Techniques for Choosing Cases 87

5
0

0
00

Techniques for Choosing Cases

c.
�
0

,,,
i I

•
11''

',,
I

:I,Ii
E 0 1,1
<O
0 1,1

"'
0
Gerring, J. (2007). “Chapter 5: Techniques for Choosing Cases”. En ii;
1'1,,

'I
0 ,ij
.Q

Case Study Research, Cambridge, Cambridge University Press: E

'
a
z i:1
pp.86-150. ;q
0
N
,:11 111

l,,!1
I, 1:1
l11li :1::! 1
0

Case study analysis focuses on a small number of cases that are expected
to provide insight into a causal relationship across a larger population
of cases. This presents the researcher with a formidable problem of case
selection. Which cases should be chosen?
In la rgc-sample research, case selection is usually handled by son1e
version of randon1ization. If a sa1nple consists of a large enough num­
ber of independent random draws, the selected cases are likely to be
fairly representative of the overall population on any given variable.
furthermore, if cases in the population are distributed hon1ogeneously
across the ranges of the key variables, then it is probable that sonic
cases will be included fro111 each i1nportant segn1cnt of those ranges,
thus providing sufficient leverage for causal analysis. (For situations in
which cases with theoretically relevant values of the variables are rare, a
stratified sample that oversamples son1e subset of the population n1ay be
employed.)
A demonstration of the fact that randon1 san1pling is likely to produce
a representative sample is shown .in Figure 5.1, a histogran1 of the 1nean
values of 500 random sa1nples, each consisting of 1,000 cases. For each
case one variable has been measured: a continuous variable that falls
som�where between zero and one. In the population, the mean value of
this variable is 0.5. How representative are the randon1 samples? One
good way of judging this is to compare the means of each of the 500
. _
random samples to the population rr1ean. As can be seen in the figure, all
of the sample means are very close to the population mean. So randon1
san1pling was a success, and each of the 500 san1ples turns 011r to he fairly
1Tp1Tsc111:1tivl' of thl' popuL1rio11.
0.0 0.2 0.4 0.6 0.8 1.0
Mean value of X
FIGURE 5.1. San1ple 111eans of large-sample draws. A histogram showing the
1nean values of one VJriable in 500 samples of 1,000 cases each. Population
1nean = 0.5.
However, in case study research the sa1nple is s1nall (by definition);
.ind this makes randomization problematic. Consider what would hap­
pen if the sa1nple size were changed from 1,000 cases to only 5 cases.
'l'he results are shown in Figure 5.2. On average, these s1nall-N ran­
dom samples produce the right answer, so the procedure culminates in
results that are unbiased. I-Iowever, many of the sample means are rather
far from the population mean, and some are quite far indeed. Hence,
even though this case-selection technique produces representative sam­
ples on average, any given sample may be wildly unrepresentative. In
statistical tern1s, the problem is that small sample sizes tend to pro­
duce estimates with a great deal of variance - sometimes referred to
;1s ;1 prohlem of precision. For this reason, random sampling is unre-
11;1 hie in small-N research. (Note that in this chapter "N" refers to
t·;1st·s, not observations.) Moreover, there is no guarantee that a few
l':ISl's, ..:hoscn randomly, will provide leverage into the research question
1h.11 ;1ni1natcs ;.111 investigation. The sa1nple 1night be representative, but
1111 ir1lt,n11:1t·!vc.
ll r:111do111 s:11npli11g is inappropri<1tc as a selection method in case
·,l1 HI�· n·o..;c;1rt·h, l11nv, 1hc11, is one 1·0 choose ,1 s:1111plc co1nprised of
iq· ',(·v,-r.il ,\',<'',:' k(·t·11 in 1ni11d ili:l! 1I H · l'.o,ds of t':l'i(" st•h•l'tion
1111,· 1
 lI. Doing Case Studies Techniques for Choosing C1ses 89
88
TABLE 5.L Techniques of case-selection

:1
"'
1. Typical
• Dc{inition: Cases (01:e or n1ore) are typical exa1nples of some cross-case
relatio:1ship.
• Cross·casc technique: A low-residual case (on-lier),
OJ"' • Uses: Hypothesis testing.
0

e" ' • Represer:Lati11eness: By Jefin.ition, the typicai ca<;e is reprcsentadve.

1. Diverse
0
0
• Definition: (:ases (two or more) ilh1minate the fu'.l range o f variation o n X 1 ) '(,

:I
e or X 1 ;Y.
��
• Cross-case technique: Diversity may be calculated by (a) categorical values of
� 1 or Y (e,g., J cwisb, Catholic, Protestant), {b) standard dcvi�tio:1s of X 1 o::: Y

--
O,D 0,2 0.4 06 .D
Mean value of X
FIGURE 5.2. San1ple 1nea;1s of small-san,ple drav;s, A histogra1n sho\ving the me:111
values of one variable in 500 samples of 5 cases each. Populat\or: tnean � 0.5.
remain the ,.;a1ne regardless of the size of the chosen samplt:. Large·N
cross�case a nalysis and case study analysis both ai1n to ldenrify cases that
reproduce the relevant causal features of a larger universe ( representative·
ness) and provide variation alo:1g the <.Ernensions of theoretical interest
(ca11sal leverage). In case study research, ho\vever, these goals n1ust be
n1et throug�1 purposive ;nonrandon1) selection procedures. These may be
enun1erated according to nine techniques, frorn \vhich \Ve derive nine case
study types: typical.. diverse, extreme_, deriant, irrfluenfii,,1}. crucial_, path­
u,ay, rnost-sitnilar, and most-di,fferent.
i1f
• cortin1m11
1 · 1 or .C1 com b'1nattcns
sn r' · o f van�cs
' 1'e.g., I)ased on cross-tabulations'
factor analysis, or discriminant analysis_'..
• i}ses: }lypothesis gene!'ating or hypothesis
• .Representativeness: Dive:.-se cases a;:e likely ::o be n.-:presen::ative in the mir:in:aI
sense of rcpresrn_ting the full variation of the popul-stion \though they ;:n ight
not tnirrot the d1$tnbution of that variation in the popu:aticn).
3. Extreme
• Definition: Cases (one or n1ore) exeTplify extren1e or unusual values on X 1 or
'{ rd,1tive to '>on1e u nivariate disrributio1� .
.. Cross�case technique; A case lying ;-,1,1ny standard deviations a\VJ.y frorn :he
mean of X 1 or Y.
• (]ses: f-fyporhesis generatJng '.open-en<led probe ofX 1 or Y).
• Representativeness; Ach:evable only in co1np,rdson ro -a larger samole of cases,
4. Deviant
• Definition.: Cases (one or rnore) deviate fro.m so1ne cross-case relationship.
• Cross�casc technique: A high-residual case (oudie:d.
• [Jses: I--Iypothesis gcat:rJting (to develop new expl�nations of Y).
• Representativeness: After the case study is conducted, i t rnav De corroborated
by :J. cross-case test, which in_clud-es a general hypothesis (a �e\v variable) baSed
on tne c:1.:,e study resean.: h. It the ;::ase .is now an on·-lier, it mav be con�idered
1

Table 5.1 su1nmarizes each type, including its general definition, a tech­
nique for ldentlfying it \vithtn a population of potential cases, its uses,
and i-c� probable representativeness. While each of these techniques is
norn1ally practiced on one or several cases {the diverse, most-si1nilar. and
n1ost-diffcrent methods require ax least rwo), all may e1nploy additiona]
cases � with the proviso that, at some point; they wiH no longer offer a�
opportunity for in··depth analysis and \vlH thus no longer be case studies
in the usual sense.
l"he main point of this chapter is co show how case-selection p:oct:dures
rest> at least irnplii.: itly, upon an analysis of a larger population of potcnri;_i 1
cases. The c:isc(s) identified fo� intensive study is chn-..cn fro1n ;l p opula·
tinn, ;111d rbt rl';1so;;s f1 H" rhi;; t:hoicc hing,· upon dn· \\',1\ 1n u·l111.. h i1 [..,
representative of the nc'.Y rela-c:ionship.
5. Influenria)
• Definition: Cases (one or more) ,vith influential configurations of the
independent vari2bles,
• (--,rossMc.ase technique: .Ha t matrix or Cook's distance.
• lJses: f--iypothesi°' testing {to verify the status of cases that n1av influence the
n:sults of ,1 cross-case aEalys1s).
• l\eprest•11!atit•e11ess: Not per�inen::1 given the goals of the in1luensial-case st·01dy.
6. (:ru.:ial
� 1 )1·/inlliou: ( .Jsvo. (,111c or 1no1y) ,1rc n�ost- or least-likely to exhibit a given
nu1t·onh'.
• ( 'n1,.., , <1�,· In lnu,1111" t ,i11.d11.nnr .1·,�1",..,0h'll! oF rctttivc (flh'.frdness.
90 II. D oing Case Studies
TABLE s.1 (continued)
• Uses: Hypothesis testing (confirmatory or disconfirmatory).
• Representativeness: Assessable by reference to prior expectations about the
case and the population.
7. Pathway
• Definiti�n: Cases (one or rnore) \.vhere X 1 , and not X2, is likely to have caused
a positive outcon1e (Y = 1).
. .
• Cross-case technique: Cross-tab (fqr categorical variables) or residual analys1s
(for continuous variables).
• Uses: Hypothesis testing (to probe causal mechanisms).
• Representativeness: May be tested by exarnining residuals for the chosen cases.
8. Most-similar
• Definition: Cases (t\VO or n1ore) are sirnilar on specified variables other than
X 1 and/or Y.
• Cross-case technique: Matching.
• Uses: Hypothesis generating or hypothesis testing.
• Representativeness: May be tested by exan1ining residuals for tbc chosen cases.
9. Most-different
• Definition: Cases (two or more) are different on specified variables other than
X 1 and Y.
• Cross-case technique: The inverse of the rnost-similar 1nethod oi large-N case
selection (see ahove).
• Uses: Hypothesis generating or hypothesis testing (eliminating deten11inistic
causes).
• Representativeness: May he tested by exan1ining residuals for the chosen cases.
situated within that population. This is the origin of the tern1inology j ust
listed - typical, diverse, extreme, and so on. It follows that case-selection
procedures in case study research may build upon prior cross-case analysis
and depend, at the very least, upon certain assumptions abont a broader
population. This, in turn, reinforces a central perspective of the book:
case study analysis does not exist, and is in1possible to conceptualize, in
isolation fro1n cross-case analysis.
To be sure, the sort of cross-case analysis that might be possible in a
is' on how 1nuch information one has about these cases, on what sort of
given research context rests on how large the population of potential cases
general model might be constructed, and with what degree of confidence
that model might be applied. In order for most quantitative (statistical)
case-selection techniques to be fruitful, several caveats must be satisfied.
First, the inference must pertain to more than several cases; otherwise,
Techniques for Choosing Cases 91
statistic-al analysis is usually problematic. Second, relevant data must be
available for that population, or a significant sample of that population,
on key variables, and the researcher must feel reasonably confident in
the accuracy and conceptual validity of these variables. Third, all the
stan<lar<l considerations of statistical research (e.g., identification, specifi­
cation, robustness) must be carefully considered and, wherever possible,
investigated. T shall not dilate further on these fan1iliar issues except to
warn the researcher against the unthinking use of statistical techniques. 1
When these requiren1ents are not 1net, the researcher must en1ploy a
qualitative approach to case selection. Thus, the point of this chapter is
not to insist upon quantitative techniques of case selection in case study
research. My purpose, rather, is to elucidate general principles that might
guide the process of case selection in case study research, ,.vhether the tech­
nique is quantitative or qualitative. Son1e of these principles are already
widely known and widely practiced. Others are less common, or less
well understood. Most of these methods are viable - indeed, are virtually
identical - in qualitative and quantitative contexts. Hence, the statistical
sections of this chapter usually simply reformulate the logic of qualitative
case-selection procedures as they might be applied to large populations
where the foregoing caveats apply.
Typical Case
In order for a focused case study to provide insight into a broader phe­
nomenon, it must be representative of a broader set of cases. It is in this
context that one n1ay speak of a typical-case approach to case selection.
The typical case exemplifies what is considered to be a typical set of values,
given some general understanding of a pheno1nenon. By construction, the
typical case is also a representative case; I en1ploy these two terms syn­
onymously.2 (The antonyn1, deviance, is discussed in a later section.)
Some typical cases serve an exploratory role. Here, the author chooses
a case based upon a set of descriptive characteristics and then probes for
causal relationships. Robert and Helen Lynd selected a single city "to be
I C11j,1rari (2003); Kennedy (2003). Interestingly, the potential of cross-case statistics in
IH'lping to choose c1st·� f-or in-deptb ;inalysis is rtcognizcd in so1nc of the earliest discus­
of tilt' Cl\t' \t11(lr 1 1 1 1 · t l 1 n , I k.g., ()ueen 1 928: 226) .
•.i1 ,w,
l'lw l.1tt(·r l ( T 1 1 1 i� 1>l11·11 1·11q 1 l111T.l 1 1 1 l lw p�vd1ol()gic;il literature (e.g., Hersen anJ Barlow
l '!"'t,: ,1,,1) .
 II. Doing Case Studies
92
as representative as possible of conte1nporary A111crican life." Specifically,
they we re looking for a city with
1) a tetnpcrate cli1nate; 2) a sufficiently rapid rate of growth to ensure the presence
of a plentiful assortn1cnt of the growing pains accornpanying contemporary social
change; ?,) an industrial culture with modern, high-speed machine production;
4) the absence of <lo,ninance of the city's industry by a single plant u_.e., not a
one-industry town); 5) a substantial local artistic life to balance its industrial
activity . . . ; and 6) the absence of any outstanding peculiarities or acute l?cal
problerns \vhich would 1nark the city off from the 1nidchannel sort of Arnencan
community. 3
After exa1nining a numbe r of options, the Lynds decided th at Muncie,
Indian a , was more representativ e th an, or at lea st as .representa tive as,
other midsize d cities in Ame ric a , thus qua lifying as a typic al case.
This is a n inductive appro a ch to case se lection. Note that typicality
m ay be understood according to the mean, median, o r mode on a par­
ticul ar dimension; there m ay be n1ultiple dimensions ( as in the foregoing
example); and e ach m ay be differently weighted (some dim e nsions may he
more irnporta nt than others). Where the selection criter ia a re multidimen­
siona l and a la rge sample of potenti al cases is in play, some for111 of factor
a n a lysis m a y be use ful in identifying the most-typic al case(s). Although
the Lynds did not employ a statistical 1nodel to evalua te potentia l cases,
it is e asy to see how they might ha ve done so, at least along the first five
criteria . (The fina l criteria would be difficult to ope ra tiona lize in a l arge
sample, since it involves "peculi arities" of any sort.)
Howev er, the more comn1on employ111ent of the typica l-c ase 1ne thod
involves a causal n1ode l of son1e phenomenon of theo retical interest. Here,
the researche r has identifie d a particular outcome (Y), and per haps a
specific X 1 /Y hypothesis, which she wishes to investiga te. In order to do
so, she looks for a typic a l exa n1ple of th a t ca usal re lationship. Intuitively,
one ima gines tha t a case se lected according to the mean v alues of all
pa rameters must be a typica l case relative to so1ne causal relationship.
However, this is bv n o 1nea ns assured.
Suppose that th-e Lynds v.rere primarily intereste d in explaining feelings
of tr ust/distrust an1ong members of different social cla sses (one of the
implicit research go a ls of the Middletown study). This outcome is like ly
to be a ffected by many factors, only some of which a re included in tbeir
six se lection critcria . So choosing cases with r espe ct to a ca usal hypothesis
' I.ynd and 1.ynd ( 1929/1956), quoted in Yin (2004: 29-.1 0 ).
Techniques for Choosing Cases 93
involves, first of a ll, ide ntifying the relevant variables. It involves, secondly,
the selection of a case tha t has "typic al " v a lues re lative to the overa ll
ca usa l model; it is we ll e xplained.
Note tha t cases with atypic al scores on a particular dimension (e .g.,
very high or very low) may still he typical exa mples of a causal relation­
ship. Indeed, they m ay be more typical than cases whose values lie close
to the mean.
Note also that because the typica l case embodies a typica l v alue on
some set of variables, the va r i ance of interest to the researcher must lie
within that case. Specifica lly, the typical case of some phenome non 1nay
be he lpful in exploring causa l mechanis1ns and in solving ide ntification
problems (e.g., endogeneity between X 1 and Y, an omitted v aria ble that
n1 ay a ccount for X1 and Y, or som e othe r spurious ca usa l association).
D epending upon the results of the case study, the author ma y confirn1
an existing hypothesis, disconfirm that hypothesis, or refra me it in a wa y
tha t is consistent with the findings of the case study.
Cross-Case Technique
How might one identify a typic a l cas� from a large popula tion of poten­
ti al cases? If the ca usa l re la tionship involves only a single independe nt
variable and if the relationship is quite strong, it m ay be possible to
identify typical cases simply by eyeballing the e vidence . A strong posi­
tive associ ation he tween X 1 and Y m ea ns th at a case with similar (high,
low, or middling) v a lues on X 1 and Y is probably a typic al case. How­
ever, there are fe,v biva ria te causal relationships in socia l science . Usually,
more tha n one causa l factor must be evalua ted, even if the additional vari­
a bles se rve only a s cont rols. Moreover, ,vithout som e over all a sse ssment
of the cross-case evide nce it may be difficult to say ,vhether the general
relationship is positive or negative , strong or we ak. Thus, in any la rge-N
sample (i.e., \iV henev er the nun1ber of pote ntial cases is great) it is advis­
able to perform a form a l cross-case analysis in order to identify "typical"
cases.
Suppose th at a n a rbitra ry case in the popula tion, denoted as case i,
has a known score on e ach of se ver al relevant v ariables. Fo r the sa ke
of economy of language, let the v ariables involved in the re lationship be
labe l e d Y1 and X 1 ,i, . . . XK,i , where Yi is the score of case i on one va ria ble and
L\H:h of the XK,/s is the score of ca se i on one of the K other variables under
L·onsidcration. 1'hus, the re L-itionship involves a tota l of K + 1 va riables.
k l·:111 hl' ;1ny intl'gl'r grT:ltl'r th:111 or equal to 1 .
94 II. Doing Case Studies Techniques for Choosing Cases 95

With these symbols, the established relationships among the variables
can be expressed mathematically. The idea is to find a function, fl), such
that the average score of y for cases with some specific set of scores on
x 1 . . . XK is equal to f(x1, . . . XK ). Thus, the function f() should be chosen
to capture the key ideas ahout the relationship of interest. A familiar
example 1nay 1nake this discussion clearer.
Often, researchers choose an additive (linear) function to play the role
of f(). Using traditional statistical notation, in which the average score of
Yi across infinite repetitions of case i is denoted hy its expectation, E(yi ),
a linear function represents a relationship in which:
E(y; ) = f3o + f3i x 1., + . . . + f3KXK.i
Each of the f3 K 's in this equation represents an unknown constant. Regres­
sion analysis allows researchers to use known information about the y and
x 1 • . . XK variables for a set of cases to estimate these unknown constants.
Estimates of /3 K will be denoted here as bK .
Using this terminology, we can now develop a formula for the degree to
which a particular case is typical in light of a given relationship. A case is
"typical" in the tenns of small-N methodology to the extent that its score
on the y variable is close to the average score on that variable for a case
with the san1e scores on the x1 . • . XK variables, as given by equation 5. 1 .
That is,
Typicality(i) = -abs[ y; - E(y; lx 1.;, . . . XK; I]
= -abs[y; - bo + b 1 x 1,, + · · · + bKxK,,]
According to this discussion, the typicality of a case with respect to a
particular relationship is simply -1 times the absolute value of that case's
error term (its residu al) in regression analysis. This measure of typicality
ranges, in theory, fro1n negative infinity to zero. When a case falls close to
the regression line, its typicality will be just below zero. When a case falls
far from the regression line, its typicality will be far below zero. Typical
cases have s1nall residuals.
In a large-N sample, there will often be many cases with high (i.e., near­
zero) typicality scores. In such situations, researchers may elect not to
focus on the cases with the highest estimated typicality, for such estimates
may not be accurate enough to distinguish among several almost-identical
cases. Instead, researchers may choose to randomly select from the set
of cases witb very high typicality, or to choose from among these cases
;1ccording to additional criteria, such as those to he discus;o.;('tl hl ' n', or hy
reason of practicality (cost, convenience, etc.). However, scholars should
try to avoid selecting from among the set of typical cases in a way that
is correlated with relevant omitted variables; such selection procedures
con1plicate the task of causal inference.
Consider the (presurnably causal) relationship hetween economic
developn1ent and level of democracy.4 Democracy is understood here as
a continuous concept along a twenty-one-point scale, from - 1 0 (most
autocratic) to + 10 (most den1ocratic).5 Economic development is mea­
sured in standard fashion by per capita GDP." Figure 5.3 displays this
relationship in the form of a bivariate scatterplot. The classical result is
strikingly illustrated: \Vealthy countries arc almost exclusively democratic.
(5.l)
(For heuristic purposes, certain simplifying assumptions are adopted. I
shall assunie, for exa1nple, that this 1neasure of democracy is continuous
and unbounded. 7 I shall assume, more importantly, that the true relation­
ship between economic development and democracy is log-linear, positive,
and causally asynnnetric, with economic development treated as exoge­
nous and democracy as endogenous.8 )
Given this general relationship, how n1ight a set of " typical" cases he
selected? Recall that the Y variable is simply the democracy score, and
there is only one independent variable: logged per capita GDP. Hence,
the simplest relevant model is:
E(Polity, ) = f3o + /3i GDP, (5.3)
(5.2)
For our purposes, the most important feature of this model is the resid­
uals for each case. Figure 5.4 shows a histogram of these residuals. Obvi­
ously, a fairly large nun1ber of cases have quite low residuals an<l therefore
might be considered typical. A higher proportion of cases fall far below
the regression line than far above it, suggesting that the model may be
·I I .ipso.:t ( 1 959). Whether economic dcvclop1nent has only the effect of n1aintaining demo­
cr;Hic regimes (Prze\vorski et al. 2000) or also of causing regime transitions (Boix and
S1okes 200.1) is not relevant to the present discussion, \\'here J assume a simple linear
1\·l.1l'i()11�hip betvYecn \Vealth and dcn1ocracy.
Thi� sniring dcrivl'-� frrnn the Polity2 variable in the Polity lV dataset (11arshall and
J:1gJ',l·r� 2005).
I l.11.1 .11·1· dr:1wn 11·"111 1 11< ' l',·1111 \\/odd 'J:1hll·s d;1tasct (Su1111ncrs :ind Heston J 991).
l'o tlt •,,.., l'r,·1cr .111,I l.1, l- 111.111 ( 1 1 )( 1 \ )
Ii l',111 •.,·,· ( . , · 1 1 11q•, , I , 1] I 1 i l f l ', I 1 1 1, I 1'1 n 1\ , 1 1 •J 1 1· 1 . 1 [ . (.11)()0).
96 II. Doing Case Studies Techniques for Choosing Cases 97

- --- . .. .- ·- - ·
--- -- - - - - --- -- --
"'0
� ���:
I · .. .. . w
00

.£
�
I.O
0
TI 0

"'
N
0 u

•
c

•
:0
0

:,;
E o
0
z
�-

•
.. •
0

- -- -�--
? 8 9 10
Logged 1995 per capita GDP
I'IGURE 5.3. The presumed relationship bet\veen economic development and
detnocracy. A scatterplot showing level of den1ocracy (on tbc vertical axis)
and level of wealth (on the horizontal axis) of all available countries in 1 995.
N = 131.
incomplete. Hopefully, within-case analysis will be able t o shed light on
the reasons for the asy1nn1etry. 9
Because of the large nun1bcr of cases \vith quite sn1all residuals, the
researcher will have a range of options for selecting typical cases. Indeed,
in this cxa1nple, twenty-six cases have a typicality score bct\veen O and - 1 .
Any or all of these might reasonably be selected a s typical cases with
respect to the n1odel described in equation 5.3.
Conclusion
Typicality responds to the first desideratum of case selection, that the
chosen case be representative of a population of cases (as defined by
the primary inference). Even so, it is in1portant to remind ourselves that
a single-minded pursuit of representatlveuess does not ensure that this
desideratum \vill be acbieved. Indeed, the issue of case representativeness
is not an issue that can ever be definitively settled in a case study for­
mat. When one refers to a " typical case" one is saying, in effect, thar the
probability of a case's representativeness is high, relative to other cases.
Note that the measure of typicality introduced here, the size of a case's
residual, can he misleading ii the statistical model is misspecified. And
0
-20 -1 5 -10 -5 0 5 10
Residual from robust regression
FIGURE 5.4. Potential typical cases. A histogram of the residuals from a rohust
regression of logged per capita GDP on level of dcn1ocracy.
it provides little insurance against errors that are purely stochastic. A
case n1dy lie directly on the regression line but still be, in so1ne in1por­
tant respect, atypical. For example, it rnight have an odd co1nbination of
values; the interaction of variables might be different fron1 that in other
cases; or unusual causal mechanisms 1night be at vvork. Most i1nportant,
an analysis of residuals does not address problems of sample bias. If the
large-N samplL is not representative of the relevent population then any
analysis based on the former is apt to be flawed. Typicality does not ensure
representativeness. For these reasons, it is important to supple111ent a sta­
tistical analysis of cases \Vith evidence drav.rn fron1 the case in question
(the case study itself) and with our general knowledge of the world. One
should never judge a case solely by its residual. Yet, all other things being
equal, a case with a lo\v residual is less likely to be unusual than a case with
;1 high residual, and to this extent the method of case selection outlined
here may be a helpful guide to case study researchers faced with a large
11un1ber of potential cases.
I >ivcrsc c:asc

'! ln this cx:1n1p le, the :1sy111111etry is p robabl y due to the failure of i-hc 111odd Ill 1:1kc inro i\ '>t"l' o 1 1 d 1." , l '><" •,t·lc1 I l ! l l l ·,! 1 , l l <" J ',Y l1:1s :1s its prin1ary objective the acbieve­
;\l"l '•H11tt d1t · r('stril·1nl Lill) ',\' of tlw dqwndl'lll V;triahk, as di�l"ll��1·d r.ll'li,·1. ttH'lll n! 1 1 1 . 1 , 1 1 1 1 1 1 1 1 1 1 . 1 1 1 . 1 1 1( 1· .ilo11)', rt'lt"v:1 1 1 1 d i 1 1 1t·11-;io11s. I refer to this ;1s a
98 Il. Doing Case Studies Techniques for Choosing Cases 99

diverse-case n1ethod. For obvious reasons, this method requires the selec­
tion of a set of cases - at n1inimu1n, two - that are intended to represent
the full range of values characterizing X 1 , Y, or son1e particular X 1 /Y
relationship. tu
Where the individual variable of interest is categorical (on/off, red/
black/blue, Jewish/Protestarrt/Catholic), the identification of diversity is
readily apparent. The investigator simply chooses one case from each
. specified, however, because researchers are usually interested only in the types that
category. For a continuous variable, the choices are not so obvious. H�w­
ever the researcher is well advised to choose both extreme values (high
and 'iow), and perhaps che mean or median as well. One may also look
for break-points in the distribution that seem to correspond to categorical
differences an1ong cases. Or one may follow a theoretical hunch about
which threshold values count - that is, which ones are likely to produce
differerrt values on Y.
1\nother sort of diverse case takes account of the values of multiple
variables (i.e., a vector) rather than a single variable. If these variables
arc categorical, the identification of causal types rests upon the inte.r­
section of each category. Two dichoto1nous variables produce a matrix
with four ccllsi three dichotomous variables produce a n1atrix of eig�t
cells, and so forth. If all variables arc deemed relevant to the analysis,
the selection of diverse cases mandates the selection of one case drawn
from within each cell. Let us say that an outcome is thought to be
affected by sex, race (black/white), and n1arital status. Here, a divcrse­
case strategy of case selectiou would identify one case within each of these
intersecting cells - a total of eight cases. Again, things become n1ore co1n­
plicated when one or more of the factors is continuous, rather . than cat­
egorical. Here, the diversity of case values do not fall nea�ly into �ells.
Rather' these cells must be created by fiat - for example, high, medium,
low.
It will be seen that where multiple variables are under consideration,
the logic of diverse-case analysis rests upon the logic of typological the­
orizing - ,,._.,here different combinations of variables are assumed to have
effects ou an outcome that vary across types. George and Bennett define
a typological theory as
a theory that specifies independent variables, delineates then1 into nominal, ordi­
nal, or interval categories, and provides not only hypotheses on how these vari­
ables operate singly, but contingent generalizations on how and under what con­
ditions they behave in specified conjunctions or configurations to produce effects
on specified dependent variables. We call specified conjunctions or configurations
of the vJriables "types." A fully specified typological theory provides hypotheses
on all of the mathe,natically possible types relating to a pheno111enon, or on the
full 'property space,' to use Lazarsfeld's term. Typological theories are rarely fully
arc relatively common or that have the greatest implications for theory-building
or policy-making. 1 1
George and Smoke, for example, wish t o explore different types o f deter­
rence failure - by "fait accompli," by "limited probe," and by "controlled
pressure." Consequently, they wish to find cases that exemplify each type
of causal mcchanism. 12
Diversity may thus refer to a range of variation on X 1 or Y, or to a
particular combination of causal factors (with or without a consideration
of the outco1ne). In each instance, the goal of case selection is to capture
the full range of variation along the dimension(s) of interest.
Cross-Case Technique
Since diversiry can mean mauy things, its en1ploy1nent in a large-N setting
is necessarily dependeut upon how it is understood. If it is understood to
pertain only to a single variable (X, or Y), then the task is fairly simple,
as we have discussed. Univariate traits are usually easy to discover in a
large-N setting through descriptive statistics or through visual inspection
of the data.
Where diversity refers to particular combinations of variables, the rele­
vant cross-case technique is so1ne version of stratified random sa1npling (in
a probabilistic setting) JJ or Qualitative Comparative Analysis (in a deter-
1ninistic setting). 14 If the researcher suspects that a causal relationship is
affected not only by con1binations of factors but also by their sequencing,
1 1 ( ;eorge and Bennett (2005: 235). See also Elman (2005) and Laz:usfcld and Barton
I II <JS II ).
ivlorc precisely, (�eorge and Smoke (1 974: S.14, 522-3fi, Chapter 1 8; see also discussion
i 1 1 C :(lllicr ;111d M,1honcy 1 996: 78) set out to investigate causal pathways and discovered,
1 .•

in 1lw ()r dwir invl'st·igation of n1any cases, these three causal types. But for our
p 1 1 r11u�t·�
10 This method has not heen given n1uch attention by qualitative methodologists; hence, the l'()llf'�c

absence of a generally recognized name. It bears son1e resem�la1: ce to J. S. MiU's Joint wh:11 i\ i 1 u pu1·1.u11 i� tli;1t rhl' fin<1[ s,unple include at least Ont' representative of
l\1cthod of Agrceinent and Difference (Mill l 843/1872), \vhich is to <;11y, 11 n1t:xture of " l \'11(', ..

1110-;t ·�iinilar ,ind 1 no-;r-(li fferl'n1 analy:;,is, :1s discu ssl'd Liter. P,1 !'1(111 (_IIHl.1 : .' 1·1) ('111 p l<1 y �
Ill I
1·.11 l l

I I l\. 1.1 ·, 1 1 1 I . ' 1 11 ll 1 1

I I \ .
,·1· ( IH ,\II ( 'I ·1

d l l · 1 1 >1H t'j'I 1>f " n, : 1 x i 11 1 1 1 1 11 v.1ri;1lion (\H'll'l'l >)',('llt'ily) �,1 t1 1 pli1 q•, . "
100 If. Doing Case Studies
then the technique of analysis n1ust incorporate temporal elements. 1 5
Thus, the method of identifying causal types rests upon whatever method
of identifying causal relationships is presumed to exist.
Note that the identification of distinct case types is intended to identify
groups of cases that are internally homogeneous (in all respects that n1ight
affect the causal relationship of interest). Thus, the choice of cases within
each group should not be problematic, and may be accomplished through
random sampling. However, if there is suspected diversity within each
category, then measures should be taken to assure that the chosen cases
are typical of each category. A case study should not focus on an atypical
member of a subgroup.
Indeed, considerations of diversity and typicality often go together.
Thus, in a study of globalization and social welfare systems, Duane Swank
first identifies three distinctive groups of welfare states: "universalistic"
(social democratic), "corporatist conservative," an<l "liberal." Next, he
looks within each group to find the most-typical cases. He decides that
the Nordic countries are more typical of the universalistic model than
the Netherlands, since the latter has "some characteristics of the occu­
pationally based program structure and a political context of Christian
Democratic-led governments typical of the corporatist conservative
nations." 16 'fhus, the Nordic countries are chosen as representative cases
within the universalistic case type, and are acco1npanied in the case-study
portion of his analysis by other cases chosen to represent the other welfare
state types (corporatist conservative and liberal) .
Conclusion
Encon1passing a full range of variation is likely to enhance the representa­
tiveness of the sample of cases chosen by the researcher. This is a distinct
advantage. Of course, the inclusion of a full range of variation may dis­
tort the actual distribution of cases across this spectrum. If there are more
"high" cases than "low" cases in a population and the researcher chooses
only one high case and one low case, the resulting sample of two is not per­
fectly representative. Even so, the diverse-case 1nethod often has stronger
claims to representativeness than any other small-N sample (including the
typical case). The selection of diverse cases has the additional advantage
of introducing vari ation on the key variahles of interest. A set of diverse
1 ' Ahhon (200 I ); Ahhoir :ind hu-rc.\t ( l lJ8(1); 1\hh<)lt ;111d I\.\\' ( .1 1 1111 1 1
II, \w:1 1 1 k ( ,1.{lll.!: 1 1 ) . s('(' ,ll\u l·\pi1w. A11dt·l'�('II ( ] '!')0).
Techniques for Choosing Cases 101
cases is, by definition, a set of cases that encompasses a range of high and
low values on relevant dimensions.
There is, therefore, n1uch to reconunend this method of case selection. I
suspect that these advantages are con1n1only understood and are applied
on an intuitive level by case study researchers. However, the lack of a
recognizable name ;:-, and an explicit methodological defense - has made it
difficult for case study researchers to identify this method of case selection,
and to explain its logic to readers.
Extreme Case
The extreme-case n1ethod selects a case because of its extreme v�lue on an
independent or dependent variable of interest.'1 7 Thus, studies of domestic
violence may choose to focus on extreme instances of abuse. 18 Studies of
altruis1n may focus on those rare individuals who risk their lives to belp
others (e.g., Holocaust rcsisters).1 9 Studies of ethnic politics may focus
on the most heterogeneous societies (e.g., Papua New Guinea) in order to
better understand the role of ethnicity in a de1nocratic setting. 20 Studies
of industrial policy often focus on the most successful countries (e.g., the
NICs), 21 and so forth. 22
Often an extreme case corresponds to a case that is considered to
he prototypical or paradigmatic of some phenon1ena of interest. This is
hccause concepts are often defined by their extre1nes, that is, their ideal
types. Ger1nan fascism defines the concept of fascism in part because
ii- offers the n1ost extreme example of that phenomenon. Ho,vever, the
111ethodological value of this case, and others like it, derives from its
('Xtren1ity (along some dimension of interest), not fro1n its theoretical
-;t·atus or its status in the literature on a subject.
The notion of "extreme" may now be defined more precisely. An
v x trcme value is an observation that lies far away fro1n the 1nean of a given
1 · !1 does not 1nake sense to apply the extre1ne-case method in a confinnatory/
di�nrnfirn1atory analysis. If a particular causal relationship is at issue, then both X 1
.ind Y n1ust be taken into account when choosing cases, as described in the various sce-
11ario� l'h:lt Follo,v. At present, therefore, ,ve shall assume that the researcher has a general
q1lt's!ion in 1ni11d, but not a specific hypothesis.
Ill J\ i"o)\V I I\' { 1 ')87 ).
I '' I\ 1 1 >lll'! l\' ( I ')')() ),
'II J{ ( ·ill l' ( .1 0IH)/.1.(I() I ) .
' I [ ) 1 · 1 1 1 ( I '!.',:' / ) ,
1 1 ,1 l 1 1 1 1 lw1 <'\,111q ,I, .. . ., , l , >11 1 , 1 ,, 11 < l l\ \.il1<111 ! ')' ( l '}'J(, }; ( ;l'dtlcs ( 1 990); ;ind Tendler
I l 'l'l ')
102 II. Doing Case Studies Techniques for Choosing Cases 103
(.

distribution. For a continuous variable, the distance from the 111ean may
be in either direction (positive or negative). For a dichotomous variable
(present/absent), I understand extreme to mean unusual. If most cases
are positive along a given dimension, then a negative case constitutes an
extreme case. If most cases are negative, then a positive case constitutes
an extreme case. All things being equal, one is concerned not only with w
cases where something "happened," but also with cases where son1ething

"'
did not. It is the rareness of the value that makes a case valuable, in
this context, not its positive or negative value.23 Thus, if one is studying �
0

state capacity, a case of state failure is probably 1nore inforn1ative than

a case of state endurance simply because the former is 1nore unusual.
Similarly, if one is interested in incest taboos, a culture where the incest

•-
taboo is absent or weak is probably more useful than a culture where it is
0

present. Fascism is more important than nonfascisn1; and so forth. There

is a good reason, therefore, why case studies of revolution tend to focus
on "revolutionary" cases. Theda Skocpol had much more to learn from
France than fro1n Austro-Hungary, since France was more unusual than r-·.·· - --- - --�-
0

Austro-Hungary within the population of nation-states that Skocpol
was concerned to explain. 24 The reason is quite simple: there are fewer
revolutionary cases than nonrevolutionary cases; thus, the variation
that one wishes to explore as a clue to causal relationships is encap­
sulated in these cases, viewed against a backdrop of nonrevolutionary
cases.
Cross-Case Technique
As stated, extreme cases lie far fron1 the mean of a variable. Extremity
(E), for the ith case, can be defined in terms of the sample mean (X) and
the standard deviation (s) for that variable:
0.0 0.5 1.0 1.5
Extremeness
FIGURE 5.5. Potential extreme cases. A histogram of the "cxtren1eness" of all coun­
tries on the dimension of dcn1ocracy, as measured by standard deviations fron1
the 1nean (absolute value).
to set an arbitrary threshold. Under assumptions of no nnality, cases with
an extremeness score s1naller than two would generally not be considered
extreme. If the researcher wishes to be more conservative in classifying
cases as extreme, a higher threshold may be employed. In general, the
d10ice of threshold is left to the researcher, to be ma'de in a way that is
appropriate to the research problem at hand.

I X; - X I
Ei = --- (5.4) The mean of our democracy variable is 2.76, suggesting that the coun­
t ries in the 1995 dataset tend to be somewhat 1nore den1ocratic than
:1uthoritarian (zero is defined as the break-point between de1nocracy and
This definition of extremity is the absolute value of the standardized
:111tocracy). The standard deviation is 6.92, implying that there is a fair
("Z") score for the ith case. Cases with a large Ei qualify as extreme.
ii 111011nt of scatter around the n1ean.
Sometimes, the only criterion is a relative one. The researcher wishes to
Figure 5 .5 shows a histogram of the extremeness scores for all countries
find the 111ost exrreme case(s) available. At other times, it may be helpful
0 1 1 ll'vl'l of den1ocracy. As can easily be seen, no cases have extremeness

',cor('s g1T;1n__T than t\VO. The t\VO countries with the highest scores are
Traditionally, 1nethodologists have conceptualized cases as having "po.\itivc" or "nq..;:1
l !,1 1 .1r ;111d S;111di i\r;1hi;1. ' l'hcsc countries, which both have a de1nocracy
23

tive" v.;ilucs (e.g., E1nigh 1 997; tvlahoney and Goertz 2004; R;1gi11 )JlllO: f,0; !C1gin 2!HH:
1 26 ) . ',1 or(· (ll I O i'nr I qq ) , .11T 1,roh;1 hly I hl· t·wo hcst candidates for extreme-
21
· Sk(K'])C)I ( l 'J/ lJ). ' , I ',t' , l 1 1 . i i \'', i •,.
104 II. Doing Case Studies Techniques for Choosing Cases 105

re gre ssion line. However, this sort of analysis m eans that the researcher
Conclusion
The extreme -case n1ethnd appears to violate the social scienc e folk wisdorn is no longer pursuing an extren1e-case method. The extreme-case n1ethod
variab l . )' Se lecting case s is purely exploratory - a way of probing possible causes of Y, or possible
warning us not to "select on the depe nde nt
25
e

of what additional factors might affect th e outcome of interest , or of

cases are effects of X 1 , in an open-ended fashion. If th e researcher has son1e notion
on the dependent variable is indeed problen1atic if a numbe of r
are all
chosen all of which lie on one end of a variable's spectrum (they
sample to what relationship the causal factor of interest has to Y, then she ought to
positiv� or negative), and if the rese archer then subj ects this 2 6
v of a popula tion. Results pursue one of the other methods explored else where in this chapter. This
cross-case analysis as if it were repre sentati e
. Moreo v r, there also implies that an extreme-case n1ethod may transform into a different
for this sort of analysis would a lmost assuredly b e biased
e

exp licitly kind of approach as a study evolve s, that is, as a rnore specific hypothesis
will be littl e variation to expl ain, since the values of ea ch case arc
constrained. comes to light. Useful " extre111 e " case s at the outset of a study 1nay prove
e-case less useful at a later stage of analysis.
Ho\vever, this is not the proper employment of the extrem
metho d.)
me thod. (It is more appropriately labe led an extrem e-samp le
s that lie
The extreme-case method refers back to a l arger sample of cas
e
Deviant Case
of variation
in the background of the analysis and provide a full range
as well as a more repre sentative picture of the popula tion. It is a self­ The deviant-case n1ethod selects the case(s) that, by reference to son1e gen­
din1 nsion of int e rest, not eral understanding of a topic (e ither a specific theory or con1111on sense),
conscious atte mpt to maximize variance on the e
throug h the
to minimize it. If this population of cases is well understood - demonstrates a surprising value. Barbara Geddes note s the importance of
, or through deviant cases in medical science, where researchers are habitually focused
author's own cross-case analysis, through the work of others
a single
common sense - then a researcher 1nay justify the selection of on that which is pathological (according to standard theory and practice) .
-case analys is. lf no t, the Th e New England Journal of Medicine, one of the premier journals o f the
case exen1p l ifying an cxtre1ne value for within
me hod (see the field, carries a regular fe ature e ntitled "Case Records of rhe Massachuse tts
researcher n1ay be well advised to fol l o-....v a diverse-case t

earlier discussion). General Hospital." The se articles be ar titles like the following: "An 80-
eness.
By way of conclusion, let us return to the problcn1 of representariv Year-Old Woman with Sudden Unilateral Blindness" or "A 76-Ye ar-Old
entativ eness refers to a case that Man with Fever, Dyspnea, Pulmonary Infiltrates, Pleural Effusions, and
In the context of causal analysis, repres
exemplifies values on X 1 and Y that conforn1 to a genera l patter n. In a Confusion." 27 Similarly, medical researchers are keen to inv e stigate thOse
is gauged
cross-case rnodel, the representativeness of an individual case rare individuals who hav e not succtunbed, despite repeated exposure, to
by the size of its residual. '[he represe ntative case is therefo re a typical the AIDS virus. 28 Why are they resistant? What is different about these
case (as will be discuss ed). It people? What can we learn about AIDS in other patients by observing
case (as already discussed), not a deviant
devian t. Ther e is people \vho have built-in resistance to this disease?
will be seen that an extreme case 111ay be typical or
a causal
simply no way to te ll, be cause the re se archer has not yet specified Case studies in psychology and sociology are often comprised of devi­
e d, we 1nay ant (in the social sense) persons or groups. In economics, case studies
proposition. Once such a causal proposition has been specifi
on is sin1ilar to some popula tion of 111ay consist of countries or businesses that overperform (e.g., Botswana,
then ask whethe r the case in questi
relatio nship of interes t). Microsoft) or underperform (e .g., Britain through most of tbe twe nti eth
case s (in all respects that might affect the X1/Y
contex t of a
It is at this point that it becoines possible to say, within th e cl'ntury; Sears in recent decades) relative to soine set of expectations. In
from, the
cross-case statistical 1nodel, whether a case lies near to, or far
( ;l'ddes (2003: l .1 1 ). For other examples of case \York fron1 the annals of 1nedicine, see
" ( :linictl Rrpor1," i11 ·1·1;,. I ,111cct; "Case Studies" in The Canadian A1cdical Association
.
25 Geddes (1990); King, Keohane, and Verba ( 1994). See also discussions in Brady :1nd /"11r11,1/; .11l<I 1".11 i. ,w, 1·.<.11,·" 1 ii tlH' J, J11n1irl of ()hstetrics and Gynecology, often devoted
Collier (2004); Collier and Mnhoney (1 996); and Rogowski ( 1 995). 1 1 ) ,·li11i,·:1l , .1·.,". 1 d 1 ·., 11·,.,·il 1 1 1 ln1i,·!'k 2.00 1 : 7). For ex;1111ples frorn the subfield of com-
-1 1' The exception wollld [w ,1 cirn1 111st:111ce i11 which the rl'\t\H,-l11·r 111t(·t11h in di ...pr()Vl' 1 ' · 1 1 . 1 1 11·,· 1" ' 1 1 1 11 ', ·. , . I·
I( 111<
l)',il ( I '/OJ , \ ) ,
d\"ll'r1nini\li l' ,H)', t1nw111 (l )i1l11 l 1J'Jl·{' ).
'!i
1111 < I il,111e 11'1 .111, I \ ' , 1 1 '"1· I,. d I i I ' 1 ' 1., I . I l.11 11< ',, l'.1111.1 In,, .111d l·,111( i ( I !)< Jr,) .
1 06 II. Doing Case Studies Techniques for Choosing Cases 107

political science, case studies 1nay focus on countries where the welfare of a formal, 1nathematical representation of the hypothesis at hand, a
state is more developed (e.g., Sweden) or less developed (e.g., the United deviant case lies as far as possible from that prediction. Referring back to
States) than one would expect, given a set of general expectations about the model developed in equation 5 .3, we can define the extent to which a
welfare state develop111ent. case deviates from the predicted relationship as follows:
In all fields, the deviant case is closely linked to the investigation of the­
Deviance(i) = abs[y, - E(y, lxu, . . . x1u)l
= abs[y, - b o + b, x u + · · · + bKXK,f ]
oretical anomalies. Indeed, to say " deviant" is to imply "anon1alous. " 29 (5.5)
Note that while extre1ne cases are judged relative to the mean of a single
distribution (the distribution of values along a single dimension), deviant
cases are judged relative to so1ne general model of causal relations. The Deviance ranges from 0 , for cases exactly on the regression line, to
deviant-case method selects cases that, by reference to some general cross­ a theoretical lin1it of infinity. Researchers will usually be interested 111
case relationship, demonstrate a surprising value. They are "deviant" in selecting fron1 the cases with the highest overall estimated deviance.
that they are poorly explained by the multivariate model. The impor­ In our running example, a two-variable model with economic devel­
tant point is that deviantness can only be assessed relative to the general opment (Xi ) and democracy (Y), the most deviant cases fall below the
(quantitative or qualitative) model employed. regression line. This can be seen in Figure 5 .4. In fact, all eight cases with
This means that the relative deviantness of a case is likely to change a deviance score of more than ten have negative residuals; their scores on
whenever the general model is altered. For example, the United States is t·hc outcon1e are lower than they "should" be, given their level of deve\­
a deviant welfare state when this outcome is gauged relative to societal opn1ent. These eight cases are Croatia, Cuba, Indonesia, Iran, Morocco,
wealth. But it is less deviant - and perhaps not deviant at all - when Singapore, Syria, and Uzbekistan. Our general mo<lel of den1ocracy does
certain additional (political and societal) factors are included in the model, not explain these cases very well. Quite possibly, we could develop a bet­
as discussed in the epilogue. Deviance is model-dependent. Thus, when ter model if we understood what - aside from GDP per capita - might
discussing the concept of the deviant case, it is helpful to ask the following he driving the choice of regime type in these polities. This is the usual
question: relative to what general model (or set of background factors) is purpose for which deviant-case analysis is employed.
Case A deviant?
The purpose of a deviant-case analysis is usually to probe for new- but < :onclusion
as yet unspecified - explanations. (If the purpose is to disprove an extant As I have noted, rhe deviant-case method is an exploratory form of anal­
theory, I shall refer to the study as a crucial case, as will be discussed later.) ysis. As soon as a researcher's exploration of a particular case has iden­
Thus, the deviant-case method is only slightly more determinate than the I ificd,a factor to explain that case, it is no longer (by definition) deviant.
extren1e-case method. It, too, is an exploratory for1n of research. The ( ! "he exception would be a circumstance in which a case's outcome is
researcher hopes that causal processes within the deviant case will illus­ dt'en1ed to be accidental or idiosyncratic, and therefore inexplicable by
trate some causal factor that is applicable to other (deviant) cases. This ;1ny general model.) If the new explanation can be accurately n1easurcd as
means that a deviant-case study usually cuhninates in a general proposi­ ,1 'iingle variable (or set of variables) across a larger sample of cases, then

tion - one that may be applied to other cases in the population. .1 new cross-case 1nodel is in order. In this fashion, a case study initially
I r;1111cd as a deviant case is l ikely to be transformed into some other sort of
.111alysis.
Cross-Case Technique
This feature of the deviant-case study also helps to resolve doubts about
ln statistical terms, deviant-case selection is the opposite of typical-case
11-; rvprl'sl'nta tiveness. Evidently, the representativeness of a deviant case is
selection. Where a typical case lies as close as possible to the prediction
pro hll'lll:ll ic, si ncv t·hl' c;1sl' in question is, by construction, atypical. How­
c·vcr, 1 liis 1)rohlt·t11 l·. 1 1 1 Ii<' 1 1 1 ! 1 ig;1tcd if the rcscarcl�er generalizes whatever
p 1 c q ()'> i t i ( ) 1 1 i•, 1 1r( 1 \·1 , l,·d l1y Ilic L' :\.',t' study to other cases. In a Jarge-N
29 For a discussion of the important roli.: of :11101nalil"\ i n thl" dl"vt·lnp111l·11l ni' .'>L"i(' 1 1 1 i l i ,
1

1 1 1 c 1di·I , 1 1 1 1', 1· . . H , 0 1 1 q d 1·. l11·,I Ii\ 1 l 1c- lT1·.1ti( )ll of ;1 v:1ri;1hh· to rcprl'Sl'llt the
rheorizing, sec Flm:111 (2003) ;111d ! .:1L11os ( ] 'J 7 8 ) . h)r ("x:111 1pl!"', "I d(·1· 1 . 1 1 1 t , .1,(' 1"(''><".11< I ,

1 1 1 · \\ l n 1 H 1t l H",I'. ! l i . 1 1 t i " , .1 .., ·. 1 1 111\' 1 1 . 1 ·. !!lc- 1 1 1 di(·d . · 1 · 1 11·, 111.1\· 1·1 ·(pi irt' ..,011 1 t ·
d!·'>i l \ll'> 1 1 \ 1\w �1H · i.tl '>l.H"lll l"'>, •,,"!' 1\ 1 1 w 1 1 1 . 1 ( 1 '!'1 1 ) : < ·"l'l'nlr .1 · ( 1 1 1 1 1 · 1 i , I , I . , , 1 , · 1 1 1 1 I 'I ' � ) :
J · 1 1 1 i ) •, l 1 { I 'l<J ' \ ; . 1 1 1 d k ,·11, Ltl I . 1 1 1 , I \\-', , 1 1 I I ''·I 11/ I ' ! 1 ·, J
 II. Doing Case Studies
108
original coding of cases (in addition to the cnsc under intensive study).
However, so long as the underlying infor111Jtion for this coding is avail­
able, it should be possible to test the new hypothesis in a cross-case model.
If the new variable is successful in explaining the studied case, it should no
longer be deviant; or, at the very least, it will be less deviant. In statistical
terms, its residual will have shrunk. lt is nov,,1 typical, or at least more
typical, and this relieves concerns about possible unrepresentativeness.
Influential Case
So1netlmes the choice of a case is motivated solely by the need to verify the
assun1ptions behind a general 1nodel of causal relations. Herc, the analyst
attempts to provide a rationale for disregarding a problematic case, or a
set of problematic cases. That is to say, she attempts to show why apparent
deviations fro1n the norn1 are not really deviant, or do not challenge the
core of the theory, once the circumstances of the special case or cases
are fully understood. A cross-case analysis may, after all, be marred by
several classes of problems, including measurement error, specification
error, errors in establishing proper boundaries for the inference (the scope
of the argun1ent), and stochastic error (fluctuations in the phenomenon
under study that arc treated as rando1n, given available theoretical and
empirical resources). If poorly fitting cases can be explained away by
reference to these kinds of problerns, tben the theory of interest is that
n1uch stronger. This sort of deviant-case analysis answers the question,
"What about Case A (or cases of Type A)? How does that (seemingly
disconfirn1ing) case fit the n1odel?"
Because its underlying purpose, as well as the appropriate techniques
for case identification, is different from that of the deviant-case study,
I offer a new term for this method. The influential case is a case tbat
appears at firsr glance to invalidate a theory, or at leas� to cast doubt
upon a theory. Possibly, upon closer inspection, it does not. Indeed, it
may end up confirming that theory - perhaps in some slightly altered
forn1. In this guise, the influential case is the "case that proves the rule."
A si1nple version of influential-case analysis involves the confirn1ation
of a key case's score on son1e critical d imension. This is essentially :1
question of measurement. Son1etin1es cases are poorly explained sin1p l y
because they are poorly understood. A close exan1ination of a particut1r
context n1ay reveal that an apparently falsifying case has been 111iscodcd.
If so, tbe initial challenge presented by that case to so1\ll' )!;l'Ilcr,11 tlH·orv
has hccn ohvi,11\'tl.
Techniques for Choosing Cases 1 09
However, the more usual employment of the influential-case 1nethod
culminates in a substantive reinterpretation of the case - perhaps even
of the general 1nodel. It is not just a question of measurement. Consider
Thon1as Ertn1an's study of state building in Western Europe. As summa­
rized by Gerardo Munck, this study argues
that the interaction of a) the type of local governn1ent during the first period of
statebuilding, with b) the timing of increases in geopolitical competition, strongly
influences the kind of regime and state that emerge. [Ertman] tests this hypoth­
esis against the historical experience of Europe and finds that 1nost countries fit
his predictions. Denmark, however, is a 1najor exception. In Denmark, sustained
geopolitical competition began relatively late and local govern1nent at the begin­
ning of the state building period was generally participatory, ,vbich should have led
the country to develop 'patrimonial constitutionalism.' But in fact, it developed
'bureaucratic absolutisrn.' Ertrnan carefully explores the process through \vhich
Denmark came to have a bureaucratic absolutist state and finds that Den1nark
had the early tnarks of a patrin1onial constitutionalist state. However, the country
was pushed off this developrnental path by the influence of Gern1an knights, who
entered Denmark and brought with thern Gennan institutions of local govern­
lTtent. Ertn1an then traces the causal process through which these i1nported insti­
tutions pushed Den,nark to develop bureaucratic absolutism, concluding that this
�leveloprnent was caused by a factor well oursidc his explanatory fra1nework.
30
Ertn1an's overall framework is confirmed insofar as he has been able to
show, by an in-depth d iscussion of Denmark, that the causal processes
stipulated by the general theory hold even in this apparently disconfirming
i.: ase. Denmark is still deviant, but it is so because of "contingent historical
l·ircumstances" that are exogenous to the theory.3 1
The reader will have noted that influential-case analysis i s sin1ilar to
deviant-case analysis. Both focus on outliers, unusual cases (relative to the
theory at hand). However, as we shall see, they focus on different kinds
1if unusual cases. Moreover, the animating goals of these t\vo research
designs are quite different. The influential-case analysis begins with the
.1i111 of confirming a general n1odel, while the deviant-case study has the
.1!111 of generating a new hypothesis that modifies an exis'ting general
rnodcl. The confusion between these two case-study types stems from
Ihl' fact that the same case study may fulfill botb objectives - qualifying a
f',1·1H-ral 1nodel and, at the sa1ne tin1e, confirming its core hypothesis.
In their study of Roberto Michels's " iron law of oligarchy," Lipset,
I 'n nv, ,111d ( :olc111:111 choose to focus on an organization- the International
1
" r,, 1 1 1 1 1 , I, ( . ' ( H J . I : 1 I 1i\ , ,, ' .,I,,,, I , 1 1 1 1. 1 1 1 1 1 ')'!7),
I I I 1 1 1 11.1 1 1 i l "' l ' \ l r , I
 II. Doing Case Studies Techniques for Choosing Cases Ii 1
110

Typographical Union - that appears to violate the central presupposi·
tion.n The ITU. as noted bv one of the authors, has "a long-term two­
party systen1 wi,th free elccrions and frequent turnov� r in �ffice': and is
thus anything but oligarchic. 3 3 Thus, it calls into question M,chels s grand
_ _
generalization about organizational b � havior. "f'he autho: s expla\n this
curious result hv rhe extraordinarily high level of educanon among the
me�bers of this.union. Thus, Michcis's law is shown to be valid for most
orgar:izations, but not all. It is vaEd with qualifications. Note that the
respecificatiou of the original model (in effect, Lipset, Trow, and Cole­
of the leverage of each case can be derived from the diagonal of ,he hat
matrix. Specifically, the leverage of case t is given by the number in the
Ui) position in the- hat rnatrix, or H i,i, 37
For any X n1atrix, the diagonal ea tries in the hat n1atrix will automat­
ically adJ up to K + 1 . Hence, interpretations of 1he leverage scores for
different cases will necessarily depend on the overall number of cases.
(]early\ any case with a score near one is. a case \Vith a great deal o.1: lever­
age. In n1ost regressiorl' situations, however, no Cd.Se has a score that high.
l'\ standard rule of thutnb is to pay close attention to cases with a Ievrrao-e b

n:ian introduce a new control variable or boundary cond1t1on) 1nvolves score higher than 2 ( K + 1) iN. Cases with a leverage score above this
the exploration of a new hypothesis. In this respect, the use of an influ­ value are good candidates for influential··case selection.
ential case to confirm :1n existing rheory is quite similar to the use of a A..n interesting feature of the hat 111atrix i s that it does not depend on
deviant case to unearth a new· theory, the values of the dependent variable. Indeed, the Y vector does not appear
in equation 5.6. This means that the measure of leverage derived from the
hat matrix is, in effect, a 111easure of potential influence. It -::elJs us how
Cross-Case Technique
s that, if counterfactuaUy much difference the case \vould make in the final estin1ate if it were to
Influential cases in regression are those case
ble, would most sub_stan­ have an unusual score on the dependent vafJablc, but it docs not tell us
assigned a different value on the dependent varia
titative n1easures ot lllflu­ hovv much difference each case actually n1ade in the final esri1nate,
tial!v change the resulting estimates. Two quan
nostics, 34 The first, often ...;\nalysts involved in selecting influential cases will sornetin1es be inter­
enc� are con1mon ly applied in regression diag
from what is called the hat L'sted in measures of potential influence, because such 1neasures are rele­
referred to as the "leverage" of a case, derives
penJent .variables. for all o : vant in selecting cases when there n1ay be so111e a priori uncertainty about
matrix.35 Suppose that the scores on the inde
tl-:e matrix X, ¥.rh1ch has N scores on the dependent variable, 11uch of the infonnation in such case
the cases in a regression are represented by
K + 1 columns (representing studies conies from a careful, in·-depth rneasi:rement of the dependent
row s (representing each of the N cases) and
a constant). Further, allow variable - which may sometimes be unkno\vn� or only approxitn01tely
the K independent varia bles and allowing for
variable for a!l of the case&, known, before the case study begins. The measure of levera�e derived
y ro represent the scores on the dependent fron1 the hat n1:1trix is appropriate for such situations because lt does not
mn.
Therefore. Y will have N rows and only one colu
ix, H, is as follows: n:quire actual scores for the dependent variable,
Using these symbols, rhc formula for the hat matr A second co1nmonly discussed n1easure of influt:'nce i n statistics is
i5.6l
H = X(XTx:- 1 x'f { 'ook's distance. This statistic is a 111easure of the extent to which the estl-
ix transpose operation, 11,atcs of the fi i parameters would change if a given case were omitted frorr1
ln this equation, the syrr1bol "T" represents a matr
rse operation.36 i\ measure l lie analysis. Because regression analysis typicaHy includes 1nore than one
and the symbol •• --- 1 represents a matrix inve
1

/f i pal'J1neter, a measure of influence requires some method of combining

'

t!H· differences in each parameter to produce an overall n1easure of a

32 Lipse:j Trow, and Colcr:nan {1956 ). r.1st•'s influence. The Cook\; distance statistic resolves this dilen1ma by
:n Lipset (1959: 70).
34 Belsey, Kuh, and Welsch (2004). _. _ . . . ,
::s Tbis somcwhar;;.urious name dcnves fr�>nt
the fact that, 1t the hat n1atr:x 1!:'. mult1phcd ,iy
the vector containing voiues of the dependent val'iable, the resu!t is the vector o_f �tt,:d
values for each c.isc. Typically, the ve(.-tor of fitted values for :he depe:1d_ent var13hk ,, 1' !'lw di,, ll'i·;io11 lwn· i11;•. , I \ ,.., i l k H�L' {if thrt hat 111atnx ::1 lincar regression. Anal ysts rnay
disti ngui:,hed from rhe acrn-al ve.:tor uf value '> on the dependent varia�lc by the uv· o! .d•.,, lw 1:jji-n",'1·,i 1 1 1 ·,Hu.111,,11•. du1 d1) 1111! rt• ..:n11h'.c linear regression pn:,blerns., e.g.,
cvs : 1 w fit,cd v.i!HI'.',, �·.t1 1 b, wl•i"t" !�w ,l1· 1,,·:1t:,· ,; 1,11 ·,1!,k · . ,111 1 ,,11,,1;1,a1� (n-,:.1\q.�nri,.:;,f. Snrncti ines, t'.:ese situati 0 ns
rbe ..�,, or "h.at" symbcL Hence, the D.:.n ni,ull'.rix. ,,.,..-b1ei1 yrndu , 111 iw.1>, \>nn111 ,,L1:, ,! 1, 1 1 l n n , jp I· ,,- , q L 1,! )•,,·11tT.d11rd liiw.1rmt><lt·i�, which includes
sni<l to put �he- !lat on the dcpc:;dl'l11 v;iri:ih ll!i,

I•\\ ', i'.1 111'1 ,I, . 1 1 1 , I ,, I, 11 "' I I \ I :'-, 1, i!ll.11 ·J1 11d l'l, ·!d1'I I 'IX')).
\;, Sec ( ;r,'\''.W l_)_ '.102) '.c1 hril't' rrv), w.
I! • >I 'I! ' ,I I! j ,I
,l

l .
112 11. Doing Case Studies
taking a weighted sum of the squ(1red par;1111eter differences associated
with deleting a specific case. Spccif-ic�1 11y, the for1nula for Cook's distance
lS:
(b_, - b)T XT X(h_, - b)
(K + l)MSE
In this forn1ula, b represents all of the paran1eter estimates from the
regression using the whole set of cases, and b_ i represents the parameter
estimates fro1n the regression that excludes the ith case. X, as above,
represents the matrix of independent variables. K is the total nun1ber of
in the formula by the use of K + 1). Finally, MSE stands for the mean
independent variables (not including the constant, which is allowed for
squared error, ,vhich is a 1neasure of the an1ount of variation in tbe depen­
dent variable not linearly associated with the independent variables. 3 8
This so1newhat intimidating n1athcmatical notation gives precise
expression to the intuitive idea, discussed earlier, of measuring influence
as a weighted su1n of the differences that result in each parameter esti-
1nate when a single case is deleted fron1 the san1ple. One disadvantage of
this forn1ula is that it requires a number of extra regressions to be run in
order to co1npute rneasures of inflnence for each case. The overall regres­
sion must of course be cotnputed, and then an additional regression, with
one case deleted, is required for each case.
Fortunately, n1atrix-algebraic manipulation de1nonstrates that the
expression for Cook's distance given in equation 5.7 is equivalent to the
follo\ving, computationally 111uch easier expression:
( K + 1)(1 - H, , )
In this expression, H;, i refers to the 111easure of leverage for the ith case,
taken fro1n the diagonal of the hat 1natrix, as already discussed. K once
again represents tbe nu1nber of independent variables. Finally, rl is a
special, modified version of the ith case's regression residual, known a.-;
the Studentized residual, which needs ro be separately co1nputed.
The Studentized residual is designed so that the residuals for all cases
v,,,jl] have the same variance. Tf the standard regression residual for case f
3 8 Spccific;i.lly, the MSE is found by srnnming thi' :-.quarl'd residuals fro111 rht· 1'1111 rt'gn·��i(ltl
:1nd then dividing by N - K - I , where N i.\ illl' 11111nlwr ()f C\St'-� ;11HI k i� dw llllllllll'1 ,ii
i11,kpt·11dt·11t v;1riahk�.
Techniques for Choosing Cases 1 13
is denoted by Ei , then the Studentized residual, rl, can be cotnputed as
follows. (All symbols in this expression are as previously defined.)
cf
(5.9)
r s=======
JccMS £ ( 1 - H,=
,=- .,l
(5.7)
As can be seen from an inspection of equations 5.8 and 5.9, Cook's dis­
tance for a case depends primarily on two quantities: the size of the regres­
sion residual for that case and the leverage for tbat case. The most influ­
ential cases are those with substantial leverage that lie significantly off the
regression line.
Cook's distance for a given case provides a su111mary of the overall
difference that the decision to include that case makes for the parameter
esti1nates. Cases witb a large Cook's distance contribute quite a lot to
the inferences drawn from the analysis. In this sense, such cases arc vital
for 1naintaining analytic conclusions. Discovering a significant 1neasure-
1nent error on the dependent variable or an i1nportant 01nitted variable
for such a case may dran1atica lly revise esti1nates of the overall relation­
ships. Hence, it 1nay be reasonable to select influential cases for in-depth
study.
To su1n1narize, three statistical concepts have been introdnced in this
section. The hat matrix provides a measure of leverage, or potential influ­
ence. Based solely on each case's scores on the independent variables, the
hat n1atrix tells us ho\v tnuch a change in (or a n1easurement error on)
the dependent variable for that case would affect the overall regression
line. Cook's distance goes further, considering scores on both the indepen­
dent and the dependent variables in order to tell us how 1nuch the overall
1:egression estimates would be affected if each case were to be dropped
(5.8) lro1n the analysis. This produces a measure of how much actual influ ence
L'ach case has on the overall regression.
Either the hat matrix or Cook's distance n1ay serve as an acceptable
nicasure of influence for selecting case studies, although the differences
just discussed must be kept in mind. In the following examples, Cook's dis­
t:1ncc \Vill be used as the primary measure of influence because our interest
[<; in whether any particular cases 1night be influencing the coefficient esti-
111;1rcs in our democracy-and-development regression. A third concept, the
."it 1idcntized residual, was introduced as a necessary ele1ncnt in con1puting
< :ook's disc111CL'. ('l ' hc h;1r n1atrix is, of course, also a necessary ingredient
I l l ( :(>Ok 's disLllll l'. )
1:iJ', lll "(' ) . (1 c, I H ,\\'', I I H · < < H , k \ ,listat1lT scores for each of the countries
1 1 1 du· I q q � 1 1('1 1 . q 1 1 1 .1 I ,l 1 1 ' . 1 1 H I d1·1 1 1( u,T,ll·�· d:1LJ.'i('l . iVl ost col1 1Hrics h J Vl'
114 II. Doing Case Studies Techniques for Choosing Cases 115

105 cases are likely to be and, hence, the less l ikely a researcher is to use hat
matrix and Cook's distance statistics for purposes of case selection. In
�
ci these instances, it may not matter very much what values individual cases
0 display. (It may of course matter for the purpose of investigating causal

"'0
ci mechanisms. However, for this purpose one would not employ influential
statistics to choose cases.)

t'l
c
© 0
0

Crucial Case
ci
0
�0
0
0 "0 74
,75 Of all the extant methods of case selection, perhaps the most storied - and
0 certainly the 111ost controversial - is the crucial-case method, introduced
to the social science ,vorld several decades ago by Harry Eckstein. In

l1, l 1, 1111 IJl 1 ,I J l,,1, 1 1 1 1 1 ,1 1 \1J1, .1 l 1,l, 1, [ , , 11J1,l1

OJ
0
0 his seminal essay, Eckstein describes the crucial case as one "that must
0
0
J , I ,,
0 20 40 60 80 100
Country code
FIGURE 5.6. Potential influential cases. The Cook's distance scores for an OLS
regression of de111ocracy on logged per capita GDP. The three numbered cases
have high Cook's distance scores.
quite lo\V scores. The three most serious exceptions to this generalization
are the numbered lines in the figure: Jamaica (74), Japan (75), and Nepal
( 1 05 ). Of these three, Nepal is clearly the most influential by a wide
margin. 1-Ience, any study of influential cases \vould want to start with an
in-depth consideration of Nepal.
Conclusion
The use of an influential-case srrategy of case selection is lin1ited to
instances in which a researcher has reason to be concerned that her results
are being driven by one or a few cases. This is most likely to be true i11
small to moderate-sized san1ples. Where N is very large - greater than
1 ,000, let us say - it is extremely unlikely that a small set of cases (much
less an individual case) \Vill play an " influential" role. Of course, then·
may be influential sets of cases - for example, countries \Vi thin a particul:1r
continent or cultural region, or persons of Irish extraction. Sets of i11f11 1
ential observations arc often problematic in a titnc-scrics cross-sccti()11
dataset -....vhere each unit (e.g., country) contains n1ultipll' ohsvrv;1 t ion,,
(through tin1e) and hence n1a y h,1 vc a strong in tl lll'llCl' 011 :1ggrrg;1 t c rv" uIt',.
Stil 1, the gl'JH.'Ll 1 rull' is: the L1 rgcr 1 l1v s;1111pli:, 1 lic I('"" i 11 q1t ,rt :111 t i 1l d i \' 1du .ti
closely fit a theory if one is to have confidence in the theory's validity, or,
conversely, must not fit equally well any rule contrary to that proposed.""
120 A case is "crucial" in a some\vhat \.\'eaker - but much more com1non -
sense when it is most, or least, likely to fulfill a theoretical prediction. A
"tnost-likely" case is one that, on all dimensions except the dimension
of theoretical interest, is predicted to achieve a certain outcome, and yet
does not. It is therefore used to disconfirm a theory. A " least-likely" case is
one that, on all dimensions except the dimension of rheoretical interest, is
predicted not to achieve a certain outcome, and yet does so. It is therefore
used to confirm a theory. In all for1nulations, the crucial case offers a
111ost-difficult test for an argument, and hence provides what is perhaps
the strongest sort of evidence possible in a nonexperimental, single-case
.'-l'tting.
Since the publication of Eckstein's influential essay, the crucial-case
;1pproach has been claimed in a multitude of studies across several social
'-l'icnce disciplines and has come to be recognized as a staple of the
1.'asc study method.40 Yet the idea of any single case playing a crucial
(or "critical") role is not \videly accepted a111ong most methodologists.41
(Fvrn its progenitor seems to have had doubts.)
l !nfortunately, discussion of this method has focused misleadingly on
\vli:1t :1rc presumed to be largely inductive issues. Are there good crucial
I '! l·t I,;: \I ('Ill ( I 'J7 ) : \ ] 8).
1c1 h,r <'-.;:1111pll', oi' IIH· lTt1,·i.tl ,·,1,(· 11wthod, see r..cnnett, Lepgold, and Unger (1994); Desch
l .1 111 l.'.); ( ;ornlt11 .11al \1n11•.111.111 (.' l lllO); kl'mp ( 1 986); and Reilly rincl Phillpot (2003).
h1r r, nu·1.il d1·., 11·.·. 1 1 1 1 1 , ·.,·,· < ,1 ' "1!'.1· .111d lk1111t'll (200)); l .t·vy (2002:1); nnd Stinchcombe
{ I 'li,S: '·I ,', I.
I I 1 , n , , - . , , , ', . -
1. 1 1 0 >11 I 1 1 1 1 1 I )

116 II. Doing Case Studies
case s out there in the empirical wor l d? Have social scientists done a g ood
job in identifying them? Yet the prac ticability of this m ethod re sts on
issues that are large ly de ductiv e in nature, as we shall see.
The Confirmatory (Least-Likely) Crucial Case
L et us begin with the confirmatory (a.lea. least-like ly) crucial case. The
implicit logic of this research design may be sumtnarized as foll o� s. Given
a set of facts w e are aske d to conte mplate the pro bability that a given the ­
ory is true. While the ,facts n1atter, to be sure, the e ffectivene ss of this ort
. �
o f re se arch also rests upo n the formal p rop ertie s of the theo ry 111 question .
Specifically, the degree to which a theory is an1enable to confir1nati on is
contingent upon how many pr edicti ons can be derived fro m the theory
and on how "risky )' each in dividual pr ed ic tion is. In Popper's word s,
Confirmations should count only if they are the result of risky predictions; that
is to say, if, unenlightened by the theory in question, we should have expected
. the d is tinction between a 111us t-.fit case and a l east-likely case - that, in
an event which ,.v as incompatible with the theory - an event which would have
refuted the theory. Every 'good' scientific theory is a prohibi tion; it forbids certain
things to happen. The more a theory forbids, the better it is. 42
A risky pre diction is therefor e o ne that is highly precise and dete rmi­
nate , an d thus unlike ly to be explainable by other causal factors (exte rnal
to the t heo ry of intere st) or thr ough sto chastic processes. A theory pro ­
duces 1nany such pre dictions if it is fully e lab orated, issuing predictions
no t only on the central outcome o f in terest but also on specific causal
1ncchanisms, and if it is broad in purvie,v. (The no tio n o f riskiness n1ay be
conceptualize d wit hin tbe Popperian lexicon as degrees of falsifiability.)
These po ints can also he ar ticulated in Bay esian terms. Co lin H owson
and Pete r Urbach explain: "The degree to which h [a hyp othe sis] is con­
firmed by e [a set o f evidence ] depends . . . on the extent t o which P( elh )
exce eds P(e), that is, o n how 1nuch more pro bable e is relative to the
hypo thesis an d backgroun d assumptions than it is re lative just to back­
.
groun d assumptions." Again, "confirmation � s � orrelated "'." 1t� � ow m� �l�
,
mo re probable the evidence is if the hyp othesis ts true than if tt ts false.
Thus, the stranger the pre dic tio n o ffer ed by a theo ry - re lative to what
we would n or1nally expect - the greater the degre e o f co nfirn1ation tb;1t
will be afforded by the evi dence. As an intuitive example, Hlnrvso n and
Urbach offer the following:
·12 Popper ( l 96J: _)6). Sec ,1lso Popper ( 1 9.14/\ 968 ) .
1 1 I h1w�(111 .1111\ l lrli:1l ·l1 ( 1 '18 <J: 81,).
Techniques for Choosing Cases lJ 7
If a soothsayer predicts that you will rneet a dark stranger somctin1e and you do in
fact, your faith in his powers of precognition would not be 1nuch enhanced: you
would prohably continue to think his predictions were just the result of guesswork.
However, if the prediction also gave the correct number of hairs on the head of
that stranger, your previous scepticis1n would no doubt be s everely shaken,44
Whil e these Poppcrian/Bayesian insights45 are relevant t o all empiri­
cal re search designs, they are es pe cially rel e vant to case study re search
de signs, for in th ese settings a single case ( or, at m ost, a small numhe r
of cases) is require d to bear a he avy burden of pro of. It shoul d be no
surprise, there fore, that Poppe r's i dea of "riskine ss" ,vas appropriated by
case study researchers like Harry Eckstein to validate the enterprise of
single-case analysis. (Although Eckste in does no t cite Popper, the intel­
le ctual lineage is clear. ) Riskiness, here, is anal ogous to what is usually
referred to as a "1nost-difficult" research design, which in a case study
research design woul d be understood as a least-like ly case. Note also that
the e ve nt, actually doe s fit the terms of a theory - is a 1natter of degree.
Cases are 1norc or less crucial for confirming th eori es. The point is that,
in som e circumstances, the riskiness of the th eory may compensate for a
paucity of empirical evi dence.
The cr ucial-case re search de sign is, perforce , a highly deductive
enterprise ; much depen ds on the quality of the th eory under investiga­
tion. It f ollows that the theories 1nost ame nable to crucial-case analysis
are th ose that arc lawlike in their precision, degree of elaboration, con­
siste ncy, and sco pe. The n1ore a theory attains the status of a causal law,
the e asier it will be t o con firn1, o r to disconfirm, with a single case.
Indeed, risky predictions are co1n1non in natural scienc e fields such
as physics, which in turn se rved as the template for th e deductive­
no mo logical ("c overing-law" ) model of sci ence that in flue nced Eckstein
and o th ers in the postwar decades.46 A fre quently cited example is the
first imp ortant empirical de monstrati on of the theory o f relativity, which
took the form of a single-event predictio n on the occasion of the May
29, 1 91 9, solar eclipse. Stephe n Van Evera describes the impact of this
prediction o n the validation of Ein ste in's theory.
l ·I lhid.
1 ' ;\ 1\iird pl>sil i < ,11 , 11· l1i, \1 1 1 1 1 1 1 H 1r 1 � 11> lw 1wid1er Popperian nor Dayesian, has been articu­
l.11nl l,y .\-1.11·1, ( I '1"1,: ( 1 1.11 1 1 1 · 1 1 , ) h 1 1n1 t hi-; 1wr�pcctivc, the sa111e idea is articulated as
. 1 111.1 1 1 1· r ,,1 " ·., · 1·,·1, · 1 , ··.1 · .
I r , C, · · , , - . , ,. , l l , -111 1 •• 1 1 1 ' 1 I '
( (
118 II. Doing Case Studies
Einstein's theory predicted that gravity would bend the pc1th of light toward a
gravity source by a specific an1ount. Hence it predicted that during a solar eclipse
stars near the sun would appear displaced - stars actually hehind the sun would
appear next to it, and stars lying next to the sun would appear farther fro1n it- and
it predicted the arnount of apparent displacen1ent. No other theory made these
predictions. The passage of this one single-case-study test brought the theory wide
acceptance because the tested predictions \Vere unique - there was no plausible
coinpeting explanation for the predicted result - hence the passed test was very
strong.47
The strength of this test is the extraordinary fit between the theory and a
set of facts found in a single case, and the corresponding lack of fit between
all other theories and this set of facts. Einstein offered an explanation of
a particular set of anomalous findings that no other existing theory could
1nake sense of. Of course, one must assume that there was no - or limited -
measurement error. And one must assume that the phenomenon of inter­
est is largely invariant; light <loes not bend differently at different times
and places (except in ways that can be understood through the theory of
relativity). And one must assume, finally, that the theory itself makes sense
on other grounds (other than the case of special interest); i t is a plausible
general theory. If one is willing to accept these a priori assumptions, then
the 1919 "case study" provides a very strong confirmation of the theory.
I t is difficult to imagine a stronger proof of the theory from within an
observational (nonexperimental) setting.
In social science settings, by contrast, one does not commonly find
single-case studies offering knock-out evidence for a theory. This is, in
my view, largely a product of the looseness (the underspecification) of
most social science theories. George and Bennett point out that while the
thesis of the democratic peace is as close to a "law" as social science
has yet seen, it cannot be confirmed (or refuted) by looking at specific
causal mechanisms because the causal pathways mandated by the theory
are multiple and diverse. Under the circumstances, no single-case test can
offer strong confirmation of the theory (though, as we shall discuss, the
theory may be disconfirmed with a single case). 48
However, if one adopts a softer version of the crucial-case method
the least-likely (most difficult) case - then possibilities abound. Lily Tsai's
investigation of governance at the village level in China employs sever.i 1
in-<lepth case studies of villages that are chosen (in part) because of tlll'i1
Van Evera (1 997: 6fi-7). Ser: also Fckstl'il1 ( 1 975) ;111d Popper ( 1 9(1 \).
47
4 H Geo1·ge ,ind lk11ncll (200.1: 209).
Techniques for Choosing Cases 119
least-likely status relative to the theory of interest. Tsai's hypothesis is that
vill�ges with greater social solidarity (based on preexisting religious or
.
famrhal networks) will develop a higher level of social trust and mutual
obligation and, as a result, will experience better governance. Crucial
cases, therefore, are villages that evidence a high level of social solidarity
but that, along other dimensions, would be judged least-likely to develop
good governance - that is, they are poor, isolated, and lack de1nocratic
institutions or accountability mechanisn1s from above. "Li Settlement "
in Fujian province, is such a case. The fact that this impoverished villa�e
nonetheless boasts an impressive set of infrastructural accomplishments
such as paved roads with drainage ditches (a rarity in rural China) sug­
gests that something rather unusual is going on here. Because her case is
carefully chosen to eliminate rival explanations, Tsai's conclusions about
the special role of social solidarity are difficult to gainsay. How else would
one e�plain this otherwise anomalous result? This is the strength of the
least-likely case, where all other plausible explanations for an outcome
have been mitigated.49
Jack Levy refers to this, evocatively, as a "Sinatra inference": if it can
make it here, it can make it anywhere. 50 Thus, if social solidarity has
the hypothesized effect in Li Settlement, it should have the same effect in
1nore pi-opitious settings (e.g., where there is greater economic surplns).
The same implicit logic informs n1any case study analyses where the iutcnt
of the study is to confirm a hypothesis on the basis of a s ingle case (with­
out extensive cross-case analysis). Indeed, I suspect tbat, implicitly, most
case study work that focuses on a single case and is not nested \vithin
a cross-case analysis relies largely on the logic of the least-likely case.
llare\y is this logic n1ade explicit, except perhaps in a passing phrase or
two. Yet the deductive logic of the "risky" prediction may in fact be cen­
tral to the case study enterprise. Whether a case study is convincing or
not often rests on the reader's evaluation of how strong the evidence for
;1n argument might be, and this in turn - v1herever cross-case evidence is
li111i�ed and no manipulated treatment can be devised - rests upon an esti-
111:1t1on of the degree of "fit" between a theory and the evidence at hand ' as
·
discussed.
l'I
T�.11. I'_()07). 1t �h1111ld lll' 1to1nl I h;H T�;ii's conclusions do not rest solelv on this crucial
v.,�r·. lnd('l'd, >II!' n11pl,,, .., .1 l>r,,.,d r;1ngt· of 111erhodological tools, en�ompassing case
111('1 I i , 1,I•,
',! i u l y 1 1 1 1 1 1 \ I < I \ ', .1 ·,1·
\II l 1 ·1· r l .'ll0 .1 . 1 1 , 1 - 1 1 ',, , ii·. . . l,l, . . 11 1 '. i l < J'I.I: · I'll;
\.1 1 , . 1 11 ( 1 (/ <) S:: •!9); ,ll\d Sh;ikr ( 1 988:
,
1 ·1 ! , ) ,
120 II. Doing Case Studies
The Disconfitmato,y (Most-Likely) Crucial Case
A cenrra1 Popper.ian insight is that it is easier to disconfirm an inference
than to confirrn that san1e inference. (Indeed, Popper doubted that any
inference could be fully confirn1ed, and for rhis reason preferred the tcrn1
�corroborate,") This is particularly true of case study research desig:1s,
where evidence is l imited ro one or several cases. The key proviso is that
the theory under investigation n1ust take a consistent {a.k.a. .invariant,
deterministic) form, even if its predictions are not terrifically precise, well
elaborated, or broad.
i\s it happens, there are a fair n ui:nber of i nvariant propositions float­
ing a;ound the social science disciplines. 5 1 In Chapter rrhree, ,ve discus11.ed
an older rheory that stipulated that political srability \'v"Ould occur only
in countries that are relatively homogeneous, or \vhere existing hetero ­
geneities are n1itigated by cross-·cL1tting cJeavages. 52 ,i\.rend lijphart's study
of the Netberland.s� a peaceful country with reinforcing social cleavages, is
com111only viewed as refuting rhis theory on the basis of a single ln-depth
case analysis. 5.l
f-Ierctofore, I h2.vc treated causal factors as dichoton1ous. Countries
have either reinforcing or cross-cutting cleavages� and they have regirnes
that are either peaceful or conflictual. Evidently, these sorts of paran1eters
are often 1nattcrs of degree. In this read:.ng of the theoryJ c2.ses are rnore
or less crucial. Accordingly, the n1ost useful - th.1t is, n1ost cruclal - case
for Lijpharfs purpose is one that has the 111ost segregated social groups
and the most peaceful and democratic track record. In these respects,
the Netherlands was a very good choice. Indeed, the degree of disconAr­
mation offered by this case study is probably grea,er rhan the degree of
disconfinnation th.1t 1night have been provided by another case, such a:,
India or Papua Ne\v Guiuea - cow1tries where social peace has not ahvays
been secure. "fhe point is rhar \vhere variables·are continuous rather than
dk:hototnous� it is po-ssible TO evaluate potential cases in ter1ns of thei1
degree of crucialness.
Note that vvhen disconfir1ning a causal argnrr1ent, background cau:.:1\
factors are irrelevant (except as they 1night affect the classification of the
case -\:v ichln the popuh:.tion of an i nference). Tt does not 1n.1tter ho"v 1hv
51 Goertz nncl Levy (forrh.;;:orrin2}; Goertz ai:o Star· (2003;.
52 Ahnond (19 Sfi); Bentley {190811 967); I ifset ( \ 960/ 1 lj;-i3): '!'r.i1;1aP ( \ 'JS I ),
s.-i Lijphart (1968). See a!so di-e,cu,;.<,:ions- i11 E,.- ks'.TiJ' \ 1 975) ,nid l ::ph:1n ( ! 91, lJ ), hl,. .1,l.!1
ti0nal e-;,,:nmpks of ,:,1sL· �1 ;1dk-� ,li�<.:011!ir11ti•;µ. g;·nvr:1! rw(>ps,\il1(111� id .1 dc1t·,·111i11,·<1:,
nature, �-\'(' l\lli:n ( [<Jr,'i;; 1.nh,·t, Trow, ,\l'.d 1 ·,)l,·111.111 ( l'l)f,): N1t,!·.ud ( !'J(JiJ); H,·1!h
(/0: )0f.1: t H i l l; :11:,l d 1,· di ,. 11 .,.1 .. 1,-. I l l J J:,,i: \ I ')'!;\) ,,!Jd H1 11·,,11\".l 1 i l ' l 'i\J
Techniques for Choosing (;ases l2l
Netherlands� India, and Papua �ew Guinea score o n other factors that
affect democracy and social peace.
Granted, it lnay be questioned \'vhether prcsurned invariant theories are
really invariant; perhaps they are better understood as probabilistic. Per··
haps, that �s, the theory of cross-cutting cleavages is still true, probabilisti­
cally, despite the apparenr Dutch exception. Or perhaps the theory is still
true, d eternllnisticatly, within a subset of case5. that does not include the
Netl1crlands. (fh i s sort of clai111 see1ns unlikely in rhi� particular instance
but it is quite plausible i n many orhers.) Or perhaps the theory is in need
of refr �n1i:1g; �t is true, deterministically� but applies only to cross-cutting
.
ethn1c/rac1aJ c.1e.1vages, not to cleavages that are prlmarily religious. One
may quibble over \vhat it n1eans to "disconfirm" a theory. �fhe point is
that tJ; e crt:cial case has, in all these circumstances, provided important
updattng ot a theoretical prior.
Conclusion
In this section; f have argued that the degree to \vhich cruc:al cases can
provide decisive confirmatio n or disconfirn1ation of a theorv' is in Iar•re
part a product of the structure of rhc rheory to be tested. It is a deductive
"
n1�tter ra �he� th �n an inductive rr1a tter, ;-;tricdy speaking. In this respect,
_ . ' .
,1 pos1t1v1st , or1entation to\vard the \vork qf social science n1av lead to a
� reater appreciation of the case study fonnat - :1.ot a denigrati�n of that
tor1nat, as is usually supposed. Those who, with Eckstein, e1nbrace the
notion of covering laws are llkely to be attracted to the idea of cases that
arc crucial. By the same token, those \\·ho are in1pressc<l by the irregularity
:1 nd com plex1ty of social behavior are unlikely to be persuaded by crucial­
i.:.i:-:e studies, except as a 111echod of disconfirrning absurdly rigid caus.aJ
I:, w,.
I have show!1, relatedly, that it is almost al\vays easier to discon.hrm
,1 theory than to confirm it with a single case. Thus, a theory that is
111H.l erstood ro be de::erminlstic 111ay be disconfirmed by a 1.:ase study, prop­
erly choser:. This is the 1nost co111n1on e1nployn1ent of rhe crucial-case
l!H'thod in soci::l science settings.
�ote- that the crucia]-ca;;e n1ethod of case selection cannot be employed
in a la rgc� N contexc. "fhis is bcca use the n1ethod of selection would render
I hr· L';l ell' st:1dx l°L'd L11HJant. ()nee OllC identifies the refevant paran1eters and
1!1c '>l'Orcs ol ;1JI t':l'><'" 011 d1nsc p�1:an;.eters, one has ln effect constructed
.1 , 1 0.s;"- t'.l l' 1nod('I ! 1 1 . 1 1 \viii, !,y it.•;clf, confinn or disconfirn1 rhe theory i n
�.
q 1 1 v.,l 1n11 . I Ii!' , .1•,c ',l n,lr 1· . dH·1;,. cl o:·1h irrrlcv;1nt, at least a s a 1neans of
1 1 11tfl1 n1.1111i11 , q d 1- , , 1 1 d 1 1 1 1 1 . 1 1 1 > 1 1 1 11 ri·111;111 1c, lii )dily rclc1.-:inr a s ;1 lllL\lllS o f
1 22 II. Doing Case Studies
exploring causal mechanisms, of course. However, because this objective
is quite different from that which is usually associated with the term, I
enlist a new tern1 for this technique.
Pathway Case
One of the most important functions of case study research is the elucida­
tion of causal mechanisms. This is well established (see Chapter Three).
But what sort of case is most useful for this purpose? Although all case
studies presun1ably shed light on causal mechanisms, not all cases are
equally transparent. In situations where a causal hypothesis is clear and
has already been confirmed by cross-case analysis, researchers are well
advised to focus on a case where the causal effect of one factor can
be isolated from other potentially confounding factors. I shall call this
a pathway case to indicate its uniquely penetrating insight into causal
mechanisms.
To clarify, the pathway case exists only in circun1stances where cross­
case covariational patterns are well studied but ,vherc the mechanisn1
linking X 1 and Y ren1ains dim. Because the pathway case builds on prior
cross-case analysis, the problem of case selection n1ust be situated within
that sample. There is no stand-alone pathway case. Thus, the following
discussion focuses on how to select one (or a few) cases from a cross-case
sample.
Cross-Case Technique with Binary Variables
The logic of the pathway case is clearest in situations of causal sufficiency -
where a causal factor of interest, X1 , is Sufficient by itself (though per­
haps uot necessary) to cause a particular outco1ne, Y, understood as a
unidirectional or asymn1etric casual relationship. The other causes of
Y, about which we need make no assumptions, are designated as J
vector, X2 .
Note that wherever various causal factors are deemed to be substi­
tutable for one another, each factor is conceptualized (individually) as
sufficient. 54 Situations of causal equifinality presu1ne causal sufficiency
on the part of each factor or set of conjoint factors. The QCA technique,
for exan1ple, presumes causal sufficiency for each of the designated c,1us;il
paths.
Techniques for Choosing Cases 123
Consider the following examples culled by Bear Braumoeller and
drawn from diverse fields of political science.55 The decision to seek an
alliance is motivated by the search for either autonomy or security.56 Con­
quest is prevented by either deterrence or defense.57 Civilian intervention
in 111ilitary affairs is caused by either political isolation or geographical
encirclement.58 War is the product of miscalcnlation or loss of control. 59
Nonvoting is caused by ignorance, indifference, dissatisfaction, or inac­
tivity.6 0 Voting decisions are influenced either by high levels of informa­
tion or by the use of candidate gender as a proxy for social information.6 1
Democratization conies about through leadership-initiated reform, a con­
trolled opening to opposition, or the collapse of an authoritarian regime. 62
These, and many other, social science arguments take the form of causal
substitutability - multiple paths to a given outcome.
For heuristic purposes, it will be helpful to pursue one of these exam­
ples in greater detail. For consistency, I focns on the last of the exemplars -
democratization. The literature, according to Braumoeller, identifies three
n1ain avenues of democratization (there may be more, but for present pur­
poses let us assume that the universe is limited to three). The case study
for1nat constrains us to analyze one at a time, so let us lin1it our scope to the
first one - leadership-initiated reform. So considered, a causal-pathway
case would be one with the following features: (a) democratization,
(b) leadership-initiated reform, (c) no controlled opening to the opposi­
tion, (d) no collapse of the previous authoritarian regime, and (e) no other
extraneous factors that 111ight affect the process of democratization. In a
case of this type, the causal mechanisms by which leadership-initiated
refor1n may lead to democratization will be easiest to study. Note that it
is not necessary to assume that leadership-initiated reform always leads
to de1nocratization; it may or may not be a deterministic cause. But it is
necessary to assume that leadership-initiated reform can sometimes lead
to den1ocratization. This covariational assumption about the relationship
1 1 Ibid. My chosen examples are limited to those that might plausibly be n1odeled ·with
dichoto111ous variables. For further discussion and additional exan1ples, see Most and
St·;ur ( 1 984) and Cioffi-Revilla and Starr (1995).
·,,, ivlorrow ( 1 99 1 ; 905).
Sl·hellin� ( l 966: 78).
",ll l't >�l'll ( 1 984: }')).
1•11 H .w:,d.il(· .111,l !{11·.I ( 1 ·1'! \ ' \ II
.,,, l l"V)' ( I ')H \ : 81,).
'" 1 l d , l ) n 1 1 1 . , 1 1 ( l ' J'I ' )
,. ' ( . , 1 , > 1 1 1 1 · 1 ( l ' ! ' l l I
124 II. Doing Case Studies
TABLE 5.2. Pathway case with dichotomous causal factors
Case
types
X1 = the variable of theoretical interest. X2 = a vector of controls
(a score of zero indicates that all control variables have a score of zero,
while a score of one indicates that all control variables have a score
of one). Y = the outco1ne of interest. A�H = case types (the N for
each case type is indeterminate}. H = path\.vay case. Sample size =
inci ctenninate.
Assumptions: (a) all variables can be coded dichoto1nously; (b) all
independent variables are positively correlated \.Vith Y in the general
case; (c) X1 is (at least sometin1es) a sufficient cause of Y.
between X1 and Y is presu1nably sustained by the cross-case evidence (if
it is nor, there is no justification for a path\.vay case study).
Now let us move from these examples to a general-purpose model.
For heuristic purposes, let us presume that all variables in that rnodel are
dichotomous (coded as zero or one) and that the model is complete (all
causes of Y are included). All causal relationships will be coded so as to
be positive: X1 and Y covary as do X2 and Y. This allovvs us to visualize
a range of possible cotnbinations at a glance.
Recall that the pathway case is always focused, by definition, on a
single causal factor, denoted X1 • (The researcher's focus n1ay shift to other
causal factors, but may focus only on one causal factor at a tinie.) In
this scenario, and regardless of how n1any additional causes of Y there
might be (denoted X2 , a vector of controls), there are only eight relevant
case types, as illustrated in Table 5.2. Identifying these case types is a
relatively si1nple 1natter, and can be accon1plished in a s1nall-N sa1nple hy
the construction of a truth table (modeled after Table 5.2) or in a largc-N
sample by the use of cross-tabs.
Note that the total nun1ber of combinations of values depends 0 1 1 t l H·
number of control variables, which we have rcprcscntl·d \.V ith ,1 sini 1, h ·
vector, X2 . If this vector consists of ,1 single v,1ri:1hll' , t hc 1 1 t l H'IT ;11-v 0 1 1 h
eight C:lSl' l")'!ll'S, If d1is Vl'l'\01" l'llllSii.;ls (Jf 1 \V O v:1 ri:d1]l'� ( \. • . , , \. • 1 , ) , t l u·n l \tc
Techniques for Choosing Cases 125
total number of possible combinations increases from eight (23 ) to sixteen
(24 ), and so forth. However, none of these co1nbinations is relevant for
present purposes except those where X2a and X2b have the same value
(zero or one). "Mixed" cases are not causal pathway cases, for reasons
that should become clear.
The pathway case, following the logic of the crucial case, is one where
the causal factor of interest, X1 , correctly predicts Y's positive value
(Y = 1) while all other possible causes of Y (represented by the vector,
X2 ) 1nake " wrong" predictions. If X 1 is - at least in son1e circumstances -
a sufficient cause of Y, then it is these sorts of cases that should be most
useful for tracing causal mechanisms. There is only one such case in
Table 5.2 - H. In all other cases, the mechanism running from X 1 to
Y would be difficult to discern, because the outcome to be explained does
not occur (Y = 0), because X 1 and Y are not correlated in the usual way
(violating the terms of our hypothesis), or because other confounding fac­
tors (X2 ) intrude. In case A, for example, the positive value on Y could be
a product of X 1 or X2 . Consequently, an in-depth examination of cases
A-G is not likely to be very revealing.
Keep in mind that because we already know from our cross-case exam­
ination what the general causal relationships are, we know (prior to the
case study investigation) whar constitutes a correct or incorrect predic­
tion. In the crucial-case n1ethod, hy contrast, these expectations are deduc­
tive rather than empirical. This is what differentiates the two methods.
i\nd this is why the causal-pathway case is useful principally for eluci­
dating causal 1nechanisms rather than for verifying or falsifying general
propositions (which are already apparent from the cross-case evidencc). 63
Now let us con1plicate n1atters a bit by in1agining a scenario in which at
least some of these substitutable causes are conjoint (a.k.a. conjunctural).
' [ 'hat is, several combinations of factors - Xa + Xb or Xe + Xd - are
sufficient to produce the outcome, Y. This is known in philosophical circles
;\s �1n INUS condition, 64 and it is the pattern of causation assumed in most
''
1
H course, we should leave open the possibility that an investigation of causal mechanisn1s
(
n_ight invalidate a general clai1n, if that claim is utterly contingent upon a specific set
1
of c:111s;1I mechanisms and the case study sho\.VS that no such n1echanisn1s are present.
1 1 ,)WL'VL' r, rhis is rather unlikely in 1nost social science settings. Usually, the resu1t of such
11 li 1 1di 1 1g will lw ;1 rcformuhtion of the causal processes hy \.vhich X 1 causes Y - or,
.1l11·r1 1 .11 ivd )·, .t n·,1li-.r: 1 l i!11 1 1 h ; 1 t t ill' c;1sc under Investigation i� aberrant (atypical of the
)',l ' I W r. i i p11p1 1 I.\I i 1 \ I I 1 11 , , l \ l ' \ ) ,
,,-I :\ 1 1 lNl I� , 1 n 1 d 1 1 11 1 1 1 1 !'1i-1 ·. I<> . 1 1 1 l 1 1 ',1illi1 i!'llt h11t NcCl'\Sary part of ,1 condition which is
11·.,·II l 1 1 1 1 1 1 ·, , . . ,,. 1 1 1 l,i 1 I ',i , 1 1 1 , 11 1 1 1 l , , i . t 1,, 1 1 1 1 < 1 1 1.ll' l'<'\lii1 Th1 1 ,,;1 wlit·n 0 1 w identifies a short
' 1 1 , 1 1 11 .I', 1 1 11 . • . 1 1 1 ., ' , , I I 1 1 1 , . , , 1 .. 1·. . 1 1 1 1 1 1·.. 1 1 1 . . t l i - , t, 1li. 1 1 1111' li r·t · \\':I', t", \ 1 1\t'd Ii} .l -;hurl
126 II. Doing Case Studies
QCA (Qualitative Comparative Analysis) models. 65 Here, everything that
has been said so far must be adjusted so that Xi refers to a set of causes
(e.g., X, + Xb) and X2 refers to a vector of sets (e.g., X, + X<l , X, +
X1, Xg + Xh, . . . ). The scoring of all these variables makes matters more
difficult than in the previous set of examples. However, the logical task IS
identical, and can be accomplished in a similar fashion, that is, in small-N
datasets with truth tables and in large-N datasets with cross-tabs. Case
H now refers to a conjunction of causes, but it is still the only possible
pathway case.
Cross-Case Technique with Continuous Variables
Finally, we must tackle tbe most complicated scenario -when all (or most)
variables of concern to the 1nodel are continuous, rather than dichoto­
mous. Here, the job of case selection is considerably more complex, for
causal "sufficiency" (in the usual sense) cannot be invoked. It is no longer
plausible to assun1c that a given cause can be entirely partitioned, that is,
that all rival factors can he eliminated. Even so, the search for a pathway
case may be viable.
What we are looking for in this scenario is a case that satisfies two
criteria: ( 1 ) it is not an outlier (or at least not an extreme outlier) in the
general model, and (2) its score on the outcome (Y) is strongly influenced
by tbe theoretical variable of interest (X 1 ), taking all other factors mto
account (X2 ). Io this sort of case it should be easiest to identify the causal
1nechanisms that lie between X1 and Y.
In a large-N sample, these two desiderata may be judged by a careful
attention to the residuals attached to each case. Recall that the question
of deviance, which we have discussed in previous sections, is a matter of
degree. Cases are more or less typical/deviant relative to a general model,
with very high residuals (e.g., standardized residual > I 2 I). for cases that
as judged by the size of their residuals. It is easy enough to exclude cases
lie closer to their predicted value, small Jifferences in the size of residuals
1nay not matter so much. But, ceteris paribus, one would prefer a case
that lies closer to the regression liue.
circuit in conjunction with some other background factors (e.g., oxygen) that were also
necessar.y to that outcome. But one is not implying that a short circuit \Vas necessary 10
that fire, v..·hich might have been (under different circu1nstances) caused by other factor�.
See :tvlackie (1965/1993).
65 R:igin (2000).
Techniques for Choosing Cases 127
Achieving the second desideratum requires a bit of manipulation. In
order to determine which (non-outlier) cases are most strongly affected
by X 1 , given all the other parameters in the model, one must con1pare the
size of the residuals (their absolute value) for each case in a reduced-form
model, Y = Constant + X2 + Resreduced, to the size of the residuals for
each case in a full model, Y = Constant + X2 + X 1 + Resfoll · The pathway
case is that case, or set of cases, that shows the greatest difference between
the residuals for the reduced-form model and the full model (1'Rcsidual).
Thus,
Pathway = IResrcduccd - ReSfulll, if I Resrcduccdl > I Resfuul (5.10)
Note that the residual for a case must be smaller in the full model than in
the reduced-form model; otherwise, the addition of the variable of interest
(Xr ) pulls the case away from the regression line. We want to find a case
where the addition of X 1 pushes the case toward the regression line, that
is, it helps to "explain" the case.
As an example, let us suppose that we are interested in exploring
the effect of n1ineral wealth on the prospects for de1nocracy in a soci­
ety. According to a good deal of work on this subject, countries with a
bounty of natural resources - particularly oil - are less likely to democra­
tize (or, once having undergone a democratic transition, are more likely to
revert to authoritarian rule).66 The cross-country evidence is robust. Yet,
as is often the case, causal mechanisn1s ren1ain rather obscure. Consider
the following list of possible causal pathways, summarized by Michael
Ross:
A 'rentier effect' . . . suggests that resources rich governments use low tax rates
anU patronage to relieve pressures for greater accountability; a 'repression
cffccr' . . . argues rhar resource wealrh rerards democrarization by enabling gov­
ernn1ents to boost their funding for inrernal security; and a 'modernization
effecr' . . . holds that grov.fth based on the export of oil and minerals fails to bring
,1bout the social and cultural changes thar tend to produce democratic govern-
111cnt. 67
Are all three causal mechanisms at work? Although Ross atten1pts to test
factors in a large-N cross-country setting, his answers re1nain rather
d1l·sc
"'' !\.1rr1) ( l 'J'J'!), 1 1111111 ' 1111 ' \ . ( . ' Oi l � ) ; Rt)�� (200 1 ).
1{1 l\', ( .1 1 1 1 l I : \ .' ' ii I
128 II. Doing Case Studies Techniques for Choosing Cases 129

speculative. 6 8 Let us see how this might be handled by a pathway-case TABLE 5.3. Possible pathway cases where variables are scalar and
approach. assumptions probabilistic
The factor of theoretical interest, oil wealth, may be operationalized
Resrcduccd Res fuu ..O..Residual
as per capita oil production (barrels of oil produced, divided by the total
Country

population of a country). 69 As previously, we measure detnocracy with a Iran -.282 -.456 .175
continuous variable coded from - 10 (1nost authoritarian) to +10 (most Turkrnenistan - 1 .220 -1.398 .178
!vlauritania -.076 -.255 .179
democratic). Additional factors in the model include GDP per capita
Turkey 2.261 2.069 .192
(logged), Muslims (as percent of the population), European language (per­ Switzerland . 1 77 -.028 .205
cent speaking a European language), and ethnic fractionalization ( 1 - Venezuela .148 .355 -.207
likelihood of two randomly chosen individuals belonging to the same Belgium .518 .310 .208
ethnic group). 70 These are regarded as background variables (X2) that Morocco -.540 -.776 .236
may affect a country's propensity to den1ocratize. The full model, li1nitcc.l Jordan .382 .142 .240
Djibouti -.451 .245
to 1995 (as in previous analyses), is as follows:
-.696
Bahrain -1.411 -1.673 .262
Luxen1bourg .559 .291 .269
Democracy = -3.71 Constant + 1.258 GDP (5.11) Singapore -1.593 -1.864 .271
+ -.075 Muslim + 1 . 843 European On1an - 1 . 2 70 -.981 -.289
Gabon -1. 743 -1.418 -.325
+-2 . 0 93 Ethnic fract + - 7.662 Oil -1.681
Saudi Arabia -1 .253 -.428
R2,<l1 = .450 (N = 149) Nonvay .315 1.285 -.971
lJnited 1\rab En1iratcs - 1.256 -.081 -l.175
The reduced-form model is identical except that the variable of theoretical Kuv..,ait -l.007 .925 -1.932
interest, Oil, is rcrnoved. Resrcduccd = the standardized residual for a case obtained frorn the reduced 1nodel
{without Oil) - equation 5.12.
Democracy = -. 8 3 1 Constant + .909 GDP (5.12) Res(u/1 = the standardized residual for a case obtained fron1 the full n1odel (,.vith
Oil) - equation S. 1 1 .
+-.086 Muslim + 2.242 European 6.Resi<lual = Res, wiu,cd - Resiul l · Listed in order of absolute value.

+ - 3.023 Ethnic fract

R2,d; = .428 (N = 149)
What does a cornparison of the residuals across equations 5. 1 1 and
5.12 reveal? Table 5.3 displays the highest t-Residual cases. Several of
68 Ross tests these various causnl mechanisn1s with cross-country data, en1ploying variot1:-.
proxies for these concepts in the benchmark model and observing the effect of these .­
presu1nably intermediary - effects on the main variable of interest (oil resources). This
is a goo<l example of how cross-case evidence can be n1ustered to shed light on cn1\,il
rnechanisn1s; one is not ]in1ited to c.1se study forrnats, .-ts discussed in Chapter Three. Still,
as Ross notes (2001: 356), these tests are by no n1l:ans definitive. Indeed, the cocffici,·111
on the key oil variable re1nains fairly constant, except in circurnst:1nces ,vhen: rhe sa111pk
is severely constrained.
69 Derived frorn l-fun1phreys (2005).
70 C DPpc (.Ll ta are fro111 World Bank (2003). tv1 11';1i1n� :111d L11rtipl', \11 l.1111•, t 1:w, 1 · .\r1· , 1 1dnl
hy l h l' ;t11tl i oL Lth 1 1 il i'r.tl·1i 1 J 1 1:il11. :11iu 1 1 j., dr·,1w 1 1 11·<1 1 1 1 i\k-, 1 11.1 t'T .11. I ' I l l ) \ 1 .
these 1nay be summarily ren1oved fro1n consideration by virtue of the
fact that I Resrecluccd l < I Resfull l· Thus, we see that the inclusion of Oil
increases the residual for Norway; this case is apparently better explained
1uithout the inclusion of the variable of theoretical interest. Needless to
say, this is not a good case to explore if we wish to exan1ine the causal
111cchan isms that lie between natural resource wealth and de1nocracy. (It
rnight, however, be a good case for model diagnostics, as discussed in the
previous section on influential cases.)
!\ n1ong cases where the residual declines from the reduced to the full
111odcl, scvcr;1\ arc clear-cut favorites as pathway cases. The United Arab
1:. n1ir,1tl'-" :111d l<u\v:1il h:1vc t h e highest .6.Residual values and also have
L1irly 111od('.., I n· ..,id11.il:-. in t ill· ftil! 111odcl ( Rcs r,i1il , signifying that these
1 , I ..,('', :tn· 1 1 0 1 c \ ! 1 i · 1 1 1 t · ( 1 1 1 1 l 1 t · 1 •, ; indeed, :ll'l'ording to the p;1r;1111ctcrs of this
·
1 11 1 1d(·I, 1 11(' I 1 1 1 1 ! ( ' 1 I :\ 1 .ii 1 I 1 1 1 1 1 . ! ! < ", \\' 0 1 i l d hi' 1 l')',,1 nlctl ;\.., .1 1 ypil·:1 I L·:1:-.c. l ' hv
130 If. Doing Case Studies
analysis suggests, therefore, that researchers seeking to explore the effect
of oil wealth on regirne type might do well to focus on the-se tv.ro cases,
since their patterns of dernocracy cannot be \Veil explained by other fac­
tors such as, economic develop111ent, religion, European influence, o r eth­
nic fractionalization. 'fhe presence of oil wealth in these countries would
appear to have a strong independent effect on the prospects for dernocra­
tization Jn these countries, an cffett that is well 111odeJed by our general
rbeory and by the available cross-case evidence. And this effect should
be interpretable in a case-study format more interpretable� at any rate,
than it would be in other cases.
Conclusion
'fhe logic of caus::J "elimination" is rnuch more compelling \vhere vari­
ables arc dichotornous and where causal sufficiency can be assumed {X1 is
sufficient by itself, at least in so1ne circumstances, to cause Y). \Vhe:re vari­
ables are continuous the strategy of the path\.vay case is more dubious, for
potentially confounding causal factors (X2 ) cannot be neatly partitioned.
Even so, rhJs discussion has sho\.vn v;,rhy the selection of a pathway case is
a logical approach to case study analysis in tnany circumstances.
The exceptions may be brlcfly noted. Sometimes, \Vhere all variables
in a n1odel are dichoto1nous, there are no pathway cases, that is, no cases
of type H (in Table 5.2). This is known as the "empty cell" problem, or a
problern of severe causal n1ulticollinearity. The universe of observational
data docs not always oblige us \Vith cases th.at allr>W us to test a given
hypotbesis independently of all others.
\X'herc variables are continnous� the analogous problem J5 that of a
causal variable of interest (X i ) that has only minimal effects on the our­
con1e of interest. That is, its role in the general model is qulte minor (as
judged by its standardized coefficient or b y F-tests comparing the reduced­
form model and the full model). In these situations, the only cases that
are strongly affected by X, - if there are any at all - may be extreme
outliers, and these sorts of cases are not properly regarded as prov1ding
confirmato:::-y evidence for a proposition, for reasons that are abundantly
clear by now,
Finally, it must be underlined that the identification of a causal-pathway
case does nor obviate the utility of exploring other cases. Ho\vCvs"r, thl<;
sort of rr:ulticase investigation moves beyond the logic of the c:1 ui.;al p;1th
way case, underlining a point that we sh.:dl ret urn to in the cn111.: l11ding
Techniques for Choosing Cases 131
section of the chapter: case-selection procedures often cornbine different
logics.
Despite the technical nature of this discussion, it should be noted that
when researchers refer to a particular case as an "'example" of a broader
pheno1nenon, they are often referring to a pathv1ay case. This sort of
case illustrates the causaI relationship of interest i n a particularly vivid
1nanner, and therefore may be regarded as a common trope a1nong case
study researchers,
Most-Similar Case
The most-similar method, unlike the previous methods, employs a mini­
mum of two cases. 71 In its puresr form, the chosen pair of cases is sia1ilar
in all respects except the variable(s) of interest.
If the study is exploratory (Le., hypothesis-gener?.ting), the reseJrcher
looks for cases that differ on the outcon1e of theoretical jnrerest but are
similar on various factors that might have contributed to that uutcon1e, as
illustrated in Table 5.4 (A). This is a common form of case selection at the
initial stage of research. Often, fruitful analysis begins with an apparent
anon1aly: two cases are apparently quite sirnilar, and yet demonstrate
surprisingly different outcomes. 'The hope is that intensive studv of these
cases will reveal one or at 1nost several - factors that differ ac;oss these
cases. These differing factors (X 1 ) are the putative causes.
So1netimes, a researcher begins with a strong hypothesis, i n which case
her re-search design is confirmatory (hypothc:;;is-tcsting) from the get�go.
'fhat is, she strives to identify cases that exhibit different scores on the
factor of interest and sin1ilar scores on all other possible causal facto.�s,
as illustrated in the second (hypothesis-testing) diagram in Ta ble 5.4 (B).
If she discovers such a case, it ls: regarded as providing confirmatory evi­
dence for the proposition, as well as fodder for an exploration of causai
1nechanisms.
The point is that the purpose of a most-similar research design,
and hence its basic set-np, may change as a researcher rnoves fro1n an
exploratory to a confirn1atory mode of analysis. However, regardless of
' I \on11,,l1!!:_'� the ni,;1\1-sim!br 1r.cthod is known as the ''method of difference," after
1h 111vn1tiir ( \1111 1 1<•1 i/ J i-:72), F(;!' bier tr\'atment5 �ee Cohen and �agcl (1 934);
h·J :,1:1 ( l '!�'I); l ,!'l' : 1 1 11•, i .1 \HJ t ; Clt.1plt'r 9); Li 1 phnrt ( 1 971, 1 975); :\1eL·kst:·i,,�tl-: ( 1 975);
l'1 t1'w1ir,,h1 .11111 Ii 1111-· ! 1 •1 "li): .11111 Sl,1H p,,l ,ind S1 ,111e� ( 1 980),
132 II. Doing Case Studies
TABLE 5.4. Most-similar analysis with two case types
(A) Hypothesis-generating (Y-centered):
(BJ Hypofhesis-testing (X1/Y-centered):
X 1 = the variable of theoretical interest. X2 = a vector of
controls. Y = the outcorne of interest.
where one begins, the results, when published, look like a hypothesis­
testing research design. Question marks have been removed: (A) becomes
(B) in Table 5.4. Consequently, the notion of a "most-similar" analysis is
usually understood as a tool for understanding a specific X 1 /Y relation­
ship.
As an example, let us consider Leon Epstein's classic study of party
cohesion, which focuses on t\VO si1nilar countries, the United States
and Canada. Canada has highly disciplined parties whose members vote
together on the floor of the House of Commons, while the United States
has weak, undisciplined parties \vhose members often defect on floor votes
in Congress. In explaining these divergent outcomes, persistent over many
years, Epstein first discusses possible causal factors that are held more or
less constant across the two cases. Both the United States and Canada
inherited English political cultures; both have large territories and het­
erogeneous populations; both are federal; and both have a fairly loose
party structures with strong regional bases and a weak center. These are
the "control" variables (X2 ). Where they differ is in one constitutional
feature: Canada is parliamentary, while the United States is presidential.
And it is this institutional difference that Epstein identifies as the differ­
entiating cause (X 1 ). 72
Several caveats apply to any most-similar analysis (in addition to
the usual set of assumptions applying to all case study analysis). firs1 ,
72 For further examples of the n1ost-simiL1r method, sn· Bn·111HT ( 1 ')7f,); I h 1 n i l l 1 )1 1
l.ipsct ( 1 968}; Migu('l (2004); r,..,Joul,kr { 1 ')77); : 1 1 1 d 1'<1\IH't' (.1 11(),1).
Techniques for Choosing Cases 133
one must code cases dichotomously (high/low, present/absent). This is
straightforward if the underlying variables are also dichotomous (e.g.,
federal/unitary) . Hov,ever, it is often the case that variables of concern
in the n1odel are continuous (e.g., party cohesion). In this setting, the
researcher n1ust "dichotomize" the scoring of cases so as to simplify the
two-case analysis. This is relatively unproblen1atic if the actual scores on
this dimension are quite different (on X 1 and Y) or virtually identical
(on X2 ) . Unfortunately, the empirical universe does not always oblige the
requirements of Millean-style analysis, and in these instances the logic of
most-similar con1parison becomes questionable.
Some flexibility is admissible on the vector of controls ( X2 ) that are
"held constant" across the cases. Nonidentity is tolerable if the deviation
runs counter to the predicted hypothesis. For example, Epstein describes
both the United States and Canada as having strong regional bases of
power, a factor that is probably n1ore significant in recent Canadian his­
tory than in recent An1erican history. However, because regional bases of
power should lead to weaker parties, rather than to stronger parties, this
el en1ent of nonidentity does not challenge Epstein's conclusions. Indeed, it
sets up a most-difficult research scenario, as discussed earlier. At the same
time, Epstein's description of Canadian and An1erican parties as "loose"
might be questioned. Arguably, An1erican parties, dominated in the latter
twentieth century by direct primaries ( open to all who declare themselves
a n1ember of a party and, in son1e states, even to those \Vho arc men1-
bers of the opposing party), are considerably more diffuse than Canadian
parties. The problem of coding continuous variables in a dichotomous
n1a nner is threatening to any most-si1nilar analysis.
In one respect, ho\vever, the requirements for case control are not so
stringent. Specifically, it is not usnally necessary to measure control vari­
ables (at least not with a high degree of precision) in order to control for
them. If two countries can be assnmed to have sin1ilar cultural heritages,
one needn't worry about constructing variables to measure that heritage.
( )ne can simply assert that, whatever they arc, they are more or less con­
.-;tant across the two cases. This is similar to the technique employed in a
r,111 don1ized experiment, where the researcher typically docs not atten1pt
t o 1nc:.1sure all the factors that n1ight affect the causal relationship of inter­
l':-.t. She assun1cs, rath er, that these unknown factors have been neutralized
, llToss the trc:1 1 11H·1H :111d co1nrol groups by randomization. This can be a
l n tgl' : 1 d v ; 1 1 1 1 ;q,,,. o\'(T l.1 r1 ·.c N lTnss -c;1sc in cthods, where each case must be
( I ')'/ '); , l ',..,i 1 •, 111·d ,I .., 1 11·1 di, ·.1 i ! l c 1 1 1 1 . 1 l l 1 c l1·v;t11l conl rol v.1 riah lcs � often a highly
q1 11",t io1 1 . i l il1· l l ! t H t · 1 l 1 1 1 1 • . 1 1 1 1 1 1 1 11t· ! 1 1.11 l l l l l '-,[ iinpo\t' .-; t rong :lS.<; l l lllpt ions
134 II. Doing Case Studies
about the shape of the underlying causal relationship (usually presumed
to be linear).
Cross-Case Technique
The most useful statistical tool for identifying cases for in-depth anal­
ysis in a most-similar setting is some variety of "matching" strategy.73
Statistical estimates of causal effects based on matching technique s have
been a major topic in quantitative methodology over the last twenty-five
years, first in statistics74 and subsequently in econometrics75 and political
science. 76
Matching techniques are based on an extension of experimental logic.
In a randomized experiment, elaborate statistical models are unneces­
sary for cansal inference because, for a large enough selection of cases,
the treatment group and the control group have a high probability of
being similar in their background characteristics (Xz ). Hence, a simple
difference-of-means test is often sufficient to analyze the effects of a treat­
ment variable (X i ) across groups.
In observational studies where the hypothesized causal factor (Xi) is
dichotomous, the situation is superficially the san1e. For purposes of dis­
cussion, we shall refer to cases with a "high" score on X1 as members of
the treatment group, and to cases with "low" scores as n1embers of the
control group. Thus are observational stndies translated into the lexicon
of experimental analysis.
However, in observatioual studies it is unusual to find cases that differ
on X 1 but not on various background characteristics (X2) that n1igbt
affect the outcome of interest. For exa111ple, countries that are strongly
democratic (or strongly authoritarian) are likely to be similar in more
than one respect. This greatly complicates the analysis of X 1 's inde pende11 1-
effcct on the outco 111e.
The traditional approach to this problem is to introduce a variabil'
for each potential confounder in a regression model of causal relation
ships. But this standard-issue technique requires a strong set of assn111p·
tions about the behavior of the various factors introduced into the model.
Matching techniques have been developed as an explicit alternative t o
73 For good introductions, see Ho et al. (2004); Morgan and Harding (2005); Rosl'1 1 h:u111 1
(2004); an<l Roscnhaun1 and Silber (200 1 ). For a discussion of lllatching pr()cnlur(·� 111
Stc1ta, see Abadie et al. (2001).
74 Rosenbau1n and Rubin ( 1 985); Roscnb,1t1rn (2004).
7 � I hihn ( l 998).
71
' ] lo L'1 :d. ( 2()(),1 ); Jin.ii ( }()()) ).
Techniques for Choosing Cases 135
the control-variable approach. This alternative begins b y identifying a set
of variables (other than the dependent variable or the main independent
variable) on which the cases are to be matched. Then, for each case in
the treatment group, the researcher identifies as many cases as possible
fro1n the control group with the exact same scores on the rnatching vari­
ables (the covariates). Finally, the re searcher looks at the difference on
the dependent variable between the cases in the treatment group and the
matching cases in the control group. If the set of rnatching variables is
broad enough to include all confounders, the average difference between
the treatn1ent-group and the matching control-group cases should pro­
vide a good estimate of the causal effect. Even in a situation in which the
set of matching variables includes some, but not all, confounders, 1natch­
ing may produce better causal inferences than regression models because
cases that match on a set of explicitly selected variables are also n1ore
likely to be similar on nnmeasured confounders.77
Unfortunately, the re latively simple matching procedure just described,
kno\.vn as exact matching, is often impossible. This procedure typically
fails for continuous variables such as wealth, age, and distance, since there
may be no two cases with the sau1e score on a continnous variable. For
example, there is no unden1ocratic country ,vith the exact same per capita
GDP as the United States. Note that the larger the number of covariates,
the lower the likelihood of finding exact matches.
In situatious where exact n1atching is infeasible, researchers may
instead en1ploy approximate matching, ,vhere cases from the control
group that are close enough to 1natching cases from the treatment group
are accepted as matches. Major weaknesses of this approach include the
fact that the definition of "close enough" is inevitably arbitrary, as well
as the fact that, for large sets of matching variables, few treatment cases
;ire likely to have even approximate matches.
"fo deal with situations in which exact matching is impossible, method­
ologists have developed an alternative procedure known as propensity­
S(nre 1natching. This approach suggests a somewhat different definition of
'ii1nilarity than the previous two. Rather than focusing on sharing scores
0 1 1 the n1atching variables, propensity-score matching focuses on shariug
., si1nilar estimated probability of having been in the treat1nent group,
· 1 11 >1\Tl'(T, 11 1 . 1 1 , li11 1 1·. 1·. , l,·.11 h· i11kritH' t 1 ) :1 1vcll-dcsig11ed and well-executed randomized
('\)ll't i11w111. ! II<' lw1u 1 1 1 ·. .,j 11 1 .11< 1 1 1 1 1 1•. t'\(1·1H I only so for ns equivalence on the varic1blcs
1 '1'1" 1 1 11' 1 1 n l1ul,-,I . 1 1 1 , l . 1 1 1 1 111111,, . ,1·-111,·d 1·.1r1.1hk� 1h.11 i'nrtuitously h;1ppen to he sitnil:-tr
.IL I "',', 1 1 11 ' , . 1 ·., ·. l',1 • , , i , 1 1 • ,I 1 " " 1 " I 1 .111,l,111111.1 1 1 1 111 l1.11 1dl, . . ,.tll \ l l l l l l(",l\llT"('( I l':Hi:1hll·�.

136
JI. Doing Case Studies
conditional on the marching va riables. In other words, wh en lo oki ng for
a match for a specific case in t he treatment group, researchers look for
cases in rhe control group that - b e fore the sco re on the indep endent vari ­
able is know n - would have been a s likely to be in the treatm ent group
as actually chosen cases. This is accomplished by a two-stag e analysis,
the first sta ge of which a pp roaches tb e key independent v ari a ble, X1 , as a
dependent va riable and the 111atching variables as i ndependen t vari a bles.
(This is similar i n spirit to selection models, wh ere a tv.,ro-sta ge approach
to causal infe ren c e is adopted.) Once this n1ode l has been cstirnated, the
co efficient e stin1ates a re disregarded. lnstea d, the secon d stag e o f the anal­
ysis employs the fitted values for each c a se, which te ll us the probability of
that c ase be ing assigned to the treatn1ent group, conditio nal on its sco res
on the matching var iables. These fitted values are ref erred to as propen­
sity score s. The final step i n the p rocess is to cho os e matche s for each cas e
in the treatn1ent group. This is accomplishe d by se lecting case s fron1 the
co ntr o l group with simila r propensity sco res.
The end re sult of thi s procedure is a se t of n1atched c a ses that c an
be compa re<l in whatev er way the rese a rcher deems app rop riate. These
are the '" n1ost-sin1ilar" cas es, return ing to the qualita tive t e nninology.
Ro senba um and Silber sumn1 ar ize the re sults of re c en t medic a l studies:
Unlike model-based adjustrnents, whe re patients vanish and a rc rep laced by the
coe fficients of a n1ode l, in matching, ostensibly comparable patte rns are compared
dire crly, one by one. Modern n1atching n1ethods invo l ve s� atistical �odel ing and
coinbinatorial al gorithJns, but the end resu lt is a coll ection of p a irs or sets ? f
peop l e ,...vho look con1parable, at least on average. ln rna�ching,_ peopl e retain
th eir integrity as peop le, so they can be exan1ined and their stories can be told
individually. 78
Matchi ng, conclude the a uthors, "facilita tes, rather than inhibits, thick
descriptio n . " 79
Indeed, the sa1nc 1natching technique s that ha ve be en used successfully
i n observational studies of medica l treatme nrs might also be adapted to the
study of n a tion-state s, politica l p artie s, citie s, or indeed any paired c as� s
. histo r ical experienc e).
i n the social science s. Suppose that, in orde r to study the relanonsh1p
betwe en wea lth an d democracy, the re searcher wishes to select a c�1st·
that is a s similar a s possible to Costa Ric a in background v ar iables, while
being a s different as possible on pe r capita GDP, the variable of thcoretic;1 l
interest, and the outco me of in te re st, de n1ocracy.
78 Rosi:nhaurn :1nd Sil her (200 1 : 22..\).
';'') !hid.
137
Techniques for Choosing Cases
In order to select rri�st-similar case s for the study o f the relationship
betwe en wealth and dernocracy, one 1nust arrive at a statistical model of
the c auses of a country's wealth. Obviously, such a propo sition is com­
plex. Since this is an illustrative exarnple, w e shall be satisfied with a
cartoon 1nodel that i ncludes only a fe w independent vari a bles. A coun­
try's wealth will be assum ed to be a function of the origin of its lega l
system (measured by dummy variable s for English legal heritage, French
leg a l heritage, socialist legal heritage, German legal heritage, and Scandi­
navian lega l heritage ) and its geographic endowments (m easured by th e
dista nc e of each country's capital city frorn the equato r ).
'fhe first step in selecting most-sin1ilar case s is to run a nonparam etric
re gression with these inde pendent variables and logged per capita GDP
(th e i ndependent v ari able of theoretic al inte re st) a s th e d ependen t vari­
abl e. The fitted value s fron1 this regressio n serve as p ropen sity scores,
and cas es \Vith si1nilar propensity scores are i n terpreted as 111atching. The
propensity score for our focus c a se, Costa Ric a, is 7.63. Exan1 i n in g the
propensity-score da ta , one see s th a t Beni n ha s a prope nsity sco re o f 7.58 -
quite si1nilar to Costa Rica 's. At the same time, Ben i n's per capita GDP
o f $1,163 is substan ti a lly different from C osta Rica's per capita GDP of
$5,486, as are their democracy score s in 1995 (Benin is much less demo­
cr a tic than Co sta Rica ). Hence, Costa Ric a an<l Ben i n may be viewed as
1nost-similar case s for testing the rel ationship hctwcen vvea lth and democ­
racy, as illustrated in Ta hle 5.5. An in -<lepth ana lysis of th e se two cases
n1ay shed light on the causal pathw a ys between eco nomic <leve lopm en t
and de mocracy. Indeed, the se t\VO c ase s arc probably more infonnative
than other two -case con1parisons precisely because the case-se lection pro­
c edure has identified countries whose other a ttributes a rc roughly equa l
i n their propensity to democracy/authoritaria nism. This means th at the
differenc es on the variable of theoretical interest (GDP per c a pita) and the
outcon1e (democ racy) can be given a caus al i nterpretation - an interprc­
t.ition that would probably not h e sugge sted by a qualitarive a ssessment
of th e se two countrie s (,.vhich are quite different i n culture, region, and
It is i n1po rtant to keep in mind that the quality of the "m a tch"
depends en ti rely on the quality of the statistic al model us ed to generate
t lil' propensity sco res. A superficial n1ode l like the one us ed here may
produce rather supt"rlicial 111:itch e s. Yet, i n a large-N context - ,.vhere
d 1 i1.v11s, i f not 1ho11s.11HI..., ( ) f casL'S vie fo r inclusion - a forn1al approa ch
In l·;1'>t ' s(·h·t 1 i c i 1 1 1 d l 1· 1 ·. ·.ii•.11dil·.1n1 ,Hlv;uH;1gcs. At the very l east, one's
.1·,·,11111p111 111·, .11 1 · 1i·11 1 k 1 1·, ! ! 1 .111·, 11.11c·111.
13 8 II. Doing Case Studies
TABLE s.s. Paired cases resulting front matching procedure
GDP PropensitJi
per capita score Denlocracy
Benin
Cases
Costa Rica
Conclusion
The most-similar method is one of the oldest recognized techniques of
qualitative analysis, harking back to J. S. Mill's classic study, System of
Logic (first published in 1834). By contrast, matching statistics are a rela­
tively new technique in the arsenal of the social sciences, and have rarely
been employed for the purpose of selecting cases for in-depth analysis.
Yet, as suggested in the foregoing discussion, there n1ay be a fruitful inter­
change between the two approaches. Indeed, the current popularity of
n1atching among statisticians - relative, that is, to garden-variety regres­
sion models - rests upon what qualitative researchers would recognize as
a "case-based" approach to causal analysis. If Rosenbaum and Silber are
correct, it may be perfectly reasonable to appropriate this large-N method
of analysis for case study purposes.
To be sure, the purpose of a case study is son1ewhat different in situa­
tions where a large-N cross-case analysis has already been conducted.
Here, the general causal relationship is usually clear. We know from
our cross-case study that GDP per capita is strongly associated with
democracy; there is a strong presu1nption of causality. Of course, the case
study analysis may give us reasons to doubt. Perhaps the causal path­
ways from economic developn1ent to regime type are difficult to identify.
Perhaps the presumed causal pathways, as identified by previous research
or theoretical hunch, are si1nply not in evidence. Even so, the usual pur­
pose of a case study analysis in this setting is to corroborate an initi,11
cross-case finding.
By contrast, if there is no prior cross-case investigation - at least llOlll'
of a formal nature - the case study performs a somewhat different roll'.
Herc, we will be more interested in the covariational patterns th�1t ;1rc
discovered between X 1 and Y. Thus, Epstein's study of An1eric�111 :111d
Canadian political parties is notable for its princip:d l--i11di11g: t h:H tlit"
underlying c.iusc of p,1rt y cohesion is to he found i11 IIH· <,lructurc ol
Techniques for Choosing Cases 139
the executive (parliamentary/presidential). Indeed, Epstein spends rela­
tively little ti1ne in this article discussing possible causal mechanisms;
his principal focus is on "scoring" the relevant variables, as discussed.
By the same token, if Epstein had already conducted a large-N cross­
case analysis prior to his case study, and if this cross�case analysis had
r�vealed a strong pattern between executive type and party cohesion,
hrs two-case analysis of the United States and Canada (cases that we
presume would have very sirnilar propensity scores) would now serve a
rather different purpose. Evidently, the function of the most-similar case
study shifts subtly but in1portantly when the case-selection procedure is,
itself, a mode of analysis, offering strong prima facie evidence of a causal
relationship.
As with other methods of case selection, the most-similar method is
prone to problems of non-representativeness. If this technique is employed
in a qualitative fashion (without a syste1natic cross-case selection strat­
egy), potential biases in the chosen cases must be addressed in a spec­
ulative way. If the researcher employs a matching technique of case
selection within a large-N sample, the problem of potential bias can be
addressed by assuring a choice of cases that are not extreme outliers, as
j udged by their residuals in the full model. Most-similar cases should also
be "typical" cases, though some scope for deviance around the regres­
sion line may be acceptable for purposes of finding a good fit among
cases.
Most-Different Cases
A final case-selection method is the reverse image of the previous method.
1 --Iere, variation on independent variables is prized, while variation on the
outcome is eschewed. Rather than looking for cases that are most-similar
(>Ile looks for cases that are most-different. Specifically, the researcher trie;
to identify cases where just one independent variable (X 1 ) , as well as the
dependent variable (Y), covary, while all other plausible factors (X20_ a)
:-.how different values. 80
Hii Tl1t· n1osr-differcnt n1cthocl is sometitnes referred to as the " inethuJ of agreernent," fol­
luwin� i(s inventor, J. S. Mill (1 843/1872). See also Defelice (1986); Gerring (2001:
-' I / I •l); l . 1jph:111 ( I < !7 1 , 1 97S); Meckstroth (1975); Przeworski and Teune (1970);
.111d 'il,(1q1ol , 1 1 1 d \<111H·1·. ( l < !:-lO). h>r t'X:1111plC's of rhis 1nethod, see Collier and Col-
11!'1" ( I 'I'! l/.1 1 H 1 .> 1 ; I "111T1·,1· . 1 1 1 . I I l11pn1x ( ] ')(,2); K:HI ( 1 997); Moore (1966); Skocpol
( l 'l '' 1 ) : . 11111 \ , 1 • . 1 1 . 11 I ' ! I l l ', ' I I I f < , 1 \ , · 1 r 1 , 1 1 111,,I , i i 1 h ( ·�(' \ t 1 1 d i t '\ :irt· desL· r i lwd as co1n/Ji11-
. .
111 1: 1 1 1 1 1 · . 1 ·. 1 1 1 1 1 1 . 1 1 1 1 1 , I "'" I ,loll, 1 , I l l I I H 1J i . ., I , ,
140 II. Doing Case Studies
TABLE 5.6. Most-different analysis with tiuo cases
X1 = the variable of theoretical interest. X2a -d = a vector
of controls. Y = the outcome of interest.
The sin1plcst form of this two-case comparison is illustrated in Table
5.6. Cases A and B arc deen1ed "most-different," though they are similar
in two essential respects - the causal variable of interest and the outcome.
As an example, I follow Marc Howard's recent work, which explores
the enduring impact of co1nmunism on civil society. 8 1 Cross-national sur­
veys show a strong correlation between forn1er co111munist regi1nes and
low social capital, controlling for a variety of possible confounders. It is a
strong result. Howard wonders \vhy this relationship is so strong and why
it persists, and perhaps even strengthens, in countries that are no longer
socialist or authoritarian. In order to answer this question, he focuses on
two most-different cases, Russia and East Gern1any. These t,vo countries
were quite different - in all ways other than their communist experience -
prior to the Soviet era, during the Soviet era, and in the post-Soviet era,
as East Germany was absorbed into West Gern1any. Yet they both score
near the bottom of various cross-national indices intended to measure the
prevalence of civic engagen1ent in the current era. Thus, Howard's case
selection procedure meets the requirements of tbe most-different research
design: variance is found on all (or most) dimensions aside from the key
factor of interest (comn1unisn1) and the outcome (civic engagement). 82
What leverage is brought to the analysis by this approach? Howard\
case studies combine evidence drawn fron1 mass surveys and from in
depth interviews of s1nall, stratified samples of Russians and East (;l'r
mans. (This is a good illustration, incidentally, of how quantitative and
qualitative evidence can be fruitfully combined in the intensive study o!
several cases.) The product of this analysis is the identification of thn·('
causal pathways that, Howard clai1ns, help to explain the l::1ggard st:11 11'>
of civil society in post-co1nmunist polities: "the n1istrust of co111n11111i"1
81 I-lowarJ (2003). In the following discu�si011 ! rrv;1I tlw !t' rlll\ "srn." i;d
society," and "civic L'll!..!,;1 ).!,L'lllt'lll" i11tlTd1<11ti•,t·.1 hly.
82 1 [! l \V;;rd (2(lll \: (, '!),
Techniques for Choosing Cases 141
organizations, the persistence of friendship networks, and the disappoint-
men! w1' th post-con1mun1sn1. .
" 8 3 s·imp1y put, Howard concludes, "a great
number of citizens in Russia and Eastern Germany feel a strong and lin­
gering sense of distrust of any kind of public organization, a general sat­
isfaction with their own personal networks (accompanied by a sense of
deteriorating relations within society overall), and disappoint1nent in the
developments of post-coinmunism." 8 4
Results obtained from the analysis of East Germany and Russia are pre­
su1ned to apply in other post-communist polities (e.g., Lithuania, Poland,
Bulgaria, Albania). Indeed, by choosing a heterogenous sample, Howard
solves potential problems of representativeness in his restricted sample.
However, this sample is not representative across the entire population of
the inference, which is intended to cover all countries, not just comn1u­
nist ones. (To argue that communis1n impedes the developn1ent of civil
society is to imply that noncommunistn stimulates the development of
civil society. The chosen sample is trnncated !censored] on the dependent
variable).
Equally problematic is the lack of variation on key causal factors of
interest - comn1unism and its putative causal pathways. For this reason, it
is generally difficult to reach conclusions about the cansal status of these
factors on the basis of the 1nost-different analysis alone. It is possible,
that is, that the three causal pathways identified by Howard also operate
within polities that have never experienced communist rule. If so, they
are not properly regarded as causal.
Nor docs it see1n possible to conclnsively eli1ninatc rival hypotheses
on the basis of this most-different analysis. Indeed, this is not Howard's
intention. He wishes 1nerely to show that whatever influence on civil soci­
l'ty might be attributed to economic, cultural, and other factors does not
exhaust this subject.
My considered judgment, based on the foregoing methodological
dil en1111as, is that the 1nost-different research design provides only minimal
insight into the proble1n of why com1nunist systems appear to suppress
ri vie engagement, years after their disappearance. Fortunately, this is not
tl il' only research design employed by Howard in his admirable study.
l ndL"ed, rhe author employs two other small-N cross-case methods, as
\\!l·ll :ls a h1rgc-N cross-country statistical analysis. In n1y opinion, these
l l l l ' t l i od s do 1110-;1 c i i' 1 hc :111:1lytic work. East (;ermany may be regarded
t·: q 1it.1I," ", 11·11
'1 1 ll,1.I., I . ' '
'' 1111. I ' 1 1 ·,
1
142 II. Doing Case Studies
as a causal-pathway case (as discussed earlier). It has all the attribmes
nor1nallv assumed to foster civic engagemenr (e.g., a gro\ving economy,
multipa�ry competition 1 civil liberties, a free press, close association with
\Vestern European culture a nd politics), but nonetheless shows little or no
improvement on this dimension during the post-transition era. 85 It is plilu­
sible to attribute this lack of change to its com1nunist past, as Howard
does. 'fhe contrast bet\veen East and West Germany provides a rnost­
sin1ilar analysis, since the two polities share virtually everything except
a communist past. This variation is also deftly exploited b y Howard. In
short, Ho\\'ard;s conclusions are justitiable, but not on the basis of rnost­
dilferent analysis,
I do not wish to dismiss rhe most-differem research method entirely.
Surely, Hov.rard's findings are stronger with the i ntensive analysis of Russia
than they would be without. Yet if one strips away the pathway case (East
Gern1a:1y) and the most-similar analysis (East/West Ger1nany), there is
little leli upon which to base an analysis of causal relations (aside from
the large-N cross-national analysis). Indeed� most scholars \\ ho e111ploy 1
the most-different merhod do so in conjunction with other methods. 86 It
ls rarely� if ever, a stand-alone rnethod. 87
Conclusion
Generalizing from this discussion of Marc Howard's work, I offer the fol·
l o\.\ring surnmary remarks on the rnost-diffe'rent n1ethod of case analysis.
(I leave aside issues faced by all case study analyses, issues that formed
the basis oi Chapter Three.)
85 Ibid., &.
};6 See, e.g., CoUier and Collier ( i991/2002); Karl (1997}; 11oore (1966); SAocpol ( 197 q::
and Ynshar (2005: 23\. Karl ( 1 997), which affect,,: to be a most-diffe:-ent systctn ana; y �r.
{20), is a r;rticu\arl; dc;.sr ex.::n1ple of tt1s. Her study, focused ostensibl y or. pc11 :,
:,:ates forates with large oil reserves), m::lkes two sorts of inferences, The Ersr co1lLTn1',
the t:.1.suall y) obstructive role of oil in political and econon1ic development. T!;e SCl"r"I
sort of inference con.:etns variation within the population of 9etro-s.ratei>, s,hov1i ng th.11
sorr.e count:ies (e.g, Norway1 Indo'1esia} p-,anage to avoid the pattlulogies hrm1gh1 • "'
1
elsewhere by oil resourct"s. Wheu attf'nlpting to explain the constnunlng rok o! ,)ii .. 11
petro-states, Karl u;uaHy relies on co1"...trasts between petro-stares and nnn·p('fl'0·�1.11,-,,
(e.g., Chapter 10). Onl y when attempting :o explain ciifferences antrn�g pctro-s:,ne'> d( I ,
she re�trict her s,1mple to p:::tro-states, In 1Ic.y opinion, very litr!e U'>e i,, ma,.k ,:f th,· 111,, ,I
different research cesign.
87 This \vas re-.'.ognized, at ie.1st implii.:iti y, by l\1ill ( I S4:1!1 872: 2 58--·'>l. SbT! ,ci<-:rn l!.r. i" ' n
f'choed by n1eth0Jo;,,gists ir, th,: int<:rveurn� yeirs (e_g., ( :(Jf1c11 ,111d '.'..1gv'. I 1 t t·1: . I , I "
c;crri:;g lO(J I : S!:s11tpnl J!ld '.)o0\\'1� I 1Jl'OL !11dcnl. ('\.r•lil'il dd,·11-.:·� pj :lw ,ldl, ' '"
'!!(l',l
l'h'1h:)d :tri· r:u,· \ht11 ,,,.,, l kL-l1u· l 'J81il.
Techniques for Choosing Cases 143
Let us begin with a methodological obstacle that is faced bv both Mil­
lean styJes of analysts - the necessity of dichoton1izlng every.variable in
the analysis. RccalJ that, as 1,vlth most-simiiar analysis, differences across
cases nJ1.1sr be sizeable enough to be interpretable in an essentially dichoto­
n1ous fashion (e.g., high/low, presentiabsent), and si1nilarities must be
close enough to be understood as essentially identical (e.g., high/high,
present/present). Otherwise the results of a Millean-style analvsis are not
interpretable. The problem of "degrees" is deadl y if the vari�bles under
consideration are by nature continuous (e.g.) GDP). This is a particu­
lar concern i n Hov1ard's analysis, where East Germa:r:y scores somewhat
higher than Russia in civtc engagement; they are both low, though Russia
is considerably lower. Howard assun1es that this divergence is minin1al
enough :o he undersrood as a difference of degree rather than of kind,
a judgment that 1nighc be questioned. In these respects, most-different
analysis is n o n1ore secure - hut also no less - than :nost-similar analysis.
In one respect, most-different analysis is superior to mos:�similar anal­
ysis. If the cocUng assumptions are sound, the n1ost··different research
desjgn 111ay be useful for eliminating necess,;ry causes. Causal factors that
do not appear across the chosen cases e.g., X2a d in Table 5.6 are evi­
dently unnecessary fo: the production of Y. lfowever, it does not follo\v
that the most-different method is the best method for eliminating neces­
sary causes. Note that the defining feature of this method is the shared
1.;Jen1ent across cases -- X i in Tnble 5.6. This feature does not help one to
l'li minate necessary causes. Indeed, if one were focused solely on eli1ninat­
ing necessary causesi one would presumably seek out cases chat register
the same outcomes and have n1axirnum diversitv on other attributes. In
Htble 5.6, this would be a set of cases that satisfy conditions X2a-:h but
not X 1 . Thus, even the presumed strength of the most-different analysis
is not so strong.
Usually, case study analysis is focused on the ide�titication (or darifi­
ration) of causal relations, not on the t�limination of possible causes. In
this setting, the most-different technique is useful� but only lf assu111p-
1inns of ""causal uniqueness" hold. By this l mean a situation in \vhich a
1-�1\'t'll outcome i s the product of only one cause: Y cannot occur except in
th1· presence of X 1 , X i is necessary, and in some situations (given certain
! 1 .1tkground conditions) sufficient) to cause y_s s
< :011!-.idtT tht· follovving hypothetical exa111ple. Suppose that a new dis-
1',l',t\ ,1hotJt \Vhi,:h little is k 11P\\.' ll, has appeared i n Country A. There are
1 44 II. Doing Case Sti,dies
hundreds: of infected persons across dozens of affected communities in
that country. In Country B, '.oca:ed at the other end of the world, several
ne\v cases of the disease surface i n :t single co1nln unity. Jn this setting,
vve can imagine two sorts oi Mil lean analyses. The first exan1incs two
similar con1111unit:es within Country A, one of which has developed the
disease aud the other of which has not. This is the most-similar style of
case comparison and focuses accordi:1gly on the i dentification of a dif­
1
ference Setween the two cases that might account for variation across the
sample. l\ second approach focuses on (highly dissi1nilJr) comrnunitJes
v:here the disease has appeared across rhe t\vo counrries and searches for
any similarities that rnight accoun:: for rhese similar outco1nes. ·rbis is the
most-different research design.
Both are plausible approaches to tbis particular problen1> and we can
i 1nagine cpide1niologists employing thern sin1ultaneously. Ho\vevcr, the
most-different desi gn demands stronger assumptions about rhe uuderly­
ing factors at \vork. It supposes that the disease arises fron1 the sarne cause
i n any T>etting. This n1ay be a reasonable operating ass-�mption when one
1s dealiug \vith certain natural p bc:1on1ena like diseases. Even so, there are
111any exceptions. Death} for example, has 1r:.any causes. for rhis reason ,
it wnuld not occur to us to Jook for n1ost-different cases of high n1ortality
arnund the world. [.1 order for the most-different research design to effec­
rively identify a causa1 factor at vvork ir. a given outcome, the researcher
1nust assu1ne that XJ - the factor held constant across rhe diverse cases -
is the only possibie cause of Y {see 'fable 5.6). This assumption rare-! y
holJs in social scientific sertings, for n1ost outco1nes of interest to antbro
pologisrs, econon1ists� political scientists, and sociologi&ts have multipi1
cause!',. There are many ways to ;,vir- an election, to build a ·•,veifare state,
to get in;:o a \var, to overthrow a government, ur - returning to M,Jr(·
Ho\vard's \.vork to build a s::rong civil society. l\nd it is for this reason
that n1ost-different analysis ls rarely applied in social science \vork a1h l ,
where applied, ts rarely convincing.
If this seems a tad severe} there is a u1ore charitable \VJ)' of approachinl'.
the most-differeut method. Arguably, this is not e1 pure "method" at all lm1
1r�ereiy a supplement, a way of incorporating diversity in the subsamp l e , d
cases that _provide the unusual outcome of interest, If the unusual outco1nc
is r,;;volution, one 111ight \vish ro encompass a v.ride va:·iety of revolutin1L
in one's analysis. If the unusual outcotne i s post-corrununi�t civil ..;o('.1·1 \ .
it seems appropriate to include a diverse set of post-cnn1rill ll:1st pol 11 ic.'-< 111
one's sample of case studies, :ls '.Vt1 rc I {nw;1rd doe-.. Fro111 this pt·r..;pvl·1 11. ('.
rhc 1 1 1 ost-diffcrvn1 nH·thod ( :-.n--c,11lcd) :11ighi ht· hc1 l!·1 Lil,c!t·,t .1 dit·1·1·.,·
(,f,\i' 1 1 1\'dHid, ,1, c , plnn,d 1\1 ri!cr.
Techniques for Choosing Cases 14 5
Conclusion
In order to be a case of son1ething broader than itself, the chosen case 1nust
�)e_ representative (in so1T1c respects) of a larger popula:ion. Otherwise - if
lt 1S purely i dios_rncratic ("nnique") it i.s 11nfnforn1ative about anything
.
other than t1selt. A study based on a uonrepresentative sample has no
�o: very Ht�le) exrer�al validity. To be sure, no phenomcuon is purely
1 dtosy11crat1c� rhe nonou of a uuique case is a n1.1tter that \Voul<l be difficult
to defiuc. One is concerned, as ahvays, with n1atters of degree. Cases are
tnore or less representative of some broader phenomenon aud. on that
�core, rnay be c�nsidercd better or worse subjects for intensive ,analysis,
\The one exception, as nored i is the int1uential case.;
_
Of all the problems besetting case study analysis, perhaps the n1osr per­
.
siste�� and the most persistently bemoaned - is the problem of sample
btas. Lisa 1vlart1n finds !hat the overemphasis of iuternational relations
scholars on a few \vel f-known cases of economic sanctions- 111ost of which
failed to elicit any change in the sa nctioned country - "has distorted an,1-
lys::s' vie\v o f the dynamics and characteristics of econon1ic sanctions. "90
llarbara Geddes charges that many analyses of industrial policy have
_
tocused exclusively on the n1ost successful cases - prin1arity the East Asian
NICs - leading to biased inferences. 91 Anna Bteman and Carolyn Shelton
,..how :bat ca �e study work on the question of structural adjust111ent is sys�
tc1nat1cally biased insofar as researchers tend to focus on disaster cases -
those where structural adjust1nent is associ.ared ,vith very poor health and
hu °: an developtncnt outco111es. These cases, often located in snb-Saharau
_
/\ tnca, are by no rneans representative of t he entire population. Conse­
quently, scholarship on the question of structural adjustn1ent is bighly-
-.,kewed in a particular ideological <l.ir;;ction (against neoliberalism). si2
, , Ii i , I i i ll\ 111,I ',: , 1, . . , . , 1 ,1 1 1 ,
146
These examples might be multiplied many times. Indeed, for many top­
ics the n1ost-studied cases are acknowledged to be less than representa­
tive. It is worth reflecting upon the fact that our knowledge of the world
is heavily colored by a few "big" (populous, rich, powerful) countries,
and that a good portion of the disciplines of economics, political science,
and sociology are built upon scholars' familiarity with the economics,
political science, and sociology of one country, the United States.93 Case
study work is particularly prone to proble1ns of investigator bias because
so 111uch rides on the researcher's selection of one case (or a few cases).
Even if the investigator is unbiased, her san1ple 1nay still be biased sin1ply
by virtue of "rando1n" error (which n1ay be understood as 1neasurement
error, error in the data-generation process, or an underlying causal feature
of the universe).
There are only two situations in which a case study researcher need
not be concerned with the representativeness of her chosen case. The first
is the influential-case research design, \Vhere a case is chosen because of
its possible influence on a cross-case model, and hence is not expected
to be representative of a larger sample. The second is the deviant-case
111ethod, where the chosen case is employed to confirn1 a broader cross­
case argument to which the case stands as an apparent exception. Yet in
tbe latter instance, the c�osen case is expected to be representative of a
broader set of cases - those, in particular, that are poorly explained by
the extant model.
In all other circutnstances, cases mnst be representative of the popula­
tion of interest in whatever ways 1night be relevant to the proposition
in question. Note tbat where a researcher is attempting to disconfir111 :1
deterministic proposition, the question of representativeness is perhaps
more appropriately understood as a question of classification: is the
chosen case appropriately classified as a member of the designated
population? If so, then it is fodder for a disconfirming case study.
If the researcher is attempting to confirm a deterministic proposition,
or to 1nake probabilistic argurneuts about a causal relationship, then tht·
problem of representativeness is of the more usual sort: is Case A u 1 1 i 1
homogenous relative to other cases in the population? This is not : 1 1 1
easy matter to test. However, in a large-N context the residual for th.ti
93 Wahlke (1979: 1 3 ) ,vrites of the failings of the "behavioralist" mode of politiL·;tl \l l r 1 1 , ,
analysis. "lt rarely .1.in1s at gcncraliz;Hion; rcsc,nch effort:-. h:1vt· lwt·n vnnfi11( ' d \'\\1· 1 1 1 1 . i l l 1
t o c1se studies of single poliric:11 systt·rns, 1111 1:-,t < 1 1' tlw1n dt·.1li1 1)', w i 1 h 1 h<' 1\1111·11, . 1 1 ,
sr:-.tt·111."
JI. Doing Case Studies Techniques for Choosing Cases 147
case (in whatever n1odel the researcher has greatest confidence in) is a
reasonable place to start. Of course, this test is only as good as the 1nodel
at hand. Any incorrect specifications or incorrect 1nodeling procedures
will likely bias the results and give an incorrect assessment of each case's
" typicality." In addition, there is the possibility of stochastic error, errors
that cannot be modeled in a general frarneworlc Given the explanatory
weight that individual cases are asked to bear in a case study analysis, it
is wise to consider more than just the residual test of representativeness.
Deductive logic and an in-depth knowledge of the case in question are
often more reliable tools than the results of a rather superficial cross-case
model.
In any case, there is no dispensing with the question. Case studies (with
the two exceptions already noted) rest upon an assumed synecdoche: the
case should stand for a population. If this is not true, or if there is reason
to doubt this assumption, then the utility of the case study is brought
severely into question.
Fortunately, there is some safety in numbers. Insofar as case study evi­
dence is combined with cross-case evidence, the issue of sample bias is
tnitigated. Indeed, the skepticism about case study work that one com-
1nonly encounters in the social sciences today is, in my viev.,, a product
of a too-literal interpretation of the case study method. A case study tout
court is thought to mean a case study tout seul. Insofar as case studies
and cross-case studies can be enlisted within the san1e investigation (either
in the san1e study or by reference to other studies of the same subject),
problems of representativeness are less worrison1e. This is the virtue of
cross-level work, a.k.a. "triangulation."
Ambiguities
Before concluding, I wish to draw attention to two ambiguities in case­
sclection strategics for case study research. The first coucerns the admix­
! un_· of several case-selection strategies. The second concerns the changing
...iatus of a case as a study proceeds.
So111e case studies follow only one strategy of case selection. They
,ire typical, diverse, extreme, deviant, influential, crucial, pathway, most­
\1111ilur, or most-different research designs, as discussed. However, many
l ,\:-.(' studies n1ix ,1nd ,natch among these case-selection strategies. Indeed,
111..,of :tr :1s :111 case st·ndics seek representative samples, they are all in search
I d " 1 ypic;1]" L·:1 scs . · l'lll 1 ", it is L·o111111on for writers to declare that their case
1·,, 1 1 11' 1•x.1 1 1 q 1lc, 1 10! 1 1 ( ' \ l rt · 1 1 11 · . 1 1 1 d 1 ypic;tl; it has ;111 cxtrc111e value on X 1
1 1 1 Y h111 I', 1 H 1 1 , 1 1 1 1 , 1 1 1 1 · 1 t 1 ", 1 w, 1 ·. , 1 d i o :-. r 1 1 r r: 1 l i L· . 'l'h tTc is not 1n11(.J1 that
148 II. Doing Case Studies Techniques for Choosing Cases 1 49

one can say about these combinations of strategies except that, where This is unfortunate, but inevitable. One cannot con struct the perfect
the ca s es allow for a variety of e1npirical strategies , there is no reason research design until (a) one has a specific hypothesis and (b) one is rea­
not to purs ue them. And where the sa1ne case legitin1ately serves several s onably certa in about what one is going to find "out there" i n th e empir­
functions at once (without further effort on the res earcher's par t), there ical world. This is parti cularly true of observational rese arch designs, but
is l i ttle cost to a multipronged approach to ca se analysis .
it als o applies to ma n y experi1nental research desi gn s : usually, there is
The second issue that deserves emphasis is the changing status of a case a "good" ( i nformati ve) finding, and a finding that i s less insightful. In
during the cours e of a researcher's investi gation-v1rhich may last for y ears, short, the perfect case study research design i s usually apparent only ex
if not decades. The problem is particularly acute when a researcher begi n s post facto.
in an exploratory 1node and then proceeds to hypothesis testing (that is, There are three \Vays to handle this. One can explai n, straightfor\vardly,
she develops a specific X1 fY proposition), or whe n the operative hypoth­
that the initial research \Vas undertaken in an exploratory fa shion, and
esis or key control variable changes (a ne\v causal factor is discovered or therefore not constructed to test the specific hypothesis that is - now - the
another outcome becomes the focus of analysis). Things change. And it is primary argument. Alternatively, one can try to red esign the study after
the n1ark of a good researcher to keep her mind open to new evidence and the new (or revised) hypothesis has been formulated. This may requi re
new insights . Too often, methodological discussions give the mi sleading additional field research or perhaps the integrati on of additional ca s es or
i mpression that hypotheses are clear and remain fixed over the course
variables that can be obtained through secondary sources or consultation
of a study's development. Nothing could be further from the truth. The of experts. A final approach is to simply jettison, or deen1phasi ze, that
unofficial transcripts of academia - accessible in inforn1al settings, where portion of the res ea rch that no longer addresses the (revi s ed) key hypoth­
researchers let their guards down (particularly if inebri ated) - are filled esis. A three-ca s e study 1nay become a two-case study, and s o forth. Lost
with stories about dead ends, unexpected findings, and dra st ically revised ti me and effort ar e the costs of this downsi zi ng.
theory chapters . It would be interesting, i n th is vei n, to con1pare published In the event, practical considerations will probably determine which
work w ith dissertation prospectuses and fellowship applications. I doubt of these three strategi es, or con1biuations of strategi es, is to b e followed.
that the corre lati on between thes e two stages of research is particularly The point to ren1en1ber is that revisi on of one's cross -case rese arch desi gn
s trong.
is normal and to be expected. Not all t\vis ts and turns on the me anderi ng
Research, after all, is about discovery, not si mply the veri ficati on or fal- trail of truth can be anticipated.
sification of existing hypotheses. That s aid, it is also true that research on
a pa rticular topic should move from hypothesis ge nerating to hypothesis
Are There Other Methods of Case Selection?
testing. This marks the progress of a field, and of a scholar's own ,vork.
!\ t the ontsct of thi s chapter, I s ummarized the task of case selecti on a s
As a rule, res earch that beg ins \v i th an open-ended (X- or Y-centered)
a n1atter of achievi ng t\VO objectives: representati veness (typicality) and
analys is should conclude witb a determ inate X1 /Y hypothes is.
v,1riation (causal leverage). Evidently, there are oth er Objectives a s well.
1-he problen1 is that research strategies that are ideal for exploratio11
l;or example, one \Vis hes to identify ca ses that are independent of each
are not al\vays ideal for confirmat ion. I dis cussed this trade-off in Chaptc1
od1er. If chos en cases are affected by each other, tbe problem (sometimes
Three a s it pertains to the cross-case/cas e study dilem1na. It also app !it",
k 11own as Galton's problem or a probl em of diffusion) n1ust be corrected
to various methods of case study analysis, as pres ented in this chap l t 'I '.
hcfore analy s is can take place. I have neglected this i ssue because i t i s
The extreme-case method is inherently exploratory, since there is no clc:1 1
us11ally apparent to the researcher and, in any case, there are no easy
caus al hypothes is; the researcher is concerned merely to explore v;1ri;1
l!·L'hniqucs that m i ght be uti l ized to correct for such biases .94
tion on a single dimensi on (X 1 or Y). Other methods can he en1p\oy l ·(I
I h:1vc also di sregarded pragmatic/logistical i ssues that mi ght affect case
i n either an open-ended (exploratory) or a hypothesi s-testing (co11fir1n.1
.,cll'L'tion. Fvidcntly, case selection is often influenced- by a researcher's
tory/di sconfirmatory) 1node. The d i fficulty is th:1t oncl' thl' 1-csl':1rchcr h.1·.
arrived ar J detern1inatc hypothesis, thl' o r i ginally chosl' ll rl's;l':lrl'h l i l ''>lJ', 1 1
l.1,·1.,, \ i 1 1 1 p i 1 w, i 1 1 g ll))!ll) C.lSl' scl,:l'l· ioll, SL' (' (;erring
n1;1y no longer ht' so \Vl'll l·o11sl!'llL' lt'd.
I
' I t > I 1 1 1 1 1 1 1 1 · 1 .11,,. I I ''" , 1 1 i,l 1 1 1 1 ·• . 1 1 1 . I " 1 1 1 , · 1

I ' I II 1 1 I '1i ii i I
 II. Doing Case Studies
150
ry, a personal entrfe into that
fa1nili arity w ith the language of a count
r fundin g that cove rs one arch ive
loc ale , special access to i mportant data, o i te
rather t han a not her. Pragma
tic considerations are often - and qu
elect ion process.
rightly - decisive in t he case-s
al prominence of a partic­
A fina l conside ration concerns the theoretic times
ect. Researche rs are some
ular case wi thin the literature on a subj
attenti on in previ­
have receive d exten sive
oblige d to study cases t hat
es or
to as "p aradigm atic " c as
ou s studies. T hese a re sometimes refer red
" exemplars. " 95 i-
ility nor t h eoret ical p rom
However, ne ither pragmatic/logistica l ut these
i n case selection . That is,
nence qua lifies as a methodological fac to r ming
lidity of the find ngs s
featu res of a case have no b earing o n the v a
i tem

p p heral
gr ant these issu es a
fr om a study. As such, it is appropriate to
eri

in the bo ok.
status in this cha pter, as I have elsewhere s­
is tradition a l to 1nake a di
One fina l caveat mus t be issued. While it se l ok
n and case u ly , l
tinc ti on between the tasks of case selectio
a a sis a c o o

e can­
1n to be nd stinct and overlapp in g. On
at these proc esses revea ls the
i i
sort of a nalysis that it mi ht be g
not choose a case w ithou t considering the
reade r should consider cho osi ng
subj ected to, and vic e versa . Thus, the
d ou t in this ch apter along with
cases by employing t he nine techni ques l ai quasi­
any considerat ions that mi
g ht be introduced hy v irt ue of a case 's
er Six) a nd its potential for
process acing tr
e xperimenta l quali ti es (Chapt
w turn.
(Ch apter Seven), subj ects to which we no

95 Flyvbjerg (2004: 427).