to back his amendment rather than the
other way round. Consider, for example, John McCain and Russ Feingold
(Wisconsin), coauthors of the Bipartisan Campaign Reform Act of 2002.
The probability that, in any interaction between these two, it is McCain
proposing the bill and Feingold backing it is
exp(λ
McCain
)
exp(λ
McCain
) + exp(λ
Feingold
)
=
exp(21.58)
(exp 21.58) + exp(19.85)
≈ 0.85.
By contrast, the same calculation for McCain and Rick Santorum yields a
probability of (very close to) 1. This is hardly surprising given that the sen-
ator from Pennsylvania did not offer an amendment in the 108th Congress,
12
According to their website, “Knowlegis staff carefully researched, sorted and con-
sidered thousands of data points to determine . . . power. . . reviewed thousands of me-
dia articles, hundreds of bills that passed out of committee . . . over a thousand amend-
ments. . . We collected data on the leadership, committee, and caucus positions of each
Member . . . researched relevant campaign contributions, and considered any characteristic
or action that could contribute to their Sizzle-Fizzle factor. . . there are more than 10,000
data points and variables that were considered in the 2006 Knowlegis Power Rankings”
source:http://www.congress.org/congressorg/power rankings/backgrounder.tt
Section 3.5 36
but it lends some validity to the estimator.
A further factor in favor of the new approach concerns the goodness of
fit: in contrast to standard methods this is at least meaningful (what is the
goodness of fit for a subjective survey?) and, in fact, respectable, at some
85% of interactions correctly predicted. This figure is within the ballpark of
similar statistics for industry standards like NOMINATE (see Poole and Rosen-
thal, 1997), though given the structure of the underlying statistical model
it is calculated somewhat differently. Appendix 3.10 reports more details.
In sum, Table 3.1 seems to be a reasonable ‘influence list’ both in terms
of its underlying statistical model and its actual contents; it tells us little,
however, about the causes of power, a subject to which we now turn.
3.5 Structured Modeling
We usually have theories about what explains the power of different individ-
uals; indeed, sometimes we treat characteristics that are causes of power as
if they were synonymous with power itself: consider, for example, the notion
that ‘the rich are powerful.’ A pleasing feature of the current approach and
estimator is that we can separate these notions and explicitly incorporate
individual specific explanatory variables as predictors of power. We estimate
λ
i
=
p

r=1
β
r
x
ir
(3.6)
Section 3.5 37
and thus we predict the power of each senator as a (linear) function of ex-
planatory variables x
i1
, x
i2
, . . . x
ip
with coefficients β
1
, β
2
. . . , β
p
(see Firth
(2005) and Springall (1973) for details). Since the approach here is essen-
tially a logistic regression, we can interpret coefficient estimates as positive
or negative, in terms of their marginal effects on power, and we have com-
mensurate standard errors. We can thus make uncertainty statements about
our predictors.
We used institutional and personal information to explain power in the
Senate. Collecting such data is straightforward: the Senate itself, the Gov-
ernment Printing Office and the United States Census Bureau provided all
the relevant variables below in electronic form. We break the findings into
four subsections dealing with ‘Party and Ideology,’ ‘Committees and Agenda
Control,’ ‘Geographic Factors’ and ‘Career Factors.’ The intention here is
not to provide an exhaustive account of power, but to demonstrate the
strength of the approach and possible avenues for future research. Before
describing the results, notice that the nature of ‘power’ now being consid-
ered is altered somewhat. In the previous section, power was an ability to
be ascribed to individual senators—it was a ‘personal’ characteristic. Now
though, power is a function of variables and is being treated in an ‘institu-
tional’ sense, separate from the individuals who wield it. Otherwise put, it is
now a maintained assumption that institutional (and other) characteristics
make individuals powerful.
13
13
Notice that a potential endogeneity concern—that individuals already endowed with
latent power ipso facto obtain important institutional advantages—is being avoided by
construction.
Section 3.5 38
3.5.1 Party and Ideology
In contrast to the ‘textbook Congress’ of Fenno (1973), there is increasing
evidence that Congressional voting is ideological and party driven (McCarty,
Poole and Rosenthal, 2006).
14
This is more true of the House than the Sen-
ate, but nonetheless it suggests some testable hypotheses. As noted above,
senators must assemble majorities to obtain their own preferences. If a ma-
jority party exists (the Republicans for the 108th Congress) we might expect
those from the majority party to be more powerful than those from the mi-
nority (Democrats). A refinement on this theme is that we expect leaders of
Congressional parties to be especially powerful: the Senate Majority Leader
(Bill Frist for the 108th Congress), for example, has the ability to schedule
debate. Other officials—which we refer to as ‘junior leaders’—such as Pol-
icy Committee Chairs have powers to design and execute policy ideas. In
Table 3.2 we give the results of the model for a Majority dummy variable
and SeniorLeadership, a dummy that denotes either the senior Senator
for each party, or the whips for each party. We also interact these variables.
The positive and significant effects of being a Republican and being part
of the Senate’s senior leadership are evident from Table 3.2. From the in-
teraction term it is evident that being a Republican and a leader adds an
extra fillip to one’s power. To be clearer here and recalling equation (3.6),
consider the power of some senator A who is a member the majority party
14
See also Huitt (1957) who discusses cross cutting tensions of ideology and party as it
applies to senators.
Section 3.5 39
and also a senior leader. By Table 3.2,
λ
A
= 0.286 ×1 + 0.205 ×1 + 0.754 ×1 = 1.245.
The power of Senator B who is a rank-and-file Republican is
λ
B
= 0.286 ×1 + 0.205 ×0 + 0.754 ×0 = 0.286.
The probability that, if these two interact, it is the Republican leader propos-
ing and B supporting is
exp(1.245)
exp(1.245) + exp(0.286)
= 0.72.
If B is a rank-and-file Democrat, then this probability rises to
exp(1.245)
exp(1.245) + exp(0)
= 0.78.
In Table 3.3 we estimate the same model with AllLeadership, a dummy
that includes all senior leaders (as in Leadership) in addition several other
categories: Conference Chairs, Party Committee Chairs, Conference Secre-
taries and Senatorial Campaign Chairs. This model fits the data as well
as the previous model (note the similar values of the Akaike Information
Criterion (AIC)). Interestingly though, general leadership status does not
confer the power that senior leadership does—notice that the coefficients for
AllLeadership and the interaction are now smaller than the commensurate
ones in Table 3.2.
Section 3.5 40
To demonstrate how American politics has become increasingly polarized,
McCarty, Poole and Rosenthal show that legislators’ ideological positions,
and the median positions for parties, have become increasingly disparate
along a liberal-conservative dimension. As politics becomes more polarized,
we might have several conflicting expectations. On the one hand, senators
who occupy the center ground—are close to the legislative median— may
find it easier to broker deals with others to their left and right, and hence
will be more powerful. On the other hand, if voting is strongly party based,
then perhaps the most ideologically extreme in each party will be able to
motivate their ‘core’ supporters into backing their preferred positions. This
will be a fortiori true for relatively radical senators from the majority party.
At base, this is one version of the ‘party versus floor median’ debate well
discussed in literature elsewhere (Cox and McCubbins, 1993; Rohde, 1991,
cf. Krehbiel, 1998).
We can measure ideological extremism via senators’ NOMINATE scores (Poole
and Rosenthal, 1997).
15
In particular, the score for the senator in one di-
mension is treated as their ‘conservatism’ (we label this ‘Conservatism’ and
the higher the score, the more conservative and less liberal they are). ‘Ex-
tremism’ is a different concept and can be ascertained by taking the absolute
value of this score (Extremism): a very high score now implies a senator very
far to the left or to the right, while a low score reflects a senator in the center
of the chamber. We add Distance from Median that records the absolute
15
In particular, DW-NOMINATE scores, available from http://voteview.com/dwnomin.htm
Section 3.5 41
value of senators’ distances from their party median in terms of NOMINATE
scores. In Table 3.4 we report the effects of conservatism (controlling for
majority party membership and distance from the median) and in Table 3.5
we report the effects of extremism.
The lessons from Tables 3.4and 3.5 are interesting. Once we control for
majority party membership, conservatism and distance from the median
have a negative impact on power. Consider, for example, Senator A who
is a moderate Republican with a NOMINATE score of 0.2 (the scale runs -1
through 1). By contrast Senator B is deeply conservative with a NOMINATE
score of 0.8. Their powers are
λ
A
= −0.591 ×0.2 + 0.844 ×1 +×−0.702 ×|0.2 −0.441| = 0.557
and
λ
B
= −0.591 ×0.8 + 0.844 ×1 +−0.702 ×|0.8 −0.441| = 0.119.
In any interaction, the probability that A proposes some coalition, while
B joins it, is around 0.61. To see more of the interplay between ideology
and power, consider Figure 3.1: here, we plot the Senator’s power estimates
against their NOMINATE scores, and then impose a solid loess curve. The
top graphic displays the plot for all senators, and the one below is the
same graphic for senators who proposed amendments in the 108th Senate.
Section 3.5 42
Notice that, in the top panel of Figure 3.1, the extremes of the Senate are
rewarded in power terms: the loess dips slightly as it crosses the middle of the
ideological spectrum. But, interestingly, in the lower panel (which represents
the most powerful senators), we see two things: first, majority party status
boosts one’s power—notice that the loess rises as it moves right. Within
the majority party (the Republicans) though, the most powerful senators
are not drawn from the far right wing: notice the high λ
i
recorded for those
with a NOMINATE score around 0.4. In Table 3.5, the positive coefficient
on Extremism suggests once again that senators towards the middle of the
chamber lack power. Interestingly, it is also evident while being in the
majority party is beneficial, it does not pay to be a right wing Republican:
rather, the powerful are from the median of the party. Figure 3.2 confirms
this idea: in the first panel, the power ratings for all senators are shown,
and in the bottom panel, the analysis is restricted to the proposers only.
Notice that the bulk of the mass occurs around 0.4 for both parties, with
the most powerful senators of both parties occurring just above and below
this scaling. Otherwise put, it does not pay to be the median of the chamber,
but it does pay to be the median of your party.
3.5.2 Committee and Agenda Control
Positive political theorists, especially those of the ‘Rochester school’ (Amadae
and Bueno de Mesquita, 1999), have suggested that it is the organization of
Congress in terms of its committees that confers power on actors. Indeed,
in a series of articles Shepsle and Weingast (1987) and Krehbiel (1991) dis-
cussed precisely why committees are powerful. Of course, not all commit-
Section 3.5 43
−0.5 0.0 0.5
0
5
1
0
1
5
2
0
NOMINATE score
P
o
w
e
r

(
λ
i
)
−0.5 0.0 0.5
1
8
1
9
2
0
2
1
2
2
NOMINATE score
P
o
w
e
r

(
λ
i
)
Figure 3.1: NOMINATE score versus power (λi); open squares are Republicans, closed
circles are Democrats; solid line is loess. Top figure is for all senators; bottom figure is
for all proposing senators.
Section 3.5 44
0.0 0.2 0.4 0.6 0.8
0
5
1
0
1
5
2
0
Extremism
P
o
w
e
r

(
λ
i
)
0.2 0.4 0.6 0.8
1
8
1
9
2
0
2
1
2
2
Extremism
P
o
w
e
r

(
λ
i
)
Figure 3.2: Extremism versus power (λi); open squares are Republicans, closed circles
are Democrats; solid line is loess. Top figure is for all senators; bottom figure is for all
proposing senators.
Section 3.5 45
tees are created equal, and they attract different memberships with different
motivations (Fenno, 1973). Nonetheless, there is general agreement that fi-
nancial committee positions—those that have ‘power of the purse’ tend to
be the most coveted spots for ambitious senators. This is not least because
such positions enable members to channel substantial pork to their home
states. For example, Charles Grassley, Chairman of the Senate’s Appropri-
ations Committee, used the 2004 appropriations bill to steer some $50m to
Iowa’s visionary complex of biomes, Earthpark (The Economist, 2006).
Committees themselves are, of course, hierarchical structures. As men-
tioned above, Grassley was the Chair of Appropriations, not simply a mem-
ber. Chairs have several de jure responsibilities and rights pertaining to
timetabling, hearings and the selection of bills to be considered. Other than
the Chair, who must be a member of the majority party, committees are
constructed of minority and majority party members. Above, we discussed
reasons why majority party senators might be powerful, and presumably
this goes mutandis mutatis for majority party committee members.
Since space is limited, we only discuss some of the committees and their
members here. In particular, in Table 3.6 we report regression coefficients
for the Appropriations, Armed Services, Commerce, Finance and Rules and
Administration committee all denoted with these names. We also include a
majority party interaction term, and a term for Chairs (of any committee)
denoted Chair. Where Table (3.6) reports positive coefficients, the Com-
mittee assignment increases the power of a senator relative to one not on
Section 3.5 46
the that committee. Majority party members get an extra boost in power
on the Armed Services, Appropriations and Finance committees, but not
on the Rules and Administration or Commerce committees. Caution is re-
quired in interpreting the findings for some committees, such as Rules and
Administration, which are notoriously weak and generally involved in less
interesting work. Indeed, senators may accept a role on such committees
as the ‘price’ for serving on more powerful committees elsewhere. Hence,
the magnitude of the coefficient maybe somewhat misleading: senators are
powerful because of their ‘good’ committee assignments which are correlated
with weak assignments, not caused by them.
16
From the perspective of a senator, the most powerful position is a role on the
Finance committee (notice that the addition of the Finance coefficient and
that for Majority×Finance is a larger number than for any other commit-
tee). Perhaps unsurprisingly, Chairs are more powerful than rank-and-file
members; in Figure 3.3 we compare chairs to Democrats and non-chairing
Republicans who propose amendments. Notice that the median power of
chairs clearly exceeds that of the other groups and, in fact, their entire
inter-quartile range is more powerful than that of Democrats.
3.5.3 Geographic Factors
One of the results of the Constitutional Convention of 1787 was the so
called ‘Connecticut Compromise.’ Dealing with the creation of the United
States’ legislative bodies, the Compromise proposed two houses: a lower
16
We are grateful to Antoine Yoshinaka for this suggestion.
Section 3.5 47
P
o
w
e
r
Dems Republicans Chairs
1
8
.
5
1
9
.
0
1
9
.
5
2
0
.
0
2
0
.
5
2
1
.
0
2
1
.
5
2
2
.
0
Figure 3.3: Boxplot showing relative power of committee chairs, non-chairing Republi-
cans and Democrats who propose amendments.
Section 3.5 48
house elected in proportion to population, and a senate, in which each state
would have two representatives, regardless of its population. The fear of
large state tyranny, though admonished as unlikely and illogical by the likes
of Madison and Hamilton, motivated small states—like Delaware, Maryland,
New Jersey and Connecticut—to seek institutional protection.
Of course, as argued with respect to the power indices approach above, the
fact that voting resources are de jure equal across states should not imply
that all senators are equally powerful. One way to examine this notion more
formally is to use the present model with geographic factors as explanatory
variables. We do not have particularly strong priors about the effect of ge-
ographic factors on a senator’s power, but some vague ideas might be as
follows.
Historically, representatives from the South were very powerful actors. Up
until the 1960s, Southern states were solidly Democratic, and possibly dis-
senting voices—from blacks and poor whites—were excluded from voting
(Key, 1949). As a result, Southern senators faced few challenges in their
home states and could use the committee seniority system—that rewarded
long service irrespective of party affiliation—to obtain powerful chairman-
ships. In the ‘post-reform’ period, this systematic concentration of power in
Southern hands was much reduced Rohde (1991). Moreover, the South is
no longer under hegemonic Democratic control as demonstrated by George
Bush with victory in every Southern state in his 2004 Presidential reelec-
tion. Nonetheless, we might still expect, for historical or other reasons, that
Section 3.5 49
senators from the South—Texas, Louisiana, Arkansas, Mississippi, Alabama,
Florida, Georgia, North Carolina, South Carolina, Virginia and Tennessee—
wield disproportionate power. We use a South dummy taking the value of
1 for senators from these states.
The expected effect of a state’s wealth on a senator’s power is arguably
ambiguous. On the one hand, senators from rich states may be able to pro-
cure greater ‘home-grown’ funding for their campaigns and causes, rendering
them more influential. On the other hand, senators from poorer places may
have more sway in Washington because they can point to underfunded pub-
lic services and crumbling infrastructures in their home states as evidence
that they have a more urgent claim to the nation’s resources. Combined
with suitably deployed rhetorical skill, we could imagine poverty may boost
a state’s representative’s powers. We measure wealth use the Census Bu-
reau’s Median Household Income statistics for each state in dollars (Median
Income).
We have similarly vague priors viz the effect of population density on power.
On the one hand, small, primarily urban, densely populated states have sen-
ators who literally represent ‘more’ citizens (we measure this with Census
Bureau’s population estimate for 2003, Population), which may aid rhetor-
ical appeals. On the other, sparsely inhabited, primarily agricultural states
may increase the power of senators who represent them, in part because,
for historical reasons, farm-based financial aid—an important component
of rural states’ federal funding—is easier to deliver than other types of sup-
Section 3.5 50
port. We use the land area of the states and denote this variable Land Area.
The significant, negative coefficients in Table 3.7 suggest that poorer,
smaller, more densely populated, non-Southern states will yield senators
with more power. Of course, this might not correspond to any particular
state. To help interpretation, in Figure 3.4, we color a map of the contigu-
ous United States according to the predicted λ
i
that a senator from that
state based on the coefficients of Table 3.7. Based only on geographical
factors, the most powerful state—in terms of its Senators—is California and
it is shaded lightest. New York and West Virginia are similarly light col-
ored. ‘Weaker’ states are dark colored. There are no particular regional
patterns discernable from Figure 3.4, except perhaps a band of states from
New York west through Missouri which appear disproportionately light (and
thus powerful) relative to their neighbors.
3.5.4 Career Factors
For most politicians, a position in the Senate is a career ambition (Brace,
1984; Rohde, 1979). But, once attained, senators have strong incentives
to seek reelection (Mayhew, 1974). As implied above, this is in part be-
cause long service is linked to promotion in terms of committee and other
assignments. Here we calculated the years of service since first entering the
Senate through 2004 and denoted this variable Service. Though not for-
mally associated with greater rights or responsibilities, longevity of service
for a particular state makes a senator the ‘senior’ representative of his con-
stituency. We use a Senior dummy to check for any extra power effect that
Section 3.5 51
Figure 3.4: Map of contiguous United States with shading proportional to ‘power’ given
by coefficients in Table 3.7: lighter states are associated with more powerful senators.
Section 3.5 52
such status confers.
From a career perspective, sociologists have long argued that being male
strengthens advantages one in the work place, and this is no less true in pol-
itics. Consider, for example, the comment and excitement drawn by Nancy
Pelosi’s ascension to the Speakership of the House in 2006.
17
For this reason,
we add a dummy variable for Male here.
The coefficients for Service and Senior status are much as we might
anticipate in Table 3.8: longer time served in the Senate, as well as senior-
ity makes for a more powerful senator. The coefficient for sex though is
perhaps not as expected. Being male is actually associated with a lower
power than being a female. In Figure 3.5 we report a boxplot for males and
females, in terms of their power. Interestingly, though the median power
of the sexes is approximately equivalent, the distributions are very differ-
ent: while male senators are counted among the most powerful, they are
also some of the weaker members of the Senate. Females, by contrast, are
heavily concentrated in the upper power ranges.
3.5.5 Summary of Findings
In summary, a senator is more powerful if the senator is:
- a member of the majority party, and has a leadership position within
17
Pelosi herself seemed well aware of her exceptionism and, in her acceptance speech
noted that “[i]t is an historic moment for the Congress, and an historic moment for the
women of this country. . . For our daughters and granddaughters, today we have broken the
marble ceiling. For our daughters and our granddaughters, the sky is the limit, anything
is possible for them.”
Section 3.5 53
P
o
w
e
r
Female Male
0
5
1
0
1
5
2
0
Figure 3.5: Boxplot showing relative power of male and female senators.
Section 3.6 54
the party;
- from the center of the ideological distribution of her party;
- a majority party member of the Finance and Appropriations commit-
tees;
- chair of a committee;
- is from a relatively poor, densely populated non-Southern state;
- is female, long serving and is the Senior Senator from her state.
3.6 Discussion
This chapter proposed a new way to measure power in structured settings.
Rather than relying on a priori metrics that arguably measure something
akin to ‘luck’ or ‘pivotality’ rather than power, we suggested an approach
that assumes the powerful have a greater capability to form coalitions in
order to pass bills that they care about. This method is straightforward to
implement, and allows the separation of ‘power’ from its ‘causes’ in a stan-
dard generalized linear model framework. Thus, we can talk of predicted
probabilities and make statements of uncertainty regarding the ‘effects’ of
certain variables. We applied the statistical model to the 108th Senate, us-
ing (personal) amendments as the bills around which Senators attempt to
form coalitions and we found that inter alia majority party membership,
moderation within a party, chairing of committees, long service records and
seniority all make for more powerful senators. These findings in and of
Section 3.6 55
themselves are perhaps not shocking, but they do demonstrate the strength
and flexibility of the approach, along with establishing some validity of the
method.
In terms of future avenues of research, there are obvious ‘cross section’ and
‘time series’ extensions to this work. Here, we choose to study the Senate:
its 100 members are a manageable number of observations for which to as-
certain biographical and other data. There is no reason why, with more
time, political scientists could not execute a similar model for the House of
Representatives. Temporally, we studied one Congress and thus the results
here are something of a ‘snapshot.’ Studying actors’ power over time may
allow a more complete picture: certainly we could imagine that changing
rules in Congress, along with the prevailing political climate, might alter
the power of say, chairs and party leaders.
Taking the model outside of the United States Congress is a possibility
too, though any such extension requires knowledge of actors’ preferences in
interactions. The United Kingdom House of Commons, for example, may
provide another test ground for the model via private members bills. Un-
fortunately, due to strong party whipping and strategic voting, deciding
actors’ preference in this circumstance is not always straightforward. There
is a similar caveat for the US Supreme Court, though in that case it is sim-
ply unclear who ‘leads’ a coalition. Outside of American and Comparative
politics, International Relations with its concentration on specifically dyadic
interactions may be amenable to such an approach. We leave this for future
Section 3.6 56
work.
Section 3.6 57
Senator Power (λi)
1 Warner, John 22.10
2 McConnell, Mitch 22.09
3 Gregg, Judd 21.84
4 Byrd, Robert 21.63
5 McCain, John 21.58
6 Frist, Bill 21.57
7 Cochran, Thad 21.44
8 Grassley, Charles 21.44
9 Boxer, Barbara 21.39
10 Graham, Lindsey 21.36
.
.
.
.
.
.
.
.
.
22 Kennedy, Edward 20.79
23 Daschle, Tom 20.79
.
.
.
.
.
.
.
.
.
29 Kerry, John 20.58
.
.
.
.
.
.
.
.
.
33 Lott, Trent 20.48
.
.
.
.
.
.
.
.
.
45 Clinton, Hillary 20.13
.
.
.
.
.
.
.
.
.
54 Feingold, Russ 19.85
57 Lieberman, Joe 19.74
.
.
.
.
.
.
.
.
.
93 Santorum, Rick -0.34
.
.
.
.
.
.
.
.
.
99 Inouye, Daniel -0.73
100 Jeffords, Jim -0.74
101 Akaka, Daniel -0.82
Table 3.1: Baseline results for model of power for US Senators in the 108th Senate.
Section 3.6 58
Estimate Std. Error
Majority 0.286
∗∗∗
0.034
SeniorLeadership 0.205
∗∗∗
0.072
Majority×SeniorLeadership 0.754
∗∗∗
0.104
Table 3.2: Effect of majority party and senior leadership status on power in the Senate.
Asterisked coefficients (***) imply p < 0.01. AIC: 11052.
Estimate Std. Error
Majority 0.304
∗∗∗
0.035
AllLeadership 0.118
∗∗
0.047
Majority×AllLeadership 0.373
∗∗∗
0.073
Table 3.3: Effect of majority party membership and any leadership status on power in
the Senate. Asterisked coefficients imply p < 0.01 (***), p < 0.05 (**). AIC: 11152.
Estimate Std. Error
Conservatism −0.591
∗∗∗
0.080
Majority 0.844
∗∗∗
0.073
Distance from Median −0.722
∗∗∗
0.122
Table 3.4: Effect of conservatism, distance from median and majority status on power
in the Senate. Asterisked coefficients imply p < 0.01 (***), p < 0.05 (**). AIC: 11169.
Section 3.6 59
Estimate Std. Error
Extremism 1.14
∗∗∗
0.107
Majority 0.783
∗∗∗
0.075
Majority×Extremism −0.9636
∗∗∗
0.168
Distance from Median −0.702
∗∗∗
0.123
Table 3.5: Effect of extremism, distance from median and majority status on power in
the Senate. Asterisked coefficients imply p < 0.01 (***), p < 0.05 (**). AIC: 11136.
Estimate Std. Error
Chair 0.596
∗∗∗
0.042
Majority −1.055
∗∗∗
0.089
Appropriations 0.245
∗∗∗
0.040
Majority×Appropriations 1.376
∗∗∗
0.100
Armed Services 0.189
∗∗∗
0.042
Majority×Armed Services 1.353
∗∗∗
0.102
Commerce 0.264
∗∗∗
0.040
Majority×Commerce −0.737
∗∗∗
0.068
Finance 0.482
∗∗∗
0.046
Majority×Finance 1.445
∗∗∗
0.104
Rules Admin 0.440
∗∗∗
0.039
Majority×Rules Admin 0.312
∗∗∗
0.068
Table 3.6: Effect of committee membership, majority party membership and chair status
on power in the Senate. Asterisked coefficients imply p < 0.01 (***), p < 0.05 (**). AIC:
9434.9.
Section 3.6 60
Estimate Std. Error
Median Income −2.199 ×10
−5
∗∗∗
2.003 ×10
−6
South −0.256
∗∗∗
3.390 ×10
−2
Land Area −1.668 ×10
−6
∗∗∗
2.135 ×10
−7
Population 2.488 ×10
−8
∗∗∗
1.723 ×10
−9
Table 3.7: Effect of incomes, southern state representation, land area (state size) and
state population on senator’s power. Asterisked coefficients imply p < 0.01 (***), p < 0.05
(**). AIC: 10954.
Estimate Std. Error
Service 0.029
∗∗∗
0.001
Senior 0.543
∗∗∗
0.029
Male −0.521
∗∗∗
0.035
Table 3.8: Effect of incomes, southern state representation, land area (state size) and
state population on senator’s power. Asterisked coefficients imply p < 0.01 (***), p < 0.05
(**). AIC: 10205.
61
Appendix to Chapter 3
3.7 The “Busy” Senator Problem
The concern is that a senator could increase his influence score by simply
proposing more amendments. That is, by being ‘busy’ in a legislative sense,
he would appear more powerful. Here we show this to be untrue.
First, consider three proposing senators A, B and C. Suppose that A and
B decide to somehow combine their efforts such that only one of them will
propose and form coalitions for all amendments jointly that they formerly
worked on separately. Whether A will have B do all the proposing and coali-
tion forming, or whether B will delegate his work to A, write the senator
who does the proposing and forming as S
AB
.
Clearly, S
AB
is busier—in that he now proposes more amendments—than ei-
ther A or B. Importantly, though, the probability that any randomly drawn
interaction involving S
AB
and C has C convincing S
AB
, will be simply the
weighted average of the former probability that C convinces A and C con-
Section 3.8 62
vinces B: there is no fillip from proposing more amendments. Hence, ‘busi-
ness’ alone cannot yield a higher power rating. To show this, let “CconS
AB

be the event that C convinces S
AB
. The probabilities for the constituent
senators are:
A
Pr =
Pr(AconC)
Pr(CconA)
and
B
Pr =
Pr(BconC)
Pr(CconB)
.
The probability for the joint, ‘busy’ senator is:
S
AB
Pr =
Pr(S
AB
conC)
Pr(CconS
AB
)
=
Pr(AconC) + Pr(BconC)
Pr(CconS
AB
)
=
Pr(CconA)
Pr(AconC)
Pr(CconA)
+ Pr(CconB)
Pr(BconC)
Pr(CconB)
Pr(CconS
AB
)
= γ
A
Pr +(1 −γ)
B
Pr .
where γ =
Pr(CconA)
Pr(CconS
AB
)
.
Thus, business cannot itself increase a senator’s influence.
3.8 Logit Model with Order Effect
We estimate the ‘order effect’ model as described in equation 3.5 and present
the results in Table 3.9. Inspection suggests that the rank order is closely
similar to that of the model without an order effect. In fact, performing a
Section 3.10 63
Spearman rank correlation test between the sorted lists yields a ρ = 0.144
with a p-value (on the null hypothesis that ρ ≤ 0) of 0.08. We estimate
δ—the advantage of proposing to someone in one’s own party—at 0.52 (and
is significant at the 1% level). That is, there is a ‘home advantage’, but it
does not disturb our earlier findings greatly when accounted for.
3.9 Amendment Proposers
Table 3.10 gives the rank ordering of senators in accordance with the number
of amendments they proposed. Clearly, simply proposing more amendments
does not make you more powerful: for example, Grassley proposes only 3
times yet is ranked in the top 10. For completeness, Figure 3.6 displays the
frequency information from Table 3.10.
3.10 Goodness-of-Fit
Recall that the method calculates a (maximum likelihood) value of λ
A
and
λ
B
for all senator pairs who interact. Above, we defined an ‘interaction’ as
an amendment in which either A proposed and B supported or B proposed
and A supported. Then, as in 3.3, the λ
i
have the interpretation that
Pr(A convinces B|A and B interact) =
exp(λ
A
)
exp(λ
A
) + exp(λ
B
)
. (3.7)
The conditioning of the probability in (3.3) implies that we need not con-
sider the counterfactual of non-proposing senators as proposers. Hence, the
matrix of votes to be predicted denoted Y
act
, is of dimension I
p
×I where
Section 3.10 64
0
5
1
0
1
5
2
0
Number of Amendments
F
r
e
q
u
e
n
c
y
1 2 3 4 5 6 7 8 9 10 15
Figure 3.6: Amendment numbers: distribution.
Section 3.10 65
|I
p
| is the number of proposing senators (74 for this case) and |I| is the
number of senators in total (here, 101, since we included President Bush
to aid interpretation). Y
act
then, is the actual interactions, and is of the
following form:
Akaka Alexander Allard . . .
Allard 1 1 0 . . .
Baucus 3 0 0 . . .
Bayh 1 0 0 . . .
Biden 4 2 1 . . .
Bingamen 9 4 2 . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
where the proposers lie to the left and all the voting senators form the
columns. So, for example, Joe Biden proposed and was backed four times
by Daniel Akaka, twice by Lamar Alexander and once by Wayne Allard.
Notice also that senators voting for their own amendments are not, in and
of themselves, counted as backers.
For any particular cell of Y
act
, there exists an associated total number
of possible interactions (in either direction) for the senators. For example,
Joe Biden and Wayne Allard interacted a total of 3 times: once when Allard
was the proposer and Biden backed him, twice when Biden was the proposer
and Allard backed him. This matrix has symmetric form when proposers
as compared to voting senators who were also proposers, but not otherwise.
Section 3.10 66
We denote this matrix Y
poss
and a similar segment to that above appears
as:
Akaka Alexander Allard . . .
Allard 1 1 0 . . .
Baucus 3 0 1 . . .
Bayh 1 0 0 . . .
Biden 4 2 2 . . .
Bingamen 9 4 3 . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
From the estimates of Table 3.1 via Equation (3.3), we can obtain predicted
probabilities of ‘convincing’ for any particular pair of senators A and B. For
example, consider the following subset of the matrix:
Akaka Alexander Allard . . .
Allard 1.00 1.00 0.50 . . .
Baucus 1.00 1.00 0.48 . . .
Bayh 1.00 1.00 0.21 . . .
Biden 1.00 1.00 0.52 . . .
Bingamen 1.00 1.00 0.74 . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
where the incidence of 1 implies that the proposer on the left is the one
convincing the voter in the column to back the amendment (rather than the
other way round) essentially with certainty. We denote this matrix O
pred
.
Since we have direct estimates of the underlying latent abilities, we do not
dichotomize this matrix to zeros and ones as may be seen in standard logis-
Section 3.10 67
tic (and probit) regression predicted probability contexts.
The expected votes matrix is the element-by-element multiplication Y
poss
×
O
pred
yielding another matrix of dimensions I
p
× I. Note that every ele-
ment of Y
poss
× O
pred
is non-zero valued. Subtracting this matrix from
Y
act
yields a matrix of both positive (the model over predicts) and negative
entries (the model under predicts). The absolute sum of these columns yields
the total number of misclassifications which is (when rounded) 2062 of some
13,869 total interactions. Hence, the percentage of interactions correctly
predicted is 1 −
2062
13,869
= 0.85.
Section 3.10 68
Senator Power (λi)
1 McConnell, Mitch 19.63
2 Warner, John 19.59
3 Gregg, Judd 19.36
4 Frist, Bill 19.10
5 McCain, John 19.09
6 Cochran, Thad 18.98
7 Byrd, Robert 18.97
8 Graham, Lindsey 18.97
9 Grassley, Charles 18.96
10 Nickels, Don 18.87
11 Boxer, Barbara 18.84
.
.
.
.
.
.
.
.
.
21 Kennedy, Edward 18.30
.
.
.
.
.
.
.
.
.
25 Daschle, Tom 18.19
.
.
.
.
.
.
.
.
.
30 Kerry, John 18.10
.
.
.
.
.
.
.
.
.
33 Lott, Trent 17.97
.
.
.
.
.
.
.
.
.
46 Clinton, Hillary 17.49
.
.
.
.
.
.
.
.
.
52 Feingold, Russ 17.30
53 Lieberman, Joe 17.25
.
.
.
.
.
.
.
.
.
94 Santorum, Rick -0.89
.
.
.
.
.
.
.
.
.
98 Akaka, Daniel -1.20
99 Jeffords, Jim -1.34
100 Inouye, Daniel -1.51
101 Johnson, Tim -1.53
Table 3.9: Baseline results for model of power for US Senators in the 108th Senate,
logit with ‘order effect’ (party advantage) estimated.
Section 3.10 69
B
y
r
d
B
i
n
g
a
m
a
n
B
o
x
e
r
D
o
d
d
F
e
i
n
s
t
e
i
n
L
a
u
t
e
n
b
e
r
g
K
e
n
n
e
d
y
1
5
1
0
1
0
9
9
9
8
M
c
C
a
i
n
D
a
s
c
h
l
e
W
a
r
n
e
r
D
o
r
g
a
n
H
a
r
k
i
n
D
u
r
b
i
n
G
r
e
g
g
8
7
7
6
6
5
5
L
e
a
h
y
M
c
C
o
n
n
e
l
l
B
i
d
e
n
B
r
e
a
u
x
C
a
n
t
w
e
l
l
C
l
i
n
t
o
n
F
r
i
s
t
5
5
4
4
4
4
4
H
u
t
c
h
i
s
o
n
L
e
v
i
n
M
i
k
u
l
s
k
i
R
e
e
d
S
p
e
c
t
e
r
B
a
u
c
u
s
C
o
c
h
r
a
n
4
4
4
4
4
3
3
C
o
n
r
a
d
C
o
r
z
i
n
e
E
d
w
a
r
d
s
F
e
i
n
g
o
l
d
G
r
a
h
a
m
(
F
L
)
G
r
a
h
a
m
(
S
C
)
G
r
a
s
s
l
e
y
3
3
3
3
3
3
3
K
e
r
r
y
K
y
l
L
i
n
c
o
l
n
N
i
c
k
l
e
s
S
c
h
u
m
e
r
T
h
o
m
a
s
B
o
n
d
3
3
3
3
3
3
2
B
u
n
n
i
n
g
D
a
y
t
o
n
H
a
t
c
h
H
o
l
l
i
n
g
s
I
n
h
o
f
e
L
a
n
d
r
i
e
u
L
i
e
b
e
r
m
a
n
2
2
2
2
2
2
2
L
o
t
t
M
u
r
r
a
y
R
e
i
d
R
o
c
k
e
f
e
l
l
e
r
S
n
o
w
e
A
l
l
a
r
d
B
a
y
h
2
2
2
2
2
1
1
B
r
o
w
n
b
a
c
k
C
a
m
p
b
e
l
l
C
a
r
p
e
r
C
o
l
l
i
n
s
D
e
W
i
n
e
E
n
s
i
g
n
E
n
z
i
1
1
1
1
1
1
1
H
a
g
e
l
K
o
h
l
M
u
r
k
o
w
s
k
i
N
e
l
s
o
n
(
F
L
)
R
o
b
e
r
t
s
S
a
r
b
a
n
e
s
S
e
s
s
i
o
n
s
1
1
1
1
1
1
1
S
m
i
t
h
S
t
a
b
e
n
o
w
V
o
i
n
o
v
i
c
h
W
y
d
e
n
1
1
1
1
T
a
b
l
e
3
.
1
0
:
N
u
m
b
e
r
o
f
a
m
e
n
d
m
e
n
t
s
p
e
r
s
e
n
a
t
o
r
.
70
Chapter 4
Rebels with a Cause?
Legislative Activity and the
Personal Vote in Britain,
1997–2005
The first duty of a member of Parliament is to do what he
thinks in his faithful and disinterested judgement is right and
necessary for the honour and safety of Great Britain. His second
duty is to his constituents, of whom he is the representative but
not the delegate. Burke’s famous declaration on this subject is
well known. It is only in the third place that his duty to party
organization or programme takes rank.
—Sir Winston Churchill, 1954
4.1 Introduction
At least since the time of Burke and his Speech to the Electors of Bristol
(1774/1975), scholars of politics have debated the proper ‘model’ for elected
representation. While much of the discussion has been normative in na-
ture, country-comparativists are able to distinguish quite different practices
Section 4.1 71
across polities. At one extreme of these stylized facts, House members in the
United States are perhaps closest to Burke’s ‘delegates’: primarily concerned
with delivering (federal spending) ‘pork’ to their constituents and, indeed,
consciously acting in line with the general ideological preferences of their
district (e.g. Fiorina, 1974; Mayhew, 1974). The nation, or even the party
with which they identify, is thought to exert a relatively weak pull on their
decision calculus (Jacobson, 1989). By contrast, akin to a ‘trustee’ concep-
tion, members of parliament in the United Kingdom House of Commons are
elected on a party ticket, are thought to have a negligible ‘personal’ vote,
and make policy according to some notion of ‘national interest’, broadly
construed. In particular, members of the governing party generally support
the executive’s bills, regardless of any constituency or personal concern.
In that they hold representatives to account, elections are identical mecha-
nisms in all systems (Riker, 1982). But contingent on the model of represen-
tation in place, the policy-making behavior for which citizens will reward or
punish their legislators should vary markedly. Analysts of American politics
have devoted considerable attention to this issue and there is a voluminous
theoretical and empirical literature that investigates the legislative determi-
nants of individual office-holders’ electoral performances (e.g. Canes-Wrone,
Brady and Cogan, 2002). Congruent with the ‘delegate’ assumption, the
findings generally suggest that more ‘maverick’ incumbents—those who vote
with their party least often—tend to do better come polling time.
Despite its Burkean vintage though, equivalent research for the United King-
Section 4.1 72
dom House of Commons—or any ‘Westminster’ system—has been much less
developed. That the study of British politics has generally neglected the fun-
damental link between constituents, their expectations of behavior and their
elected agents is unfortunate, but unsurprising. It is unfortunate because
ignoring Britain (and Westminster systems more generally) has resulted in
an asymmetrical development of political science understanding of represen-
tation and accountability. While we know much about the United States
where parties are weak, we can say little concerning polities where parties
and party voting are strong forces.
The lacuna is unsurprising: with some important exceptions (Kam, 2007;
Pattie, Fieldhouse and Johnson, 1994), scholars have assumed away the po-
tential importance of legislative activity, and looked elsewhere for personal
vote effects (see, for example, Cain, Ferejohn and Fiorina, 1987; Norton and
Wood, 1993). In any case, with over a thousand votes to consider in a given
parliamentary session, and no convenient ideal point summary available (see
Spirling and McLean, 2007), scholars have not been able to conduct system-
atic studies that analyze MP legislative behavior as a whole—that is, every
decision on every vote. It has not helped that public opinion studies point
to seemingly inconsistent preferences and contradictory expectations, with
citizens claiming that both party unity and ‘independent-mindedness’ mat-
ter (Johnson and Rosenblatt, 2007).
We seek to rectify this theoretical and empirical deficit in the current chap-
ter: using a new data set that incorporates some 2200 parliamentary votes
Section 4.2 73
cast by around 200 backbench Labour MPs between 1997 and 2005, with
commensurate constituency controls, we explain ballot box performance as
a function of personal voting decisions in the House of Commons. We show
that, all else equal, MPs that disrupt important government business are
punished by their constituency voters while those that show freedom of
thought and action on less important matters tend to be rewarded for their
independence. This effect is worth up to 2.5 percentage points at election
time. In pursuing our study we suggest a novel theoretical model of parti-
san voter behavior in Britain and new non-parametric, statistical methods
as yet unseen in political science. Our findings extend the growing literature
that emphasizes the importance of seeing MPs as individual and account-
able representatives, at least partly responsible for their own electoral fates.
We proceed as follows: in Section 4.2 we give the substantive background
for our study, and report on previous efforts in the literature. In Section
4.3, we give a simple formal expression describing the actions of (partisan,
Labour) voters in response to different actions on various roll call types at
Westminster. In Section 4.4 we describe and use a “random forest” classifi-
cation technique that allows us to include essentially every vote as a possible
independent variable, along with suitable controls, for predicting subsequent
electoral performance. In Section 4.5 we report results consistent with our
model, and show that several bills have a small but significant effect on
performance with signs as predicted. Section 4.6 concludes.
Section 4.2 74
4.2 Members, Constituents and the Personal Vote
in Britain
Party voting by the electorate, either based on national Cabinet—i.e. executive—
policy positions and performance (see Budge, 1999; Crewe, 1983; Norris,
2001) or on cruder partisan loyalties (see Butler and Stokes, 1974) is thought
to be the norm for Westminster systems like Britain. Congruent with this
principal, almost all legislative business is directly controlled by the execu-
tive, and party pressure or ‘whipping’ is commonly used by the government
to herd its ‘backbench’ MPs into supporting the Cabinet’s programmatic
agenda when required (e.g. Cowley, 2002, 3–7).
1
By conventional wisdom
then, since voters don’t (and have no reason to) care about parliamentary be-
havior, there is little an individual backbencher has to offer her constituents
aside of her party label at election time.
Interestingly though, when the public was recently asked to name the qual-
ities that an MP ought to possess, while two-fifths claim that a represen-
tative should “be loyal” to her party, some 60 percent also claim that she
should “be independent-minded” (Johnson and Rosenblatt, 2007, 165–167).
In Figure 4.1 we show the changing nature of public opinion on this subject:
clearly freedom of thought seems increasingly important, while loyalty is
falling away as a requirement. This is a fortiori true for the period in which
Labour has held large majorities of seats in government i.e. post-1997.
1
Whipping typically includes offers or threats related to ministerial promotion: see
Kam (2006) for a discussion of the Canadian case.
Section 4.2 75
1985 1990 1995 2000 2005
4
0
4
5
5
0
5
5
Year
I
m
p
o
r
t
a
n
t

C
h
a
r
a
c
t
e
r
i
s
t
i
c

(
%
)
’Independent−minded’
’Loyal to the party’
Figure 4.1: Percentage of respondents naming different qualities as “important for an
MP to have”: solid line is percentage naming ‘independent minded’, broken line is per-
centage naming ‘loyal to party’. Multiple choice, more than one choice permissible; survey
dates are 1983, 1994, 1996, 2000, 2005. Source: Johnson and Rosenblatt (2007, 166, Table
1).
On the one hand, such a diversity of views might imply that voters fail to
understand the nature of parliamentary government and the importance of
party dominance and discipline therein (e.g. Paltzelt, 2000). On the other
hand, it is not impossible to rationalize both preferences, especially for vot-
ers who lean towards the governing party: united governments are more
efficient, and able to respond rapidly to crises (see, e.g. Alesina and Drazen,
Section 4.2 76
1991; Tsebelis, 2002); they also exhibit greater longevity and thus are more
able to deliver benefits to their supporters. On the other hand, perhaps
‘maverick’ representatives are more able and willing to impose their con-
stituents’ views and concerns on the government at large when the attention
of national politics—and national parties—is elsewhere (see, e.g., Canes-
Wrone, Brady and Cogan, 2002; Pattie, Fieldhouse and Johnson, 1994), or
perhaps they avoid dangerous ‘group think’ (in the sense of Janis, 1972)
behind the scenes. But whether different voters hold different preferences,
or voters are internally divided, it is not difficult to see the potentially tricky
balancing act this presents for MPs involved in Commons roll calls.
Despite—or perhaps because of—the complicated dynamics this suggests,
previous work has tended to examine other MP-voter links. A series of
studies, for example, examine constituency service as both a goal of mem-
bers and its effects at election time (see Cain, Ferejohn and Fiorina, 1987;
Norton and Wood, 1993; Searing, 1994). Others look to local campaign-
ing activity (e.g. Denver and Hands, 1997; Pattie, Johnston and Fieldhouse,
1995), with some suggesting that, whatever the truth may be, MPs increas-
ingly believe they enjoy a personal vote dependent on their appearances,
statements and legislative behavior (Cowley, 2005).
When scholars have examined parliamentary votes as a predictor of per-
formance, they have tended to be limited in scope: in the case of Pattie,
Fieldhouse and Johnson (1994), they deal with just six ‘key’ votes with var-
ious levels of attendant whipping, while Kam (2007) uses ‘dissent’ on all
Section 4.3 77
whipped bills (in the aggregate) to predict voter familiarity with their MP’s
name.
2
These strategies are not unreasonable given the specific questions
that were to be addressed, and the studies are important in that they are
rare breaks with the orthodox thinking on voting in Britain. Nonetheless, if
we could be agnostic about the potential bills of interest, and thus include
all votes as possible predictors, we would garner the best of both investi-
gations: we would have a survey of parliamentary behavior that is deeper
in terms of both the number of votes in the sample and broader in terms
of their content and the pressure applied in each case. We would thus cap-
ture the totality of legislative behavior, rather than a potentially misleading
small sample. Before doing so, we need to be clear about what voters desire
and expect from their representatives. We outline our intuitions in the next
section.
4.3 Formalizing Intuitions
As with other scholars in the field, our work here will concentrate on the
behavior of government (i.e. Labour) party MPs: the ‘power’ of the gov-
ernment relative to other parties in a Westminster system means that these
representatives receive more media attention, and are more likely to be held
accountable for the state of the polity at election time. Thus these are the
members for whom the stakes are highest—for whom party unity and/or
independence matter most.
2
Though for New Zealand, rather than Britain, Kam (2007, 135) does show that dissent
is an effective vote-winning strategy.
Section 4.3 78
Our intuition is that on almost all issues (whipped or unwhipped) debated
in the House of Commons, voters are uninformed. The bills and proposals
are too technical and it is too time-consuming to review and understand
their contents. Thus, voters rely on heuristics. For the reasons cited above,
party unity is prized on issues that are most important to the government’s
legislative agenda. Notice that this effect will be strongest for bills which
exhibit “government-versus-opposition” voting: that is, on roll calls where
the non-government parties attempt to foil the executive for the sake of op-
position, rather than because they actually have some different and sincere
policy preference (see Spirling and McLean, 2007). The reason is that these
votes most starkly test the Cabinet’s ability to govern; indeed losing on such
a roll call would typically trigger a vote of confidence.
3
By contrast, when votes are relatively free of party pressure, precisely be-
cause they do not endanger a government agenda, maverick, independent-
minded voting is welcomed. Indeed here, ‘dissent’ from the MP’s party as
a whole is a ‘good thing.’ However, on the rare occasion that free votes
arise whose contents and consequences are simple to grasp for the public at
large, voters reward MPs on a spatial basis. That is, left-leaning partisan
voters want their MPs to plump for the more ‘socialist’ option while right
partisans prefer the more ‘conservative’ choice.
3
Such votes dangerously undermine Prime Ministerial authority: John Major called
such a vote in 1993 in response to sustained rebellion over a series of European inte-
gration votes. Major was compelled to resign his position as party leader—though he
subsequently won reelection—in response to the continual sniping. Fifteen years earlier,
James Callaghan was forced to call a general election (which he subsequently lost) after
suffering defeat in Commons in March 1979.
Section 4.3 79
More formally, suppose there are 1, . . . , i, . . . N voters in constituency j,
for which MP
j
is the incumbent governing party member of parliament. For
any particular bill (or proposal, or amendment) b that is to be voted upon,
there are two possible responses by MP
j
—‘aye’ or ‘no’. We will write the
choice of MP
j
as a point o
j
in p-dimensional space: o
j
∈ R
p
. For a bill that
is essential to government business, denote the Cabinet’s preferred outcome
as o
g
. There are q
1
such bills of a total of B that come to be voted upon in
House of Commons. Since the government essentially never loses on parlia-
mentary business it cares about, in a slight abuse of notation, we will use
o
g
to also refer to the outcome voted for by the majority of the governing
party. This extends to the B −q
1
bills that are not essential to government
business. Finally, denote the median voter of constituency j as having an
outcome preference on any particular roll call as o
i
∈ R
p
. Notice that we
assume she voted for the incumbent at the previous election. We can now
write her utility from MP
j
’s voting record in the Commons as
U
i
(·) = −w
1
whipped
¸ .. ¸
_
q
1

b=1
d(o
j
, o
g
)
_
+w
2
unwhipped
¸ .. ¸
_
_
q
2

q
1
+1
d(o
j
, o
g
)
_
_
. ¸¸ .
voter uninformed
−(1 −w
1
−w
2
)
unwhipped
¸ .. ¸
_
_
B

q
2
+1
d(o
j
, o
i
)
_
_
. ¸¸ .
voter informed
(4.1)
where 0 ≤ w
1
, w
2
≤ 1 refer to the ‘weights’ that voter i places on each
legislative activity component and d(·, ·) is the Euclidean distance between
Section 4.3 80
two points in p dimensions.
4
With no loss of generality we assume that voter
i compares the utility from (4.1) with some constant c—say, zero—and casts
her ballot for MP
j
at the subsequent election if U
i
(·) ≥ c, else she abstains
or votes for a competitor. Several simple comparative statics, consistent
with our reasoning above, follow:
• for (heavily) whipped government business in Equation (4.1), so long
as w
1
> 0,
∂U
i
(·)
∂d(o
j
,og)
< 0. In words: for bills that deal with impor-
tant government business, (partisan, government party) voters will
disapprove of fractious, rebellious MP behavior. Moreover, the fur-
ther away the MP’s vote (o
j
) places him relative to the government
(o
g
), the worse he does.
• for unwhipped roll calls on which voters are uninformed, so long as
w
2
> 0,
∂U
i
(·)
∂d(o
j
,og)
> 0. In words: on these bills, voters reward ‘in-
dependent thinking’ insofar as it manifests itself as choosing options
different (far from) the party majority option.
• for unwhipped roll calls for which voters are informed, so long as w
1
+
w
2
< 1,
∂U
i
(·)
∂d(o
j
,o
i
)
< 0. That is, the voter’s utility is decreasing the
further the MP’s choice is from her ideal (outcome) point.
For reasons that we will discuss below, it is not possible to ‘test’ this theory
directly. Nonetheless, this model will give us a structured way to think about
our findings. Before leaving the theoretical discussion here, it is helpful to
4
Recall that the Euclidean distance between two points X = (x
1
, x
2
, . . . , x
n
) and Y =
(y
1
, y
2
, . . . , y
n
) is simply

n
i=1
(x
i
−y
i
)
2
. It is thus always positive, and caters for points
in any number of dimensions in space.
Section 4.4 81
clarify three issues. First, this is an as if model: we are not contending that
most voters directly examine voting records to ascertain their representa-
tive’s performance. Rather, they receive media and elite reports summariz-
ing these patterns. What matters then is the types of MPs and the types of
voting behavior in a broad sense. Second, there is a potential contradiction
in the voter calculus: punishing a rebellious MP also hurts the party by po-
tentially denying them an admittedly unreliable seat in the Commons. We
do not doubt that such reasoning occurs, but we will assume that at some
point an MP’s actions are sufficient to merit voting against them, even if it
superficially damages the national party in the short term.
5
Third, to the
extent that a time trend exists, Figure 4.1 implies that independent thought
ought to be increasingly rewarded, in addition to (or perhaps in place of)
loyalty. Otherwise put: (relatively) free votes should be better predictors of
performance in 2005 than in 2001.
4.4 Data and Method
Ideally, we would test our predictions from Section 4.3 with constituency
level data recording citizens’ preferences for their MP’s specific legislative
actions, in addition to partisan, government party voters’ views on the desir-
ability of party unity and favored positions on various free-vote issues. We
would compare these predilections to their representative’s actual record
5
As a case in point, consider the fate of Neil Hamilton, Conservative MP for Tatton in
Cheshire. Accused by The Guardian newspaper of taking bribes to ask questions in the
House of Commons, Hamilton proclaimed his innocence, sued for libel and went on to fight
for his seat in the 1997 General Election, the then 5th safest Tory district in the country
with a previous majority of 16,000 votes (27.6 percent). Hamilton lost to an independent
challenger, by some 11,000 votes on a 18 percent swing away.
Section 4.4 82
and calculate the effect that this divergence had on the MP’s subsequent
electoral performance. No such data exist so, instead, we will try to infer
micro-behavior from aggregate level findings.
In the Americanist context, scholars have summarized Congress members’
legislative behavior using ideal point estimation, and then contrasted this
with metrics purportedly capturing the underlying political orientation of
citizens in their home district. Thus ‘divergence’ and the degree to which a
member is ‘out-of-step’ with his constituents may be assessed (e.g. Canes-
Wrone, Brady and Cogan, 2002). But such a strategy is fraught with dif-
ficulty for the United Kingdom case: the sheer preponderance of strategic,
‘government-versus-opposition’ voting in the House of Commons makes in-
terpreting industry-standard tools for summary very difficult (Spirling and
McLean, 2007).
The analysis of British parliamentary voting in the context of our theory
is hampered by two further considerations. First, it is not publicly obvious
how much whipping is applied on individual roll calls. While parties cer-
tainly acknowledge that strong pressure is present at Westminster, they do
not typically publish details of its intensity and, in any case, whipping may
take the form of informal threats and offers.
6
Moreover, the fact that a roll
call is ostensibly ‘free’ is no guarantee that it is completely undirected.
7
So,
6
In practice, party whips contact party members with a schedule concerning upcom-
ing divisions. They literally underline the requirement that members attend and vote
appropriately according to the importance the party places on the roll call in question.
7
See McLean, Spirling and Russell (2003) for a specific incidence of this phenomenon.
Section 4.4 83
though we will use terms ‘whipped’ and ‘unwhipped’ below, they are meant
only as broad-brush shorthand for the general environment in which the bill
is being considered: this is especially true when the parties differ in their
application of their whip. The second problem is that, within the parlia-
mentary record, Hansard, missingness is endemic and unexplained. Since
there is no way to formerly record absences, there is no way of knowing if a
missed vote was due to indifference, principled abstention, illness, ministerial
business, etc.
4.4.1 Data: 1997–2001, 2001–2005
For each data set, our dependent variable is the difference in electoral suc-
cess (vote percentage in their home constituency) for each of the MPs in
the study between the respective elections. As noted above, we restrict the
analysis to government party—i.e. Labour party—MPs for whom we think
the theory predicts the strongest effect. Further, we deal solely with MPs
who hold seats in England and who did not hold ministerial office at any
time during the parliamentary terms. The location restriction avoids in-
cluding MPs whose electoral fates may be the product of nationalist party
dynamics that are difficult to control for. The next avoids the potential en-
dogeneity problems inherent in ‘good behavior’, front-bench promotion and
constituency popularity. Simply put, MPs are given ministerial responsibil-
ities contingent on toeing the party line,
8
but that same status makes them
much more recognizable in their home districts, increasing the possibility
8
See Benedetto and Hix (2007) for a discussion of this issue and its consequences for
discipline.
Section 4.4 84
their subsequent vote share is influenced by factors other than their roll call
record. For obvious reasons, we also drop MPs who choose not to fight
the subsequent election, or who joined the parliament mid-way through via
by-elections (and thus for whom we are missing some dependent variable
information). These data cuts leave some 225 backbench Labour MPs for
1997–2001, and 194 for 2001–2005.
Our approach is to predict, i.e. to regress, the vote shares on every single
roll call vote the MP casts in the relevant parliament, data which was com-
piled from (online) Hansard reports.
9
This way, we know that the totality
of MP legislative behavior is being captured. We have also collected numer-
ous (standard) controls which include the difference in electoral spending
by Labour between the elections (a reasonable proxy for campaign ‘effort’—
see Pattie, Johnston and Fieldhouse (1995)), the difference in turnout, and
the proportion of constituency residents falling into several traditionally
marginalized groups that might be disproportionately sympathetic towards
a Labour government.
10
4.4.2 Estimation Problem: p > n
The dependent variable of interest is (essentially) continuous. Hence, ordi-
nary least squares (OLS) might seem attractive as an estimation procedure.
9
See Firth and Spirling (2006) for a discussion of this data. We impose the requirement
that at least 50 percent of all the MPs in the sample voted on that roll call for it to enter
the data. This avoids including divisions where the majority of MPs were missing.
10
Compiled from the Census 2001 Report for Parliamentary Constituencies. In particu-
lar, the percentage of pensioners, of non-whites, of individuals living with a limiting long
term illness (which includes disability), of unemployed, and the proportion of households
headed by a lone parent.
Section 4.4 85
Unfortunately, on the ‘right hand side’ there are around 1300 variables—
predominantly parliamentary divisions—which we wish to use as regressors
(or ‘predictors’). That is, the number of parameters to be estimated, p far
exceeds the number of observations, n: we have negative ‘degrees of free-
dom’, p > n. The consequence is that the standard OLS solution,
β = (x

x)
−1
x

y, (4.2)
is unavailable; moreover, the equivalent of Equation (4.2) for essentially any
parametric model—discussed in more detail below—is elusive.
One unsatisfactory solution is to re-specify the regression problem with fewer
variables, but this requires an entirely arbitrary a priori restriction on the
model. A better option is to employ systematic exploration methods that
uncover structure and reduce the data down to its ‘important’ component
variables. Moreover, because our theory is somewhat vague, we would like
a flexible approach that assumes as little as possible about the precise func-
tional form of the relationship between the y and x we alluded to above.
4.4.3 Tree-based Regression and Random Forests
OLS is a special case of ‘regression’ which we can write more generally
as y = m(x) + where m() describes an unknown model for the data we
observe. Depending on the assumptions we make about m()—in particu-
lar, what we believe we ‘know’ about the function that it represents—we
commit to a parametric or nonparametric approach. In parametric models,
Section 4.4 86
the functional form of the relationship between the dependent and indepen-
dent variables is assumed known, but it contains parameters whose values
are unknown and to be estimated. Political scientists are most familiar
with generalized linear models (GLM) where y is linearly associated with
x profiles but in which the functional relationship between them can be
non-linear.
11
Examples include logit or Poisson regression. When analysts
have a strong grasp of the underlying physical causal process (and thus,
for example, the potentially subtle ways that variables interact), paramet-
ric models are a good choice. They are of limited use, however, when the
data are not well understood, or theory is not easily forthcoming—especially
when the number of independent variables is large and thus the causal story
complex. This is precisely because the assumptions that make parametric
models straightforward to fit also restrict the variety of relationships that
they can adequately describe: and if the model is ‘incorrect’ for the data
(e.g. linearity is assumed when non-linear relationships are more apt), re-
sults will be less than satisfactory. Notice that the goodness of fit criteria
may be misleading in these circumstances. In any case, for problems in
which we have p > n, standard parametric models cannot be fit to the data.
In contrast to parametric models, nonparametric models make very few
assumptions about the functional form between the dependent and inde-
pendent variables. While parameters of the model are estimated, they do
not necessarily have a clear interpretation in terms of the physical causal
11
That is, E(y|x) = G(βx) where E() is the expectations operator, G() is some function
and β are the parameters to be estimated.
Section 4.4 87
process thought to underlie the data: instead the concern is with function
estimation and good prediction.
A simple nonparametric method, popular in biostatistics, is that of regres-
sion trees (Ripley, 1996, 213). Within this literature, the task of optimizing
predictive performance is performed by classifying cases according to their
independent variable values via a series of logical splits. Typically, at each
split, or ‘node’, the algorithms divide one independent variable into (bi-
nary) disjoint subsets into which the cases can be placed. Further splits
are then proposed for any cases that are not in terminal nodes (known as
‘leaves’ of the tree). There are lots of different rules that may be used to
‘cut’ the independent variables (Ripley, 1996, 216–220), and hence several
different ways to assess how well the tree classifies the problem in question.
The general idea is to split the data so that cases within the subsets are
as like each other as possible; in regression contexts, this literally means
choosing to split a ‘parent’ node such that the resulting two ‘child’ nodes
have the minimum possible within node variance in terms of their values of
the dependent variable.
12
By seeing which variables—when split—best pre-
dict the cases, we obtain some notion of ‘importance’ in a regression context.
To see how this process might work, consider the following toy example:
we have data on the 2004 Presidential election, in terms of George Bush’s
percentage of the vote in any given state. We also have ‘explanatory’ vari-
ables for each state, including its population density, its median household
12
More details may be found in Appendix 4.7.
Section 4.4 88
income, and the proportion of its population that are black. In Figure 4.2
we report a tree grown for this data using the split criterion defined above.
13
At the top of the tree, we see the first split was on population density, with
those having a value of lower than 206.8 sent ‘to the left’ and those that
are more dense than this cutoff sent ‘to the right.’ On the left, these rela-
tively rural states were once again split on population, with the least dense
(< 35.2) sent to the left. The next split for those states occurs in terms
of their black population. If it was less than 1.5 percent (0.015 to the left
of the figure), they were dropped in the bin of states with a mean value of
65 percent for Bush. Literally, these are large, rural, predominantly white
states such as Utah and Wyoming. Each leaf (end value) in the figure corre-
sponds with a different bin, and states are placed there contingent on their
variable values.
Several comments are in order here. First, the variables are not all of the
same ‘importance’: different variables at different nodes do better in re-
ducing the variation in the subsequent (split) data sets than others (here,
population density does particularly well). Second, there is no guarantee
that ‘important’ variables will be statistically significant in a regression con-
text or vice versa: classification and/or tree-regression are not synonymous
with OLS. Third, we made a judgement call as to when the tree would stop
growing. Clearly, there are 50 states, but only 8 bins as terminal nodes—
65.00, 57.17, 46.00, 51.00, 59.78, 54.60, 48.20, 42.29. In this case, the tree
13
Implemented in the R (R Development Core Team, 2006) statistical language and
environment using the tree package: see Ripley (1996).
Section 4.4 89
|
popn.density < 206.848
popn.density < 35.2142
black.popn < 0.015 black.popn < 0.065
black.popn < 0.025 median.income < 40418.5
popn.density < 361.186
65.00 57.17
46.00 51.00
59.78 54.60
48.20 42.29
Figure 4.2: Example of a classification tree for American states, predicting (categorizing)
the percentage of vote for George Bush in the 2004 Presidential election.
Section 4.4 90
|
Figure 4.3: Example of a tree that uniquely classifies American states, in terms the
percentage of vote for George Bush in the 2004 Presidential election. Node labels removed
for clarity of presentation.
is grown until no more than 10 cases reside in a terminal node. In this toy
instance, we could have, in fact, insisted on a perfect fit to the data: 50
terminal nodes with one for each state. Such a tree is given—without the
distraction of node labels—in Figure 4.3. Its complexity should be obvious.
For most purposes, fitting a large data set with lots of independent variables
in a way that resembles Figure 4.3 won’t be helpful: we will ‘overfit.’ That
Section 4.4 91
is, we will essentially reproduce the complexity of the data by fitting its
random (‘noisy’) features which are not part of the causal process in which
we are interested. Such trees have high variance, in the sense that altering
the data on which they were grown (by, perhaps, adding new points out
of the sample) would result in very different findings. This is undesirable
given that we want to know about the general structure of the data at hand.
To reduce this variance, we can ‘prune’ the tree back after it is grown, and
there are various methods for doing so (Ripley, 1996, 221–231). Alterna-
tively, we can ‘stop’ growing at some pre-defined point. But less complex
trees introduce a new problem of bias: simpler models may not be good
enough to correctly represent the subtleties of the data we are interested
in. Though we can make the selection of ‘right-sized’ trees automatic using
various cross-validation procedures, there may be a better way to proceed.
One solution is to grow multiple trees—each potentially classifying the cases
differently—and find some way to aggregate the results from all of them. A
popular and powerful way to do this is Breiman’s (2001) random forests
algorithm.
14
The key idea is that each tree will be unpruned and thus have
low bias. And, while the individual trees will have high variance, for reasons
noted below, there will be a low correlation between them. Combining the
trees will give the best of both worlds: low bias and low variance.
Random forests actually builds on an earlier technique—also proposed by
14
Implemented in R by Liaw and Wiener (2002).
Section 4.4 92
Breiman (1996)—called ‘bagging’.
15
Bagging is an abbreviation of bootstrap
aggregating and, like most tree methods, it begins by splitting the data into
a ‘training’ and ‘test’ set. In random forests, we do not need to make this
split. We proceed as follows:
1. take a bootstrap sample from the data, D. In practice, this means
taking a sample of size N—with replacement—from D (see Efron and
Tibshirani, 1993, for a general discussion). Denote this sample as D
k
and notice that the same observations can therefore appear multiple
times in D
k
, and others from D might not appear at all. In fact,
approximately one third of the cases will not appear in the sample
(called ‘out of bag’).
2. train the predictor using D
k
. That is, grow one tree from that sample
until the cases reside in sufficiently small nodes (commonly, five cases).
These trees are not pruned.
3. repeat steps 1 and 2 1, . . . , b, . . . B (say, 500) times.
4. aggregate the results of the B trees. In a regression context, we average
across the trees.
Random forests get their name from the randomness employed at each node
in each tree: rather than employing the best split among all variables to con-
tinue to grow the tree, a random forest splits a randomly chosen subset of all
the predictors; in practice this subset is approximately one-third of the total
number of independent variables (around 350 here). This strategy results in
15
See Ruger et al. (2004) for a political science application of an earlier contribution by
Breiman et al. (1984).
Section 4.4 93
low correlation between the trees, and yet the predictive performance gener-
ally exceeds that of related techniques—such as neural nets (see Beck, King
and Zeng, 2000; de Marchi, Gelpi and Grynaviski, 2004)—without overfit-
ting. A schematic diagram describing random forests is given in Figure 4.4.
We will be interested in several quantities from our models: first, the good-
ness of fit, which is calculated in a similar way to the R
2
for a standard
regression with the difference that we explicitly attempt to predict data
that was not involved in fitting the trees. Second, we want to know how
‘important’ different variables are; to do this, we will inspect a residual sum
of squares measure for trees where the variable is involved, relative to those
where it is not. Third, the marginal effect of any given variable on y will be
a concern. Broadly, we obtain this by comparing the predicted y when the
variable is in the (aggregate) random forest model, to when it is not. Ap-
pendix 4.8 gives more technical details. Notice that because random forests
are not attempting to estimate regression coefficients in the usual, GLM,
way there is no requirement that n > p.
If random forest classification is successful, we expect observations which
are ‘alike’ to end up in the same terminal nodes across trees. We can use
this fact to impute missing observations, and Appendix 4.9 explains how.
4.4.4 Estimation Plan
Given the above explication, our investigation now proceeds in several steps,
which combine both quantitative and qualitative methods:
Section 4.4 94
Data D
1 2 3 4 5
6 7 8 9 10
4 10 6
4 9 4 1
7 3 10
6 2 3
10 9 8 8
2 4 6
10 2 5
7 5 8 3
7 9 7
6 8 1
2 5 5 10
7 8 7
2 5 8 1 5 7 1 4 8 3 4 9 ’out−of−bag’
’in bag’
tree 1 tree 2 tree B−1 tree B
bag 1 bag 2 bag B − 1 bag B
Figure 4.4: Random forest schematic. Here the data set consists of ten observations
labeled 1 through 10. These are ‘bootstrap sampled’ into bags (1 thorough B). For each
bag, some observations are ‘left out.’ A tree is grown from each bag, and the results
aggregated.
Section 4.5 95
1. impute values for the missingness in the divisions using a random forest
classification
2. use imputed data and random forest to calculate relative ‘importance’
of different controls and roll calls in predicting electoral success.
3. using 2 as a guide, obtain and report the (random forest) marginal
effect of some important variables (roll calls)
4. use narrative histories (such as Cowley (2002) and Cowley (2005)) to
establish nature of bills—whipped/unwhipped, etc.
As a further comparison we will also report OLS coefficients for the variables
from 2 (using each variable with all controls to predict performance). To
reiterate, we will use data classification techniques to find the most ‘impor-
tant’ variables (which will involve a judgement call on our part) and then
see if and how they fit our theory. Because the method fits by maximizing
prediction, concern will focus on the types of roll calls that matter (‘heavily
whipped’, ‘free’, ‘issue of conscience’, ‘programmatic concern’, etc.), rather
than their precise content. In other words, we will identify those bills that
best predict the dependent variable in question and then describe their broad
characteristics in terms of our interest in party discipline versus indepen-
dence. Because essentially all votes are used, there is thus no danger that
we are looking in the ‘wrong place’ when predicting electoral performance
from legislative behavior.
Section 4.5 96
4.5 Results
The results are in two parts, first for 1997–2001 and then, separately, for
2001–2005. Of course, for both data sets, we follow the stages set out in the
previous section.
4.5.1 First term: 1997–2001
For our first data set, recall that the dependent variable is the difference
in vote share that the backbench Labour MP’s garnered in their home con-
stituency for the 1997 versus 2001 election. In fact, both events were land-
slides for the Labour party as a whole, and Labour lost just six seats (of
a 179 seat majority) in the latter. Nonetheless, there was quite significant
variation in terms of the performance differences of individual MPs with a
mean change of −1.59 percent (variance of 14.77), and a maximum loss of
-12.5 percent recorded by Khabra Piara (Ealing Southhall), and a maximum
gain of 12.7 percent recorded by Michael Jabez Foster (Hastings and Rye).
In Figure 4.5, we graph the most important 30 variables from the random
forest regression where, x-axis scale refers to the decrease in the average
residual sum of squares when the variable is in the tree, as opposed to when
it is left out. The (mean) pseudo-R
2
goodness of fit measure for the model
was around 0.06.
Note first that the controls are seemingly essential and all contribute more
than any roll call in accounting for differences in the dependent variable. In
particular, the percentage of citizens in a district with a debilitating illness
Section 4.5 97
div302.001023
div063.990210
div071.000210
div134.010227
div191.990520
div290.000725
div243.990714
div078.971103
div079.990301
div311.980622
div046.990125
div300.001023
div299.001023
div294.980608
div012.001219
div111.971210
div186.000509
div303.001023
div307.001023
div281.991102
div312.980622
div162.010321
div323.001107
Unemploy
LoneParent
Pension
Spend
Ethnic
DiffT
Illness
0 100 200 300 400
IncNodePurity

i
m
p
o
r
t
a
n
t


r
o
l
l

c
a
l
l
s
c
o
n
t
r
o
l
s
Figure 4.5: Variable importance plot for 1997–2001 data. Controls and ‘important’ roll
calls used in subsequent analysis shown as subsets of all variables. x-axis gives the increase
in the regression performance from including each variable.
Section 4.5 98
(Illness in the graphic) is most important, followed by turnout differential
(DiffT), the proportion of non-whites (Ethnic), the proportion of Pension-
ers (Pension), the proportion of lone parent households (LoneParent), the
proportion of citizens unemployed (Unemploy) and then the electoral spend-
ing differential (Spend). The following points on the graphic refer to the
importance of the roll calls and have labels beginning with div. Deciding
which of these bill votes are worthy of further inspection is essentially a
judgement call, and here we consider the top 9, below which there is a slight
drop off in importance.
In Table 4.1 we report our results: for each variable, we give the marginal
effect from the random forest: literally, this is the difference in the predicted
constituency performance when an MP votes with the minority of his party
as opposed to when he votes with the majority. We also report an OLS coef-
ficient (with p-value) from including that predictor in a regression—with all
controls. In both cases, a positive number means the MP did better at elec-
tion time while a negative one means he was worse off. Since the controls are
not the focus of the study, they are left out of the table. As noted above, it
can be quite difficult to know for sure how much party pressure was applied
in the various votes, but column 3 gives a casual author-based coding of
the discipline imposed for that vote, ascertained from a qualitative reading
of the relevant House of Commons debate and surrounding literature; the
whipping is one of four ordinal levels: ‘free’ (essentially no whipping), ‘weak’
(little whipping, perhaps timetabling concerns), ‘moderate’ (some pressure,
government has preference for outcome), ‘strong’ (tight whipping, govern-
Section 4.5 99
ment has strong, programmatic concerns).
Beginning first with bills which Cowley (2002) discusses as central to the
Government’s post-1997 legislative agenda, we note that the signs are in the
expected direction. That is, rebeling on the Transport Bill (row 7) leads
to MPs being punished at election time—i.e. a negative coefficient—that is
statistically significant. This roll call dealt with the Government’s attempts
to privatize the National Air Traffic System. The rebels on this roll call
were backing an (non-Government) amendment that would have substan-
tially delayed the denationalization.
The vote on the Crime and Disorder Bill (rows 3) was concerned with lower-
ing the age of homosexual consent from 18 to 16, such that it would be on a
parity with heterosexual consent. They were ostensibly unwhipped, yet this
belies some potential party pressure on MPs. For this bill, voting for an
amendment to enforce a minimum age of at least 18 in certain circumstances
was punished by the electorate. Within the realms of the model we advanced
above this suggests two possibilities: first, that voters were sympathetic to
such social equality issues, and the MPs that vote for them. For gay rights,
such preferences seem to be the case: some 93% of British respondents in
a recent survey supported laws to protect gay people from discrimination
(Cowan, 2007), while the legalization of same-sex civil partnerships appears
to have occasioned much less heated debate in the UK than their potential
introduction in the US. Certainly, those taking the more conservative posi-
tion here did worse. Second, note that though the government was formally
Section 4.5 100
C
o
d
e
N
a
m
e
D
i
s
c
i
p
l
i
n
e

R
e
b
e
l

E

e
c
t
(
R
F
)

R
e
b
e
l

E

e
c
t
(
O
L
S
)
p
-
v
a
l
u
e
d
i
v
3
2
3
.
0
0
1
1
0
7
P
r
o
g
r
a
m
m
i
n
g
o
f
B
i
l
l
s
w
e
a
k
+
0
.
1
7
9
+
1
.
1
4
4
0
.
0
3
8
d
i
v
1
6
2
.
0
1
0
3
2
1
W
e
i
g
h
t
s
a
n
d
M
e
a
s
u
r
e
s
(
S
.
I
.
)
w
e
a
k
+
0
.
2
9
1
+
0
.
8
2
7
0
.
2
5
6
d
i
v
3
1
2
.
9
8
0
6
2
2
C
r
i
m
e
a
n
d
D
i
s
o
r
d
e
r
B
i
l
l
f
r
e
e
-
0
.
0
1
1
-
0
.
3
9
1
0
.
4
6
0
d
i
v
2
8
1
.
9
9
1
1
0
2
C
i
t
y
o
f
L
o
n
d
o
n
(
W
a
r
d
E
l
e
c
t
i
o
n
s
)
B
i
l
l
m
o
d
e
r
a
t
e
-
0
.
1
7
6
-
0
.
8
4
5
0
.
0
9
2
d
i
v
3
0
7
.
0
0
1
0
2
3
E
l
e
c
t
i
o
n
o
f
S
p
e
a
k
e
r
f
r
e
e
+
0
.
2
1
3
+
1
.
7
4
3
0
.
0
0
6
d
i
v
3
0
3
.
0
0
1
0
2
3
E
l
e
c
t
i
o
n
o
f
S
p
e
a
k
e
r
f
r
e
e
+
0
.
1
4
2
+
1
.
1
6
5
0
.
0
7
5
d
i
v
1
8
6
.
0
0
0
5
0
9
T
r
a
n
s
p
o
r
t
B
i
l
l
s
t
r
o
n
g
-
0
.
2
7
1
-
1
.
2
8
7
0
.
0
3
1
d
i
v
1
1
1
.
9
7
1
2
1
0
D
o
c
t
o
r
A
s
s
i
s
t
e
d
D
y
i
n
g
f
r
e
e
-
0
.
1
4
7
-
0
.
4
7
2
0
.
3
6
0
d
i
v
0
1
2
.
0
0
1
2
1
9
H
u
m
a
n
F
e
r
t
i
l
i
s
a
t
i
o
n
a
n
d
E
m
b
r
y
o
l
o
g
y
f
r
e
e
+
0
.
1
7
1
+
0
.
8
2
2
0
.
1
3
2
T
a
b
l
e
4
.
1
:

M
o
s
t
i
m
p
o
r
t
a
n
t

p
a
r
l
i
a
m
e
n
t
a
r
y
v
o
t
e
s
f
o
r
p
r
e
d
i
c
t
i
n
g
e
l
e
c
t
o
r
a
l
p
e
r
f
o
r
m
a
n
c
e
b
y
b
a
c
k
b
e
n
c
h
L
a
b
o
u
r
M
P
s
i
n
2
0
0
1
g
e
n
e
r
a
l
e
l
e
c
t
i
o
n
.
F
i
r
s
t
c
o
l
u
m
n
i
s
d
a
t
a
c
o
d
e
f
o
r
t
h
e
d
i
v
i
s
i
o
n
w
i
t
h
d
a
t
e
a
f
t
e
r
p
o
i
n
t
:
s
o
,
d
i
v
3
2
3
.
0
0
1
1
0
7
i
s
d
i
v
i
s
i
o
n
3
2
3
t
a
k
i
n
g
p
l
a
c
e
o
n
N
o
v
e
m
b
e
r
7
,
2
0
0
0
.
N
a
m
e
o
f
b
i
l
l
i
n
c
o
l
u
m
n
2
;
r
a
n
d
o
m
f
o
r
e
s
t

c
o
e

c
i
e
n
t

i
n
c
o
l
u
m
n
3
f
o
r
b
e
i
n
g
a

r
e
b
e
l

(
i
n
t
h
e
m
i
n
o
r
i
t
y
a
n
d
/
o
r
f
r
u
s
t
r
a
t
i
n
g
t
h
e
G
o
v
e
r
n
m
e
n
t

s
a
m
b
i
t
i
o
n
s
)
;
O
L
S
c
o
e

c
i
e
n
t
i
n
c
o
l
u
m
n
6
;
p
-
v
a
l
u
e
f
o
r
t
h
a
t
c
o
e

c
i
e
n
t
i
n
c
o
l
u
m
n
7
.
Section 4.5 101
uncommitted to the amendment, it allowed it time for consideration and,
when the House of Lords rejected it, the government reintroduced it via the
separate Sexual Offences (Amendment) Bill. Indeed, when the Lords once
again refused its assent, the government threatened to enact it directly over
the wishes of the upper chamber. In short then, the executive was clearly
committed and sympathetic to this issue—suggesting that rebels might be
considered at least ‘out of step’ if not ‘troublemakers’ per se.
Votes that concerned the election of a new House of Commons Speaker
occupy rows 5 and 6 of the table. Unlike, say, the United States House of
Representatives, the Speaker of the House of Commons is a non-partisan
referee of parliamentary debate. Selected from back-bench MPs, Speaker
candidates do not usually receive official endorsements though MPs from
the same party often vote similarly in the elections. Ultimately then, these
are free votes, and the whip is not present. Moreover, these are votes for
which voters are uninformed: the candidates are not typically a priori well
known, nor do they have particular ‘policies’ beyond their general personal
reputations. We would predict then, that rebeling—in the sense that MPs
vote in a way that contrasts with the majority of their party—would be a
boon to electoral performance. In fact, both coefficients—OLS and random
forest—are in the correct direction (positive), and they are statistically sig-
nificant.
The other roll calls in the table have a slightly more complicated status in
terms of the party pressure applied to the vote. For example, consider the
Section 4.5 102
June 2000 “Programming of Bills” motion that concerned the Modernisation
Committee’s recommendations for changes to the way that parliamentary
business be conducted. Usually, ‘House business’ issues are seen as free
votes, not least because they emanate from non-partisan Select Committees
that are designed to hold the government to account. Interestingly, and as
a break with standard practice, the Modernisation Committee was chaired
by the Leader of the House, a cabinet member. Both in committee and
during the ensuing Commons debate, MPs accused the Government—via
the Leader of the House—of promoting its own agenda (see Kelso, 2007,
140–141). In particular, several MPs felt that the proposed reforms made
it more difficult for parliament to oppose the executive and hold it to ac-
count. Thus, members backing an amendment to these plans that would
have extended debating time were de facto frustrating Government ambi-
tions. There are thus two consequences: first, the vote was ‘free’ for Labour
backbenchers but pressure was surely present and second, the ‘rebels’ were
actually the majority of the party on this vote (note that the Leader of the
House opposed the amendment). With all this in mind, the sign is essen-
tially as predicted: those showing independent-mindedness on a bill that
was not an explicit part of the Government’s agenda did better at the polls.
The City of London (Ward Election) Bill appears in row 4: the coefficient
is significant. By this private bill, the City of London intended to reform
its self-governing arrangements as a municipality. Unusually for an author-
ity in a democratic state, the City gives businesses votes, and the proposed
legislation would have extended this right to more businesses. For many left-
Section 4.5 103
leaning MPs in the Labour party, such a plan was anathematic. It is unclear
exactly how or even if the Government asserted pressure on its MPs though.
Certainly, the Leader of the House claimed it was an explicitly free vote,
though Salter (2004) argues that the Government may have pressured MPs
to at least pass something that could be considered ‘reforming’, just to be
done with the issue. In July of 1999, the ideological components of the plans
came under attack, with Labour backbenchers in the minority condemning
them. In November of the same year, the bill was not yet passed and there
was lively debate as to whether or not the proposals should be reconsidered
in the next parliamentary session which, without explicit Government in-
volvement they would not have been. To the extent that the rebels were
now holding up procedural Government business, it is perhaps understand-
able that the sign is negative and they were punished for their intransigence.
The votes concerning “Weight and Measures” (row 2), “Human Fertilisa-
tion and Embryology” (row 9) have signs as predicted, though not at signif-
icant levels. The first vote dealt with a particular “statutory instrument”,
a method of delegated legislation that allows a minister to make changes
to rules and laws without the requirement that the proposals be passed as
a separate parliamentary bill. In this particular case, the vote dealt with
how long Imperial measurement units would be allowed to continue as a
‘supplement’ to the metric units required by European Union legislation. It
should be clear that such a vote should not fall into the category of essen-
tial governmental programmatic business as defined above—and we see that
‘rebels’ did better when they opposed this vote. The Human Fertilisation
Section 4.5 104
vote pertained to cloning as a medical research practice. Again, we see rebels
garnering a (minor) uptick in support for their ‘independent-mindedness’ on
a free vote for which the public is presumably fairly ill-informed.
The final case to consider concerns a vote on “Doctor-Assisted Dying” (row
8). Euthanasia is not generally a salient issue for most British voters at the
ballot box (unlike, say, abortion for US voters), and the debate has been
characterized by its cross-cutting dimensions. For example, conservative,
religious groups, the British Medical Association (the main doctors’ trade
union) and several prominent left-wing MPs oppose doctor assisted dying.
It is thus not precisely clear where the average Labour voter places herself
on the issue, so it is hard to know what to make of the fact that ‘rebels’
(those who voted in favor of euthanasia) did worse at election time. In any
case, the effect is apparently not significant.
4.5.2 Second term: 2001–2005
The 2005 election saw most Labour MPs do worse than in 2001, with a
mean loss of 7 percent in our data set (variance was 12.32). Given that
Labour entered the election with 30 seats where the majority was less than
7 percent, this was clearly a more competitive election. The national swing
was a commensurate 6 percent away from the governing party, though their
overall performance was sufficient to garner a workable majority of 66 seats.
The biggest (Labour) loser of the election was Roger Godsiff who suffered a
reduction in his vote total of some 21 percent (primarily to a candidate from
the new anti-war Respect party) and the biggest gains were made by Shaun
Section 4.5 105
Woodward, a Conservative defector who left his original party in 1999. In
Figure 4.6 we graph the most important variables for this regression from
the random forest technique. The (mean) pseudo-R
2
goodness of fit measure
for the model was around 0.04.
Once again, we note that the controls are certainly important, though their
relative ranking has been permutated from that in Figure 4.5. And, again,
we make a judgement call as to which of the bills are worthy of further
inspection. Here, it is just 10, and we report their effects in Table 4.2. As
in the 1997–2001 session, we see that the ‘Programming of Bills’ is an ap-
parently important predictor, occupying rows 1 and 8 of the table. In both
cases, ‘rebeling’—in the sense of voting with the minority—was beneficial for
MPs (and this effect is statistically significant). These votes were explicitly
free though, once more, knowing precisely how much pressure was applied
in practice is difficult to gauge. With this in mind, as with our previous
findings, our theory predicts the positive sign we see in practice.
Unlike the programming matters of the previous paragraph, the “Children
Bill” was part of the Government’s legislative plans for the session and was
therefore whipped appropriately. The vote in row 2 of the table actually con-
cerned a backbench amendment (from David Hinchliffe, Chair of the Health
Select Committee) that would have struck out part of the Government’s
bill that allowed parents to use “reasonable physical chastisement” in disci-
plining their children. The result would be a bill that essentially outlawed
all such spanking. Somewhat unusually—given that it proffered a change
Section 4.5 106
div244.020516
div239.020514
div307.041103
div241.030618
div059.040225
div316.020724
div311.041108
div237.020514
div180.030508
div093.011129
div045.050125
div281.041026
div336.021029
div240.020514
div048.050126
div279.041026
div011.010705
div358.031110
div208.040628
div283.041026
div044.011025
div306.041102
div278.041026
DiffT
LoneParent
Spend
Illness
Unemploy
Ethnic
Pension
0 50 100 150 200 250
IncNodePurity

i
m
p
o
r
t
a
n
t


r
o
l
l

c
a
l
l
s
c
o
n
t
r
o
l
s
Figure 4.6: Variable importance plot for 2001–2005 data. Controls and ‘important’ roll
calls used in subsequent analysis shown as subsets of all variables. x-axis gives the increase
in the regression performance from including each variable.
Section 4.5 107
C
o
d
e
N
a
m
e
D
i
s
c
i
p
l
i
n
e

R
e
b
e
l

E

e
c
t
(
R
F
)

R
e
b
e
l

E

e
c
t
(
O
L
S
)
p
-
v
a
l
u
e
d
i
v
2
7
8
.
0
4
1
0
2
6
P
r
o
g
r
a
m
m
i
n
g
o
f
B
i
l
l
s
f
r
e
e
+
0
.
5
3
5
+
1
.
6
6
0
0
.
0
0
5
d
i
v
3
0
6
.
0
4
1
1
0
2
C
h
i
l
d
r
e
n
B
i
l
l
:
R
e
a
s
o
n
a
b
l
e
P
u
n
i
s
h
m
e
n
t
m
o
d
e
r
a
t
e
-
0
.
1
3
0
-
1
.
1
1
3
0
.
0
3
9
d
i
v
0
4
4
.
0
1
1
0
2
5
S
o
c
i
a
l
S
e
c
u
r
i
t
y
w
e
a
k
+
0
.
2
5
6
+
0
.
9
3
1
0
.
1
3
0
d
i
v
2
8
3
.
0
4
1
0
2
6
R
e
m
o
v
a
l
o
f
R
e
f
e
r
e
n
c
e
s
t
o
S
t
r
a
n
g
e
r
s
f
r
e
e
+
0
.
2
8
8
+
1
.
2
9
7
0
.
0
4
3
d
i
v
2
0
8
.
0
4
0
6
2
8
H
u
m
a
n
T
i
s
s
u
e
B
i
l
l
:
c
o
n
s
e
n
t
f
o
r
r
e
m
o
v
a
l
w
e
a
k
/
m
o
d
+
0
.
2
4
0
+
1
.
3
3
5
0
.
0
2
4
d
i
v
3
5
8
.
0
3
1
1
1
0
F
l
u
o
r
i
d
a
t
i
o
n
o
f
W
a
t
e
r
S
u
p
p
l
i
e
s
f
r
e
e
+
0
.
1
4
1
+
1
.
2
3
3
0
.
0
1
3
d
i
v
0
1
1
.
0
1
0
7
0
5
M
e
m
b
e
r
s

A
l
l
o
w
a
n
c
e
s
,
I
n
s
u
r
a
n
c
e
f
r
e
e
+
0
.
1
0
0
+
0
.
2
4
3
0
.
6
3
0
d
i
v
2
7
9
.
0
4
1
0
2
6
P
r
o
g
r
a
m
m
i
n
g
o
f
B
i
l
l
s
f
r
e
e
+
0
.
2
5
0
+
1
.
2
6
3
0
.
0
4
6
d
i
v
0
4
8
.
0
5
0
1
2
6
M
o
d
e
r
n
i
z
a
t
i
o
n
o
f
t
h
e
H
o
u
s
e
f
r
e
e
+
0
.
0
4
7
+
2
.
5
1
7
0
.
6
3
4
d
i
v
2
4
0
.
0
2
0
5
1
4
L
i
a
i
s
o
n
C
o
m
m
i
t
t
e
e
:
P
o
w
e
r
t
o
T
a
k
e
E
v
i
d
e
n
c
e
f
r
e
e
+
0
.
0
8
2
+
0
.
5
8
6
0
.
1
9
7
T
a
b
l
e
4
.
2
:

M
o
s
t
i
m
p
o
r
t
a
n
t

p
a
r
l
i
a
m
e
n
t
a
r
y
v
o
t
e
s
f
o
r
p
r
e
d
i
c
t
i
n
g
e
l
e
c
t
o
r
a
l
p
e
r
f
o
r
m
a
n
c
e
b
y
b
a
c
k
b
e
n
c
h
L
a
b
o
u
r
M
P
s
i
n
2
0
0
5
g
e
n
e
r
a
l
e
l
e
c
t
i
o
n
.
F
i
r
s
t
c
o
l
u
m
n
i
s
d
a
t
a
c
o
d
e
f
o
r
t
h
e
d
i
v
i
s
i
o
n
w
i
t
h
d
a
t
e
a
f
t
e
r
p
o
i
n
t
.
N
a
m
e
o
f
b
i
l
l
i
n
c
o
l
u
m
n
2
;
r
a
n
d
o
m
f
o
r
e
s
t

c
o
e

c
i
e
n
t

i
n
c
o
l
u
m
n
3
f
o
r
b
e
i
n
g
a

r
e
b
e
l

(
i
n
t
h
e
m
i
n
o
r
i
t
y
a
n
d
/
o
r
f
r
u
s
t
r
a
t
i
n
g
t
h
e
G
o
v
e
r
n
m
e
n
t

s
a
m
b
i
t
i
o
n
s
)
;
O
L
S
c
o
e

c
i
e
n
t
i
n
c
o
l
u
m
n
6
;
p
-
v
a
l
u
e
f
o
r
t
h
a
t
c
o
e

c
i
e
n
t
i
n
c
o
l
u
m
n
7
.
Section 4.5 108
to government legislation—the vote was ‘free’ in theory. Nonetheless, the
Government clearly had a preference on the vote, and we see a negative
and significant coefficient implying that ‘rebels’ did worse after ‘defying’ the
executive. The other piece of Government legislation in the table is the “Hu-
man Tissue Bill” (row 5) but, once again, the substantive circumstances of
the vote are quite odd by Westminster standards. In particular, the roll call
concerned a Liberal Democrat amendment to a government bill that would
have made organ donation automatic unless the deceased had previously
registered an objection. While this was a free vote for both the main Oppo-
sition parties, it was whipped by the government, on the philosophical basis
that such an amendment would, by presuming consent, be denying individ-
uals the right to exercise their conscience. Clearly then, this bill was not a
cut and dried “government-versus-opposition” case and the low turnout adds
fuel to this argument. With this in mind, the significant and positive coeffi-
cient for rebel behavior is perhaps less anomalous than it might first appear.
Several of the remaining entries in Table 4.2 are free votes, primarily that
dealt with internal House issues. For example, the “Removal of References
to Strangers” (row 4) dealt with changes to parliamentary language, “Mem-
bers’ Allowances, Insurance” referred to various MP compensation concerns,
“Modernization of the House” (row 9) referred to the sitting hours of the
Commons and “Liaison Committee: Power to Take Evidence” (row 10) dealt
with procedural issues for cross-party committees. In all cases, we see the
sign is positive, as predicted for ‘rebeling’ on such unwhipped votes, though
it is only significant in one case.
Section 4.5 109
The last two votes to consider from Table 4.2 are quite obscure. The first,
on “Social Security” (row 3) actually refers to a statutory instrument that
regulates unemployment services in the UK. The Conservatives abstained
on the issue and, though not technically a free vote, it was not a crucial part
of the Government’s agenda. Thus the sign for ‘rebeling’ is in the correct
direction (i.e. it is positive) but it is not significant. Finally, row 6 deals
with “Fluoridation of Water Supplies”, a perennial debate in the Commons.
The free vote concerned whether or not local authorities could force water
providers to add fluoride to residential and other supplies. As predicted,
the sign is positive—rewarding the minority on a free vote—and, in fact,
significant.
4.5.3 Summary of Results
We found that controlling for various other factors,
- For both 2001 and 2005, MPs rebeling against important government
legislative business (‘trouble-makers’) were punished by voters, and
this effect was sometimes statistically significant.
- For both 2001 and 2005, MPs dividing with the minority on free votes
(being ‘mavericks’) which involved issues on which voters were ill-
informed tended to be rewarded by voters
- For 2001, Labour MPs voting against popular social justice legislation—
in particular, gay rights—in free votes were generally punished by the
electorate
Section 4.6 110
- the OLS effects, when significant, were on the order of 1 to 2 percent
of the vote differential from the previous election. The random forest
(marginal) effects were generally of a much smaller magnitude, though
the signs generally agree.
4.6 Discussion
It seems that old certainties about British politics are changing: govern-
ment party MPs are now more rebellious than they have been in a gen-
eration (Cowley, 2002, 2005), and there is increasing evidence that local
factors matter at election time (Denver and Hands, 1997; Pattie, Johnston
and Fieldhouse, 1995). The current paper was concerned with connecting
the two trends and systematically assessing the evidence for voter reac-
tion to patterns of MP activity in Parliament. Using new non-parametric
techniques and roll call records in their totality, we found that (Labour) vot-
ers do indeed seem to respond to overall patterns of activity: they punish
‘trouble-making’ rebellion on heavily whipped government business, while
rewarding more maverick legislators when party unity is not per se at stake.
Moreover, this finding makes sense of seemingly inconsistent public opin-
ion preferences for both party loyalty and free thinking, and adds a layer
of intuitive, though subtle, depth to our understanding of accountability in
Westminster systems. This perhaps allows some ‘catch up’ with American-
ist efforts, while producing new directions for future research.
As was explicitly noted above, the model of voter behavior was ‘as if’ in
Section 4.6 111
nature: we would not contend citizens have a sufficiently advanced knowl-
edge of policy making or parliament to directly reward or punish legislators
on certain bills. Rather, we think that voters receive signals from a possibly
media-based elite, who digest and then describe overall patterns of MP be-
havior as ‘rebellious’, ‘loyal’, ‘fractious’, ‘conforming’ and so on. So, while
our data and method tapped the generalities of this dynamic, it could not
trace the causation path precisely. Further work in this area might thus
profitably gather media or other data on the messages that voters receive
in order to disentangle what exactly are the triggers that change the labels
MPs receive.
Finally, this paper has potentially important ramifications for institutional
reform in the United Kingdom. At the time of writing, the House of Com-
mons Modernisation Committee is engaged in an attempt to strengthen the
role of backbenchers in terms of their dealings with the executive, and is
soliciting expert opinion on these matters (see, e.g., Cowley, 2006). It is
hoped that this will resurrect public engagement in politics after two gen-
eral election turnouts of around 60 percent (very low numbers by Western
European standards). We have suggested that voters are already either het-
erogenous or discerning in their responses to backbench behavior, and fur-
ther independence for MPs will presumably make them more so. Whether
MPs supporting these reforms understand that their electoral fortunes will
be increasingly tied to their legislative performance and choices—and not
the party’s overall popularity—is unclear. Given that institutional norms—
whipping and unity—appear to both affect and be affected by ‘what voter’s
Section 4.6 112
want’, changing those norms ought to have interesting and complex conse-
quences.
113
Appendix to Chapter 4
4.7 Regression Tree Split Criteria
Consider a ‘parent’ node t containing some number of cases N(t) that is to
be split. We measure the ‘impurity’ of t via the least squares deviation (also
called the ‘deviance’), written R(t), which is
R(t) =

i=t
(y
i
−y(t))
2
(4.3)
where y
i
is the dependent variable value for the ith case at node t, and y(t)
is the mean of all the cases in node t. Notice that Equation (4.3) is simply
the residual sum of squares at t.
We will split the cases from t into two distinct subsets such that the subse-
quent ‘child’ nodes have summed deviations that are as small as possible. In
particular, we require that our split s (on some value of one of the covariates)
maximizes
R(t) −R(t
L
) −R(t
R
) (4.4)
Section 4.8 114
where R(t
L
) is the within node variance of the left child and R(t
R
) is the
within node variance of the right child. This is equivalent to choosing a split
maximizing the reduction in deviance.
As an example, suppose our parent node t has a variance of 20. Two possi-
ble splits in the covariates, s
1
and s
2
are proposed. These may be different
points in the same variable, or different splits in different variables. Suppose
s
1
yields two children: one with a deviance of 11, the other with a deviance
of 19. For s
2
, the left and right child both have a deviance of 15. These
splits are considered equivalent since both result in an equal value of (4.4).
4.8 Model Fit, Variable Importance and Effect
Determining goodness of fit for random forests regression proceeds as follows.
Notice from step 1 in Section 4.4 above that in any given bootstrap sample
(i.e. in any given tree grown) around one third of the cases from the original
training set will not appear there. These are called “out-of-bag” (oob) data.
Every time a particular case i is in the oob data with respect to a tree,
it is ‘run down’—that is, regressed—according to that tree and an (oob)
prediction produced. Take the average of all those oob predictions for i (that
is, one for each tree where it is oob) and denote that value as ˆ y
oob
i
. Akin to
linear regression, we would like to know the percentage of the variance in y
that is explained by our model. For the current technique, this is computed
as
1 −
MSE
oob
ˆ σ
2
y
Section 4.9 115
where
MSE
oob
=
1
n
n

i=1
{y
i
− ˆ y
oob
i
}
2
where MSE is the mean square error (Liaw and Wiener, 2002, 20).
The importance of variable v is estimated by considering the decrease in
node ‘impurities’ when v is split (i.e. it is part of a tree), averaged out over
all the trees in the forest. These impurities are measured via the residual
sum of squares (RSS)—a formula we given in Appendix 4.7. We compare
the (average) value of RSS when v is in x and when it is not. Intuitively,
if the variable is ‘important’, it will be responsible for a better regression
performance: it will generally occur more often in the ‘best’ split and/or the
reduction in deviance when it occurs will be greater.
Finally, we will typically want to know the marginal effect that a variable v
has on y. Liaw and Wiener (2002) suggest estimating the function
˜
f(v) =
1
n
n

i=1
m(v, x
iC
)
where v is the variable of interest and x
iC
are all the other variables. The
element m(v, x
iC
) is the (aggregated) random forest regression function (the
predicted y): the idea being that we vary the values of v (‘aye’ or ‘no’ in the
case of a roll call) to see the (predicted) effect on constituency performance.
Section 4.9 116
4.9 Imputing Missingness
Suppose that there is missingness in the underlying data (in the x). Begin
by filling missingness in each column with the median (if continuous) or
modal (if discrete) category for that column. Then run a random forest
regression procedure and compute the proximity matrix: that is the fraction
of times (across all the trees) that each i, j pair of cases ends up in the
same terminal node. For continuous missing data, the imputed value is
the weighted average of the non-missing data for that column, with the
proximities as the weights. For categorical missingness—the case for our
work here—the category with the largest average proximity (which will be
either ‘aye’ or ‘no’) is the imputed value. As a toy example consider a
variable x = [0, 0, NA, 1]. Observation 3 is first filled with a 0 (since it is the
most common category). Suppose that the resulting proximity matrix was
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
1 2 3 4
1| 0.5 0.6 0.1
2| 0.5 0.4 0.1
3| 0.6 0.4 0.2
4| 0.1 0.1 0.2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
Observation 3 is with Observations 1 and 2
1
2
of the time, but with Obser-
vation 4 only
1
5
of the time: the largest average proximity is thus zero and
we impute this for the observation.
117
Chapter 5
Discussion
The preceding dissertation was concerned with three quite separate substan-
tive areas, involving three somewhat different methods. What draws the
chapters together is the focus on finding new solutions to ‘old’ problems:
identifying ‘turning points’ in conflicts; measuring the ‘power’ of political
actors; describing the ‘effects’ of roll call voting on electoral support. The
essays all suggested thinking about these issues in new ways. In the first
chapter, we employed a change point identification technique that was no-
table for its flexibility: the number of parameters was, itself, a parameter
to be estimated. In the second chapter, we argued that ‘power’ makes little
sense in the absence of ‘contests’ in which actors compete to garner finite
resources. Our statistical model reflected our (more realistic) assumptions
on this matter. In the third chapter, we argued that a very large number
of variables (i.e. roll calls) might collectively demarcate members of par-
liament in terms of their appeal to voters. Rather than employing an ad
hoc ‘thinning out’ of variables such that parametric models could be fit, we
Section 5.0 118
decided to take a more systematic non-parametric approach.
In the individual chapters we made some comments about future uses of
these techniques within their substantive areas, but it may be helpful to
think more generally about applications here. Reversible jump Markov
chain monte carlo, which we introduced in the first chapter, seems an in-
triguing tool for ‘cluster analysis’ in parliaments. Historically, ideal point
estimation (frequentist or Bayesian) has been a popular choice for analyz-
ing legislatures (Clinton, Jackman and Rivers, 2004; Poole and Rosenthal,
1997). A model of legislator behavior is posited in which elected members
compare the utility of the status quo relative to that the policy changed
potentially introduced by passing a new bill. A distribution for the error
terms is assumed—typically Type-1 extreme value or Guassian—to produce
a (parametric, random-utility) statistical model. On occasions, such models
produce misleading or unhelpful results. As noted in Chapter 4, Spirling and
McLean (2007) show that strong whipping in Britain’s House of Commons
produces rank ordering on an ‘agreement with the government’ dimension,
rather than an ideological spectrum. Clustering algorithms, which search for
groupings, are an alternative way to proceed. Some of these techniques—
optimal partitioning methods being among them (Venables and Ripley, 2002,
316)— require an a priori commitment to a fixed number of clusters K to
which cases are then assigned. By contrast, other (typically ‘hierarchical’)
methods fit numerous models (one for each value of k total clusters) which
must then be compared. Since Bayesian procedures are philosophically ap-
pealing in terms of model comparison (see Gill, 2002, 199–224), they seem
Section 5.0 119
a natural choice here. Moreover, because we would like to be flexible about
the number of clusters, RJMCMC would seem a particular good option.
The second essay (Chapter 3) made the case that we can consider politi-
cal actors as players competing in contests. This is an idea with mileage
outside the current context. For example, at least since the time of the
ancient Greeks, scholars have studied the power of argument to convince
citizens. The earliest accounts were primarily interested in rhetoric as an
art (e.g. Aristotle, 322BC/1991), though in more recent times, psycholo-
gists and economists have studied the consequences of ‘framing’ effects, es-
pecially with respect to risk-taking behavior (e.g. Tversky and Kahneman,
1981). In these cases, ‘arguments’ essentially ‘compete’ with respondents
either suggesting which they prefer directly, or voting commensurate with
their views. Using the Bradley-Terry model in the second chapter would
be one way forward here. Indeed, if one could properly separate competing
arguments in, say, an election campaign, analysts could begin to model why
certain candidates win or lose. As constructed, the standard Bradley-Terry
model assumes that contests involve two players, with one losing and one
winning. Over a series of (independent) contests, we make the reasonable
assumption that success is binomially distributed. Of course, contests in
politics are often more complicated. For example, US presidential primaries
and caucuses are typically three- or four-way competitions, at least in the
early stages. The same is true of British electoral contests, where Labour,
Conservative and Liberal Democrat candidates vie for success in parliamen-
tary constituencies. Moreover, there are not always clear ‘winners’ and
Section 5.0 120
‘losers’: outcomes are often ‘tied.’ Paradoxically, this idea often holds in
majoritarian electoral systems where there is one winner and all others, in
effect, lose ‘equally’: for example, when candidates run for President of the
France. An economic version of this holds when firms bid for public works
contracts. With these cases in mind, it would be helpful to think more gen-
erally about the Bradley-Terry model, perhaps facilitating the estimation of
abilities when contests have several players and when ties are possible.
The third essay (Chapter 4) claimed that, when the number of variables
to consider was greater than the number of parameters to be estimated,
one helpful way to proceed was the employment of a ‘random forests’ clas-
sification procedure. King (1989, 3) makes the case that political science
“outdistances most other social sciences” in terms of data collection. For
data exploration purposes then, the model matrices are potentially very
large. Moreover, translation from ‘theory’ to statistical model is typically
not direct in political science: for example, a scholar may believe that ‘cor-
ruption’ has some systematic effect on citizen-to-citizen trust in a city , or
that ‘more proportional’ electoral systems give rise to fiercer intra-party
competition. In such cases, there are presumably several indicators and
ways of measuring the attributes of the independent variables and possibly
no way to a priori decide between them. Since non-parametric techniques
allow one to include literally hundreds more predictors than observations,
these operationalization decisions can be refined within the (first) estimation
‘round’ rather than on some ad hoc basis with knock-on effects for the study
as a whole. As data becomes increasingly available in massive amounts—
Section 5.0 121
especially via the internet—such work would seem important and valuable.
Of course, non-parametric fitting should not be considered the final stage
in any investigation of political actor behavior: nonetheless, it can help us
think more clearly about possible micro-mechanisms.
All told then, this dissertation has argued that our statistical solutions
should fit our problems, and has suggested several realms in which either
current practices might be improved, or where new profitable avenues of in-
vestigation exist. It would remiss of the author, however, to fail to address
a criticism of quantitative techniques that arises particularly in political
methodology. Phrased in various ways, typically implicit and casually sug-
gested, the claim of some is that methodologists fetishize complexity in their
models; that is, much of political methodology is ‘complicated for the sake
of being complicated.’ This argument is imprecise, incorrect and unfair. It
is imprecise insofar as it can be applied—in its most basic form—to any
data analysis technique: logistic regression is ‘more complicated’ than OLS,
which is ‘more complicated’ than cross-tabs, which are ‘more complicated’
than frequency summaries. Otherwise put, it is unclear as to where the
analyst should draw the line: what extra utility must be garnered in order
to approve of a more complex technique? Moreover, such standards clearly
differ from scholar to scholar: reversible-jump Markov chain monte carlo
approaches might seem unaccessible to someone who know little of Bayesian
methods, but are a helpful and important extension for an analyst already
familiar with MCMC techniques. The claim is incorrect because it funda-
mentally misunderstands the goal of political methodology. The aim is to
Section 5.0 122
provide tools that help us understand political behavior; to the extent that
progress towards this goal remains incomplete (few if any would disagree
with this contention), ‘more’ or ‘better’ theory, methods and data are the
way forward. It can surely never hurt to refine our techniques and our un-
derstanding of their capabilities. Note also that, in computational terms,
the costs involved with ‘complicated’ techniques have fallen precipitously
over time and will continue to do so. Finally, the claim is unfair because
it cynically assumes that methodologists either deliberately dissemble (and
are thus dishonest in describing) their work, else they are simply incompe-
tent in solving the problems that the discipline presents. Yet, one of the
strengths of work in methodology is that assumptions of models and tech-
niques are both clearly laid out, and written in a language common to its
practitioners. It is certainly in this spirit that this dissertation is intended
as a contribution to the discipline.
123
Bibliography
Achen, Christopher H. 2002. “Toward a New Political Methodology: Micro-
foundations and ART.” Annual Review of Political Sience 5:423–50.
Agresti, Alan. 2002. Categorical Data Analysis. 2 ed. New York: John Wiley
and Sons.
Alesina, Alberto and Allan Drazen. 1991. “Why are Stabilizations Delayed?”
American Economic Review 81(5):1170–1188.
Amadae, S. M. and Bruce Bueno de Mesquita. 1999. “The Rochester School:
The Origins of Positive Political Theory.” Annual Review of Political Sci-
ence 2:269–295.
Aristotle. 322BC/1991. Rhetoric. New York: Oxford University Press
(translation by George A. Kennedy).
Bachrach, Peter and Morton Baratz. 1962. “Two Faces of Power.” American
Political Science Review 56:947–952.
Banzhaf, John F. 1965. “Weighted Voting Doesn’t Work: A Mathematical
Analysis.” Rutgers Law Review 19:317–343.
Section 5.0 124
Barry, Brian. 1991. Is it Better to be Powerful or Lucky? In Democracy and
Power. Oxford: Clarendon.
Beck, Nathaniel, Gary King and Langche Zeng. 2000. “Improving Quantita-
tive Studies of International Conflict: A Conjecture.” American Political
Science Review 94(1):21–36.
Benedetto, Giacomo and Simon Hix. 2007. “The Rejected, the Ejected, and
the Dejected: Explaining Government Rebels in the 2001-2005 British
House of Commons.” Comparative Political Studies 40:755–781.
Black, Duncan. 1948. “On the Rationale of Group Decision-making.” Jour-
nal of Political Economy 56:23–34.
Brace, Paul. 1984. “Progressive Ambition in the House: A Probabilistic
Approach.” The Journal of Politics 46(2):556–571.
Bradley, R A and M E Terry. 1952. “Rank Analysis of Incomplete Block
Designs I: The Method of Paired Comparisons.” Biometrika 39:324–345.
Breiman, L., J. H. Friedman, R. A. Olshen and C. J. Stone. 1984. Classifica-
tion and Regression Trees. Monterey, CA: Wadsworth and Brooks/Cole.
Breiman, Leo. 1996. “Bagging predictors.” Machine Learning 24(2):123–140.
Breiman, Leo. 2001. “Random forests.” Machine Learning 45:5–32.
Brooks, Stephen P. 1997. “Markov chain Monte Carlo Method and its ap-
plications.” The Statistician 47(1):67–100.
Section 5.0 125
Budge, Ian. 1999. Critical Elections: British Parties and Voters in Long-
term Perspective. London: Sage Publications.
Burke, Edmumd. 1774/1975. Speech to the Electors of Bristol. In Edmund
Burke: On Government, Politics and Society, ed. B.W. Hill. London:
Fontana.
Burnham, Gilbert, Riyadh Lafta, Shannon Doocy and Les Roberts. 2006.
“Mortality after the 2003 invasion of Iraq: a cross-sectional cluster sample
survey.” The Lancet (online) 10.1016/(S0140-6736(06)69491-9).
Butler, David and Donald Stokes. 1974. Political Change in Britain. 2 ed.
Macmillan.
Cain, Bruce E., John A. Ferejohn and Morris P. Fiorina. 1987. The Personal
Vote: Constituency Service and Electoral Independence. Cambridge, MA:
Harvard University Press.
Canes-Wrone, Brandice, David W. Brady and John F. Cogan. 2002. “Out
of Step, Out of Offce: Electoral Accountability and House Members’ Vot-
ing.” American Political Science Review 96(1).
Carlin, Bradley, Alan Gelfand and Adrian Smith. 1992. “Hierarchi-
cal Bayesian Analysis of Changepoint Problems.” Applied Statistics
41(2):389–405.
Chib, Siddhartha. 1998. “Estimation and comparison of multiple change
point models.” Journal of Econometrics 86:221–241.
Section 5.0 126
Clinton, Joshua, Simon Jackman and Doug Rivers. 2004. “The Statistical
Analysis of Roll Call Data.” American Political Science Review 98.
Cowan, Katherine. 2007. “Living Together: British Attitudes to Lesbian
and Gay People.” Stonewall Poll by YouGov.
Cowley, Philip. 2002. Revolts and Rebellions: Parliamentary Voting Under
Blair. London: Politico’s Press.
Cowley, Philip. 2005. The Rebels: How Blair Mislaid His Majority. London:
Politico’s Publishing.
Cowley, Philip. 2006. “Strengthening the role of the backbencher.” Written
submission to the Select Committee on Modernisation of the House of
Commons Session 2006–07, no. 1.
Cox, Gary W. and Mathew D. McCubbins. 1993. Legislative Leviathan:
Party Government in the House. Berkeley: University of California Press.
Crewe, Ivor. 1983. The Electorate: Partisan Dealignment 10 Years On. In
Change in British Politics, ed. Hugh Berrington. London: Frank Cass
pp. 183–215.
Dardagan, Hamit and John Sloboda. 2006. “Iraq Body Count.”.
URL: http://www.iraqbodycount.net/
de Marchi, Scott, Christopher Gelpi and Jeffrey D. Grynaviski. 2004. “Un-
tangling Neural Nets.” American Political Science Review 98(2):371–378.
Denver, David and Gordon Hands. 1997. Modern Constituency Electioneer-
ing. Portland, Oregon: Frank Cass.
Section 5.0 127
Dowding, Keith. 1996. Power. Minneapolis: University of Minnesota Press.
Downs, Anthony. 1987. An Economic Theory of Democracy. New York:
Harper.
Efron, Bradley and Robert Tibshirani. 1993. An Introduction to the Boot-
strap. New York: Chapman & Hall.
Felsenthal, Dan S and Mosh´e Machover. 1998. The Measurement of Vot-
ing Power: Theory and Practice, Problems and Paradoxes. Cheltenham:
Edward Elgar.
Felsenthal, Dan S. and Mosh´e Machover. 2004. “A Priori Voting Power:
What Is It All About?” Political Studies Review 2:1–23.
Fenno, Richard F. 1973. Congressmen in Committees. Boston, MA: Little,
Brown and Company.
Filson, Darren and Suzanne Werner. 2004. “Bargaining and Fighting: The
Impact of Regime Type on War Onset, Duration, and Outcomes.” Amer-
ican Journal of Political Science 48(2):296–313.
Fiorina, Morris. 1974. Representatives, Roll Calls and Constituencies. Lex-
ington, MA: Lexington Books.
Firth, David. 2005. “Bradley-Terry Models in R.” Journal of Statistical
Software 12.
Firth, David and Arthur Spirling. 2006. “tapiR and the Public Whip: Re-
sources for Westminster Voting.” The Political Methodologist 14(2).
Section 5.0 128
Fowler, James H. 2006. “Connecting the Congress: A Study of Cosponsor-
ship Networks.” Political Analysis 14(4):456–487.
Garrett, Geoffrey and George Tsebelis. 1999. “Why Resist the Temptation
to Apply Power Indices to the EU.” Journal of Theoretical Politics 11:291–
308.
Gill, Jeff. 2002. Bayesian Methods: A Social and Behavioral Sciences Ap-
proach. New York: Chapman & Hall/CRC.
Green, Peter J. 1995. “Reversible Jump MCMC computation and Bayesian
model determination.” Biometrika 82(3):711–732.
Groseclose, Tim and Charles Stewart. 1998. “The Value of Committee
Seats in the House 1947–1991.” American Journal of Political Science
42(2):453–474.
Hastie, David. 2005. Towards Automatic Reversible Jump Markov Chain
Monte Carlo PhD thesis University of Bristol.
Holler, Manfred and Mika Widgren. 1999. “Why Power Indices for Assess-
ing European Union Decision-Making?” Journal of Theoretical Politics
11:321–330.
Huitt, Ralph K. 1957. “The Morse Committee Assignment Controversy: A
Study in Senate Norms.” American Political Science Review 51(2):313–
329.
Jackman, Simon. 2004. “Bayesian Analysis for Political Research.” Annual
Review of Political Science 7:483–505.
Section 5.0 129
Jacobson, Gary. 1989. “Strategic Politicians and the Dynamics of House
Elections, 1946-86.” American Political Science Review 83:773–93.
Janis, Irving L. 1972. Victims of Groupthink. Boston: Houghton Mifflin
Company.
Johnson, Catherine and Gemma Rosenblatt. 2007. “Do MPs have the ‘Right
Stuff’ ?” Parliamentary Affairs 60:164–169.
Kam, Christopher. 2006. “Demotion and Dissent in the Canadian Liberal
Party.” British Journal of Political Science 36:561–74.
Kam, Christopher. 2007. On Division: Cohesion, Discipline, and Intra-
party Politics. University of British Columbia: Book manuscript under
review.
Kass, Robert E. and Adrian E. Raftery. 1995. “Bayes Factors.” Journal of
the American Statistical Association 90(430):773–795.
Kelso, Alexandra. 2007. “The House of Commons Modernisation Com-
mittee: Who Needs It?” British Journal of Politics and International
Relations 9:138–157.
Key, V. O. 1949. Southern Politics in State and Nation. New York: Knopf.
King, Gary. 1989. Unifying Political Methodology: The Likelihood Theory of
Statistical Inference. Ann Arbor: University of Michigan Press.
King, Gary. 2001. “Proper Nouns and Methodological Propriety: Pool-
ing Dyads in International Relations Data.” International Organization
55(2):497–507.
Section 5.0 130
Krehbiel, Keith. 1991. Information and Legislative Organization. Ann Ar-
bor: University of Michigan Press.
Krehbiel, Keith. 1998. Pivotal Politics: A Theory of U.S. Lawmaking.
Chicago: University of Chicago Press.
Lane, Jan-Erik and Sven Berg. 1999. “Relevance of Voting Power.” Journal
of Theoretical Politics 11:309–320.
Lasswell, Harold. 1936. Who Gets What, When and How. New York: Whit-
tlesey House.
Leech, Dennis. 2002a. “Computation of Power Indices.” Warwick Economic
Research Papers (644).
Leech, Dennis. 2002b. “An Empirical Comparison of the Performance of
Classical Power Indices.” Political Studies 50(1):1–22.
Leonte, Daniela, David J. Nott and William T. M. Dunsmuir. 2003.
“Smoothing and Change Point Detection for Gamma Ray Count Data.”
Mathematical Geology 35(2).
Liaw, Andy and Matthew Wiener. 2002. “Classification and Regression by
randomForest.” R News 2(3).
Marx, Karl. 1967. Das Kapital. New York: International Publishers. English
Translation.
Mayhew, David R. 1974. Congress : the Electoral Connection. New Haven:
Yale University Press.
Section 5.0 131
McCarty, Nolan, Keith Poole and Howard Rosenthal. 2006. Polarized Amer-
ica: The Dance of Ideology and Unequal Riches. Cambridge, MA: MIT
Press.
McLean, Iain, Arthur Spirling and Meg Russell. 2003. “None of the Above:
The UK House of Commons Votes on Reforming the House of Lords,
February 2003.” The Political Quarterly 74:298–310.
Norris, Pippa. 2001. New Labour and Public Opinion: The Third Way as
Centrism? London: Macmillan pp. 32–44.
Norton, Philip and David M. Wood. 1993. Back from Westminster: British
Members of Parliament and their Constituents. Lexington: University of
Kentucky Press.
Paltzelt, W. 2000. What Can an Individual MP Do in German Parliamen-
tary Politics. In The Uneasy Relationship Between Parliamentary Mem-
bers and Leaders. London: Frank Cass.
Park, Jong Hee. 2006. “Modeling Structural Changes: Bayesian Estimation
of Multiple Changepoint Models and State Space Models.” Prepared for
American Political Science Association Meeting August, 2006, Philadel-
phia.
Pattie, Charles, Edward Fieldhouse and R. J. Johnson. 1994. “The Price of
Conscience: The Electoral Correlates and Consequences of Free Votes and
Rebellions in the British House of Commons, 19897–92.” British Journal
of Political Science 24:359–80.
Section 5.0 132
Pattie, Charles J., Ronald J. Johnston and Edward A. Fieldhouse. 1995.
“Winning the Local Vote: The Effectiveness of Constituency Campaign
Spending in Great Britain, 1983-1992.” American Political Science Review
89(4):969–983.
Poole, Keith and Howard Rosenthal. 1997. Congress: A Political Economic
History of Roll Call Voting. New York: Oxford University Press.
R Development Core Team. 2006. R: A Language and Environment for Sta-
tistical Computing. Vienna, Austria: R Foundation for Statistical Com-
puting. ISBN 3-900051-07-0.
URL: http://www.R-project.org
Riker, William. 1962. The Theory of Political Coalitions. New Haven: Yale
University Press.
Riker, William. 1964. “Some Ambiguities in the Notion of Power.” American
Political Science Review 58(2):341–349.
Riker, William. 1982. Liberalism Against Populism. San Francisco: W. H.
Freeman.
Riker, William. 1986. The Art of Political Manipulation. New Haven: Yale
University Press.
Ripley, Brian D. 1996. Pattern Recognition and Neural Networks. Cam-
bridge: Cambridge University Press.
Rohde, David W. 1979. “Risk-Bearing and Progressive Ambition: The Case
Section 5.0 133
of Members of the United States House of Representatives.” American
Journal of Political Science 23(1):1–26.
Rohde, David W. 1991. Parties and Leaders in the Postreform House.
Chicago: University of Chicago Press.
Ruger, Theodore W., Pauline T. Kim, Andrew D. Martin and Kevin M.
Quinn. 2004. “Competing Approaches to Predicting Supreme Court De-
cision Making.” Perspectives on Politics Symposium 2.
Salter, Lee. 2004. “Parliament and The Parliamentarians: The Worrying
Case of the City of London (Ward Elections) Bill.” Political Quarterly
75:155–167.
Searing, Donald. 1994. Westminster’s World: Understanding Political Roles.
Cambridge, MA: Harvard University Press.
Shapley, L.S. and Martin Shubik. 1954. “A Method for Evaluating the Dis-
tribution of Power in a Committee System.” American Political Science
Review 48:787–92.
Shepsle, Kenneth. 1979. “Institutional arrangements and equilibrium in
multidimensional voting models.” American Journal of Political Science
23:27–60.
Shepsle, Kenneth and Barry Weingast. 1987. “The institutional foundations
of committee power.” American Political Science Review 81:85–104.
Spirling, Arthur and Iain McLean. 2007. “UK OC OK? Interpreting Opti-
Section 5.0 134
mal Classification Scores for the United Kingdom House of Commons.”
Political Analysis 15(1).
Springall, A. 1973. “Response surface fitting using a generalization of the
Bradley-Terry paired comparisons model.” Applied Statistics 22:59–68.
Stam, Allan C. and D. Scott Bennett. 2006. “Predicting the Length of the
2003 U.S.-Iraq War.” Foreign Policy Analysis 2(2):101–116.
Stigler, S. 1994. “Citation Patterns in the Journals of Statistics and Prob-
ability.” Statistical Science 9:94–108.
Stuart-Fox, Devi M., David Firth, Adnan Moussalli and Martin J. Whiting.
2006. “Multiple Signals in Chameleon Contests: Designing and Analysing
Animal Contests as a Tournament.” Animal Behaviour 71:1263–1271.
The Economist. 2006. “Indoor rainforests: Rejecting Eden.” Print Edition
Jul 13.
Tsebelis, George. 2002. Veto Players: How Political Institutions Work.
Princeton, N.J.: Princeton University Press.
Tversky, Amos and Daniel Kahneman. 1981. “The Framing of Decisions
and the Psychology of Choice.” Science 211:453–458.
Venables, William and Brian Ripley. 2002. Modern Applied Statistics with
S. New York, New York: Springer.
Wawro, Gregory. 2000. Legislative Entrepreneurship in the US House of
Representatives. Ann Arbor: University of Michigan Press.
Section 5.0 135
Western, Bruce and Meredith Kleykamp. 2004. “A Bayesian Change Point
Model for Historical Time Series Analysis.” Political Analysis 12:354–374.
Winston Churchill, Sir. 1954. “Duties of a Member of Parliament.” Parlia-
mentary Affairs 8.

ii

For my mother and for my father.

iii

Curriculum Vitae
The author was born in Carshalton in the London Borough of Sutton, United Kingdom, on August 26, 1979. He attended the London School of Economics from 1997 to 2001 and graduated with a Bachelor of Science degree in 2000. He received a Master of Science degree in 2001 from the same institution, and then attended Nuffield College, Oxford University, from 2001 to 2003. He came to the University of Rochester in the Fall of 2003 and began graduate studies in Political Science. He pursued his research in political methodology under the direction of Professor Signorino and received the Master of Arts degree from the University of Rochester in 2007.

Emphasizing the ‘grand plan’ as well as the details. Curt has taught me a great deal . all things are attainable. my advisor. Preferring to gently steer rather than direct. work and emotional well-being over these last several years. and the travails of getting published.iv Acknowledgements Social reformer Sir Thomas Fowell Buxton once modestly observed that “with ordinary talents and extraordinary perseverance. Working and coauthoring with both these individuals has taught me much about our related interests. allowed me the flexibility and support to explore new avenues of methods research without the micromanagement that would have pained us both. Curtis Signorino. At Rochester. the story begins at Nuffield College. Oxford. where Iain McLean was a sage source of counsel and helped me garner admission to the University of Rochester in 2003. there are several individuals who I wish to thank for their contribution to my education. which coincided with the stress of the market.” Those things apparently include my completion of a dissertation in political science at the University of Rochester. and has been a constant source of helpful advice since that time. Curt provided much needed pastoral care when I felt particularly frustrated and anxious—especially in the last year of my dissertation writing. With this in mind. although even Sir Thomas would be surprised to learn quite how ordinary one’s talents can be in practice. Chronologically. David Firth tried patiently to teach me some statistics.

who managed the star lab during the writing phase of my dissertation. Tony. Our projects together have been the most challenging and rewarding experiences of my career to date. Though not on my committee.v about how to ‘play the game’ of our discipline. and his diligence in reading and commenting on my papers is second-tonone. . Tasos Kalandrakis is one of the hardest working men in the business. but he has become an important member of my team in the meantime. ran a very tight ship and ensured that both the hardware and software I had access to was in the best possible working condition. All these people kindly spent time helping me improve my papers with lengthy and valuable comments. I am very grateful to Jeff Gill and Andrew Martin at Wash U. I thank Matthew Platt. Larry Rothenberg stepped forward when I needed an extra body on my committee for bureaucratic reasons. Michael Herron at Dartmouth and Kevin Quinn at Harvard. Michael Epstein. and I expect our cooperative efforts will continue for some time yet. David Carter. I am very grateful to Michael Peress and Tony Lenzo who both supported my work in quite different ways. I thank him for his time and advice. My comrades in the department were constant sources of amusement and encouragement throughout the writing of this dissertation. Arnd Plagge and Subhashish Ray. Michael’s deep understanding of econometrics allowed him to provide valuable technical feedback on my papers. Also serving on my committee. Peter Loewen spent an all too short visit at Harkness. In the broader methods community. in my cohort.

I remain truly grateful that I have found such a wonderful woman to spend my life with. I know only that my outlook—academic or otherwise—would be the poorer. my father.vi during which time we became firm friends. They were constant in their support. Antoine Yoshinaka and Kris Ramsay were a strong influence on my outlook and behavior as a graduate student. repay these debts. Finally. affection and financial assistance. my sister and my brother. . I wish to convey a very special thanks to my family that loved me whether my research was progressing or stagnating: my mother. I hope I can. one day. I thank Sonia Tasbas: without her love and kindness. In the early days. Justin Fox. perhaps most of all.

‘rebeling’ on government bills apparently hurts them. leadership rank. service length and gender all affect their ability to influence others in the chamber. inter alia senators’ party identification. The first is an application of reversible jump Markov chain monte carlo. We use both an unstructured and covariate based model to show that. In particular. ideological tendency. since the number of parameters (the number of roll calls) far exceeds the number of observations (constituency performances) to be predicted. . The second chapter uses the Bradley-Terry model for pairwise contests to estimate the ‘power’ of actors in legislatures. The third chapter shows that British Members of Parliament’s voting decisions on roll calls can affect their constituency performance at election time. American politics and Comparative politics. We apply RJMCMC to the current Iraq conflict in order to identify change points in terms of civilian casualty numbers. random forest algorithm to estimate the relationship. while voting independently on less important matters tends to benefit them. We use a non-parametric. We apply this generalized linear model to the United States Senate. a more general form of MCMC popular for model search problems in statistics.vii Abstract This political methodology dissertation consists of several distinct essays that apply three new estimation techniques to all three subfields of the discipline—international relations.

. . . . . . Estimation Problem . . . . . . . . . . . . . . . . . . . . . . . . . .5 Introduction . . . . .3 2. . . . . . . . . . Results . . . . . . . . . 1 3 2 ‘Turning Points’ in the Iraq Conflict: Reversible Jump Markov Chain Monte Carlo in Political Science 2.1 Plan . . . 6 6 8 10 15 18 Discussion . . . . . . . . .4 2. . . . . . . . . . . . . . . . . . . . . . . . . .5 Introduction . . Problems with Traditional Approaches to Measuring Power . . . .4 3. . . . . . . . . . . . . . . . . . .viii Contents 1 Introduction: Purpose and Plan 1. . . . . . . . . 3 Measuring Power in Political Science: A New Method with Application to the Senate 3. . . . . . . . . . .1 2. . . . . . . . . . . . . . Statistical Theory and Model . . . . . . . . . . . . .1 3. . .2 3. . . . . .2 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 22 24 26 32 36 . . . . . . . . . . . . Structured Modeling . . . . . . . . . . . .3 3. . . . . . . . Background and Data . . . . . . . . . . . . . . . . . . . . . Application: 108th Senate .

. . .1 4. . . . . . . . . . Logit Model with Order Effect . . . .7 3. .3 4. .5.5. . . . . . . . . . . . .3 3. . . . 1997–2005 4. . .9 The “Busy” Senator Problem . Committee and Agenda Control . . . . . . . . . . . . . . . . . . . . . . . .4. . . . . . . . . . . Career Factors . .2 4. . .5. . .4. . . . . . . . . . . . . . . . .ix 3. . . . . . .2 . . . Constituents and the Personal Vote in Britain . . . . . 4 Rebels with a Cause? Legislative Activity and the Personal Vote in Britain. . . . . . . . . . . . . . . . 38 42 46 50 52 54 61 61 62 63 63 Discussion . . . . Tree-based Regression and Random Forests . . . . . . . .2 3. . . . . . . Formalizing Intuitions . . . . . . . . . . . . . .2 4. . . . . . . . . . . . . . .5. . . . . . . . . . . . .1 3. . . . . . .5. . . . 70 70 74 77 81 83 84 85 93 96 96 Data: 1997–2001. .10 Goodness-of-Fit . .4 3. . . . . .4 Introduction . . . . . . . . . . . . . . . . 104 . . . . Estimation Problem: p > n . . .5. . . . . . . . . .3 4. . . Summary of Findings . . . . . . . . . . . . . . . . . . . . . . . Data and Method 4. . . . . . . .1 4. . . . . . . . . . . . . .5 Results 4. . . . . Estimation Plan . . . . .4 4. . Geographic Factors . . . . . Second term: 2001–2005 . 3. . . Amendment Proposers . . . . . . . . . . . . . . . . . . .8 3. . . . .4. . . . . . . .4. . . . . . . . . . . .5 3. . . . . . . Members. . Appendix to Chapter 3 3. . . . . . . . . . . . First term: 1997–2001 . . . . .5. . . . . . . . . . . . . . . . . .6 Party and Ideology . . . 2001–2005 . . . . . . . . . . . . . . . . . . . . . .1 4. . . . . . .

. . . .8 4. . 114 Imputing Missingness . .6 Summary of Results . . . . . . . . . . . . . . . . . . . . .9 Regression Tree Split Criteria . . .5. 113 Model Fit. . . . . . . . . . . . . . . . . . . . 116 117 122 5 Discussion Bibliography . .x 4. . . . . . . . . . .7 4. 110 113 Appendix to Chapter 4 4. . . . . . . . . . . . . . . . . . . . . 109 Discussion . . . . . . . . . . . . . Variable Importance and Effect . . .3 4. .

. . . . . . . . Asterisked coefficients imply p < 0. . . . . Comments note political and other activities approximately contemporaneous to break. . . . . . . p < 0. 58 3. . .01 (***). Asterisked coefficients imply p < 0. . . . . 58 3. . . 57 3. AIC: 11136. . . . . . Asterisked coefficients (***) imply p < 0. . Asterisked coefficients imply p < 0. . . .4 Effect of conservatism.01 (***). . .2 Effect of majority party and senior leadership status on power in the Senate. . . . . . . . . . . . . . . . . .05 (**). . .5 Effect of extremism.3 Effect of majority party membership and any leadership status on power in the Senate. . . . . p < 0. . . . . . . p < 0. distance from median and majority status on power in the Senate.1 Change point median dates [90% HPD] in Iraq casualty data. .xi List of Tables 2. . . . . . . . . . . . . with rates between breaks. . . .1 Baseline results for model of power for US Senators in the 108th Senate. . . . . . . . . . AIC: 11052. . . . . . .05 (**). . . . . . 21 3. 59 . .01 (***). . . . . . .01. . . . . . . . . . . . . distance from median and majority status on power in the Senate. 58 3. AIC: 11152. AIC: 11169. . . . . . .05 (**).

. . land area (state size) and state population on senator’s power.9 Baseline results for model of power for US Senators in the 108th Senate. . 68 69 3. . First column is data code for the division with date after point: so. .6 Effect of committee membership. p < 0. 59 3.01 (***). Asterisked coefficients imply p < 0. . AIC: 10205. . .05 (**). . . . . 4. . . . . . . . .9. OLS coefficient in column 6. . . . .8 Effect of incomes. . . .7 Effect of incomes. . .05 (**).01 (***). .001107 is division 323 taking place on November 7. . . . . .xii 3. . . . . southern state representation. . . p < 0. . land area (state size) and state population on senator’s power.05 (**). .10 Number of amendments per senator. 60 3. Asterisked coefficients imply p < 0. . 60 3. majority party membership and chair status on power in the Senate. . p < 0. . . div323. . AIC: 9434. . . . . . . . . . . random forest ‘coefficient’ in column 3 for being a ‘rebel’(in the minority and/or frustrating the Government’s ambitions). . southern state representation. p-value for that coefficient in column 7. . . . . . . . Asterisked coefficients imply p < 0. .1 ‘Most important’ parliamentary votes for predicting electoral performance by backbench Labour MPs in 2001 general election. . . . . . 100 . .01 (***). . logit with ‘order effect’ (party advantage) estimated. . . Name of bill in column 2. . . AIC: 10954. . 2000.

. . random forest ‘coefficient’ in column 3 for being a ‘rebel’(in the minority and/or frustrating the Government’s ambitions) . .2 ‘Most important’ parliamentary votes for predicting electoral performance by backbench Labour MPs in 2005 general election. . 107 . . Name of bill in column 2. . p-value for that coefficient in column 7. . . OLS coefficient in column 6. .xiii 4. . First column is data code for the division with date after point.

.2 Extremism versus power (λi ). . . various dates of interest demarcated. . . . 47 . . 19 3. . . . . . . Change in rate of casualty incidents over time: rate on y-axis. . . closed circles are Democrats. . . . . . . May 2003–May 2007. . Dates correspond to breaks found. bottom figure is for all proposing senators. actual incidents as jittered points. . . . . . closed circles are Democrats. Iraq casualty data. . . Top figure is for all senators.3 Boxplot showing relative power of committee chairs. . . . Top figure is for all senators. . . . . . solid line is loess. solid dots are jittered incident occurrences. . . . . . . open squares are Republicans. . . . . 11 16 2. non-chairing Republicans and Democrats who propose amendments. . . Solid line represents cumulative counts. .1 NOMINATE score versus power (λi ). . 43 3. .xiv List of Figures 2. open squares are Republicans.1 Iraq Casualty Incidents. . . . solid line is loess.2 2. . . .3 Posterior of k: number of change points. . bottom figure is for all proposing senators. . . 44 3. . .

. . .3 Example of a tree that uniquely classifies American states. . . Source: Johnson and Rosenblatt (2007. .5 Variable importance plot for 1997–2001 data. . . . . . . . . . Node labels removed for clarity of presentation. .7: lighter states are associated with more powerful senators. 75 4. 2005. . . 90 4. . . . . 51 53 64 3. . . For each bag. .1 Boxplot showing relative power of male and female senators. . . . . . Multiple choice. . . . . . . . . Table 1).2 Example of a classification tree for American states. . . survey dates are 1983. . . . . . . . . . . Controls and ‘important’ roll calls used in subsequent analysis shown as subsets of all variables. . . . . . . . . . . Here the data set consists of ten observations labeled 1 through 10. . . . . . x-axis gives the increase in the regression performance from including each variable.4 Random forest schematic. . . . . . 166. . . . . . . Amendment numbers: distribution. . . These are ‘bootstrap sampled’ into bags (1 thorough B). . . 97 . . . . . . . . . . . . . . .xv 3. . . . . . . . broken line is percentage naming ‘loyal to party’.’ A tree is grown from each bag. . . more than one choice permissible. . . . . . . . . . . predicting (categorizing) the percentage of vote for George Bush in the 2004 Presidential election. . . . .5 3. 89 4. . 2000. Percentage of respondents naming different qualities as “important for an MP to have”: solid line is percentage naming ‘independent minded’. . some observations are ‘left out. in terms the percentage of vote for George Bush in the 2004 Presidential election.6 4. . 1996. . . . . . 94 4. . . . . and the results aggregated. 1994. .4 Map of contiguous United States with shading proportional to ‘power’ given by coefficients in Table 3. . . . . . . . . .

. . . . . . . . x-axis gives the increase in the regression performance from including each variable. . . . . . . 106 .xvi 4.6 Variable importance plot for 2001–2005 data. . . . . . . . . . Controls and ‘important’ roll calls used in subsequent analysis shown as subsets of all variables. . . . . .

and the questions that are actually answered. the substantive topics of investigation. while . the following dissertation is composed of three self-contained essays within the broad rubric of ‘political methodology. one of the papers is avowedly Bayesian. the essays develop tools that other political scientists can use to analyze the data scenarios they face. The stress on ‘quantitative’ is important since the chapters below all make claims that attributes of political events and actors can be measured and compared. while two are not. this term refers to the use of quantitative techniques that can give scholars insights into political phenomena. the dissertation is eclectic: in terms of the models of inference. Though these themes are consistent threads. In line with this.1 Chapter 1 Introduction: Purpose and Plan As befits its title. Two papers utilize parametric methods. To this end. the nature of the techniques used.’ For current purposes.

” For Achen (2002. patternrecognition and market research literatures. argues that “we have yet to give most of our statistical procedures legitimate theoretical microfoundations. The model might emerge . the chapters are not ‘solutions in search of applications’. an argument is made for the absence of an existing method to solve the specific problem addressed. The current dissertation is more sanguine about such endeavors: it borrows techniques more commonly seen in bio-statistics. The papers cover the three broad areas of substance seen in the discipline: International Relations. as will be explored. 1989. American and Comparative politics respectively. King (1989). for example. In any case. for example. communication and the cumulation of knowledge within the discipline. 3).0 2 the third uses non-parametric techniques. he warns against techniques which are “imported intact and without adaptation” from other disciplines (King. Moreover. Pursuing such a diverse agenda in one work is not uncontroversial. A second concern for work in political methodology is that of specifying appropriate ‘theory’. the particular problems here are neither uniquely rare nor trivial—that is. in reviewing the technical literature in each essay. maximum likelihood) in political science that would facilitate the application of consistent standards. if assumptions (and thus limitations) are clearly described. argues for a ‘unified’ methodology (in practice. Achen (2002. The contention is that. readers may judge for themselves whether or not the model is appropriate. 437). With good reason. 424). a micro-foundation is “a formal model of the behavior of the political actors under study.Section 1.

as argued above. introducing new techniques to the discipline where current ones simply cannot be used is a fruitful exercise.” In statistics. Then the statistical setup is derived from the model. especially as they relate to polity building. or some other formalism. game theory. These are arguably congruent with insurgent responses to widely publicized events including the arrest of Saddam and the . it is a thin line between ‘data mining’ and ‘data dredging’ but these essays make a good faith attempt to avoid the latter. with no further ad hoc adjustments.’ Of course. 1. rather than establishing “reliable empirical generalizations.Section 1. In any case. We find four ‘turning points’—all of which correspond to increased civilian casualties.1 Plan The first essay (Chapter 2) uses a reversible jump Markov chain monte carlo approach for examining structural breaks in the current Iraq conflict. In the first and third chapters it is not.1 3 from decision theory.” In the second chapter. such endeavors are common and their usefulness is in part derived from the fact that they simplify and describe (for the analyst) an otherwise complex picture of information. this prescription is broadly followed: a model of human behavior is suggested and an estimator derived directly. As a first pass on the data they are helpful: previously hidden regularities emerge and help scholars (including the present author) to think more systematically about what is driving behavior ‘under the hood. The interest is in identifying patterns of violence. That is in part because these other essays are more concerned with data exploration and summary.

1 4 holding of elections. Methodologically. empirically. contrary to some expectations.Section 1. This pa- . we show that. Thus the chapter makes a contribution to the systematic study of international conflict and comparative political development. In contrast to a priori pivotality indices (of which Shapley-Shubik and Banzhaf are most famous) which have been criticized for their measurement of ‘luck’ rather than ‘power’. this essay introduces a data-driven Bradley-Terry approach that proceeds directly from a theoretical model of (pairwise) actor behavior. chamber centrists lack power relative to party medians. female senators tend to be as powerful (on average) as males. The third essay (Chapter 4) uses a non-parametric regression approach to examine the relationship between a member of the British parliament’s legislative activity and their constituency electoral performance. we are able to demonstrate that. The second essay (Chapter 3) suggests a new way of measuring ‘power’ for actors in structured settings like legislatures and courts. In contrast to assumptions of no effect. with individuals such as John McCain. we show evidence of a subtle reward/punishment dynamic. contingent on the nature of the roll calls in question. In its unstructured form. We apply the model to the 108th Senate. covariate model is used. When a structured. Bill Frist and Robert Byrd appearing in the power top ten. we are able to produce a valid rank ordering of senators. Moreover. the chapter adds to the political science toolkit by suggesting a more general Bayesian approach to change point detection that is philosophically appealing.

Section 1. Methodologically.1 5 per combines quantitative and qualitative research on roll calls in the years 1997–2001. . and for which the causal story is potentially complex and non-linear. We conclude in Chapter 5 with a discussion of future applications and extensions to the work here. we describe a new ‘random forests’ approach for political scientists interested in problems for which the number of parameters far exceeds the number of observations. along with some comments on the role and future of political methodology more generally.

the current situation in Iraq provides both a testing ground for theories on the duration and termination of different types of conflicts (e. In part due to its contem- . military and economic event in progress. 2006).6 Chapter 2 ‘Turning Points’ in the Iraq Conflict: Reversible Jump Markov Chain Monte Carlo in Political Science President Bush believes that the region is at a true turning point. As an international conflict that increasingly resembles a civil war. National Security Advisor (2003) 2. 2004. This is quite separate from its obvious importance as a political. Filson and Werner.g. Stam and Bennett. —Condoleezza Rice.1 Introduction The study of inter – and intra– state conflict is a mainstay of political science. He believes that the people of the Middle East have a real chance to build a future of peace and freedom and opportunity. as well as a rich source of data for empirical work.

in undertaking our study we justify and adopt a novel (to political science) approach that uses a more general form of Markov chain monte carlo (MCMC) techniques. and negotiations to form. the utility of any data is only as good as the way it is explored and analyzed. Here. (4) the assumption of security and some military responsibilities by the Iraqi . daily recorded.1 7 poraneous nature. the first full-term Iraqi government (the early months of 2006). In keeping with the increasing acceptance and popularity of Bayesian methods in political science. These breaks are approximately contemporaneous with (1) the capture of Saddam. Of course. Examining civilian casualty data from the official cessation of hostilities (May 2003) to May 2007. (2) the installation of the Iraqi Interim Government. and the emergence of the Abu Graib scandal (late 2003 to Spring 2004). 1995). political scientists have access to carefully. we suggest that a fruitful approach for political scientists lies in examining the time series for (potentially multiple) structural breaks and their effects. well-known to statisticians as ‘reversible jump’ MCMC (Green. (3) the legislative elections for.Section 2. military and civilian casualty information: an unusual and excitingly fine level of detail. we find evidence of four change points. For scholars of American politics and public policy. and the subsequent handover of power to the Iraqi Transitional Government (Summer 2004 to early 2005). the way that these change points correspond with administration statements on the progress of the war may be particularly intriguing. We do so primarily for computational reasons. This notion extends to Comparative institutions scholars interested in the potentially pacifying effect of various post-war ‘state-building’ activities.

(2006) claim up to 600. on March 20th.000 deaths).Section 2. 2003. Baghdad and Saddam Hussein’s home region of Tikrit was under allied control—bringing a de facto end to the war. By mid-April.000 civilian fatalities since military operations began (sources are http://icasualties.2 Background and Data The United States and allied forces attacked Iraq with aerial bombardments. is the continued loss of life after the Iraqi army was formally defeated. some 3.iraqbodycount. the capture of Saddam (December. followed by a land invasion. on May 3rd. Violence has not yet abated despite the passing of some presumably important landmarks in what some characterize as the development of Iraq’s polity and stability: for example. President Bush declared that allied combat operations would now officially cease. Burnham et al. What marks the Second Iraq War though. National Assembly elections (January 2005). 2003. Iraq’s capital city. the killing of Saddam’s sons. In every case. 2. Some studies have placed the number of civilian fatalities at a much higher number. the drafting (December 2003–March 2004) and subsequent referendum approval (October 2005) of a .org/ respectively. the frequency with which such incidents occur is increasing after the break. As with all conflicts. Uday and Qusay (July 2003). At the time of writing. 2003).org/oif/ and http://www. A fortnight later. the war has not been costless.000 coalition force members had died in addition to at least 57. the placing of the former dictator on trial from ‘crimes against humanity’ (July 2004) and his execution (December 2006).2 8 government (August/September 2006) . For example.

thought responsible for planning many terrorist attacks (June 2006). the data base reports a range of possible death numbers from a ‘minimum’ to a ‘maximum. Our data are drawn from iraqbodycount. our study will enable us. the election of a new president (April 2005) and the forming of a governing coalition (May 2006). 2006).g.Section 2. insurgent and terrorist attacks)” (Dardagan and Sloboda. We are interested in violence for the post(official) war period: although we certainly cannot make firm causal claims. The data in raw form record deaths at the day level. from January 2003 through to the present and are compiled from (primarily Western) media reports and other sources. Since uncertainty often exists on precise numbers. Hence our study focuses on May 3rd 2003 through to the present time of writing (May 2007). The count includes civilian deaths caused by coalition military action and by military or paramilitary responses to the coalition presence (e.2 9 constitution. especially when different agencies have conflicting figures for the same incident. so we use the ‘minimum’ and define a ‘casualty incident’ as involving five deaths or more (our findings below are similar when we define . Abu Musab al-Zarqawi.org a (online) data base that records civilian deaths in Iraq “that have resulted from the 2003 military intervention by the USA and its allies.’ Potential ‘over-counting’ is a concern. to make statements about the plausibility of various events as “turning points” and allows us to pass some exploratory comments on how new democratic institutions and state apparatus developments are effecting Iraqis. the assumption of security responsibility by the Iraqi government (September 2006). the execution of an Al-Qaeda ringleader. for example.

That treatment is similar to the (hierarchical) presentation given by Carlin.Section 2. yT ) is a vector of observations of the random variable Y (casualty incidents) over time and let f and g be unknown densities in the same parametric family with yi ∼ f (Y |λ1 ). . rather than their size (above our minimum). Although univariate time series work is not regularly encountered in political science. k.3 Estimation Problem The single change point problem. and we graph their occurrence in Figure 2. We also report various dates that may of interest and to give readers a sense of timing perspective. we focus on the (changing) frequency of attacks.1. terrorists were able to control how often they planned to inflict casualties. . There were 1682 such incidents in our time series. . . In part this is a behavioral assumption: we would contend that. the solid dots are simply jittered incident occurrences (for which the y-axis is not the scale). T . i = 1. . . i = k + 1. We wish to estimate k the (single) change point which takes (discrete) values . it is valuable in the current context as a ‘first glance’ exploration before covariate information becomes available. We think that such work helps to prompt both theorizing and data gathering for more nuanced and sophisticated analysis. . the solid line is the cumulative incident count. For the purposes of this paper. estimated using Markov chain monte carlo techniques. 2. . . Gelfand and Smith (1992): suppose y = (y1 . . yi ∼ g(Y |λ2 ). has been discussed for and by political scientists elsewhere (see Western and Kleykamp.3 10 the incident threshold at ten or twenty deaths). there. rather than how many. at least initially. 2004). .

2003) Figure 2.1: Iraq Casualty Incidents.3 11 1500 First full term govt takes power Cumulative Count 1000 Constitution writing finished 500 Elections for transitional govt Transitional Admin Law Saddam captured Uday. May 2003–May 2007. . Qusay killed 0 0 500 1000 1500 Time (days after May 03.Section 2. various dates of interest demarcated. Solid line represents cumulative counts. solid dots are jittered incident occurrences.

where the observations are (assumed) drawn from latent state variables. Gill 2002. A Bayesian approach proceeds by placing a prior τ (k) on the change point. 486).1) requires optimization in a space that is not continuous (recall that k is discrete) which.Section 2. . . 2. suggested by Chib (1998) and applied to American politics by Park (2006). 1998). . there are profound computational difficulties in generating proposals for situations where we suspect there are more than a couple of change points. for example. Here. . 2003). treats the change point model as a type of time series Markov mixture model. we are interested in exploring multiple change point and such work (Bayesian or otherwise) is much less common in political science.1) to obtain k and the parameters λ1 and λ2 (which for the count case are arrival rates for a Poisson) if they are of interest. with respect to the logic above. In part this is because. One approach. (c) missingness in y is handled systematically. 1–6 and Jackman 2004. say. Notice that this approach requires separate Markov chain monte carlo runs for the different numbers of change points hypothesized (Leonte. . T }. This is quite apart from the philosophical appeal of Bayesian approaches of which political scientists are increasingly aware (see.3 in {1. (b) the resultant non-nested models may be straightforwardly compared using Bayes factors (Chib. Gibbs sampling does not. Nott and Dunsmuir. A frequentist approach proceeds by maximizing k T 12 L(y) = i=1 f (yi |λ1 ) i=k+1 g(yi |λ2 ) (2. There are computational advantages of a Bayesian MCMC approach here since (a) maximizing (2.

θk )p(k. between and after them. K} candidate models of our data generating process. λ1 and λ2 ). θk )p(θk |k)p(k). Otherwise put.3 13 An alternative solution is to use reversible jump Markov chain monte carlo which allows us to complete the computational operations in one ‘go’ as well as allowing us to be a priori agnostic over the number of parameters to be estimated. . each with a different number of parameters. we need to estimate 2k + 1 parameters—the change points themselves and then parameters of the densities before. the number of parameters is. (2. . the k th model in Mk has associated parameter vector θk which contains nk parameters such that θk ∈ Rnk . y) = p(y|k. of itself. θk . Now consider a very different scenario which arises for an unknown number of k change points: for every possible k. we have a set of Mk = {1. Typically when MCMC is used in political science the parameter vector θ has a known number of components. That is. a parameter. . denoted n.Section 2. More formally. .2) . Continuing to denote our data vector y. θk ) = p(y|k. For the single change point problem n = 3 (these are k. the joint distribution becomes: p(k.

θk ) . but to also ‘jump’ from space to space (from model to model) depending on the k in question. Although well known to statisticians. Setting up a Markov chain to do this may be difficult though.3 14 Since we have a constant of proportionality we can rearrange and reexpress (2. known as reversible jump MCMC (RJMCMC). We wish to generate samples from (2. θk ) is simply the likelihood.3) Notice that p(y|k. This type of problem is given a general formulation by Green (1995).3). in particular the efficiency of proposals.Section 2. posterior prior(s) likelihood (2. because it is required not simply to move around the parameter space for any particular θk . the details somewhat technical. of which standard MCMC algorithms are special cases. given a particular data generating process and p(k) is the prior on the model itself. and readers are guided to Brooks (1997) who gives an accessible overview for political scientists.2) into the more familiar p(k. The implementation of RJMCMC. Green explicitly discusses a Poisson count change point problem and we followed his approach for our application (though we varied the priors somewhat to ensure that our results were robust to such alternative specifications). can be problematic in practice and Hastie (2005) devotes considerable . while p(θk |k) is the prior for the parameter vector. θk |y) ∝ p(k)p(θk |k) p(y|k.

when did they occur in the data?” .4 Results There are three sets of (posterior) distributions that interest us here: 1. More intuitively. The first two stages of the sampler produce a Normal mixture distribution for every possible value of k. .Section 2. the standardized vector is transformed using the mean and the covariance matrix of the mixture component lk . θk ) with some specified acceptance probability. 2. this enables us to answer the question “given a particular number of change points. along with a commensurate (new) mixture which has component lk . . assuming the Markov chain is currently in state (k. 202–203). the posterior of k: this enables us to answer the question “how many change points in the data?” This will have support k = 1. A particularly pleasing feature of this software is that issues such as burn in and the requisite number of post-burn iterations are handled automatically. . Automix allocates the parameter vector θk to a component lk of the mixture and uses it to standardize θk . θk ). drawing on Hastie (2005. it is instructive to summarize the way that the model of the data generating process is selected. Then a new model k is proposed. . kmax where kmax = 10. We used his Automix sampler (with a maximum of ten possible change points) for our estimation. In the third stage. Automix then accepts the proposed state (k . To obtain the new state vector θk . Though the full details are somewhat technical.4 15 attention to designing a technique to do this. the posterior of change point positions conditional on some estimated k. 2.

1 0. 3.2: Posterior of k: number of change points. the number of change points. the posteriors of the rates for each period. Iraq casualty data.2 we display the posterior for k. The strongest evidence (in the sense of Kass and Raftery (1995)) is for k = 4 and we will explore this possibility exclusively. we can answer “what were the effects of the change points?” In Figure 2. given the number of change points. The first was the capture and arrest of Saddam at a farmhouse near Tikrit in December 2003. in late January 2004 occurs between incidents that may be of import.5 0. In Table 2.0 0.2 0.6 1 2 3 4 5 6 7 8 9 10 k Figure 2.4 0. and when they occurred.Section 2. The first break.3 0.4 16 0. conditional on some estimated k: that is.1 we summarize the results for k = 4 model in a way that answers questions 2 and 3 above. The subsequent months saw both an insurgency uprising lead by rebel Shia cleric Muqtada al-Sadr in Baghdad and the diffusion of abusive photographs taken at the Abu Graib prison where coali- .

doubling from one incident every four days. The break marks a sharp increase in the casualty rate. in February 2006. As part of this offensive. occurs not long after the elections for the first full term Iraqi government (December 2005) and at around the time of the protracted negotiations to form a new coalition government. The third break itself. This new entity.Section 2. under the Premiership of Iyad Allawi was subsequently recognized as the legitimate sovereign government of Iraq by both the United Nations and the Arab League (an important regional player). arguably.4 17 tion forces were holding Iraqi detainees. This change point saw an increase in violence from one incident every two days. to four incidents every five days. a little while after the Iraqi Interim Government assumed power from the Coalition Provisional Authority (in June 2004). leader of the Islamic Dawa . the Iraqi Interim Government began to censor the critical reports of media outlet al Jazeera. The political fallout of the latter was profound. to one every two days. and criticism of the Bush administration by allied. lasting from December through to April of the following year. This event. Arab and other politicians was widespread. Jawad al-Maliki. The second break occurs in August of 2004. rallied and spurred sectarian hatreds and violence. Allawi quickly announced new security measures to tackle insurgency forces and was criticized by some for their draconian nature. Included in the highest posterior density interval for this break is the January 2005 democratic elections for the Iraqi Transitional Government. These talks were deadlocked for some time.

violence surged after this point with. to tame insurgency (with coalition logistical and medical support).5 Discussion Our study—to our knowledge the first that uses RJMCMC in a political science context—suggests that violence is increasing and that important state-building activities. Apart from this rather grim substantive conclusion. In Figure 2.5 18 Party would become Prime Minister after the original candidate Ibrahim al-Jaafari proved unacceptable to the Sunni and Kurdish representatives in parliament. For reference. like democratic elections. . The first specific task of the Iraqi Security Forces was. and is.Section 2. are contemporaneous with upticks in casualties. We hope that our discussion here will encourage others in political science to consider such methods in future. five incidents occurring every three days.3 we summarize our findings in a different way: the open circles represent the median incident rate between the relevant breaks which are demarcated by the broken lines. a time when the Iraqi government assumed control of national security for approximately 70 percent of the country. we again draw the jittered incidents themselves on the plot. Once again. The fourth and final break occurs in September of 2006. By now incident rates were approaching three per day. 2. on average. then reversible jump techniques seem most helpful. we found that investigating time series on violence to be an interesting and fruitful exercise. If a Bayesian approach is pursued.

.Section 2. Dates correspond to breaks found.5 19 3 Sep 2006 λ 2 Feb 2006 1 Aug 2004 Jan 2004 0 0 500 1000 1500 Date Figure 2.3: Change in rate of casualty incidents over time: rate on y-axis. actual incidents as jittered points.

1 resembles a exponential curve of form F (t) = aert + (we are indebted to an anonymous TAS referee for this observation). This suggests some interesting avenues for future research: for example. we do not establish causation in any sense: the events we noted were simply occurring at around the same time as the breaks in the time series and it is speculative that they may be of direct importance.5 20 As noted above. On this point. We leave this for future research. Modeling—both theoretical and statistical—of this apparent pattern would allow us to think about the development of the conflict in Iraq in a more systematic way. one possibility is that increasing violence is a product of an increasingly organized insurgency.Section 2. . notice that the solid line in Figure 2.

2004–May 4. 2006 [Feb 16.476 Aug 7. Comments note political and other activities approximately contemporaneous to break.5 Period 0.Section 2. 2005] 1 incident every 4 days Capture of Saddam (Dec 2003) Abu Graib scandal (Apr 2004) 1 incident every 2 days Iraqi Interim Government assumes power (Jun 2004) elections for Iraqi Transitional Govt (Jan 2005) Time Rate (λ) Comments before break 1 break 1 btwn break 1 and 2 break 2 btwn break 2 and 3 Feb 22. 2006–Mar 24.283 Jan 26. 2003–Apr 30.814 4 incidents every 5 days elections for first full term Iraqi government (Dec 2005) break 3 btwn break 3 and 4 5 incidents every 3 days Iraqi govt assumes control of counter-insurgency operations for much of country (Sept 2006) 8 incidents every 3 days break 4 after break 4 Table 2. 2006 [Aug 23. 2006] 1. 2006] 2. 21 .660 Sept 4. 2006–Sept 20. 2004 [Jul 18. with rates between breaks. 2004] 0.1: Change point median dates [90% HPD] in Iraq casualty data.850 0. 2004 [Jul 10.

when and how” (Lasswell. —Honor´ de Balzac e 3. Later theorists—particularly those associated with the ‘Rochester school’ (Amadae and Bueno de Mesquita. Downs. Early positive political theory made the case that median voters—be they in the electorate or in parliaments—were particularly decisive and hence ‘powerful’ (Black. 1987. . To determine “who gets what. but by striking true. 1948. Riker.22 Chapter 3 Measuring Power in Political Science: A New Method with Application to the Senate Power is not revealed by striking hard or often. 1996. 1962). political scientists have invoked notions of ‘power’ (Dowding. 1936). 1–8).1 Introduction Scholars have long conceptualized politics as a process of conflict over resources.

Section 3. Krehbiel (1991)). A strength of this approach is that the definition of power is theoretically removed from factors that influence how powerful individuals actually are.1 23 1999)—have focused on the role of institutional rules and norms in giving ‘players’ advantages over one another in settings like Congress. we suggest an actor-based. 1986) and others (e. institutional arrangements. enormous. which includes regressors that are thought to covary with actors’ power. We demonstrate that this trait is straightforwardly uncovered by studying easily available voting records. geography and . Following Dowding (1996). by now. the empirical investigation and measurement of power—even in structured settings like parliaments. committees and courts—has proved difficult and contentious. we treat ‘power’ as a latent variable or ‘ability’ possessed to varying extents by actors like Congressmen (Senators in particular). Methodologically. 1998. 2004). Thus. and includes foundational work by Riker (1982. our contribution to the discipline is the introduction of a new version of the Bradley-Terry (Bradley and Terry. This literature is. This paper is an attempt to address this problem and to bridge a gap between the theoretical treatments of power familiar to positive political theorists and the empirical work of political methodologists. Shepsle and Weingast (1987).g. data driven approach. concentrating on a priori metrics that are removed from the data that we have regarding agents’ actual actions and decisions (Felsenthal and Machover. Shepsle (1979). We show the utility of both the conceptual and empirical suggestions by studying ‘power’ in the 108th United States Senate where we show the role that ideology. Despite this theoretical work. 1952) model for pairwise comparisons.

Section 3. To wit. committee assignment.3 we suggest.2 24 personal factors play in determining the power of its members. Marx and Das Kapital.b).1 Beginning with Shapley and Shubik’s seminal 1954 contribution. inter alia. 2002a. with extensions most famously by Banzhaf (1965). in Section 3. in Section 3. in Section 3. Second. when voters have different weights.2 we discuss the problems with traditional voting index measures of power. 2002a. 3.5 we show how the power can be predicted by its proposed ‘causes. Third.” 1 . different indices yield different results for the same data and. there Social scientists have discussed ‘power’ in various ways for well over a century: at least since.4 we apply the model to the 108th Congress and produce a ‘power list’ of Senators therein. First. the resulting ambiguity cannot be resolved because there is no objective evidence on the actual distribution of power (Leech.2 Problems with Traditional Approaches to Measuring Power Scholars have been interested in measuring the power of actors in structured settings for over half a century. party. in Section 3.6 we conclude and suggest future avenues for research. The American Political Science Association felt it was sufficiently important and unresolved even in 2006 to devote their annual meeting to “Power Reconsidered. generally.’ and we discuss the impact of. derive and discuss a new data driven model. this branch of positive theory has failed to find widespread acceptance in political science for at least three reasons. for institutions with large numbers of actors. In Section 3. for example. geographic and personal factors. there are computational difficulties with performing the requisite calculations (Leech.b).

3 See also Krehbiel (1998) for a development of ‘pivotality’ with respect to US legislatures.’ A new method is needed which. are not ideal tools for studying what political scientists conventionally think of as ‘power. Thinking this way enables political scientists to break away from some nonsensical statements such that the power of the Chief Justice in the Supreme Court is ( 1 ). or that the Chairman of the Finance 9 Committee in the United States Senate is is equal in power terms to a fresh1 man senator who has never held a committee post ( 100 ). 1964). inter alia. that is. Moving on from this notion.2 25 is much sceptism on the question of whether the indices actually measure ‘power’ at all (Barry. they measure something akin to ‘luck’ or ‘decisiveness’ rather than power. Barry (1991) argues that.Section 3. 1996. since power indices are based on the probability that an actor is pivotal. power indices tell us about the distribution of power before we consider the preferences of actors or their ability to set an agenda (see Felsenthal and Machover (1998) for a comprehensive review. Dowding. 2 .3 The implication is that one’s power index score is simply one of several independent variables explaining or ‘causing’—but theoretically distinct from—an individual’s ‘power’ (which might be thought of as a latent variable).2 The point is that any definition or measurement of power must incorporate the notion of getting one’s preference in the face of resistance from other actors. Dowding (1996) considers pivotality as a resource among many. See also: Lane and Berg (1999) and Holler and Widgren (1999) in their response to Garrett and Tsebelis (1999)). 1991. The third criticism is the central issue here. Riker. A natural defence is that power indices are intended for a priori analysis. In sum: power indices whatever their hue.

. J index the bills on .3 Statistical Theory and Model ‘Power. estimates an unobservable trait. allows for explicit comparison of the effects of different predictors. members require majorities to pass bills and an actor’s power in these settings will turn on his capability to form legislative coalitions for the issues he cares most about. precisely because they stand to lose from its passage. 3. 5. we show one way to achieve these aims. 2. . Since we cannot measure it directly. . . . . treats power as a (latent) capability. defines power as the capacity to get what an actor wants in the face of resistance. estimates an actor’s power as a function of independent variables. In the case of legislatures.Section 3. in the form of actors who vote against the bill.’ as considered here. . . . and outputs a metric. 3. power is an actor’s ability to obtain his preferred outcome. . After introducing a little notation. incorporates commensurate statements of uncertainty and. senators’—capabilities to achieve their preferences in such settings Let i = 1. . . In the next section.3 26 1. . in the broadest sense. 4. is a latent variable. I index the senators and j = 1. which are our concern here. we need a statistical model that takes observable data. . There is resistance. We made the case above that. we now provide a way to systematically measure actors —in fact. . .

but this is not strictly required here. In any session of the Senate. Notice that for any particular bill j proposed by i. since all bills are proposed by someone. there will be not more j than J such individuals. is assumed to seek a coalition j to vote for the bill and we denote that coalition cip which has |cip | memj j bers (one of whom is the proposer). It is natural to assume that ip prefers j a coalition that is a majority of all senators voting on j.3 27 which they vote. we suppose that the benefits of passing j—be they pork. there is some simple opportunity cost of the time they take to vote. For any particular j. In attempting to form the coalition. perhaps because the total size of the pork or publicity ‘pie’ available is now decreased for them and their home districts. The proposer of bill j. We could think of this as increasing the probability that j is passed. say j + 1. argue. The j nature of the convincing could take many forms: the proposer may attempt to reason. denoted ip . Else. We say that j+1 j two senators ‘interact’ in a particular session of Congress if one proposes a bill and the other joins its backing coalition. credit.4 For now. or else the inducement may be implicit. the situation may well be reversed with the previous proposer now joining a coalition ci+1p backing a senator i + 1 who had been a member of cip .Section 3. j We assume rather that a proposer is coalition size maximizing: his utility is increasing in |cip |. ip . but not all senators propose. we say that ip ‘convinces’ a senator to join cip if she subsequently votes in favor of the j j bill (and hence in favor of ip ’s ideal point relative to the status quo). In subsequent votes. there are cip − 1 interactions. threaten or bargain with his colleagues explicitly. there exists a senator who proposes the bill. news headlines—will accrue only to ip and that j other legislators pay some cost in supporting the bill. j 4 .

(3. For any given interaction (any particular bill) between a pair of senators. π BA 1 − π AB αB (3. the statistical model we discuss below will determine a score λi such that if we know only that Senator A and Senator B are involved in an interaction. Since an interaction must involve either A convincing B or B convincing A it must be true that πAB + πBA = 1. 1) be the probability that the interaction involves Senator A convincing Senator B.5 Write the odds that A convinces B as a function of their latent powers. .2) 5 An objection may be that party or some other factor compels senators to vote one way or the other. the probability that it is A convincing B (i. The derivation of the estimator begins by assuming that there exists a latent and hence unobservable power for each senator denoted αi .4 we ameliorate these concerns somewhat but see also our ‘order effect’ specification below.Section 3. such that π AB π AB αA = = . in which case this equality will not hold.3 28 For each senator. is the difference of the scores λA − λB : logodds(A convinces B|A and B interact ) = λA − λB . let πAB ∈ (0. By studying personal amendments in Section 3.e.1) The interpretation is then straightforward: the greater the value of λi relative to other senators. the more powerful we hold that senator to be. B is backing A’s proposal) rather than B convincing A.

exp(λA ) + exp(λB ) (3. The problem of estimating αi can be approached via a logistic regression. 6 . To see how.4) We can then compare the relative size of λA and λB to see which senator is more powerful. 7 Importantly.Section 3. we have a logit model that can be estimated via maximum likelihood: logit[Pr(A convinces B)] = λA − λB . πAB ) distribution. and the same probability πAB applies to each interaction. Then. in any given interaction. That is.3). then A is also more powerful than C. nAB has a binomial (NAB .2). the model does not require that the matrix of interactions is ‘complete’ in the sense that every senator interacts at some point with every other.3) Suppose that A and B interact a total of NAB times.3 29 This formulation has the obvious consequence that. If we are also willing to assume that all the other interactions between the other senators are also independent. A is more likely to be forming the coalition (and gaining his preferred outcome) than B is. first let αi = exp(λi ). let nAB —where nAB ≤ NAB — be the number of times that A convinces B.7 αA αA From Equation (3. substitute αB for 1 and αB note that αi = exp(λi ) to obtain (3. some rearrangement yields6 π AB = exp(λA ) . (3. Then. for any given interaction between A and B. πAB > πBA . and B is more powerful than C. if αA > αB . it is obvious that π AB (1 + αB ) = αB . The model implicitly assumes transitivity: if Senator A is more powerful than B. so long as the NAB interactions are independent of one another. Of these interactions.

whatever the proposers power. Notice that our method here is very different to that of previous endeavors to measure power. Otherwise put. in contrast to ‘panel of expert’ surveys. Unlike index methods. unlike subjective rankings it is objective and relies on a clearly defined definition of power. obviates the need to ‘convince’ other legislators in any real sense. party pressures and thus a priori similar preferences between senators. Hence. and the λi will reflect the ability absent this advantage. that our findings (below) are exactly replicable by any other political scientist who chooses so to do. some senators are disposed to supporting others like them. We will tackle this issue head-on by estimating a version of our logistic regression as logit[Pr(A convinces B)] = λA − λB + δw (3. in fact. it is the Bradley-Terry (Bradley . This estimator is not new to this paper.Section 3.’ Second. Hence the terms ‘data driven’ and ‘actor based. it is much cheaper and faster to calculate. it is an a posteriori in the sense that we infer ‘power’ after observing actor’s actual decisions—rather than before. Moreover. This means.3 30 A concern with this formulation might be that. δ will capture the ‘natural advantage’ of proposing to someone in a senator’s party.5) where w is an indicator taking the value 1 if the proposer’s party is the same as that of an individual he convinces and −1 otherwise.

journal citation patterns (Stigler. 2002).Section 3.. 1994) and competition for mates in the biological sciences (Stuart-Fox et al. and Fowler (2006) who studies the ‘connectedness’ of legislators in both chambers—though not in an explicitly pairwise way.9 The paper is perhaps closest in spirit (though not in execution) to work by Wawro (2000) who studies legislative entrepreneurship in the House. R (R Development Core Team. like Wawro we provide scholars with a metric that may be used for further research.8 This is not the first paper in the discipline to suggest thinking of some political phenomena as pairwise interactions: for example. Wawro shows that partybased career prospects are strongly linked to a representative’s records of actively introducing legislation. 8 Such models can be fitted with many standard statistical packages. we build on this work by presenting congruent findings—though with a different causal direction: in particular. that party and committee advancement aids senators in their quest to pass (potentially controversial) legislation. This approach is well known and well studied by statisticians working in fields as diverse as sport team rankings (Agresti. 2006). 2005). 1952) pairwise comparison (with equation 3. 9 See also King (2001) who discusses some possible extensions of standard international relations models that assume dyadic interactions. Also. Groseclose and Stewart (1998) study the value of Congressional committee assignments using (dyadic) transfers of representatives. For this paper. .5 representing an ‘order effect’ specification) method in a novel setting.4 31 and Terry. Below. Designing a novel “entrepreneurship scales score” based on five observables of behavior. 2006) was the environment of choice in conjunction with the BradleyTerry library (Firth.

Section 3.4

32

3.4

Application: 108th Senate

We use data from the 108th Senate, that met between January 2003 and December 2004. The universe of cases for the present analysis is the 255 amendments, by 74 different senators, proposed and voted upon during this time. Amendments are preferred to bills as a whole since they “tend to reflect more specific changes to a bill that are less susceptible to deviations from the sponsor’s original intent” (Fowler, 2006, 9). That is, they are more amenable to the notion of personal coalitions formed to achieve an individual’s goals. Studying the Senate in particular has the further benefit that, subject to some constraints negotiated via Unanimous Consent Agreements, its members may propose amendments at essentially any time, in any order, without permission from a Committee of the Whole. This means that, unlike the House, the status quo is afforded less protection—an ongoing concern for those studying power (see Bachrach and Baratz, 1962, for example). There are 101 actors in the current data set: 74 proposing senators, 26 non-proposing but voting senators and the President (where his views on the amendments are known). Of the amendments, 112 were winning, 139 were losing, and four were tied votes. Recall that the number of observations is the number of interactions, and is thus well in excess of the number of amendments: otherwise put, each amendment corresponds to multiple interactions and hence multiple observations.

One concern might be that senators vote strategically: against their first preferences for an alternative that is a priori less preferred. They might

Section 3.4

33

also propose strategically, to ensure their preferred alternative is selected by majority rule (Riker, 1982, 1986). However, it is not immediately apparent that strategic (let alone ‘killer’) amendments are common. Second, strategic voting on amendments, as opposed to bills, should be relatively rare: we can generally assume that those voting for an amendment want it to pass. Third, strategic voting is presumably a product of the organizational structure—like the timetabling procedures—of the Senate: but this is precisely the sort of consideration that ought to be included in the calculation of our metric. Lastly, votes in the Senate are part of the public record, so we expect senators to treat their vote seriously and think through its consequences.10

In Table 3.1, we report a selection of the power estimates for this specification. Recall that there is no sense in which the power measure is absolute, and all the coefficients are scaled relative to zero which is assigned to President Bush.

Examining the table, we note that the long serving (since 1979) John Warner of Virginia (senior senator) is ranked first. Second is Mitch McConnell of Kentucky (senior senator). In the 108th Senate he was elected as the majority whip by the Republicans and, at the time of writing, was the leader of Republicans in the Senate. Judd Gregg, the senior (Republican) senaNotice that simply proposing (and forming coalitions) for more amendments cannot in and of itself increase a senator’s power in this model: see Appendix 3.7 and Appendix 3.9.
10

Section 3.4

34

tor from New Hampshire is at third. Gregg chaired the Health, Education, Labor and Pensions Committee in the 108th Congress, and became chair of the Budget Committee in the 109th. Robert Byrd, the longest serving senator in history is the first Democrat on this power list. From West Virginia, Byrd is, at the time of writing, the President Pro Tempore—a role he had previously held in the 1980s and 1990s. John McCain of Arizona, then chair of the Commerce committee, is followed by Bill Frist of Tennessee, the Senate Majority leader for the Congress in question. Thad Cochran and Charles Grassley, chairs of the Appropriations and Finance Committees respectively (in the 109th Congress) hold the 7th and 8th spots. Barbara Boxer, the junior senator from California is the first woman to feature on the list. At the bottom of the ranking, we see both senators from Hawaii (Daniel Akaka and Daniel Inouye) along with Jim Jeffords (Republican to 2001, Independent thereafter).11 Fitting an ‘order effect’ model such that the ‘natural advantage’ of proposing to like-minded partisans is controlled for—as specified in equation 3.5—makes little appreciable difference to this rank order: see Appendix 3.8 for more details.

Subjective rankings—published in magazines such as Time and by consultancy groups like Knowlegis—reach very similar conclusions to ours. The Knowlegis power list for 2005 for example, one session after the 108th, places Thad Cochran at 1 (7 in our ranking), Mitch McConnell at 4 (2), Charles Grassley at 7 (8), John McCain at 8 (5), Bill Frist at 10 (6), Arlen SpecImportantly, the rank ordering is very different to that which might be garnered from looking solely at the number of amendments proposed by each senator. See Appendix 3.9 for this rank ordering.
11

is based upon a vast and expensive survey incorporating what is theoretically much more information. for example. According to their website. and considered any characteristic or action that could contribute to their Sizzle-Fizzle factor. The Knowlegis list. .tt 12 . the same calculation for McCain and Rick Santorum yields a probability of (very close to) 1. by contrast. The probability that. researched relevant campaign contributions. reviewed thousands of media articles.58) + exp(19. model-based voting metric that is relatively straightforward to compute.4 35 tor at 9 (15) and Orrin Hatch at 2 (19). We collected data on the leadership. conditioned on two senators interacting. sorted and considered thousands of data points to determine . . Consider. it is McCain proposing the bill and Feingold backing it is exp(λMcCain ) exp(21. there are more than 10. . over a thousand amendments. in any interaction between these two. This similarity to our ranking is remarkable given that (a) they are discussing a different Congress and that (b) our approach is based on a simple. .congress. .000 data points and variables that were considered in the 2006 Knowlegis Power Rankings” source:http://www. .12 As noted above.Section 3. . hundreds of bills that passed out of committee . coauthors of the Bipartisan Campaign Reform Act of 2002. power. committee. exp(λMcCain ) + exp(λFeingold ) (exp 21. John McCain and Russ Feingold (Wisconsin).85) By contrast. . . This is hardly surprising given that the senator from Pennsylvania did not offer an amendment in the 108th Congress. a pleasing feature of the current estimator is that the power estimates λi have meaning outside of a simple rank ordering. . and caucus positions of each Member .85. .org/congressorg/power rankings/backgrounder.58) = ≈ 0. “Knowlegis staff carefully researched. Recall that λA − λB is the (anti-logged) probability that. . it is A that convinces B to back his amendment rather than the other way round.

1997). This figure is within the ballpark of similar statistics for industry standards like NOMINATE (see Poole and Rosenthal. in fact.5 but it lends some validity to the estimator. the notion that ‘the rich are powerful. for example. indeed. however. In sum.Section 3. it tells us little. at some 85% of interactions correctly predicted.’ A pleasing feature of the current approach and estimator is that we can separate these notions and explicitly incorporate individual specific explanatory variables as predictors of power. though given the structure of the underlying statistical model it is calculated somewhat differently.10 reports more details. respectable. Table 3.6) . Appendix 3. We estimate p λi = r=1 βr xir (3. 36 A further factor in favor of the new approach concerns the goodness of fit: in contrast to standard methods this is at least meaningful (what is the goodness of fit for a subjective survey?) and.1 seems to be a reasonable ‘influence list’ both in terms of its underlying statistical model and its actual contents. sometimes we treat characteristics that are causes of power as if they were synonymous with power itself: consider. 3. a subject to which we now turn. about the causes of power.5 Structured Modeling We usually have theories about what explains the power of different individuals.

power is a function of variables and is being treated in an ‘institutional’ sense.’ ‘Committees and Agenda Control. We used institutional and personal information to explain power in the Senate. we can interpret coefficient estimates as positive or negative.’ The intention here is not to provide an exhaustive account of power. in terms of their marginal effects on power. .13 Notice that a potential endogeneity concern—that individuals already endowed with latent power ipso facto obtain important institutional advantages—is being avoided by construction. the Government Printing Office and the United States Census Bureau provided all the relevant variables below in electronic form. We break the findings into four subsections dealing with ‘Party and Ideology. notice that the nature of ‘power’ now being considered is altered somewhat. xip with coefficients β1 . Before describing the results. Otherwise put. Collecting such data is straightforward: the Senate itself. We can thus make uncertainty statements about our predictors. it is now a maintained assumption that institutional (and other) characteristics make individuals powerful. .’ ‘Geographic Factors’ and ‘Career Factors. . βp (see Firth (2005) and Springall (1973) for details). 13 .5 37 and thus we predict the power of each senator as a (linear) function of explanatory variables xi1 . β2 . xi2 . In the previous section.Section 3. . separate from the individuals who wield it. . Now though. but to demonstrate the strength of the approach and possible avenues for future research. power was an ability to be ascribed to individual senators—it was a ‘personal’ characteristic. . Since the approach here is essentially a logistic regression. and we have commensurate standard errors.

In Table 3. a dummy that denotes either the senior Senator for each party. We also interact these variables.Section 3.2 we give the results of the model for a Majority dummy variable and SeniorLeadership. A refinement on this theme is that we expect leaders of Congressional parties to be especially powerful: the Senate Majority Leader (Bill Frist for the 108th Congress).5.5 38 3.2. for example. If a majority party exists (the Republicans for the 108th Congress) we might expect those from the majority party to be more powerful than those from the minority (Democrats). 2006). Other officials—which we refer to as ‘junior leaders’—such as Policy Committee Chairs have powers to design and execute policy ideas. The positive and significant effects of being a Republican and being part of the Senate’s senior leadership are evident from Table 3. there is increasing evidence that Congressional voting is ideological and party driven (McCarty.14 This is more true of the House than the Senate. 14 .6). Poole and Rosenthal. consider the power of some senator A who is a member the majority party See also Huitt (1957) who discusses cross cutting tensions of ideology and party as it applies to senators. From the interaction term it is evident that being a Republican and a leader adds an extra fillip to one’s power. As noted above. or the whips for each party. To be clearer here and recalling equation (3.1 Party and Ideology In contrast to the ‘textbook Congress’ of Fenno (1973). but nonetheless it suggests some testable hypotheses. senators must assemble majorities to obtain their own preferences. has the ability to schedule debate.

Section 3. 39 λA = 0.2.286. exp(1.205 × 1 + 0.205 × 0 + 0. then this probability rises to exp(1.754 × 1 = 1. Conference Secretaries and Senatorial Campaign Chairs. .245.2. Party Committee Chairs.286 × 1 + 0. This model fits the data as well as the previous model (note the similar values of the Akaike Information Criterion (AIC)).3 we estimate the same model with AllLeadership. a dummy that includes all senior leaders (as in Leadership) in addition several other categories: Conference Chairs. By Table 3. it is the Republican leader proposing and B supporting is exp(1. The probability that. exp(1.5 and also a senior leader.245) = 0.72. general leadership status does not confer the power that senior leadership does—notice that the coefficients for AllLeadership and the interaction are now smaller than the commensurate ones in Table 3.286 × 1 + 0. if these two interact. Interestingly though.78.286) If B is a rank-and-file Democrat.245) + exp(0) In Table 3.245) = 0.245) + exp(0. The power of Senator B who is a rank-and-file Republican is λB = 0.754 × 0 = 0.

available from http://voteview. this is one version of the ‘party versus floor median’ debate well discussed in literature elsewhere (Cox and McCubbins. DW-NOMINATE scores.Section 3. if voting is strongly party based. McCarty. On the one hand. 1993. we might have several conflicting expectations. and the median positions for parties. At base. while a low score reflects a senator in the center of the chamber. senators who occupy the center ground—are close to the legislative median— may find it easier to broker deals with others to their left and right. and hence will be more powerful. Krehbiel. Rohde.com/dwnomin. Poole and Rosenthal show that legislators’ ideological positions. As politics becomes more polarized. We add Distance from Median that records the absolute 15 In particular.15 In particular.htm . 1997). the score for the senator in one dimension is treated as their ‘conservatism’ (we label this ‘Conservatism’ and the higher the score. On the other hand. have become increasingly disparate along a liberal-conservative dimension. cf. We can measure ideological extremism via senators’ NOMINATE scores (Poole and Rosenthal. 1991. 1998). then perhaps the most ideologically extreme in each party will be able to motivate their ‘core’ supporters into backing their preferred positions. This will be a fortiori true for relatively radical senators from the majority party. the more conservative and less liberal they are).5 40 To demonstrate how American politics has become increasingly polarized. ‘Extremism’ is a different concept and can be ascertained by taking the absolute value of this score (Extremism): a very high score now implies a senator very far to the left or to the right.

119.441| = 0.844 × 1 + −0.2 + 0.5 are interesting.557 and λB = −0.1: here.441| = 0. consider Figure 3. Consider.2 − 0.591 × 0.702 × |0.4and 3.8 + 0.591 × 0. is around 0. and the one below is the same graphic for senators who proposed amendments in the 108th Senate.5 we report the effects of extremism. and then impose a solid loess curve. The top graphic displays the plot for all senators.4 we report the effects of conservatism (controlling for majority party membership and distance from the median) and in Table 3. Their powers are λA = −0. By contrast Senator B is deeply conservative with a NOMINATE score of 0. Once we control for majority party membership. the probability that A proposes some coalition. for example.702 × |0.8. conservatism and distance from the median have a negative impact on power. while B joins it. The lessons from Tables 3. In any interaction.5 41 value of senators’ distances from their party median in terms of NOMINATE scores. .61.8 − 0. we plot the Senator’s power estimates against their NOMINATE scores. To see more of the interplay between ideology and power.2 (the scale runs -1 through 1). In Table 3. Senator A who is a moderate Republican with a NOMINATE score of 0.844 × 1 + × − 0.Section 3.

Within the majority party (the Republicans) though. it does not pay to be a right wing Republican: rather. Interestingly.2 Committee and Agenda Control Positive political theorists. interestingly. 1999). not all commit- . the most powerful senators are not drawn from the far right wing: notice the high λi recorded for those with a NOMINATE score around 0. majority party status boosts one’s power—notice that the loess rises as it moves right. the power ratings for all senators are shown. and in the bottom panel. the analysis is restricted to the proposers only. Otherwise put. especially those of the ‘Rochester school’ (Amadae and Bueno de Mesquita. it does not pay to be the median of the chamber.Section 3. the extremes of the Senate are rewarded in power terms: the loess dips slightly as it crosses the middle of the ideological spectrum.5.2 confirms this idea: in the first panel. In Table 3. But. but it does pay to be the median of your party. with the most powerful senators of both parties occurring just above and below this scaling.5. in a series of articles Shepsle and Weingast (1987) and Krehbiel (1991) discussed precisely why committees are powerful. Notice that the bulk of the mass occurs around 0.5 42 Notice that. the positive coefficient on Extremism suggests once again that senators towards the middle of the chamber lack power. we see two things: first. Figure 3. Of course.4 for both parties. the powerful are from the median of the party.1. in the top panel of Figure 3. it is also evident while being in the majority party is beneficial. have suggested that it is the organization of Congress in terms of its committees that confers power on actors. 3. Indeed. in the lower panel (which represents the most powerful senators).4.

1: NOMINATE score versus power (λi ).0 NOMINATE score 0.5 Figure 3.Section 3.5 0. closed circles are Democrats. solid line is loess. bottom figure is for all proposing senators.0 NOMINATE score 0.5 18 19 20 21 22 Power (λ i ) −0.5 43 Power (λ 0 5 10 15 20 i ) −0.5 0. . open squares are Republicans. Top figure is for all senators.

6 0.5 44 Power (λ 0 0.8 Figure 3.2: Extremism versus power (λi ).4 Extremism 0. bottom figure is for all proposing senators. open squares are Republicans.6 0. closed circles are Democrats. .0 5 10 15 20 i ) 0.8 18 19 20 21 22 Power (λ i ) 0. Top figure is for all senators.Section 3. solid line is loess.2 0.2 0.4 Extremism 0.

6) reports positive coefficients. used the 2004 appropriations bill to steer some $50m to Iowa’s visionary complex of biomes. Above. In particular. This is not least because such positions enable members to channel substantial pork to their home states. Finance and Rules and Administration committee all denoted with these names. hierarchical structures. and they attract different memberships with different motivations (Fenno. For example. Charles Grassley. Grassley was the Chair of Appropriations.5 45 tees are created equal.6 we report regression coefficients for the Appropriations. 1973). we only discuss some of the committees and their members here. Where Table (3. Since space is limited. Other than the Chair. We also include a majority party interaction term. Chairs have several de jure responsibilities and rights pertaining to timetabling. who must be a member of the majority party. in Table 3. there is general agreement that financial committee positions—those that have ‘power of the purse’ tend to be the most coveted spots for ambitious senators. Commerce.Section 3. and presumably this goes mutandis mutatis for majority party committee members. Nonetheless. of course. Earthpark (The Economist. 2006). we discussed reasons why majority party senators might be powerful. Committees themselves are. the Committee assignment increases the power of a senator relative to one not on . and a term for Chairs (of any committee) denoted Chair. As mentioned above. committees are constructed of minority and majority party members. hearings and the selection of bills to be considered. not simply a member. Chairman of the Senate’s Appropriations Committee. Armed Services.

16 From the perspective of a senator. Chairs are more powerful than rank-and-file members. Indeed.3 we compare chairs to Democrats and non-chairing Republicans who propose amendments. which are notoriously weak and generally involved in less interesting work. the Compromise proposed two houses: a lower 16 We are grateful to Antoine Yoshinaka for this suggestion. not caused by them. Majority party members get an extra boost in power on the Armed Services. Caution is required in interpreting the findings for some committees. their entire inter-quartile range is more powerful than that of Democrats.3 Geographic Factors One of the results of the Constitutional Convention of 1787 was the so called ‘Connecticut Compromise. . such as Rules and Administration.Section 3. Hence. in fact. senators may accept a role on such committees as the ‘price’ for serving on more powerful committees elsewhere. the magnitude of the coefficient maybe somewhat misleading: senators are powerful because of their ‘good’ committee assignments which are correlated with weak assignments. Perhaps unsurprisingly.’ Dealing with the creation of the United States’ legislative bodies. but not on the Rules and Administration or Commerce committees.5.5 46 the that committee. in Figure 3. the most powerful position is a role on the Finance committee (notice that the addition of the Finance coefficient and that for Majority×Finance is a larger number than for any other committee). Appropriations and Finance committees. Notice that the median power of chairs clearly exceeds that of the other groups and. 3.

non-chairing Republicans and Democrats who propose amendments.5 47 Power 18.0 Dems Republicans Chairs Figure 3.5 22. .0 19.3: Boxplot showing relative power of committee chairs.0 20.5 20.Section 3.0 21.5 21.5 19.

the South is no longer under hegemonic Democratic control as demonstrated by George Bush with victory in every Southern state in his 2004 Presidential reelection. Moreover. in which each state would have two representatives. 1949). Historically. we might still expect. Southern senators faced few challenges in their home states and could use the committee seniority system—that rewarded long service irrespective of party affiliation—to obtain powerful chairmanships.Section 3. this systematic concentration of power in Southern hands was much reduced Rohde (1991). motivated small states—like Delaware. as argued with respect to the power indices approach above. Of course. and possibly dissenting voices—from blacks and poor whites—were excluded from voting (Key. One way to examine this notion more formally is to use the present model with geographic factors as explanatory variables. We do not have particularly strong priors about the effect of geographic factors on a senator’s power. the fact that voting resources are de jure equal across states should not imply that all senators are equally powerful. representatives from the South were very powerful actors. Up until the 1960s. that .5 48 house elected in proportion to population. Nonetheless. Southern states were solidly Democratic. but some vague ideas might be as follows. and a senate. As a result. New Jersey and Connecticut—to seek institutional protection. The fear of large state tyranny. regardless of its population. In the ‘post-reform’ period. Maryland. for historical or other reasons. though admonished as unlikely and illogical by the likes of Madison and Hamilton.

sparsely inhabited. Mississippi. North Carolina. densely populated states have senators who literally represent ‘more’ citizens (we measure this with Census Bureau’s population estimate for 2003. for historical reasons. South Carolina. farm-based financial aid—an important component of rural states’ federal funding—is easier to deliver than other types of sup- . On the one hand. We measure wealth use the Census Bureau’s Median Household Income statistics for each state in dollars (Median Income). On the other hand. We have similarly vague priors viz the effect of population density on power. senators from poorer places may have more sway in Washington because they can point to underfunded public services and crumbling infrastructures in their home states as evidence that they have a more urgent claim to the nation’s resources. we could imagine poverty may boost a state’s representative’s powers. Arkansas. primarily agricultural states may increase the power of senators who represent them. rendering them more influential.5 49 senators from the South—Texas. primarily urban. On the other. senators from rich states may be able to procure greater ‘home-grown’ funding for their campaigns and causes. Population). in part because. Florida. Alabama. small. which may aid rhetorical appeals. Combined with suitably deployed rhetorical skill. Georgia. Louisiana. Virginia and Tennessee— wield disproportionate power.Section 3. We use a South dummy taking the value of 1 for senators from these states. The expected effect of a state’s wealth on a senator’s power is arguably ambiguous. On the one hand.

Section 3. a position in the Senate is a career ambition (Brace. in Figure 3. As implied above. Based only on geographical factors. 1984. But. this might not correspond to any particular state. negative coefficients in Table 3. once attained. Here we calculated the years of service since first entering the Senate through 2004 and denoted this variable Service. we color a map of the contiguous United States according to the predicted λi that a senator from that state based on the coefficients of Table 3. 3.4. The significant. ‘Weaker’ states are dark colored. To help interpretation. Of course. New York and West Virginia are similarly light colored.4. non-Southern states will yield senators with more power.4 Career Factors For most politicians. Though not formally associated with greater rights or responsibilities. this is in part because long service is linked to promotion in terms of committee and other assignments. We use the land area of the states and denote this variable Land Area.7 suggest that poorer. There are no particular regional patterns discernable from Figure 3. more densely populated. longevity of service for a particular state makes a senator the ‘senior’ representative of his constituency.7.5 50 port. senators have strong incentives to seek reelection (Mayhew. 1979). the most powerful state—in terms of its Senators—is California and it is shaded lightest. except perhaps a band of states from New York west through Missouri which appear disproportionately light (and thus powerful) relative to their neighbors. smaller.5. 1974). Rohde. We use a Senior dummy to check for any extra power effect that .

5 51 Figure 3.4: Map of contiguous United States with shading proportional to ‘power’ given by coefficients in Table 3. .7: lighter states are associated with more powerful senators.Section 3.

. as well as seniority makes for a more powerful senator. For our daughters and granddaughters. the comment and excitement drawn by Nancy Pelosi’s ascension to the Speakership of the House in 2006. we add a dummy variable for Male here. today we have broken the marble ceiling. though the median power of the sexes is approximately equivalent. for example.5. 52 From a career perspective. and an historic moment for the women of this country. they are also some of the weaker members of the Senate. anything is possible for them. Interestingly. in terms of their power. For our daughters and our granddaughters.a member of the majority party. The coefficient for sex though is perhaps not as expected. the distributions are very different: while male senators are counted among the most powerful.5 such status confers.8: longer time served in the Senate. sociologists have long argued that being male strengthens advantages one in the work place. Being male is actually associated with a lower power than being a female. In Figure 3. the sky is the limit. The coefficients for Service and Senior status are much as we might anticipate in Table 3. and this is no less true in politics.5 we report a boxplot for males and females. in her acceptance speech noted that “[i]t is an historic moment for the Congress.5 Summary of Findings In summary. a senator is more powerful if the senator is: .” . . Consider. and has a leadership position within 17 Pelosi herself seemed well aware of her exceptionism and.Section 3. by contrast. 3. are heavily concentrated in the upper power ranges.17 For this reason. Females.

.5 53 Power 0 5 10 15 20 Female Male Figure 3.Section 3.5: Boxplot showing relative power of male and female senators.

6 the party.from the center of the ideological distribution of her party.6 Discussion This chapter proposed a new way to measure power in structured settings.is from a relatively poor. 3. . These findings in and of .a majority party member of the Finance and Appropriations committees. using (personal) amendments as the bills around which Senators attempt to form coalitions and we found that inter alia majority party membership. This method is straightforward to implement. .is female. we can talk of predicted probabilities and make statements of uncertainty regarding the ‘effects’ of certain variables. densely populated non-Southern state. moderation within a party. long serving and is the Senior Senator from her state. long service records and seniority all make for more powerful senators.Section 3. chairing of committees.chair of a committee. Thus. We applied the statistical model to the 108th Senate. . . Rather than relying on a priori metrics that arguably measure something akin to ‘luck’ or ‘pivotality’ rather than power. 54 . and allows the separation of ‘power’ from its ‘causes’ in a standard generalized linear model framework. we suggested an approach that assumes the powerful have a greater capability to form coalitions in order to pass bills that they care about.

though in that case it is simply unclear who ‘leads’ a coalition. In terms of future avenues of research. deciding actors’ preference in this circumstance is not always straightforward.Section 3. political scientists could not execute a similar model for the House of Representatives. but they do demonstrate the strength and flexibility of the approach. chairs and party leaders. though any such extension requires knowledge of actors’ preferences in interactions. along with the prevailing political climate. Temporally. Outside of American and Comparative politics. we choose to study the Senate: its 100 members are a manageable number of observations for which to ascertain biographical and other data. for example. might alter the power of say.6 55 themselves are perhaps not shocking. There is a similar caveat for the US Supreme Court.’ Studying actors’ power over time may allow a more complete picture: certainly we could imagine that changing rules in Congress. may provide another test ground for the model via private members bills. along with establishing some validity of the method. There is no reason why. The United Kingdom House of Commons. International Relations with its concentration on specifically dyadic interactions may be amenable to such an approach. Taking the model outside of the United States Congress is a possibility too. with more time. We leave this for future . we studied one Congress and thus the results here are something of a ‘snapshot. there are obvious ‘cross section’ and ‘time series’ extensions to this work. Unfortunately. Here. due to strong party whipping and strategic voting.

Section 3.6 work. 56 .

John McConnell. . Trent . -0. Bill Cochran.48 . Judd Byrd.6 57 Senator 1 2 3 4 5 6 7 8 9 10 . . Charles Boxer. Kerry.36 . 20. Kennedy. . 22 23 . . . Mitch Gregg. . . John Frist. . . Santorum.74 . .10 22.39 21. . 99 100 101 Warner. 54 57 . Inouye. Lott. . Clinton. Lindsey . . . -0.58 . .84 21. . . . .44 21. .13 .58 21. John . Joe . 20. .85 19. . Russ Lieberman. .34 .79 . .73 -0. Rick . .44 21. 45 . . . 20.79 20.63 21. Feingold.09 21. . . . . 20.82 Table 3. Robert McCain. . . 33 . Hillary . 93 . 29 . . . . . . 19.Section 3.74 -0. Jim Akaka. . Daniel Jeffords. Thad Grassley.57 21.1: Baseline results for model of power for US Senators in the 108th Senate. . Barbara Graham. Tom . . Daniel Power (λi ) 22. . . Edward Daschle.

754∗∗∗ Std.047 0.373∗∗∗ Std. Asterisked coefficients imply p < 0. Error 0. Error 0.118∗∗ 0.2: Effect of majority party and senior leadership status on power in the Senate.304∗∗∗ 0.722∗∗∗ Std. Estimate Conservatism Majority Distance from Median −0.05 (**). p < 0.Section 3.01 (***).05 (**).104 Table 3.01.286∗∗∗ 0. Asterisked coefficients imply p < 0.01 (***).3: Effect of majority party membership and any leadership status on power in the Senate.205∗∗∗ 0.4: Effect of conservatism. AIC: 11169.844∗∗∗ −0.6 58 Estimate Majority SeniorLeadership Majority×SeniorLeadership 0.080 0. Error 0. Estimate Majority AllLeadership Majority×AllLeadership 0. AIC: 11052.073 0.034 0.035 0. p < 0.122 Table 3. AIC: 11152.591∗∗∗ 0. . distance from median and majority status on power in the Senate.073 Table 3.072 0. Asterisked coefficients (***) imply p < 0.

Section 3.6

59

Estimate Extremism Majority Majority×Extremism Distance from Median 1.14∗∗∗ 0.783∗∗∗ −0.9636∗∗∗ −0.702∗∗∗

Std. Error 0.107 0.075 0.168 0.123

Table 3.5: Effect of extremism, distance from median and majority status on power in
the Senate. Asterisked coefficients imply p < 0.01 (***), p < 0.05 (**). AIC: 11136.

Estimate Chair Majority Appropriations Majority×Appropriations Armed Services Majority×Armed Services Commerce Majority×Commerce Finance Majority×Finance Rules Admin Majority×Rules Admin 0.596∗∗∗ −1.055∗∗∗ 0.245∗∗∗ 1.376∗∗∗ 0.189∗∗∗ 1.353∗∗∗ 0.264∗∗∗ −0.737∗∗∗ 0.482∗∗∗ 1.445∗∗∗ 0.440∗∗∗ 0.312∗∗∗

Std. Error 0.042 0.089 0.040 0.100 0.042 0.102 0.040 0.068 0.046 0.104 0.039 0.068

Table 3.6: Effect of committee membership, majority party membership and chair status
on power in the Senate. Asterisked coefficients imply p < 0.01 (***), p < 0.05 (**). AIC: 9434.9.

Section 3.6

60

Estimate Median Income South Land Area Population −2.199 × 10−5 −0.256∗∗∗ −1.668 × 10−6 2.488 × 10−8
∗∗∗ ∗∗∗

Std. Error 2.003 × 10−6 3.390 × 10−2 2.135 × 10−7 1.723 × 10−9

∗∗∗

Table 3.7: Effect of incomes, southern state representation, land area (state size) and
state population on senator’s power. Asterisked coefficients imply p < 0.01 (***), p < 0.05 (**). AIC: 10954.

Estimate Service Senior Male 0.029∗∗∗ 0.543∗∗∗ −0.521∗∗∗

Std. Error 0.001 0.029 0.035

Table 3.8: Effect of incomes, southern state representation, land area (state size) and
state population on senator’s power. Asterisked coefficients imply p < 0.01 (***), p < 0.05 (**). AIC: 10205.

61

Appendix to Chapter 3
3.7 The “Busy” Senator Problem

The concern is that a senator could increase his influence score by simply proposing more amendments. That is, by being ‘busy’ in a legislative sense, he would appear more powerful. Here we show this to be untrue.

First, consider three proposing senators A, B and C. Suppose that A and B decide to somehow combine their efforts such that only one of them will propose and form coalitions for all amendments jointly that they formerly worked on separately. Whether A will have B do all the proposing and coalition forming, or whether B will delegate his work to A, write the senator who does the proposing and forming as SAB .

Clearly, SAB is busier—in that he now proposes more amendments—than either A or B. Importantly, though, the probability that any randomly drawn interaction involving SAB and C has C convincing SAB , will be simply the weighted average of the former probability that C convinces A and C con-

business cannot itself increase a senator’s influence. let “CconSAB ” be the event that C convinces SAB . ‘busy’ senator is: SAB Pr = = = Pr(SAB conC) Pr(CconSAB ) Pr(AconC) + Pr(BconC) Pr(CconSAB ) conC) Pr(CconA) Pr(AconC) + Pr(CconB) Pr(BconB) Pr(C conA) Pr(C Pr(CconSAB ) A B = γ Pr +(1 − γ) Pr .8 62 vinces B: there is no fillip from proposing more amendments. Pr(C conA) Pr(C conSAB ) . performing a . To show this. The probabilities for the constituent senators are: Pr = and Pr = B A Pr(AconC) Pr(CconA) Pr(BconC) .Section 3. Hence. ‘business’ alone cannot yield a higher power rating.9. Inspection suggests that the rank order is closely similar to that of the model without an order effect. In fact.8 Logit Model with Order Effect We estimate the ‘order effect’ model as described in equation 3. Pr(CconB) The probability for the joint. 3. where γ = Thus.5 and present the results in Table 3.

Figure 3.10. exp(λA ) + exp(λB ) (3. the matrix of votes to be predicted denoted Yact . We estimate δ—the advantage of proposing to someone in one’s own party—at 0.10 63 Spearman rank correlation test between the sorted lists yields a ρ = 0. Clearly. Hence. Then. there is a ‘home advantage’.10 gives the rank ordering of senators in accordance with the number of amendments they proposed. is of dimension I p × I where . That is.9 Amendment Proposers Table 3. Above.6 displays the frequency information from Table 3.3. we defined an ‘interaction’ as an amendment in which either A proposed and B supported or B proposed and A supported. but it does not disturb our earlier findings greatly when accounted for. 3.08. simply proposing more amendments does not make you more powerful: for example. For completeness.10 Goodness-of-Fit Recall that the method calculates a (maximum likelihood) value of λA and λB for all senator pairs who interact. Grassley proposes only 3 times yet is ranked in the top 10.144 with a p-value (on the null hypothesis that ρ ≤ 0) of 0. the λi have the interpretation that Pr(A convinces B|A and B interact) = exp(λA ) . as in 3.Section 3. 3.52 (and is significant at the 1% level).3) implies that we need not consider the counterfactual of non-proposing senators as proposers.7) The conditioning of the probability in (3.

Section 3. .6: Amendment numbers: distribution.10 64 Frequency 0 1 5 10 15 20 2 3 4 5 6 7 8 9 10 15 Number of Amendments Figure 3.

. . For any particular cell of Yact . This matrix has symmetric form when proposers as compared to voting senators who were also proposers. since we included President Bush to aid interpretation). Joe Biden and Wayne Allard interacted a total of 3 times: once when Allard was the proposer and Biden backed him. Notice also that senators voting for their own amendments are not. . there exists an associated total number of possible interactions (in either direction) for the senators. and is of the following form: Akaka Alexander Allard . . .10 65 |I p | is the number of proposing senators (74 for this case) and |I| is the number of senators in total (here.... . So. For example. . Allard Baucus Bayh Biden Bingamen .. . . twice when Biden was the proposer and Allard backed him. twice by Lamar Alexander and once by Wayne Allard. 1 0 0 2 4 . 101.. counted as backers. . 1 3 1 4 9 . for example. .. but not otherwise. . 0 0 0 1 2 ... . ... . Yact then.. where the proposers lie to the left and all the voting senators form the columns. Joe Biden proposed and was backed four times by Daniel Akaka. . in and of themselves. . .Section 3. is the actual interactions.

consider the following subset of the matrix: Akaka Alexander Allard . . .. .. .00 1. 0.. .00 1. .00 1. 1. Allard Baucus Bayh Biden Bingamen ..... . we can obtain predicted probabilities of ‘convincing’ for any particular pair of senators A and B.. ... 1 0 0 2 4 .3).00 1...00 1. . Allard Baucus Bayh Biden Bingamen . . .. . We denote this matrix Opred .50 0.00 . . where the incidence of 1 implies that the proposer on the left is the one convincing the voter in the column to back the amendment (rather than the other way round) essentially with certainty. . 1. .00 1. ..00 . 0 1 0 2 3 .. For example. . . From the estimates of Table 3.. we do not dichotomize this matrix to zeros and ones as may be seen in standard logis- .. ..21 0. . . . . Since we have direct estimates of the underlying latent abilities. . . 1 3 1 4 9 . .1 via Equation (3. .Section 3. .10 66 We denote this matrix Yposs and a similar segment to that above appears as: Akaka Alexander Allard .00 1. . . . .52 0...00 1.74 .. .48 0. ..

869 = 0. the percentage of interactions correctly predicted is 1 − 2062 13. 67 The expected votes matrix is the element-by-element multiplication Yposs × Opred yielding another matrix of dimensions I p × I.10 tic (and probit) regression predicted probability contexts. The absolute sum of these columns yields the total number of misclassifications which is (when rounded) 2062 of some 13. Subtracting this matrix from Yact yields a matrix of both positive (the model over predicts) and negative entries (the model under predicts).85. Hence. Note that every element of Yposs × Opred is non-zero valued.869 total interactions. .Section 3.

John Cochran. Lindsey Grassley.84 . . 30 . .59 19. 18. .49 .10 . 18. . . . . Clinton.9: Baseline results for model of power for US Senators in the 108th Senate. . Edward .19 . -0.98 18. Judd Frist. 52 53 .Section 3. Hillary . . 33 . . Rick .30 . .97 . . . Tim Power (λi ) 19. . 17. . .20 -1. 21 . Thad Byrd. . . . 17. . Bill McCain. Akaka. Feingold.30 17. . John .25 . . Russ Lieberman.97 18. Kennedy.34 -1. 98 99 100 101 McConnell. John Gregg. .10 19. Don Boxer. . Robert Graham. . Jim Inouye. Kerry. 17.63 19. Lott. . . . . .97 18. . . . . . Barbara . 94 .53 Table 3. 25 . Mitch Warner. Daniel Johnson.87 18.36 19. .09 18. . Tom . . . Charles Nickels. Trent . . Santorum.10 68 Senator 1 2 3 4 5 6 7 8 9 10 11 .96 18. . Daschle.89 . . 18. . logit with ‘order effect’ (party advantage) estimated. 46 . . . . -1.51 -1. . . Daniel Jeffords. . Joe .

69 .10 Byrd 15 McCain 8 Leahy 5 Hutchison 4 Conrad 3 Kerry 3 Bunning 2 Lott 2 Brownback 1 Hagel 1 Smith 1 Bingaman 10 Daschle 7 McConnell 5 Levin 4 Corzine 3 Kyl 3 Dayton 2 Murray 2 Campbell 1 Kohl 1 Stabenow 1 Boxer 10 Warner 7 Biden 4 Mikulski 4 Edwards 3 Lincoln 3 Hatch 2 Reid 2 Carper 1 Murkowski 1 Voinovich 1 Dodd 9 Dorgan 6 Breaux 4 Reed 4 Feingold 3 Nickles 3 Hollings 2 Rockefeller 2 Collins 1 Nelson (FL) 1 Wyden 1 Feinstein 9 Harkin 6 Cantwell 4 Specter 4 Graham (FL) 3 Schumer 3 Inhofe 2 Snowe 2 DeWine 1 Roberts 1 Lautenberg 9 Durbin 5 Clinton 4 Baucus 3 Graham (SC) 3 Thomas 3 Landrieu 2 Allard 1 Ensign 1 Sarbanes 1 Kennedy 8 Gregg 5 Frist 4 Cochran 3 Grassley 3 Bond 2 Lieberman 2 Bayh 1 Enzi 1 Sessions 1 Table 3.10: Number of amendments per senator.Section 3.

While much of the discussion has been normative in nature. His second duty is to his constituents. of whom he is the representative but not the delegate. scholars of politics have debated the proper ‘model’ for elected representation. It is only in the third place that his duty to party organization or programme takes rank. Burke’s famous declaration on this subject is well known. 1997–2005 The first duty of a member of Parliament is to do what he thinks in his faithful and disinterested judgement is right and necessary for the honour and safety of Great Britain. 1954 4. —Sir Winston Churchill.70 Chapter 4 Rebels with a Cause? Legislative Activity and the Personal Vote in Britain.1 Introduction At least since the time of Burke and his Speech to the Electors of Bristol (1774/1975). country-comparativists are able to distinguish quite different practices .

indeed. House members in the United States are perhaps closest to Burke’s ‘delegates’: primarily concerned with delivering (federal spending) ‘pork’ to their constituents and. Analysts of American politics have devoted considerable attention to this issue and there is a voluminous theoretical and empirical literature that investigates the legislative determinants of individual office-holders’ electoral performances (e. is thought to exert a relatively weak pull on their decision calculus (Jacobson. The nation. 1989). 1982). Congruent with the ‘delegate’ assumption. akin to a ‘trustee’ conception. consciously acting in line with the general ideological preferences of their district (e. regardless of any constituency or personal concern.Section 4.1 71 across polities. 2002). equivalent research for the United King- . In that they hold representatives to account. By contrast. Fiorina. Despite its Burkean vintage though.g. At one extreme of these stylized facts. broadly construed. members of parliament in the United Kingdom House of Commons are elected on a party ticket. Canes-Wrone. the policy-making behavior for which citizens will reward or punish their legislators should vary markedly. 1974. 1974). But contingent on the model of representation in place. and make policy according to some notion of ‘national interest’. elections are identical mechanisms in all systems (Riker. Brady and Cogan. the findings generally suggest that more ‘maverick’ incumbents—those who vote with their party least often—tend to do better come polling time. In particular. members of the governing party generally support the executive’s bills. Mayhew. are thought to have a negligible ‘personal’ vote.g. or even the party with which they identify.

2007). Norton and Wood. their expectations of behavior and their elected agents is unfortunate.Section 4. The lacuna is unsurprising: with some important exceptions (Kam. In any case. It is unfortunate because ignoring Britain (and Westminster systems more generally) has resulted in an asymmetrical development of political science understanding of representation and accountability. 2007). 1993). While we know much about the United States where parties are weak. 2007.1 72 dom House of Commons—or any ‘Westminster’ system—has been much less developed. with over a thousand votes to consider in a given parliamentary session. 1994). we can say little concerning polities where parties and party voting are strong forces. and no convenient ideal point summary available (see Spirling and McLean. every decision on every vote. Pattie. Fieldhouse and Johnson. 1987. for example. We seek to rectify this theoretical and empirical deficit in the current chapter: using a new data set that incorporates some 2200 parliamentary votes . and looked elsewhere for personal vote effects (see. but unsurprising. Ferejohn and Fiorina. scholars have not been able to conduct systematic studies that analyze MP legislative behavior as a whole—that is. That the study of British politics has generally neglected the fundamental link between constituents. scholars have assumed away the potential importance of legislative activity. with citizens claiming that both party unity and ‘independent-mindedness’ matter (Johnson and Rosenblatt. It has not helped that public opinion studies point to seemingly inconsistent preferences and contradictory expectations. Cain.

Section 4.2

73

cast by around 200 backbench Labour MPs between 1997 and 2005, with commensurate constituency controls, we explain ballot box performance as a function of personal voting decisions in the House of Commons. We show that, all else equal, MPs that disrupt important government business are punished by their constituency voters while those that show freedom of thought and action on less important matters tend to be rewarded for their independence. This effect is worth up to 2.5 percentage points at election time. In pursuing our study we suggest a novel theoretical model of partisan voter behavior in Britain and new non-parametric, statistical methods as yet unseen in political science. Our findings extend the growing literature that emphasizes the importance of seeing MPs as individual and accountable representatives, at least partly responsible for their own electoral fates.

We proceed as follows: in Section 4.2 we give the substantive background for our study, and report on previous efforts in the literature. In Section 4.3, we give a simple formal expression describing the actions of (partisan, Labour) voters in response to different actions on various roll call types at Westminster. In Section 4.4 we describe and use a “random forest” classification technique that allows us to include essentially every vote as a possible independent variable, along with suitable controls, for predicting subsequent electoral performance. In Section 4.5 we report results consistent with our model, and show that several bills have a small but significant effect on performance with signs as predicted. Section 4.6 concludes.

Section 4.2

74

4.2

Members, Constituents and the Personal Vote in Britain

Party voting by the electorate, either based on national Cabinet—i.e. executive— policy positions and performance (see Budge, 1999; Crewe, 1983; Norris, 2001) or on cruder partisan loyalties (see Butler and Stokes, 1974) is thought to be the norm for Westminster systems like Britain. Congruent with this principal, almost all legislative business is directly controlled by the executive, and party pressure or ‘whipping’ is commonly used by the government to herd its ‘backbench’ MPs into supporting the Cabinet’s programmatic agenda when required (e.g. Cowley, 2002, 3–7).1 By conventional wisdom then, since voters don’t (and have no reason to) care about parliamentary behavior, there is little an individual backbencher has to offer her constituents aside of her party label at election time.

Interestingly though, when the public was recently asked to name the qualities that an MP ought to possess, while two-fifths claim that a representative should “be loyal” to her party, some 60 percent also claim that she should “be independent-minded” (Johnson and Rosenblatt, 2007, 165–167). In Figure 4.1 we show the changing nature of public opinion on this subject: clearly freedom of thought seems increasingly important, while loyalty is falling away as a requirement. This is a fortiori true for the period in which Labour has held large majorities of seats in government i.e. post-1997.
Whipping typically includes offers or threats related to ministerial promotion: see Kam (2006) for a discussion of the Canadian case.
1

Section 4.2

75

’Independent−minded’ ’Loyal to the party’ 55 Important Characteristic (%) 40 45 50

1985

1990

1995 Year

2000

2005

Figure 4.1: Percentage of respondents naming different qualities as “important for an
MP to have”: solid line is percentage naming ‘independent minded’, broken line is percentage naming ‘loyal to party’. Multiple choice, more than one choice permissible; survey dates are 1983, 1994, 1996, 2000, 2005. Source: Johnson and Rosenblatt (2007, 166, Table 1).

On the one hand, such a diversity of views might imply that voters fail to understand the nature of parliamentary government and the importance of party dominance and discipline therein (e.g. Paltzelt, 2000). On the other hand, it is not impossible to rationalize both preferences, especially for voters who lean towards the governing party: united governments are more efficient, and able to respond rapidly to crises (see, e.g. Alesina and Drazen,

it is not difficult to see the potentially tricky balancing act this presents for MPs involved in Commons roll calls. MPs increasingly believe they enjoy a personal vote dependent on their appearances. Johnston and Fieldhouse. they have tended to be limited in scope: in the case of Pattie. Fieldhouse and Johnson. 1993. CanesWrone. Searing. examine constituency service as both a goal of members and its effects at election time (see Cain. On the other hand. whatever the truth may be. When scholars have examined parliamentary votes as a predictor of performance. 2002. Pattie. previous work has tended to examine other MP-voter links. Pattie.g. Norton and Wood. 2005). Despite—or perhaps because of—the complicated dynamics this suggests. Denver and Hands. But whether different voters hold different preferences. perhaps ‘maverick’ representatives are more able and willing to impose their constituents’ views and concerns on the government at large when the attention of national politics—and national parties—is elsewhere (see. statements and legislative behavior (Cowley.. 1994). or perhaps they avoid dangerous ‘group think’ (in the sense of Janis. with some suggesting that. Ferejohn and Fiorina. 1994). they deal with just six ‘key’ votes with various levels of attendant whipping.2 76 1991. e. Brady and Cogan. 1972) behind the scenes. 2002). or voters are internally divided. Tsebelis. for example. Fieldhouse and Johnson (1994). while Kam (2007) uses ‘dissent’ on all .g. they also exhibit greater longevity and thus are more able to deliver benefits to their supporters.Section 4. Others look to local campaigning activity (e. A series of studies. 1997. 1987. 1995).

if we could be agnostic about the potential bills of interest. Thus these are the members for whom the stakes are highest—for whom party unity and/or independence matter most. rather than a potentially misleading small sample. Nonetheless.e. Before doing so. 2 Though for New Zealand.3 Formalizing Intuitions As with other scholars in the field. 4. We would thus capture the totality of legislative behavior. rather than Britain. we would garner the best of both investigations: we would have a survey of parliamentary behavior that is deeper in terms of both the number of votes in the sample and broader in terms of their content and the pressure applied in each case. Labour) party MPs: the ‘power’ of the government relative to other parties in a Westminster system means that these representatives receive more media attention. and thus include all votes as possible predictors.Section 4. We outline our intuitions in the next section. 135) does show that dissent is an effective vote-winning strategy. . and are more likely to be held accountable for the state of the polity at election time.3 77 whipped bills (in the aggregate) to predict voter familiarity with their MP’s name.2 These strategies are not unreasonable given the specific questions that were to be addressed. we need to be clear about what voters desire and expect from their representatives. and the studies are important in that they are rare breaks with the orthodox thinking on voting in Britain. Kam (2007. our work here will concentrate on the behavior of government (i.

on roll calls where the non-government parties attempt to foil the executive for the sake of opposition. maverick. James Callaghan was forced to call a general election (which he subsequently lost) after suffering defeat in Commons in March 1979. Such votes dangerously undermine Prime Ministerial authority: John Major called such a vote in 1993 in response to sustained rebellion over a series of European integration votes. left-leaning partisan voters want their MPs to plump for the more ‘socialist’ option while right partisans prefer the more ‘conservative’ choice. 2007). That is. Major was compelled to resign his position as party leader—though he subsequently won reelection—in response to the continual sniping. party unity is prized on issues that are most important to the government’s legislative agenda. on the rare occasion that free votes arise whose contents and consequences are simple to grasp for the public at large. precisely because they do not endanger a government agenda. voters rely on heuristics. For the reasons cited above. indeed losing on such a roll call would typically trigger a vote of confidence.3 78 Our intuition is that on almost all issues (whipped or unwhipped) debated in the House of Commons. when votes are relatively free of party pressure. voters reward MPs on a spatial basis. Indeed here. The reason is that these votes most starkly test the Cabinet’s ability to govern. rather than because they actually have some different and sincere policy preference (see Spirling and McLean.’ However. Fifteen years earlier.3 By contrast. Notice that this effect will be strongest for bills which exhibit “government-versus-opposition” voting: that is. independentminded voting is welcomed. Thus.Section 4. ‘dissent’ from the MP’s party as a whole is a ‘good thing. The bills and proposals are too technical and it is too time-consuming to review and understand their contents. 3 . voters are uninformed.

or amendment) b that is to be voted upon. There are q1 such bills of a total of B that come to be voted upon in House of Commons. . suppose there are 1.Section 4. og ) +w2  q1 +1 d(oj . . . Since the government essentially never loses on parliamentary business it cares about. Finally. denote the Cabinet’s preferred outcome as og . in a slight abuse of notation. oi ) voter uninformed voter informed (4. This extends to the B − q1 bills that are not essential to government business. . . w2 ≤ 1 refer to the ‘weights’ that voter i places on each legislative activity component and d(·. N voters in constituency j. there are two possible responses by MPj —‘aye’ or ‘no’. For a bill that is essential to government business. . for which MPj is the incumbent governing party member of parliament. we will use og to also refer to the outcome voted for by the majority of the governing party. We will write the choice of MPj as a point oj in p-dimensional space: oj ∈ Rp . . og ) − (1 − w1 − w2 )  q2 +1 d(oj . i.1) where 0 ≤ w1 . For any particular bill (or proposal. We can now write her utility from MPj ’s voting record in the Commons as whipped q1  unwhipped q2   unwhipped B  Ui (·) = − w1 b=1 d(oj . Notice that we assume she voted for the incumbent at the previous election. ·) is the Euclidean distance between .3 79 More formally. denote the median voter of constituency j as having an outcome preference on any particular roll call as oi ∈ Rp .

. ∂Ui (·) ∂d(oj .og ) > 0. Before leaving the theoretical discussion here.1) with some constant c—say. ∂Ui (·) ∂d(oj . .3 80 two points in p dimensions. so long as w1 > 0. consistent with our reasoning above. voters reward ‘in- dependent thinking’ insofar as it manifests itself as choosing options different (far from) the party majority option. so long as w1 + w2 < 1. . 4 . . • for unwhipped roll calls for which voters are informed. follow: • for (heavily) whipped government business in Equation (4. In words: on these bills. so long as w2 > 0.Section 4.4 With no loss of generality we assume that voter i compares the utility from (4. this model will give us a structured way to think about our findings. the worse he does.og ) < 0. Moreover. y2 . and caters for points in any number of dimensions in space. . x2 . In words: for bills that deal with impor- tant government business. yn ) is simply i=1 (xi − yi ) . xn ) and Y = n 2 (y1 . . . the further away the MP’s vote (oj ) places him relative to the government (og ).oi ) < 0. • for unwhipped roll calls on which voters are uninformed. rebellious MP behavior. the voter’s utility is decreasing the further the MP’s choice is from her ideal (outcome) point. Several simple comparative statics. That is. It is thus always positive.1). government party) voters will disapprove of fractious. For reasons that we will discuss below. ∂Ui (·) ∂d(oj . else she abstains or votes for a competitor. it is not possible to ‘test’ this theory directly. it is helpful to Recall that the Euclidean distance between two points X = (x1 . (partisan. Nonetheless. . zero—and casts her ballot for MPj at the subsequent election if Ui (·) ≥ c.

We do not doubt that such reasoning occurs.000 votes on a 18 percent swing away. but we will assume that at some point an MP’s actions are sufficient to merit voting against them. consider the fate of Neil Hamilton. Rather.4 Data and Method Ideally. First. in addition to partisan. We would compare these predilections to their representative’s actual record As a case in point. Hamilton lost to an independent challenger. to the extent that a time trend exists. there is a potential contradiction in the voter calculus: punishing a rebellious MP also hurts the party by potentially denying them an admittedly unreliable seat in the Commons. the then 5th safest Tory district in the country with a previous majority of 16.000 votes (27.3 with constituency level data recording citizens’ preferences for their MP’s specific legislative actions.5 Third. sued for libel and went on to fight for his seat in the 1997 General Election. even if it superficially damages the national party in the short term. Figure 4. government party voters’ views on the desirability of party unity and favored positions on various free-vote issues. 4. this is an as if model: we are not contending that most voters directly examine voting records to ascertain their representative’s performance. we would test our predictions from Section 4. Second. Conservative MP for Tatton in Cheshire. by some 11. Accused by The Guardian newspaper of taking bribes to ask questions in the House of Commons. What matters then is the types of MPs and the types of voting behavior in a broad sense. they receive media and elite reports summarizing these patterns.Section 4. Hamilton proclaimed his innocence.4 81 clarify three issues.6 percent). Otherwise put: (relatively) free votes should be better predictors of performance in 2005 than in 2001. 5 . in addition to (or perhaps in place of) loyalty.1 implies that independent thought ought to be increasingly rewarded.

2002).4 82 and calculate the effect that this divergence had on the MP’s subsequent electoral performance. it is not publicly obvious how much whipping is applied on individual roll calls. 7 See McLean. But such a strategy is fraught with difficulty for the United Kingdom case: the sheer preponderance of strategic. 6 In practice. ‘government-versus-opposition’ voting in the House of Commons makes interpreting industry-standard tools for summary very difficult (Spirling and McLean. instead.Section 4. party whips contact party members with a schedule concerning upcoming divisions.6 Moreover. No such data exist so. 2007). CanesWrone. Thus ‘divergence’ and the degree to which a member is ‘out-of-step’ with his constituents may be assessed (e. we will try to infer micro-behavior from aggregate level findings. scholars have summarized Congress members’ legislative behavior using ideal point estimation. Brady and Cogan. In the Americanist context.7 So. Spirling and Russell (2003) for a specific incidence of this phenomenon. they do not typically publish details of its intensity and. First. in any case. While parties certainly acknowledge that strong pressure is present at Westminster. whipping may take the form of informal threats and offers. The analysis of British parliamentary voting in the context of our theory is hampered by two further considerations. They literally underline the requirement that members attend and vote appropriately according to the importance the party places on the roll call in question.g. the fact that a roll call is ostensibly ‘free’ is no guarantee that it is completely undirected. . and then contrasted this with metrics purportedly capturing the underlying political orientation of citizens in their home district.

we restrict the analysis to government party—i. illness.1 Data: 1997–2001. they are meant only as broad-brush shorthand for the general environment in which the bill is being considered: this is especially true when the parties differ in their application of their whip. 4. Labour party—MPs for whom we think the theory predicts the strongest effect. within the parliamentary record. etc. there is no way of knowing if a missed vote was due to indifference. 8 .Section 4. Since there is no way to formerly record absences.4. As noted above. 2001–2005 For each data set.e. principled abstention. Simply put. missingness is endemic and unexplained.8 but that same status makes them much more recognizable in their home districts. Further. we deal solely with MPs who hold seats in England and who did not hold ministerial office at any time during the parliamentary terms. The location restriction avoids including MPs whose electoral fates may be the product of nationalist party dynamics that are difficult to control for. increasing the possibility See Benedetto and Hix (2007) for a discussion of this issue and its consequences for discipline. The next avoids the potential endogeneity problems inherent in ‘good behavior’. Hansard.4 83 though we will use terms ‘whipped’ and ‘unwhipped’ below. The second problem is that. our dependent variable is the difference in electoral success (vote percentage in their home constituency) for each of the MPs in the study between the respective elections. ministerial business. front-bench promotion and constituency popularity. MPs are given ministerial responsibilities contingent on toeing the party line.

of non-whites. We have also collected numerous (standard) controls which include the difference in electoral spending by Labour between the elections (a reasonable proxy for campaign ‘effort’— see Pattie. of unemployed.4 84 their subsequent vote share is influenced by factors other than their roll call record.10 4. the difference in turnout. and 194 for 2001–2005. data which was compiled from (online) Hansard reports. These data cuts leave some 225 backbench Labour MPs for 1997–2001. i. or who joined the parliament mid-way through via by-elections (and thus for whom we are missing some dependent variable information).e. 9 See Firth and Spirling (2006) for a discussion of this data.9 This way.2 Estimation Problem: p > n The dependent variable of interest is (essentially) continuous. of individuals living with a limiting long term illness (which includes disability). In particular. 10 Compiled from the Census 2001 Report for Parliamentary Constituencies. This avoids including divisions where the majority of MPs were missing. the vote shares on every single roll call vote the MP casts in the relevant parliament. We impose the requirement that at least 50 percent of all the MPs in the sample voted on that roll call for it to enter the data. Our approach is to predict. ordinary least squares (OLS) might seem attractive as an estimation procedure. to regress.4. . Hence. we also drop MPs who choose not to fight the subsequent election. the percentage of pensioners. we know that the totality of MP legislative behavior is being captured. Johnston and Fieldhouse (1995)). and the proportion of households headed by a lone parent. For obvious reasons.Section 4. and the proportion of constituency residents falling into several traditionally marginalized groups that might be disproportionately sympathetic towards a Labour government.

2) is unavailable. 4.4. . we would like a flexible approach that assumes as little as possible about the precise functional form of the relationship between the y and x we alluded to above. Depending on the assumptions we make about m()—in particular.Section 4. In parametric models. but this requires an entirely arbitrary a priori restriction on the model. p far exceeds the number of observations. The consequence is that the standard OLS solution. the number of parameters to be estimated. the equivalent of Equation (4. moreover. That is. β = (x x)−1 x y. n: we have negative ‘degrees of freedom’. because our theory is somewhat vague. p > n. what we believe we ‘know’ about the function that it represents—we commit to a parametric or nonparametric approach. on the ‘right hand side’ there are around 1300 variables— predominantly parliamentary divisions—which we wish to use as regressors (or ‘predictors’). (4. One unsatisfactory solution is to re-specify the regression problem with fewer variables. A better option is to employ systematic exploration methods that uncover structure and reduce the data down to its ‘important’ component variables.4 85 Unfortunately.3 Tree-based Regression and Random Forests OLS is a special case of ‘regression’ which we can write more generally as y = m(x) + where m() describes an unknown model for the data we observe.2) for essentially any parametric model—discussed in more detail below—is elusive. Moreover.

G() is some function and β are the parameters to be estimated. Notice that the goodness of fit criteria may be misleading in these circumstances. Political scientists are most familiar with generalized linear models (GLM) where y is linearly associated with x profiles but in which the functional relationship between them can be non-linear. however.Section 4. When analysts have a strong grasp of the underlying physical causal process (and thus. While parameters of the model are estimated. 11 . In contrast to parametric models. they do not necessarily have a clear interpretation in terms of the physical causal That is. standard parametric models cannot be fit to the data. E(y|x) = G(βx) where E() is the expectations operator. for problems in which we have p > n.11 Examples include logit or Poisson regression. This is precisely because the assumptions that make parametric models straightforward to fit also restrict the variety of relationships that they can adequately describe: and if the model is ‘incorrect’ for the data (e. the potentially subtle ways that variables interact). In any case. when the data are not well understood.4 86 the functional form of the relationship between the dependent and independent variables is assumed known. linearity is assumed when non-linear relationships are more apt). or theory is not easily forthcoming—especially when the number of independent variables is large and thus the causal story complex. They are of limited use. parametric models are a good choice.g. but it contains parameters whose values are unknown and to be estimated. nonparametric models make very few assumptions about the functional form between the dependent and independent variables. for example. results will be less than satisfactory.

Typically. 1996. is that of regression trees (Ripley. at each split. Further splits are then proposed for any cases that are not in terminal nodes (known as ‘leaves’ of the tree). popular in biostatistics. 1996. 213). its median household 12 More details may be found in Appendix 4. The general idea is to split the data so that cases within the subsets are as like each other as possible. We also have ‘explanatory’ variables for each state.Section 4. or ‘node’. There are lots of different rules that may be used to ‘cut’ the independent variables (Ripley. this literally means choosing to split a ‘parent’ node such that the resulting two ‘child’ nodes have the minimum possible within node variance in terms of their values of the dependent variable.7. we obtain some notion of ‘importance’ in a regression context.12 By seeing which variables—when split—best predict the cases. the algorithms divide one independent variable into (binary) disjoint subsets into which the cases can be placed.4 87 process thought to underlie the data: instead the concern is with function estimation and good prediction. 216–220). A simple nonparametric method. the task of optimizing predictive performance is performed by classifying cases according to their independent variable values via a series of logical splits. in terms of George Bush’s percentage of the vote in any given state. and hence several different ways to assess how well the tree classifies the problem in question. To see how this process might work. including its population density. Within this literature. in regression contexts. . consider the following toy example: we have data on the 2004 Presidential election.

78.015 to the left of the figure).00. but only 8 bins as terminal nodes— 65. In this case. they were dropped in the bin of states with a mean value of 65 percent for Bush.60. If it was less than 1. predominantly white states such as Utah and Wyoming.Section 4. with the least dense (< 35. these relatively rural states were once again split on population. population density does particularly well). The next split for those states occurs in terms of their black population. 13 . these are large. there are 50 states. Literally. 46.5 percent (0. and the proportion of its population that are black. 2006) statistical language and environment using the tree package: see Ripley (1996). the tree Implemented in the R (R Development Core Team. 51.29. Third.4 88 income.8 sent ‘to the left’ and those that are more dense than this cutoff sent ‘to the right.20. we made a judgement call as to when the tree would stop growing.13 At the top of the tree. rural. Each leaf (end value) in the figure corresponds with a different bin. 57. and states are placed there contingent on their variable values.17. the variables are not all of the same ‘importance’: different variables at different nodes do better in reducing the variation in the subsequent (split) data sets than others (here. there is no guarantee that ‘important’ variables will be statistically significant in a regression context or vice versa: classification and/or tree-regression are not synonymous with OLS. Several comments are in order here. In Figure 4.’ On the left. 59. Clearly. 54.00. 48. First. with those having a value of lower than 206. we see the first split was on population density. Second.2) sent to the left. 42.2 we report a tree grown for this data using the split criterion defined above.00.

20 42.186 48.2: Example of a classification tree for American states. .2142 popn.00 51.Section 4.popn < 0.00 57.popn < 0.density < 35.848 | popn.065 65.29 black.025 46.60 Figure 4.17 black.popn < 0.015 black.00 median.income < 40418.78 54.4 89 popn. predicting (categorizing) the percentage of vote for George Bush in the 2004 Presidential election.5 59.density < 206.density < 361.

For most purposes. is grown until no more than 10 cases reside in a terminal node. in terms the percentage of vote for George Bush in the 2004 Presidential election. in fact.3. insisted on a perfect fit to the data: 50 terminal nodes with one for each state. Node labels removed for clarity of presentation.Section 4. Such a tree is given—without the distraction of node labels—in Figure 4. In this toy instance.’ That . fitting a large data set with lots of independent variables in a way that resembles Figure 4.3: Example of a tree that uniquely classifies American states. Its complexity should be obvious.4 90 | Figure 4. we could have.3 won’t be helpful: we will ‘overfit.

Though we can make the selection of ‘right-sized’ trees automatic using various cross-validation procedures. we can ‘stop’ growing at some pre-defined point. for reasons noted below. And.4 91 is. Alternatively. we will essentially reproduce the complexity of the data by fitting its random (‘noisy’) features which are not part of the causal process in which we are interested. One solution is to grow multiple trees—each potentially classifying the cases differently—and find some way to aggregate the results from all of them. we can ‘prune’ the tree back after it is grown. 221–231). perhaps. there may be a better way to proceed. But less complex trees introduce a new problem of bias: simpler models may not be good enough to correctly represent the subtleties of the data we are interested in. while the individual trees will have high variance. A popular and powerful way to do this is Breiman’s (2001) random forests algorithm. Combining the trees will give the best of both worlds: low bias and low variance. To reduce this variance.Section 4. Random forests actually builds on an earlier technique—also proposed by 14 Implemented in R by Liaw and Wiener (2002). there will be a low correlation between them. . and there are various methods for doing so (Ripley. adding new points out of the sample) would result in very different findings. Such trees have high variance. in the sense that altering the data on which they were grown (by. 1996.14 The key idea is that each tree will be unpruned and thus have low bias. This is undesirable given that we want to know about the general structure of the data at hand.

1993. In fact.15 Bagging is an abbreviation of bootstrap aggregating and. In a regression context. in practice this subset is approximately one-third of the total number of independent variables (around 350 here). 4. . five cases). b. 15 . this means taking a sample of size N —with replacement—from D (see Efron and Tibshirani. We proceed as follows: 1. Denote this sample as Dk and notice that the same observations can therefore appear multiple times in Dk . like most tree methods. . approximately one third of the cases will not appear in the sample (called ‘out of bag’). 2. 500) times. a random forest splits a randomly chosen subset of all the predictors. (2004) for a political science application of an earlier contribution by Breiman et al. (1984). In random forests. . and others from D might not appear at all. This strategy results in See Ruger et al. take a bootstrap sample from the data. That is. 3. we average across the trees. repeat steps 1 and 2 1. . . grow one tree from that sample until the cases reside in sufficiently small nodes (commonly. we do not need to make this split. for a general discussion). B (say. These trees are not pruned. In practice. D. train the predictor using Dk .Section 4.4 92 Breiman (1996)—called ‘bagging’. . Random forests get their name from the randomness employed at each node in each tree: rather than employing the best split among all variables to continue to grow the tree. it begins by splitting the data into a ‘training’ and ‘test’ set. . aggregate the results of the B trees.

We will be interested in several quantities from our models: first. to do this. 2004)—without overfitting. to when it is not. and yet the predictive performance generally exceeds that of related techniques—such as neural nets (see Beck. Second. we obtain this by comparing the predicted y when the variable is in the (aggregate) random forest model. we want to know how ‘important’ different variables are. Notice that because random forests are not attempting to estimate regression coefficients in the usual. Appendix 4. de Marchi.Section 4. we will inspect a residual sum of squares measure for trees where the variable is involved. 2000. If random forest classification is successful. our investigation now proceeds in several steps. We can use this fact to impute missing observations.9 explains how.4 93 low correlation between the trees. relative to those where it is not. 4.4 Estimation Plan Given the above explication. GLM. way there is no requirement that n > p. the goodness of fit. Gelpi and Grynaviski.4. the marginal effect of any given variable on y will be a concern. and Appendix 4. Third. which is calculated in a similar way to the R2 for a standard regression with the difference that we explicitly attempt to predict data that was not involved in fitting the trees. which combine both quantitative and qualitative methods: .4.8 gives more technical details. King and Zeng. Broadly. A schematic diagram describing random forests is given in Figure 4. we expect observations which are ‘alike’ to end up in the same terminal nodes across trees.

For each bag. Here the data set consists of ten observations labeled 1 through 10.4 94 Data D 1 2 3 4 5 6 7 8 9 10 ’out−of−bag’ 2 5 8 ’in bag’ 4 10 6 1 5 7 6 2 3 1 4 8 10 2 5 7 5 8 3 7 9 7 3 4 9 6 8 1 2 5 5 10 7 8 7 4 9 4 1 10 9 8 8 7 3 10 2 4 6 bag 1 tree 1 bag 2 tree 2 bag B − 1 tree B−1 bag B tree B Figure 4. some observations are ‘left out. . and the results aggregated.4: Random forest schematic. These are ‘bootstrap sampled’ into bags (1 thorough B).Section 4.’ A tree is grown from each bag.

there is thus no danger that we are looking in the ‘wrong place’ when predicting electoral performance from legislative behavior. use imputed data and random forest to calculate relative ‘importance’ of different controls and roll calls in predicting electoral success. rather than their precise content. we will use data classification techniques to find the most ‘important’ variables (which will involve a judgement call on our part) and then see if and how they fit our theory. 3. we will identify those bills that best predict the dependent variable in question and then describe their broad characteristics in terms of our interest in party discipline versus independence. using 2 as a guide. Because essentially all votes are used. ‘issue of conscience’. etc.5 95 1. To reiterate. use narrative histories (such as Cowley (2002) and Cowley (2005)) to establish nature of bills—whipped/unwhipped. impute values for the missingness in the divisions using a random forest classification 2. Because the method fits by maximizing prediction. etc. ‘programmatic concern’. obtain and report the (random forest) marginal effect of some important variables (roll calls) 4. ‘free’.). As a further comparison we will also report OLS coefficients for the variables from 2 (using each variable with all controls to predict performance). concern will focus on the types of roll calls that matter (‘heavily whipped’.Section 4. . In other words.

Nonetheless.5. both events were landslides for the Labour party as a whole. separately. The (mean) pseudo-R2 goodness of fit measure for the model was around 0.06. we follow the stages set out in the previous section.5 Results The results are in two parts. In fact.5.77).7 percent recorded by Michael Jabez Foster (Hastings and Rye). first for 1997–2001 and then. and a maximum loss of -12. and a maximum gain of 12. for both data sets. Note first that the controls are seemingly essential and all contribute more than any roll call in accounting for differences in the dependent variable. and Labour lost just six seats (of a 179 seat majority) in the latter. 4.5 96 4.5 percent recorded by Khabra Piara (Ealing Southhall). for 2001–2005.1 First term: 1997–2001 For our first data set. In Figure 4. x-axis scale refers to the decrease in the average residual sum of squares when the variable is in the tree. as opposed to when it is left out. recall that the dependent variable is the difference in vote share that the backbench Labour MP’s garnered in their home constituency for the 1997 versus 2001 election. Of course. the percentage of citizens in a district with a debilitating illness . In particular.Section 4.59 percent (variance of 14. we graph the most important 30 variables from the random forest regression where. there was quite significant variation in terms of the performance differences of individual MPs with a mean change of −1.

991102 div307.001023 div046.001023 div303.001219 div294.990210 div302.5 97 Illness DiffT Ethnic Spend Pension LoneParent Unemploy div323.971103 div243.001023 div186.Section 4.990301 div078.990714 div290.5: Variable importance plot for 1997–2001 data.001023 0 100 ’important’ roll calls 200 IncNodePurity controls 300 400 Figure 4.000725 div191.010321 div312.971210 div012.990520 div134.001107 div162.980608 div299.990125 div311.010227 div071. Controls and ‘important’ roll calls used in subsequent analysis shown as subsets of all variables.980622 div079. x-axis gives the increase in the regression performance from including each variable.001023 div300.980622 div281. .000210 div063.000509 div111.

govern- .1 we report our results: for each variable. we give the marginal effect from the random forest: literally. We also report an OLS coefficient (with p-value) from including that predictor in a regression—with all controls. below which there is a slight drop off in importance. followed by turnout differential (DiffT). In both cases. the proportion of Pensioners (Pension). government has preference for outcome). they are left out of the table. the proportion of citizens unemployed (Unemploy) and then the electoral spending differential (Spend).5 98 (Illness in the graphic) is most important. the proportion of lone parent households (LoneParent). perhaps timetabling concerns).Section 4. and here we consider the top 9. but column 3 gives a casual author-based coding of the discipline imposed for that vote. ‘weak’ (little whipping. ‘moderate’ (some pressure. The following points on the graphic refer to the importance of the roll calls and have labels beginning with div. ascertained from a qualitative reading of the relevant House of Commons debate and surrounding literature. it can be quite difficult to know for sure how much party pressure was applied in the various votes. this is the difference in the predicted constituency performance when an MP votes with the minority of his party as opposed to when he votes with the majority. a positive number means the MP did better at election time while a negative one means he was worse off. the proportion of non-whites (Ethnic). Since the controls are not the focus of the study. In Table 4. the whipping is one of four ordinal levels: ‘free’ (essentially no whipping). Deciding which of these bill votes are worthy of further inspection is essentially a judgement call. ‘strong’ (tight whipping. As noted above.

That is. For this bill. rebeling on the Transport Bill (row 7) leads to MPs being punished at election time—i. that voters were sympathetic to such social equality issues. Within the realms of the model we advanced above this suggests two possibilities: first.Section 4. we note that the signs are in the expected direction. while the legalization of same-sex civil partnerships appears to have occasioned much less heated debate in the UK than their potential introduction in the US.5 ment has strong. Second. note that though the government was formally . such that it would be on a parity with heterosexual consent.e. and the MPs that vote for them. programmatic concerns). Certainly. For gay rights. They were ostensibly unwhipped. yet this belies some potential party pressure on MPs. This roll call dealt with the Government’s attempts to privatize the National Air Traffic System. such preferences seem to be the case: some 93% of British respondents in a recent survey supported laws to protect gay people from discrimination (Cowan. voting for an amendment to enforce a minimum age of at least 18 in certain circumstances was punished by the electorate. a negative coefficient—that is statistically significant. 99 Beginning first with bills which Cowley (2002) discusses as central to the Government’s post-1997 legislative agenda. 2007). those taking the more conservative position here did worse. The rebels on this roll call were backing an (non-Government) amendment that would have substantially delayed the denationalization. The vote on the Crime and Disorder Bill (rows 3) was concerned with lowering the age of homosexual consent from 18 to 16.

291 +0.360 0.142 free +0. 2000.971210 div012.287 -0.144 p-value 0.391 -0.075 0.001107 Weights and Measures (S.827 Name Programming of Bills Discipline weak ‘Rebel’ Effect (RF) +0.147 +0.038 0.001219 Table 4. div323.000509 div111.031 0.001023 div303.176 free -0.460 0. random forest ‘coefficient’ in column 3 for being a ‘rebel’(in the minority and/or frustrating the Government’s ambitions).I.001023 div186.1: ‘Most important’ parliamentary votes for predicting electoral performance by backbench Labour MPs in 2001 general election.822 weak +0.165 -1.980622 div281. 100 .743 +1.001107 is division 323 taking place on November 7. OLS coefficient in column 6.179 ‘Rebel’ Effect (OLS) +1.Section 4.011 -0. Name of bill in column 2.010321 div312.006 0. First column is data code for the division with date after point: so.) Crime and Disorder Bill City of London (Ward Elections) Bill Election of Speaker Election of Speaker Transport Bill Doctor Assisted Dying Human Fertilisation and Embryology free free strong -0.472 +0. p-value for that coefficient in column 7.271 -0.092 0.256 0.991102 div307.132 div162.845 +1.171 free +0.213 moderate -0.5 Code div323.

Unlike. say. the United States House of Representatives. when the House of Lords rejected it. Indeed. consider the . Speaker candidates do not usually receive official endorsements though MPs from the same party often vote similarly in the elections.5 101 uncommitted to the amendment. We would predict then. Ultimately then. both coefficients—OLS and random forest—are in the correct direction (positive). the Speaker of the House of Commons is a non-partisan referee of parliamentary debate. and they are statistically significant. In fact. these are free votes. In short then. For example. that rebeling—in the sense that MPs vote in a way that contrasts with the majority of their party—would be a boon to electoral performance. and the whip is not present. these are votes for which voters are uninformed: the candidates are not typically a priori well known. the government threatened to enact it directly over the wishes of the upper chamber. the executive was clearly committed and sympathetic to this issue—suggesting that rebels might be considered at least ‘out of step’ if not ‘troublemakers’ per se. nor do they have particular ‘policies’ beyond their general personal reputations. it allowed it time for consideration and. Votes that concerned the election of a new House of Commons Speaker occupy rows 5 and 6 of the table.Section 4. The other roll calls in the table have a slightly more complicated status in terms of the party pressure applied to the vote. Selected from back-bench MPs. the government reintroduced it via the separate Sexual Offences (Amendment) Bill. Moreover. when the Lords once again refused its assent.

the City gives businesses votes. the Modernisation Committee was chaired by the Leader of the House. Unusually for an authority in a democratic state. Usually. Thus.Section 4. and as a break with standard practice. The City of London (Ward Election) Bill appears in row 4: the coefficient is significant. and the proposed legislation would have extended this right to more businesses. the sign is essentially as predicted: those showing independent-mindedness on a bill that was not an explicit part of the Government’s agenda did better at the polls. several MPs felt that the proposed reforms made it more difficult for parliament to oppose the executive and hold it to account. 2007. a cabinet member. members backing an amendment to these plans that would have extended debating time were de facto frustrating Government ambitions. There are thus two consequences: first. In particular. For many left- . ‘House business’ issues are seen as free votes. the vote was ‘free’ for Labour backbenchers but pressure was surely present and second. MPs accused the Government—via the Leader of the House—of promoting its own agenda (see Kelso. Both in committee and during the ensuing Commons debate.5 102 June 2000 “Programming of Bills” motion that concerned the Modernisation Committee’s recommendations for changes to the way that parliamentary business be conducted. the City of London intended to reform its self-governing arrangements as a municipality. not least because they emanate from non-partisan Select Committees that are designed to hold the government to account. 140–141). By this private bill. the ‘rebels’ were actually the majority of the party on this vote (note that the Leader of the House opposed the amendment). Interestingly. With all this in mind.

though Salter (2004) argues that the Government may have pressured MPs to at least pass something that could be considered ‘reforming’. In this particular case. The Human Fertilisation . it is perhaps understandable that the sign is negative and they were punished for their intransigence. the vote dealt with how long Imperial measurement units would be allowed to continue as a ‘supplement’ to the metric units required by European Union legislation. Certainly.Section 4. In November of the same year. a method of delegated legislation that allows a minister to make changes to rules and laws without the requirement that the proposals be passed as a separate parliamentary bill. The first vote dealt with a particular “statutory instrument”. without explicit Government involvement they would not have been. It should be clear that such a vote should not fall into the category of essential governmental programmatic business as defined above—and we see that ‘rebels’ did better when they opposed this vote. It is unclear exactly how or even if the Government asserted pressure on its MPs though. with Labour backbenchers in the minority condemning them. just to be done with the issue. “Human Fertilisation and Embryology” (row 9) have signs as predicted. the Leader of the House claimed it was an explicitly free vote. the ideological components of the plans came under attack. the bill was not yet passed and there was lively debate as to whether or not the proposals should be reconsidered in the next parliamentary session which. though not at significant levels.5 103 leaning MPs in the Labour party. To the extent that the rebels were now holding up procedural Government business. such a plan was anathematic. In July of 1999. The votes concerning “Weight and Measures” (row 2).

say. Given that Labour entered the election with 30 seats where the majority was less than 7 percent.Section 4.5. The final case to consider concerns a vote on “Doctor-Assisted Dying” (row 8). It is thus not precisely clear where the average Labour voter places herself on the issue. Euthanasia is not generally a salient issue for most British voters at the ballot box (unlike. conservative. though their overall performance was sufficient to garner a workable majority of 66 seats. the effect is apparently not significant. In any case. and the debate has been characterized by its cross-cutting dimensions. abortion for US voters). The national swing was a commensurate 6 percent away from the governing party. the British Medical Association (the main doctors’ trade union) and several prominent left-wing MPs oppose doctor assisted dying. 4.2 Second term: 2001–2005 The 2005 election saw most Labour MPs do worse than in 2001. religious groups.5 104 vote pertained to cloning as a medical research practice. For example. so it is hard to know what to make of the fact that ‘rebels’ (those who voted in favor of euthanasia) did worse at election time. we see rebels garnering a (minor) uptick in support for their ‘independent-mindedness’ on a free vote for which the public is presumably fairly ill-informed. this was clearly a more competitive election. with a mean loss of 7 percent in our data set (variance was 12. The biggest (Labour) loser of the election was Roger Godsiff who suffered a reduction in his vote total of some 21 percent (primarily to a candidate from the new anti-war Respect party) and the biggest gains were made by Shaun .32). Again.

Section 4.5

105

Woodward, a Conservative defector who left his original party in 1999. In Figure 4.6 we graph the most important variables for this regression from the random forest technique. The (mean) pseudo-R2 goodness of fit measure for the model was around 0.04.

Once again, we note that the controls are certainly important, though their relative ranking has been permutated from that in Figure 4.5. And, again, we make a judgement call as to which of the bills are worthy of further inspection. Here, it is just 10, and we report their effects in Table 4.2. As in the 1997–2001 session, we see that the ‘Programming of Bills’ is an apparently important predictor, occupying rows 1 and 8 of the table. In both cases, ‘rebeling’—in the sense of voting with the minority—was beneficial for MPs (and this effect is statistically significant). These votes were explicitly free though, once more, knowing precisely how much pressure was applied in practice is difficult to gauge. With this in mind, as with our previous findings, our theory predicts the positive sign we see in practice.

Unlike the programming matters of the previous paragraph, the “Children Bill” was part of the Government’s legislative plans for the session and was therefore whipped appropriately. The vote in row 2 of the table actually concerned a backbench amendment (from David Hinchliffe, Chair of the Health Select Committee) that would have struck out part of the Government’s bill that allowed parents to use “reasonable physical chastisement” in disciplining their children. The result would be a bill that essentially outlawed all such spanking. Somewhat unusually—given that it proffered a change

Section 4.5

106

Pension Ethnic Unemploy Illness Spend LoneParent DiffT div278.041026 div306.041102 div044.011025 div283.041026 div208.040628 div358.031110 div011.010705 div279.041026 div048.050126 div240.020514 div336.021029 div281.041026 div045.050125 div093.011129 div180.030508 div237.020514 div311.041108 div316.020724 div059.040225 div241.030618 div307.041103 div239.020514 div244.020516 0 50 100

’important’ roll calls 150 IncNodePurity

controls

200

250

Figure 4.6: Variable importance plot for 2001–2005 data. Controls and ‘important’ roll
calls used in subsequent analysis shown as subsets of all variables. x-axis gives the increase in the regression performance from including each variable.

Section 4.5

Code Programming of Bills Children Bill: Reasonable Punishment Social Security Removal of References to Strangers Human Tissue Bill: consent for removal Fluoridation of Water Supplies Members’ Allowances, Insurance Programming of Bills Modernization of the House Liaison Committee: Power to Take Evidence free free free free +0.100 +0.250 +0.047 +0.082 free +0.141 weak/mod +0.240 free +0.288 weak +0.256 moderate -0.130 -1.113 +0.931 +1.297 +1.335 +1.233 +0.243 +1.263 +2.517 +0.586 free +0.535 +1.660

Name

Discipline

‘Rebel’ Effect (RF)

‘Rebel’ Effect (OLS)

p-value 0.005 0.039 0.130 0.043 0.024 0.013 0.630 0.046 0.634 0.197

div278.041026

div306.041102

div044.011025

div283.041026

div208.040628

div358.031110

div011.010705

div279.041026

div048.050126

div240.020514

Table 4.2: ‘Most important’ parliamentary votes for predicting electoral performance by backbench Labour MPs in 2005 general

election. First column is data code for the division with date after point. Name of bill in column 2; random forest ‘coefficient’ in column 3 for being a ‘rebel’(in the minority and/or frustrating the Government’s ambitions) ; OLS coefficient in column 6; p-value for that coefficient in column 7.

107

the substantive circumstances of the vote are quite odd by Westminster standards. the Government clearly had a preference on the vote. on the philosophical basis that such an amendment would. “Members’ Allowances. “Modernization of the House” (row 9) referred to the sitting hours of the Commons and “Liaison Committee: Power to Take Evidence” (row 10) dealt with procedural issues for cross-party committees.2 are free votes. though it is only significant in one case. by presuming consent. In all cases. and we see a negative and significant coefficient implying that ‘rebels’ did worse after ‘defying’ the executive. the significant and positive coefficient for rebel behavior is perhaps less anomalous than it might first appear. Clearly then. we see the sign is positive. as predicted for ‘rebeling’ on such unwhipped votes.5 108 to government legislation—the vote was ‘free’ in theory. the roll call concerned a Liberal Democrat amendment to a government bill that would have made organ donation automatic unless the deceased had previously registered an objection.Section 4. the “Removal of References to Strangers” (row 4) dealt with changes to parliamentary language. . While this was a free vote for both the main Opposition parties. Nonetheless. be denying individuals the right to exercise their conscience. it was whipped by the government. For example. In particular. Insurance” referred to various MP compensation concerns. this bill was not a cut and dried “government-versus-opposition” case and the low turnout adds fuel to this argument. With this in mind. The other piece of Government legislation in the table is the “Human Tissue Bill” (row 5) but. once again. primarily that dealt with internal House issues. Several of the remaining entries in Table 4.

The first. The Conservatives abstained on the issue and.2 are quite obscure. MPs rebeling against important government legislative business (‘trouble-makers’) were punished by voters. though not technically a free vote. Labour MPs voting against popular social justice legislation— in particular. the sign is positive—rewarding the minority on a free vote—and.3 Summary of Results We found that controlling for various other factors. row 6 deals with “Fluoridation of Water Supplies”. 4. Finally.5. As predicted. in fact.For both 2001 and 2005.For 2001.5 109 The last two votes to consider from Table 4. gay rights—in free votes were generally punished by the electorate . The free vote concerned whether or not local authorities could force water providers to add fluoride to residential and other supplies. significant. and this effect was sometimes statistically significant. it was not a crucial part of the Government’s agenda. MPs dividing with the minority on free votes (being ‘mavericks’) which involved issues on which voters were illinformed tended to be rewarded by voters . . .For both 2001 and 2005. on “Social Security” (row 3) actually refers to a statutory instrument that regulates unemployment services in the UK.e.Section 4. a perennial debate in the Commons. it is positive) but it is not significant. Thus the sign for ‘rebeling’ is in the correct direction (i.

Moreover. were on the order of 1 to 2 percent of the vote differential from the previous election. 2005). this finding makes sense of seemingly inconsistent public opinion preferences for both party loyalty and free thinking. Using new non-parametric techniques and roll call records in their totality. As was explicitly noted above. we found that (Labour) voters do indeed seem to respond to overall patterns of activity: they punish ‘trouble-making’ rebellion on heavily whipped government business.the OLS effects. depth to our understanding of accountability in Westminster systems. The current paper was concerned with connecting the two trends and systematically assessing the evidence for voter reaction to patterns of MP activity in Parliament. 1997.Section 4. Pattie. while rewarding more maverick legislators when party unity is not per se at stake. 4.6 Discussion It seems that old certainties about British politics are changing: government party MPs are now more rebellious than they have been in a generation (Cowley. and adds a layer of intuitive. This perhaps allows some ‘catch up’ with Americanist efforts. the model of voter behavior was ‘as if’ in . though the signs generally agree. though subtle. while producing new directions for future research. 1995). Johnston and Fieldhouse. 2002. The random forest (marginal) effects were generally of a much smaller magnitude.6 110 . when significant. and there is increasing evidence that local factors matter at election time (Denver and Hands.

the House of Commons Modernisation Committee is engaged in an attempt to strengthen the role of backbenchers in terms of their dealings with the executive..Section 4. we think that voters receive signals from a possibly media-based elite. Further work in this area might thus profitably gather media or other data on the messages that voters receive in order to disentangle what exactly are the triggers that change the labels MPs receive. it could not trace the causation path precisely. We have suggested that voters are already either heterogenous or discerning in their responses to backbench behavior. who digest and then describe overall patterns of MP behavior as ‘rebellious’. Given that institutional norms— whipping and unity—appear to both affect and be affected by ‘what voter’s . Cowley. Rather. It is hoped that this will resurrect public engagement in politics after two general election turnouts of around 60 percent (very low numbers by Western European standards). ‘conforming’ and so on. e. ‘loyal’. Finally. 2006). and further independence for MPs will presumably make them more so. and is soliciting expert opinion on these matters (see. ‘fractious’. At the time of writing. this paper has potentially important ramifications for institutional reform in the United Kingdom. while our data and method tapped the generalities of this dynamic. So.6 111 nature: we would not contend citizens have a sufficiently advanced knowledge of policy making or parliament to directly reward or punish legislators on certain bills.g. Whether MPs supporting these reforms understand that their electoral fortunes will be increasingly tied to their legislative performance and choices—and not the party’s overall popularity—is unclear.

changing those norms ought to have interesting and complex consequences.6 112 want’. .Section 4.

3) is simply the residual sum of squares at t. In particular.3) where yi is the dependent variable value for the ith case at node t. we require that our split s (on some value of one of the covariates) maximizes R(t) − R(tL ) − R(tR ) (4. Notice that Equation (4. written R(t).113 Appendix to Chapter 4 4.4) . We measure the ‘impurity’ of t via the least squares deviation (also called the ‘deviance’). We will split the cases from t into two distinct subsets such that the subsequent ‘child’ nodes have summed deviations that are as small as possible. and y(t) is the mean of all the cases in node t.7 Regression Tree Split Criteria Consider a ‘parent’ node t containing some number of cases N (t) that is to be split. which is (yi − y(t))2 i=t R(t) = (4.

8 114 where R(tL ) is the within node variance of the left child and R(tR ) is the within node variance of the right child. Take the average of all those oob predictions for i (that is. For the current technique. one for each tree where it is oob) and denote that value as yi . it is ‘run down’—that is. s1 and s2 are proposed. in any given tree grown) around one third of the cases from the original training set will not appear there. Two possible splits in the covariates. This is equivalent to choosing a split maximizing the reduction in deviance. These are called “out-of-bag” (oob) data. the other with a deviance of 19. For s2 . we would like to know the percentage of the variance in y that is explained by our model. Suppose s1 yields two children: one with a deviance of 11. Every time a particular case i is in the oob data with respect to a tree. Akin to ˆoob linear regression. this is computed as 1− MSEoob σy ˆ2 . As an example.Section 4.e. Variable Importance and Effect Determining goodness of fit for random forests regression proceeds as follows. the left and right child both have a deviance of 15.4). or different splits in different variables. 4. Notice from step 1 in Section 4. These may be different points in the same variable. regressed—according to that tree and an (oob) prediction produced.8 Model Fit. suppose our parent node t has a variance of 20. These splits are considered equivalent since both result in an equal value of (4.4 above that in any given bootstrap sample (i.

Section 4. if the variable is ‘important’. Liaw and Wiener (2002) suggest estimating the function ˜ f (v) = 1 n n m(v. The element m(v. Finally. it will be responsible for a better regression performance: it will generally occur more often in the ‘best’ split and/or the reduction in deviance when it occurs will be greater. . The importance of variable v is estimated by considering the decrease in node ‘impurities’ when v is split (i. These impurities are measured via the residual sum of squares (RSS)—a formula we given in Appendix 4. xiC ) is the (aggregated) random forest regression function (the predicted y): the idea being that we vary the values of v (‘aye’ or ‘no’ in the case of a roll call) to see the (predicted) effect on constituency performance. We compare the (average) value of RSS when v is in x and when it is not. 2002.7. it is part of a tree). 20). xiC ) i=1 where v is the variable of interest and xiC are all the other variables. averaged out over all the trees in the forest. Intuitively. we will typically want to know the marginal effect that a variable v has on y.9 where MSEoob = 1 n 115 n {yi − yi }2 ˆoob i=1 where MSE is the mean square error (Liaw and Wiener.e.

.1 0.5 0.1 1|     2| 0.1     3| 0. Observation 3 is first filled with a 0 (since it is the most common category).9 116 4.2 Observation 3 is with Observations 1 and 2 vation 4 only 1 5 1 2 of the time. Suppose that the resulting proximity matrix was   1 2 3 4      0.6 0. j pair of cases ends up in the same terminal node.5 0.1 0. As a toy example consider a variable x = [0. For continuous missing data. Begin by filling missingness in each column with the median (if continuous) or modal (if discrete) category for that column.2     4| 0.Section 4. 1].6 0. but with Obser- of the time: the largest average proximity is thus zero and we impute this for the observation.4 0. 0. Then run a random forest regression procedure and compute the proximity matrix: that is the fraction of times (across all the trees) that each i. NA.9 Imputing Missingness Suppose that there is missingness in the underlying data (in the x). with the proximities as the weights. For categorical missingness—the case for our work here—the category with the largest average proximity (which will be either ‘aye’ or ‘no’) is the imputed value.4 0. the imputed value is the weighted average of the non-missing data for that column.

we . a parameter to be estimated. In the third chapter. What draws the chapters together is the focus on finding new solutions to ‘old’ problems: identifying ‘turning points’ in conflicts. we argued that a very large number of variables (i. In the second chapter. roll calls) might collectively demarcate members of parliament in terms of their appeal to voters. Rather than employing an ad hoc ‘thinning out’ of variables such that parametric models could be fit. itself. The essays all suggested thinking about these issues in new ways. we employed a change point identification technique that was notable for its flexibility: the number of parameters was.e. Our statistical model reflected our (more realistic) assumptions on this matter. describing the ‘effects’ of roll call voting on electoral support. involving three somewhat different methods. measuring the ‘power’ of political actors.117 Chapter 5 Discussion The preceding dissertation was concerned with three quite separate substantive areas. we argued that ‘power’ makes little sense in the absence of ‘contests’ in which actors compete to garner finite resources. In the first chapter.

other (typically ‘hierarchical’) methods fit numerous models (one for each value of k total clusters) which must then be compared. A model of legislator behavior is posited in which elected members compare the utility of the status quo relative to that the policy changed potentially introduced by passing a new bill. 2002. 2002. such models produce misleading or unhelpful results. which search for groupings. Clustering algorithms. but it may be helpful to think more generally about applications here. By contrast. 316)— require an a priori commitment to a fixed number of clusters K to which cases are then assigned. Since Bayesian procedures are philosophically appealing in terms of model comparison (see Gill. seems an intriguing tool for ‘cluster analysis’ in parliaments. 1997). rather than an ideological spectrum. Spirling and McLean (2007) show that strong whipping in Britain’s House of Commons produces rank ordering on an ‘agreement with the government’ dimension. 199–224). Poole and Rosenthal. random-utility) statistical model. ideal point estimation (frequentist or Bayesian) has been a popular choice for analyzing legislatures (Clinton.Section 5. they seem . 2004.0 decided to take a more systematic non-parametric approach. Reversible jump Markov chain monte carlo. Historically. Some of these techniques— optimal partitioning methods being among them (Venables and Ripley. As noted in Chapter 4. 118 In the individual chapters we made some comments about future uses of these techniques within their substantive areas. A distribution for the error terms is assumed—typically Type-1 extreme value or Guassian—to produce a (parametric. On occasions. which we introduced in the first chapter. are an alternative way to proceed. Jackman and Rivers.

Over a series of (independent) contests. RJMCMC would seem a particular good option. For example. The same is true of British electoral contests. 1981). ‘arguments’ essentially ‘compete’ with respondents either suggesting which they prefer directly. or voting commensurate with their views. US presidential primaries and caucuses are typically three. psychologists and economists have studied the consequences of ‘framing’ effects. at least since the time of the ancient Greeks. because we would like to be flexible about the number of clusters.g. As constructed. with one losing and one winning. Moreover. Conservative and Liberal Democrat candidates vie for success in parliamentary constituencies.Section 5. though in more recent times. there are not always clear ‘winners’ and . where Labour. analysts could begin to model why certain candidates win or lose. Of course. an election campaign. at least in the early stages. Tversky and Kahneman. Using the Bradley-Terry model in the second chapter would be one way forward here. The earliest accounts were primarily interested in rhetoric as an art (e.g. we make the reasonable assumption that success is binomially distributed. For example.or four-way competitions. Moreover. The second essay (Chapter 3) made the case that we can consider political actors as players competing in contests. In these cases. 322BC/1991). contests in politics are often more complicated. especially with respect to risk-taking behavior (e. Indeed. the standard Bradley-Terry model assumes that contests involve two players. Aristotle. This is an idea with mileage outside the current context. scholars have studied the power of argument to convince citizens. say.0 119 a natural choice here. if one could properly separate competing arguments in.

these operationalization decisions can be refined within the (first) estimation ‘round’ rather than on some ad hoc basis with knock-on effects for the study as a whole. With these cases in mind. or that ‘more proportional’ electoral systems give rise to fiercer intra-party competition.Section 5.0 120 ‘losers’: outcomes are often ‘tied. there are presumably several indicators and ways of measuring the attributes of the independent variables and possibly no way to a priori decide between them. translation from ‘theory’ to statistical model is typically not direct in political science: for example. In such cases. An economic version of this holds when firms bid for public works contracts. a scholar may believe that ‘corruption’ has some systematic effect on citizen-to-citizen trust in a city . 3) makes the case that political science “outdistances most other social sciences” in terms of data collection. perhaps facilitating the estimation of abilities when contests have several players and when ties are possible. when candidates run for President of the France. it would be helpful to think more generally about the Bradley-Terry model.’ Paradoxically. As data becomes increasingly available in massive amounts— . Moreover. Since non-parametric techniques allow one to include literally hundreds more predictors than observations. this idea often holds in majoritarian electoral systems where there is one winner and all others. The third essay (Chapter 4) claimed that. one helpful way to proceed was the employment of a ‘random forests’ classification procedure. the model matrices are potentially very large. when the number of variables to consider was greater than the number of parameters to be estimated. lose ‘equally’: for example. in effect. For data exploration purposes then. King (1989.

however. which is ‘more complicated’ than cross-tabs. incorrect and unfair. Phrased in various ways. Of course. The claim is incorrect because it fundamentally misunderstands the goal of political methodology. to fail to address a criticism of quantitative techniques that arises particularly in political methodology.0 121 especially via the internet—such work would seem important and valuable. Otherwise put. and has suggested several realms in which either current practices might be improved. typically implicit and casually suggested.’ This argument is imprecise. or where new profitable avenues of investigation exist. such standards clearly differ from scholar to scholar: reversible-jump Markov chain monte carlo approaches might seem unaccessible to someone who know little of Bayesian methods. much of political methodology is ‘complicated for the sake of being complicated. It is imprecise insofar as it can be applied—in its most basic form—to any data analysis technique: logistic regression is ‘more complicated’ than OLS. the claim of some is that methodologists fetishize complexity in their models. but are a helpful and important extension for an analyst already familiar with MCMC techniques. The aim is to . it is unclear as to where the analyst should draw the line: what extra utility must be garnered in order to approve of a more complex technique? Moreover.Section 5. which are ‘more complicated’ than frequency summaries. All told then. this dissertation has argued that our statistical solutions should fit our problems. non-parametric fitting should not be considered the final stage in any investigation of political actor behavior: nonetheless. that is. it can help us think more clearly about possible micro-mechanisms. It would remiss of the author.

‘more’ or ‘better’ theory.Section 5. in computational terms. the claim is unfair because it cynically assumes that methodologists either deliberately dissemble (and are thus dishonest in describing) their work. methods and data are the way forward. Note also that. else they are simply incompetent in solving the problems that the discipline presents. Finally. the costs involved with ‘complicated’ techniques have fallen precipitously over time and will continue to do so.0 122 provide tools that help us understand political behavior. . Yet. one of the strengths of work in methodology is that assumptions of models and techniques are both clearly laid out. to the extent that progress towards this goal remains incomplete (few if any would disagree with this contention). It is certainly in this spirit that this dissertation is intended as a contribution to the discipline. It can surely never hurt to refine our techniques and our understanding of their capabilities. and written in a language common to its practitioners.

S. 2 ed. and Bruce Bueno de Mesquita.” Annual Review of Political Sience 5:423–50. 1965. 1962. Amadae. Alesina. “Weighted Voting Doesn’t Work: A Mathematical Analysis. 2002.123 Bibliography Achen. “Two Faces of Power. Kennedy). Bachrach.” Annual Review of Political Science 2:269–295. Christopher H.” American Political Science Review 56:947–952. Categorical Data Analysis. Banzhaf. John F. Alan. 2002. New York: Oxford University Press (translation by George A. 1991. New York: John Wiley and Sons. “The Rochester School: The Origins of Positive Political Theory. 322BC/1991. “Toward a New Political Methodology: Microfoundations and ART. Aristotle. Agresti. 1999. M.” Rutgers Law Review 19:317–343. “Why are Stabilizations Delayed?” American Economic Review 81(5):1170–1188. . Alberto and Allan Drazen. Peter and Morton Baratz. Rhetoric.

Oxford: Clarendon. A. Friedman. Breiman. Brian. Stephen P. 1996. L. Breiman. “On the Rationale of Group Decision-making.” Journal of Political Economy 56:23–34. 1997. R.” Machine Learning 45:5–32. Bradley. Classification and Regression Trees. Paul.” Comparative Political Studies 40:755–781. Beck. J.. and the Dejected: Explaining Government Rebels in the 2001-2005 British House of Commons. CA: Wadsworth and Brooks/Cole.” Machine Learning 24(2):123–140. Giacomo and Simon Hix. H. Monterey.Section 5. 1952. Olshen and C. “Random forests. Nathaniel. “Bagging predictors. 1984. 2000. Brooks. the Ejected.” Biometrika 39:324–345.” American Political Science Review 94(1):21–36. Leo.0 124 Barry. Breiman. Duncan. Gary King and Langche Zeng. J. 2001. Leo. “Improving Quantitative Studies of International Conflict: A Conjecture. “Rank Analysis of Incomplete Block Designs I: The Method of Paired Comparisons. Brace. Benedetto. 1948. Black. 1984. Is it Better to be Powerful or Lucky? In Democracy and Power.” The Journal of Politics 46(2):556–571.” The Statistician 47(1):67–100. “The Rejected. 1991. . 2007. Stone. R A and M E Terry. “Markov chain Monte Carlo Method and its applications. “Progressive Ambition in the House: A Probabilistic Approach.

Cambridge.1016/(S0140-6736(06)69491-9). Gilbert. Politics and Society. B.” The Lancet (online) 10.” Applied Statistics 41(2):389–405. Canes-Wrone. “Mortality after the 2003 invasion of Iraq: a cross-sectional cluster sample survey. 1774/1975. Butler.. Burnham. David and Donald Stokes. Brandice. Brady and John F. 1992. Speech to the Electors of Bristol. Political Change in Britain. Chib. Riyadh Lafta. 1987. “Out of Step. Shannon Doocy and Les Roberts. Alan Gelfand and Adrian Smith. 1974. MA: Harvard University Press. Fiorina. John A. 2002. In Edmund Burke: On Government. 1999. 1998. Ian. Carlin. 2 ed. 2006. Macmillan.W. The Personal Vote: Constituency Service and Electoral Independence. Cogan.Section 5. Siddhartha. Cain. London: Fontana. David W. Ferejohn and Morris P. Hill. . Bruce E. Critical Elections: British Parties and Voters in Longterm Perspective. ed. Edmumd.0 125 Budge. Bradley.” American Political Science Review 96(1). Burke. “Estimation and comparison of multiple change point models. Out of Offce: Electoral Accountability and House Members’ Voting.” Journal of Econometrics 86:221–241. “Hierarchi- cal Bayesian Analysis of Changepoint Problems. London: Sage Publications.

0 126 Clinton. Christopher Gelpi and Jeffrey D.iraqbodycount. Berkeley: University of California Press. Legislative Leviathan: Party Government in the House.” American Political Science Review 98. “Living Together: British Attitudes to Lesbian and Gay People. 1993. In Change in British Politics.”. “Iraq Body Count. 2006. Cowley. 2002. Scott. Philip. London: Frank Cass pp. Crewe. David and Gordon Hands. Portland. London: Politico’s Publishing. 2004. Joshua. Hugh Berrington. and Mathew D. Cox. 1997. “Untangling Neural Nets. 2005. URL: http://www. McCubbins. Cowan.” Written submission to the Select Committee on Modernisation of the House of Commons Session 2006–07. Revolts and Rebellions: Parliamentary Voting Under Blair. 2004. “Strengthening the role of the backbencher. 2006. Ivor. Denver. Dardagan. Philip.net/ de Marchi. Philip.” Stonewall Poll by YouGov. Cowley.Section 5.” American Political Science Review 98(2):371–378. . Simon Jackman and Doug Rivers. no. The Electorate: Partisan Dealignment 10 Years On. 1. Oregon: Frank Cass. Modern Constituency Electioneering. 183–215. 2007. ed. Katherine. “The Statistical Analysis of Roll Call Data. Gary W. Cowley. London: Politico’s Press. 1983. The Rebels: How Blair Mislaid His Majority. Hamit and John Sloboda. Grynaviski.

“Bargaining and Fighting: The Impact of Regime Type on War Onset. Anthony. 1996. Brown and Company. Representatives. An Introduction to the Bootstrap. 1973. The Measurement of Vote ing Power: Theory and Practice. “tapiR and the Public Whip: Resources for Westminster Voting. Power. Problems and Paradoxes. . Downs.” American Journal of Political Science 48(2):296–313. Firth. Cheltenham: Edward Elgar. Felsenthal. Roll Calls and Constituencies. “A Priori Voting Power: e What Is It All About?” Political Studies Review 2:1–23. 1993.0 127 Dowding. Keith. Filson. 2004. Boston. New York: Chapman & Hall.” Journal of Statistical Software 12. Minneapolis: University of Minnesota Press. MA: Lexington Books. New York: Harper. Darren and Suzanne Werner. Dan S. Dan S and Mosh´ Machover. David and Arthur Spirling. Felsenthal. Firth. 1974. Richard F. 2005. Bradley and Robert Tibshirani. 1998. Fenno. Morris. and Outcomes. David. Congressmen in Committees.Section 5. “Bradley-Terry Models in R. Lexington. MA: Little. 1987. An Economic Theory of Democracy.” The Political Methodologist 14(2). Fiorina. 2006. 2004. Efron. Duration. and Mosh´ Machover.

Geoffrey and George Tsebelis.” Political Analysis 14(4):456–487. 1999. 1998. “Bayesian Analysis for Political Research. Ralph K.” Journal of Theoretical Politics 11:291– 308. “Why Power Indices for Assessing European Union Decision-Making?” Journal of Theoretical Politics 11:321–330. . David.0 128 Fowler. “The Morse Committee Assignment Controversy: A Study in Senate Norms. 1999. Jackman. Gill. Jeff. “The Value of Committee Seats in the House 1947–1991. 2002. “Why Resist the Temptation to Apply Power Indices to the EU. Peter J.” Biometrika 82(3):711–732. Green. Holler.” American Political Science Review 51(2):313– 329. “Reversible Jump MCMC computation and Bayesian model determination.” Annual Review of Political Science 7:483–505. Bayesian Methods: A Social and Behavioral Sciences Approach. James H. 2005. Towards Automatic Reversible Jump Markov Chain Monte Carlo PhD thesis University of Bristol. 1957. Groseclose. “Connecting the Congress: A Study of Cosponsorship Networks.” American Journal of Political Science 42(2):453–474. Tim and Charles Stewart. 2004.Section 5. New York: Chapman & Hall/CRC. Simon. 2006. 1995. Hastie. Huitt. Garrett. Manfred and Mika Widgren.

“Bayes Factors. New York: Knopf. Raftery. Johnson. “Do MPs have the ‘Right Stuff’ ?” Parliamentary Affairs 60:164–169. Gary.” Journal of the American Statistical Association 90(430):773–795. 1949. “Strategic Politicians and the Dynamics of House Elections. Southern Politics in State and Nation. Discipline. Irving L. Unifying Political Methodology: The Likelihood Theory of Statistical Inference. 1972. “Proper Nouns and Methodological Propriety: Pooling Dyads in International Relations Data. . 2006. and Adrian E. Kelso. Boston: Houghton Mifflin Company. 1946-86. Robert E. 2007. O. Christopher. Janis. King.” International Organization 55(2):497–507. 2001. On Division: Cohesion.” British Journal of Political Science 36:561–74. V. “The House of Commons Modernisation Committee: Who Needs It?” British Journal of Politics and International Relations 9:138–157.Section 5. Gary. Kam.0 129 Jacobson. “Demotion and Dissent in the Canadian Liberal Party. 1989. Catherine and Gemma Rosenblatt. Gary. University of British Columbia: Book manuscript under review. Ann Arbor: University of Michigan Press. Victims of Groupthink. 2007. Christopher. and Intraparty Politics. Kass. 2007. 1989. King. 1995. Alexandra.” American Political Science Review 83:773–93. Kam. Key.

1998.” Mathematical Geology 35(2). M. Andy and Matthew Wiener. “Relevance of Voting Power. Leech.S. Ann Arbor: University of Michigan Press. “Smoothing and Change Point Detection for Gamma Ray Count Data. Karl. “Classification and Regression by randomForest. Liaw. Jan-Erik and Sven Berg. Daniela. Chicago: University of Chicago Press. . Dennis. 1967.” Warwick Economic Research Papers (644).” Journal of Theoretical Politics 11:309–320. Das Kapital. Keith. Information and Legislative Organization. David R. Dunsmuir. Mayhew. Congress : the Electoral Connection. Krehbiel.” Political Studies 50(1):1–22. “An Empirical Comparison of the Performance of Classical Power Indices. David J. Lawmaking. English Translation. 2003. 1999.Section 5. When and How. Marx. Keith. Pivotal Politics: A Theory of U. 1936. Lane. Who Gets What. 1991. Lasswell. Nott and William T. 2002a. New Haven: Yale University Press. 1974. Leonte. “Computation of Power Indices. 2002.0 130 Krehbiel. Dennis. 2002b. Harold. New York: Whittlesey House.” R News 2(3). New York: International Publishers. Leech.

Section 5. Jong Hee. Philadelphia. 1993. Park. .0 131 McCarty. Lexington: University of Kentucky Press. Iain. Keith Poole and Howard Rosenthal. Cambridge. Johnson. Philip and David M. 2000. London: Frank Cass. Pattie. Arthur Spirling and Meg Russell. Paltzelt. What Can an Individual MP Do in German Parliamentary Politics. “Modeling Structural Changes: Bayesian Estimation of Multiple Changepoint Models and State Space Models. 2006. 2006.” British Journal of Political Science 24:359–80. W. Back from Westminster: British Members of Parliament and their Constituents. New Labour and Public Opinion: The Third Way as Centrism? London: Macmillan pp. 32–44. In The Uneasy Relationship Between Parliamentary Members and Leaders. 1994. 19897–92.” The Political Quarterly 74:298–310. MA: MIT Press. Norton.” Prepared for American Political Science Association Meeting August. Norris. Edward Fieldhouse and R. 2006. “The Price of Conscience: The Electoral Correlates and Consequences of Free Votes and Rebellions in the British House of Commons. Charles. Wood. J. “None of the Above: The UK House of Commons Votes on Reforming the House of Lords. February 2003. Polarized America: The Dance of Ideology and Unequal Riches. Pippa. 2003. Nolan. 2001. McLean.

Vienna. H. 1964. Freeman. New Haven: Yale University Press. Riker. Johnston and Edward A. Rohde. 1983-1992. The Theory of Political Coalitions. 1996. “Risk-Bearing and Progressive Ambition: The Case .0 132 Pattie. “Winning the Local Vote: The Effectiveness of Constituency Campaign Spending in Great Britain. Charles J. Ronald J. ISBN 3-900051-07-0. URL: http://www. 2006.” American Political Science Review 89(4):969–983. David W. The Art of Political Manipulation. R Development Core Team. 1995. William. Fieldhouse. William. New York: Oxford University Press. 1986. New Haven: Yale University Press.Section 5..” American Political Science Review 58(2):341–349. William. Austria: R Foundation for Statistical Computing. San Francisco: W.org Riker. R: A Language and Environment for Statistical Computing. Pattern Recognition and Neural Networks. Brian D. “Some Ambiguities in the Notion of Power. Cambridge: Cambridge University Press. Ripley. William. 1982. 1997.R-project. Poole. Keith and Howard Rosenthal. 1979. Liberalism Against Populism. Riker. Riker. Congress: A Political Economic History of Roll Call Voting. 1962.

Kenneth and Barry Weingast. Shepsle. Searing. Shapley. Westminster’s World: Understanding Political Roles. David W.” Perspectives on Politics Symposium 2. Shepsle. Kim. Ruger.” American Journal of Political Science 23:27–60. “UK OC OK? Interpreting Opti- . Chicago: University of Chicago Press. Andrew D. Salter. 1987. MA: Harvard University Press. Parties and Leaders in the Postreform House. “A Method for Evaluating the Distribution of Power in a Committee System. Rohde. “The institutional foundations of committee power. Cambridge. Donald. “Institutional arrangements and equilibrium in multidimensional voting models.S. 1979. 2007. L. Lee. 1991.” American Journal of Political Science 23(1):1–26. 1954.” American Political Science Review 48:787–92. Spirling.Section 5.” Political Quarterly 75:155–167. “Parliament and The Parliamentarians: The Worrying Case of the City of London (Ward Elections) Bill. 2004. 2004. Quinn. Pauline T. Arthur and Iain McLean. Theodore W. Kenneth. Martin and Kevin M.0 133 of Members of the United States House of Representatives..” American Political Science Review 81:85–104. and Martin Shubik. 1994. “Competing Approaches to Predicting Supreme Court Decision Making.

Allan C. S.” Political Analysis 15(1). Whiting. Tversky. 1981. 1973. Gregory. 2006. Springall. N.0 134 mal Classification Scores for the United Kingdom House of Commons.S. 2002. Modern Applied Statistics with S. Adnan Moussalli and Martin J.” Applied Statistics 22:59–68. Venables.” Science 211:453–458. . George. New York: Springer. A. “Citation Patterns in the Journals of Statistics and Probability. Veto Players: How Political Institutions Work. New York. “Predicting the Length of the 2003 U. Princeton. Tsebelis. Stuart-Fox.: Princeton University Press. William and Brian Ripley. Stam. Wawro.-Iraq War. Ann Arbor: University of Michigan Press. “Multiple Signals in Chameleon Contests: Designing and Analysing Animal Contests as a Tournament.” Print Edition Jul 13.” Foreign Policy Analysis 2(2):101–116. Devi M. Stigler.Section 5. 1994. 2006. 2000. and D. The Economist.” Statistical Science 9:94–108. “Response surface fitting using a generalization of the Bradley-Terry paired comparisons model. David Firth. 2006. Legislative Entrepreneurship in the US House of Representatives. Scott Bennett. Amos and Daniel Kahneman.” Animal Behaviour 71:1263–1271. “Indoor rainforests: Rejecting Eden. 2002..J. “The Framing of Decisions and the Psychology of Choice.

2004.” Political Analysis 12:354–374.0 135 Western. Winston Churchill.Section 5. “Duties of a Member of Parliament.” Parliamentary Affairs 8. Sir. 1954. . “A Bayesian Change Point Model for Historical Time Series Analysis. Bruce and Meredith Kleykamp.

Sign up to vote on this title
UsefulNot useful