Professional Documents
Culture Documents
doi:10.1093/jopart/muw041
Article
Advance Access publication June 28, 2016
Article
© The Author 2016. Published by Oxford University Press on behalf of the Journal of Public Administration Research and Theory, Inc. 163
All rights reserved. For permissions, please e-mail: journals.permissions@oup.com.
164 Journal of Public Administration Research and Theory, 2017, Vol. 27, No. 1
as interviews, document analysis, and process tracing networks. Then, we specify the exponential random
(Connick and Innes 2003; Margerum 2011; Ulibarri graph models (ERGMs) used to model networks and
2015b) or survey instruments and statistical modeling test our hypotheses. Finally, we assess our findings and
(Berardo 2014; Berardo and Scholz 2010; Schneider discuss their implications.
et al. 2003; Lubell, Henry, and McCoy 2010; Scott
and Thomas 2015; Thomson, Perry, and Miller 2009;
Ulibarri 2015a). The effort required to collect and ana- Background
lyze these data present a significant research design We examine this question using the case of the Federal
barrier. Perhaps more importantly, these modes of Energy Regulatory Commission’s (FERC) process for
data collection do not readily facilitate cross-case com- licensing hydropower facilities. To obtain a license for
parisons, longitudinal analysis, or replication. Survey continued operation of a hydropower dam, electric
instruments or interview scripts are typically custom- utilities, water districts, and other dam owners/opera-
ized for the network context at hand, for instance, using tors (hereafter referred to as “utilities”) undertake a
a specific type of “name generator” (Henry, Lubell, and 5-year process to identify potential project impacts,
the broader structure supported collaboration). These particularly interested in what might be expected in
were then used to measure collaboration in eight each case. In other words, what should a network
recently completed relicensing processes across the around hydropower relicensing look like when col-
country. For more information, see Ulibarri (2015b). laboration is high versus when collaboration is low?
In this analysis, we consider three distinct relicens- The policy network literature informs our expectations
ings, shown in Table 1. All three processes took place regarding what structural characteristics should—or
after FERC’s shift to a more collaborative relicensing should not—be prominent in processes with a given
approach. These three cases were selected because level of collaboration, as different network configu-
they represent three distinct levels of collaboration. In rations distinctly affect the ability of a collaborative
Washington (high collaboration), stakeholders were network to solve the problem it was created to solve
engaged via deliberation and negotiation throughout (Berardo 2014; Bodin and Crona 2009).
the relicensing process, such that both large and small
decisions incorporated the full set of interests in the Network Density
relicensing. In Missouri (medium collaboration), the Network density refers to the average value of a ran-
Note: Overall collaboration is aggregated from principled engagement, shared motivation, and capacity for joint action. See Ulibarri (2015b)
for more information.
166 Journal of Public Administration Research and Theory, 2017, Vol. 27, No. 1
presents an opportunity for communication and joint redundant—and thus inefficient—in terms of informa-
action (Bodin and Crona 2009). However, two net- tion sharing, network structures that are minimally
works can exhibit similar density statistics and yet be redundant can impede performance when credibility is
structured very differently. Networks in which interac- at a premium (Lazer and Friedman 2007). Reciprocity
tions are highly concentrated among a limited subset can enhance the credibility of actors’ commitment to
of actors, with more isolated actors and fewer paths of one another and thus support collaboration (Berardo
communication, can have high overall density due to and Scholz 2010).
these dominant relationships even if there is a lack of
H3:
Higher collaboration processes will exhibit a
overall connectivity. Accordingly, one might anticipate
stronger tendency for reciprocal ties than low-
that networks characterized by a stronger bimodal dis-
collaboration processes.
tribution of high-value ties and empty ties might not
facilitate joint action.
Network Cohesion: Efficiency, Credibility, and
Specifically, overall density (tie presence and magni-
Inclusiveness
tude) and nonzero density (tie presence) will differ to
Although reciprocated ties might be important for fos-
and Heikkila 2012; Karkkainen 2002]) and exclusion search costs are a more significant transaction cost bar-
(facilitating quick action [Imperial 2005; Margerum rier than credibility issues in networks constituted by
2011; Wondolleck and Yaffee 2000], reducing exter- fragmented authority. The case at hand, hydropower
nal decision costs [Feiock 2013], and excluding actors relicensing, exhibits a great deal of fragmented author-
who lack sufficient expertise [Day and Gunton 2003; ity because it involves local, state, and federal actors
Lasker and Weiss 2003] or resources [Wondolleck in different policy subfields of energy, fish and wildlife,
and Yaffee 2003] to contribute meaningfully). These water resources, and land use. In light of such frag-
dueling forces place a premium on collaborating with mentation, we anticipate that network effectiveness
the right people and advantage cases where there are (in terms of whether the relicensing network is able to
fewer “wrong” people involved. successfully collaborate and produce quality outputs)
One way that networks exhibit tendency toward more requires greater overall connectivity and network-wide
or less cohesiveness is through transitivity. Transitivity cohesion (as opposed to more redundant structures
reflects the “a friend of a friend is my friend” axiom, in that maximize credibility and commitment).
that networks with high transitivity are those where if
H4:
High-collaboration processes will exhibit lower
minutes we code provide direct evidence of dialogue does potentially affect our analysis. ERGMs rest on
and deliberation among stakeholders, demonstrating the assumption that all ties are observed; by sampling,
which participants present or discuss key issues. If a we knowingly omit some ties. However, these omitted
given participant gives a presentation or substantively ties are missing randomly (given our sampling proto-
engages in discussion at a given meeting, we code a col) and therefore should represent the full network
tie value of 1 from the discussant to all other meeting we would observe were we to code all ties. Moreover,
attendees.2 If a meeting attendee does not participate the unsampled meetings are not substantively differ-
in discussion, ties from said attendee to other meet- ent from those included in the sample—whether by
ing attendees are given a value of 0. Because of the topic, location, or timing over the course of the reli-
one-way nature of this interaction (one individual is censing—so we have no reason to assume that a sub-
talking to—not necessarily with—another individual) stantially different pattern of attendance or dialogue
these are directed ties that originate at the discussant took place at those meetings. Finally, unsampled meet-
and terminate with audience members. Thus, in each ings—and resultant missing edges—are less problem-
network, a tie from one individual to another is not atic in this cross-sectional analysis since we are simply
network tie influences the presence of others, thus 2015). This parameter reflects the expected value
violating the standard independence assumptions of of a tie between I and j based upon the value of all
most regression models (Robins, Lewis, and Wang ties observed in the network. Network density is
2012). We use ERGMs that explicitly model interde- computed as:
pendence among network ties (Lubell et al. 2012) by
modeling every tie conditionally based upon all other g (y ) = ∑ y , (1)ij
observed ties (Lusher, Koskinen, and Robins 2013). (ij )∈Y
To compare reciprocity and transitivity across the intuition here is that in networks with high transitivity,
three relicensing networks, we fit a third series of mod- the final value for this term should be higher on aver-
els that include terms for mutual ties and transitive tri- age since more triangles tend to be “closed.”
ads. The mutual term (Handcock et al. 2015) models Table 3 summarizes how each term is computed.
the prevalence of reciprocated ties between nodes for
nonbinary network ties:
Results and Analysis
g (y ) = ∑ min (y , y ) (3)
ij ji
Although basic network density can be computed
(ij ) ∈Y
algebraically simply by dividing the number of total
by recording the minimum observed tie value between possible ties by the total summed value of all ties to
each pair of actors. This reflects how the observed tie compute average tie value, as noted above, it is impor-
value from node i to node j influences the expected tant to account for the number of meetings that a
value of a tie from node j to node i. If network reci- given actor attends. We thus estimate the sum param-
procity is high, then this statistic will be higher because eter (described above) for each case within a GERGM
Table 4. Baseline Models with Density Parameter and Control for Attendance
Lastly, we turn to Table 6, which presents the results are based upon meeting attendance and participation,
for network models that include all terms of interest. one would expect that meeting attendance is positively
The high-collaboration case (Washington) exhibits by associated with tie values independent of any under-
far the highest degree of mutuality between network lying network drivers. As expected, each meeting an
participants. As described above, the mutuality param- actor attends significantly increases the expected tie
eter models the extent to which the tie value observed value between that actor and every other network
between nodes A and B (yAB) corresponds to the tie participant, although the magnitude of this increase
value from B to A (yBA). In other words, it models the is small: 5% (exp (0.02) = 1.05) for Washington,
reciprocal tendency of a network. In the Washington 16% (exp (0.15) = 1.16) for Georgia, and 15%
network, for each one-unit increase in yij the predicted (exp (0.14) = 1.15) for Missouri. The magnitude this
value of yji increases by 458% (exp (1.72) = 5.58) . parameter is small in each model partly because it
This might at first seem like an outlandish increase, models the general increase in expected tie value to any
but bear in mind that the baseline expected tie value network member; thus, although meeting attendance
between any two randomly selected nodes is very, very should increase the expected value of a network tie,
followed by medium-collaboration Missouri, with the greatest degree of mutuality by far (although mutual-
low-collaboration Georgia having the lowest density ity is also higher in low-collaboration Georgia than
of nonzero ties. In summary, the case with highest col- in medium-collaboration Missouri). Even though
laboration exhibits the lowest overall participation relicensing process participants come from many dif-
density and has many more isolated actors. ferent organizations, the strength of reciprocal ties
What might account for the unexpected result in the Washington case evidences that many partici-
that the most collaborative network is also the most pants engaged in repeated, two-way dialogue with one
exclusive of the three nominally collaborative relicens- another. Such interactions can help create strong con-
ing processes? Bodin and Crona (2009) hypothesize nections that foster shared understanding and enable
that because network interactions represent potential joint action (Emerson and Nabatchi 2015a).
opportunities for developing common ground and ini- The final structural attribute, transitivity, is shown
tiating joint action (i.e., dialogue and discourse create to be strongest in the network with the lowest observed
more opportunities for collaboration to occur), net- collaboration (Georgia). This supports Hypothesis 4’s
works with more ties should be the most collabora- expectation that high-collaboration processes would
network analysis; the network results instead more for a large number of cases would require substantial
closely reflect the overall level of collaboration (which manpower, using a larger sample size would allow for
includes very low levels of trust from the Missouri trends to be ascertained with more statistical certainty.
survey, drawing the final ranking downward). This Second, examining the evolution of collaborative gov-
suggests that the network analysis was effective in cap- ernance networks over time is a critical way to build
turing the overall level of collaboration but may not upon current work. Neither collaboration nor net-
mirror individual dynamics as effectively. work structures are static phenomena but vary over
the course of a decision-making process. With frequent
meetings occurring over multiple years, participants
Conclusion have the opportunity to get to know one another very
In this study, we asked whether network structure well, learn to cooperate, and possibly even adopt one
metrics reflect the quality and extent of collaborative another’s beliefs (Leach et al. 2013; Lubell 2004b). The
decision-making. Using meeting attendance and partici- network and associated collaborative actions might
pation records from three FERC hydropower relicens- start by building connections across relatively dispa-
Beierle, Thomas C., and J. Jerry Cayford. 2002. Democracy in practice: Public Gerber, Elisabeth R., Adam D. Henry, and Mark Lubell. 2013. Political homo-
participation in environmental decisions. Washington, DC: Resources for phily and collaboration in regional planning networks. American Journal
the Future. . of Political Science 57:598–610.
Berardo, Ramiro. 2014. Bridging and bonding capital in two-mode collabora- Gerlak, Andrea K., Mark Lubell, and Tanya Heikkila. 2012. The promise
tion networks. Policy Studies Journal 42:197–225. and performance of collaborative governance. In Oxford Handbook of
Berardo, Ramiro, Tanya Heikkila, and Andrea K. Gerlak. 2014. US Environmental Policy, ed. Sheldon Kamieniecki and Michael E. Kraft,
Interorganizational engagement in collaborative environmental manage- 413–34. Oxford, UK: Oxford Univ. Press.
ment: Evidence from the South Florida Ecosystem Restoration Task Force. Gray, Barbara. 2000. Assessing inter-organizational collaboration: Multiple
Journal of Public Administration Research and Theory 24:697–719. conceptions and multiple methods. In Cooperative strategy: Economic,
Berardo, Ramiro, and John T. Scholz. 2010. Self-organizing policy networks: business, and organizational issues, ed. David Faulkner and Mark De
Risk, partner selection, and cooperation in estuaries. American Journal of Rond, 7–32. New York, NY: Oxford Univ. Press.
Political Science 54:632–49. Hahn, Thomas, Per Olsson, Carl Folke, and Kristin Johansson. 2006. Trust-
Bodin, Orjan, and Beatrice Crona. 2009. The role of social networks in natural building, knowledge generation and organizational innovations: The role
resource governance: What relational patterns make a difference? Global of a bridging organization for adaptive comanagement of a wetland land-
Environmental Change: Human and Policy Dimensions 19:366–74. scape around Kristianstad, Sweden. Human Ecology 34:573–92.
Bodin, Örjan, and Christina Prell, eds. 2011. Social networks and natural Handcock, Mark S., David R. Hunter, Carter T. Butts, Steven M. Goodreau,
resource management: Uncovering the social fabric of environmental gov- Pavel N. Krivitsky, and Martina Morris. 2015. statnet: Software tools
Korfmacher, Katrina Smith, and Tomas M. Koontz. 2003. Collaboration, Pretty, Jules, and Hugh Ward. 2001. Social capital and the environment. World
information, and preservation: The role of expertise in farmland preserva- Development 29:209–27.
tion task forces. Policy Sciences 36:213–36. Provan, K. Keith G., and Patrick Kenis. 2008. Modes of network governance:
Kosnik, Lea. 2010. Balancing environmental protection and energy production Structure, management, and effectiveness. Journal of Public Administration
in the federal hydropower licensing process. Land Economics 86:444–66. Research and Theory 18:229–52.
Krivitsky, Pavel N. 2012. Exponential-family random graph models for valued Putnam, Robert D. 2000. Bowling alone. New York, NY: Simon & Schuster.
networks. Electronic Journal of Statistics 6:1100–28. Putnam, Robert, Robert Leonardi, and Raphaëlle Nanetti. 1993. Making
Krivitsky, Pavel N., and Carter T. Butts. 2013. Modeling valued networks with democracy work: Civic traditions in modern Italy. Princeton, NJ:
Statnet. The Statnet Development Team. https://statnet.csde.washington. Princeton Univ. Press
edu/ (accessed April 15, 2016). Robins, Garry, Jenny M. Lewis, and Peng Wang. 2012. Statistical network
Lasker, Roz D., and Elisas Weiss. 2003. Broadening participation in commu- analysis for analyzing policy networks. Policy Studies Journal 40:375–401.
nity problem solving: A multidisciplinary model to support collaborative Sandström, Annica, and Lars Carlsson. 2008. The performance of policy net-
practice and research. Journal of Urban Health 80:14–47. works: The relation between network structure and network performance.
Lazer, David, and Allan Friedman. 2007. The network structure of exploration Policy Studies Journal 36:497–524.
and exploitation. Administrative Science Quarterly 52:667–94. Schneider, Mark, John Scholz, Mark Lubell, Denisa Mindruta, and
Leach, William D., and Paul A. Sabatier. 2005. Are trust and social capital the Matthew Edwardsen. 2003. Building consensual institutions: Networks
keys to success? Watershed partnerships in California and Washington. and the national estuary program. American Journal of Political Science
to the distribution of possible network graph con- estimation technique that samples from the distribu-
figurations, with each hypothetical graph weighted tion of possible network configurations. Sampling is
based upon similarity to the observed configuration. weighted based upon similarity to the observed net-
Comparing the observed network to the distribution of work in terms of descriptive characteristics such as
possible networks provides a probabilistic assessment density and structural attributes such as the number of
of whether a particular driver is more or less prevalent triangular-patterned ties in order to generate a distri-
in the observed network (by showing whether a given bution of network graphs that are a suitable basis for
structure occurs more or less in the observed network comparison.
than if ties are distributed at random; Kolaczyk 2009; Basically, the MCMC procedure iteratively pro-
Lusher, Koskinen, and Robins 2013). poses a single tie change to the network and then
For nontrivial networks, a vast number of possible compares the likelihood of the pre-change and post-
network configurations exist, and it is infeasible to ana- change network graphs. Every time the proposed
lyze all possible graphs. Thus, GERGMs use a Markov change increases the likelihood of the simulated net-
chain Monte Carlo (MCMC) maximum likelihood work graph, the MCMC process makes the proposed
Figure A2. Tracing MCMC parameter estimates for Missouri network model.
178 Journal of Public Administration Research and Theory, 2017, Vol. 27, No. 1
Figure A4. Density of MCMC parameter estimates for Washington network model.
change and then repeats the same steps. When the Krivitsky (2012) specifies the likelihood function for a
proposed change decreases the likelihood (shown zero-modified Poisson-reference GERGM as:
below) of the network graph, the MCMC process
only elects to make the proposed change a fixed per-
centage of the time (e.g., 50%; Lusher, Koskinen, and
Prθ ;h,η, g (Y = y ) ∝ ∏ exp (θ y 1 ij + θ 2 1yij >0 ) y ! (A1)
ij
(i , j ) ∈ Y
Robins 2013). This iterative procedure searches the
parameter space to (hopefully) converge on a station- with the reference distribution represented by h(y):
ary distribution of network graphs that resemble the
observed network. h (y ) = 1 / ∏y ij ! (A2)
(i , j ) ∈ Y
Social networks are often sparse in that actors who
interact often interact multiple times and many other Equation A1 specifies the model sans any additional
actors have zero ties. This means that the dyad-wise model terms as the observed baseline distribution
tie distribution is zero inflated relative to a stand- of network ties. Additional terms are then added
ard Poisson distribution (Krivitsky and Butts 2013). to the model to account for endogenous structural
Journal of Public Administration Research and Theory, 2017, Vol. 27, No. 1 179
Figure A6. Density of MCMC parameter estimates for Georgia network model.
characteristics and exogenous attributes of participat- Butts 2013).5 Each model in this article is fit via a four-
ing organizations. parallel sampling chains. Every chain discards the first
10,000 networks to ensure that samples used for esti-
Model Goodness-of-Fit mation purposes are drawn from a stable distribution
When fitting a GERGM, it is important to ensure that (Lusher, Koskinen, and Robins 2013). After this “burn
the MCMC estimation process does not exhibit degen- in” period, each chain takes 160,000 samples (with
eracy; a degenerate MCMC cascades to a completely each sample being a different possible network graph).
full or completely empty model (i.e., all high-value
ties or all no-value ties) and weights these unrealis- 5 Note that the goodness-of-fit plots shown in Appendix A refer to
tic network graphs too highly (Handcock et al. 2008; the unrestricted models presented in Table 6; we also performed
Kolaczyk 2009; Lusher, Koskinen, and Robins 2013). goodness-of-fit analysis for the restricted models shown in Tables 4
and 5. Each model mixes adequately, exhibits consistent, unimodal
Figures A1–A3 show traceplots for each param-
estimation within and across MCMC chains, and faithfully replicated
eter across iterations of each MCMC chain to dem- the structural characteristics of the observed networks. Because the
onstrate that the model is mixing sufficiently and not unrestricted models in Table 6 are more difficult to fit, we focus on their
straying outside of the parameter space (Krivitsky and goodness-of-fit in this appendix.
180 Journal of Public Administration Research and Theory, 2017, Vol. 27, No. 1
So that all draws are not taken from the same small converge and that the distribution of estimates gener-
area, each chain is thinned by discarding 1,000 indi- ated by each chain are unimodal and approximately
vidual perturbations between each sample. Figures A1, normally distributed. Figures A4–A6 show density
A2, and A3, respectively, thus trace the maximum like- plots for each model parameter for Washington,
lihood parameter estimates for each MCMC chain and Georgia, and Missouri, respectively.
parameter in the corresponding model. The seemingly One advantage of the parallel modeling approach
random, back-and-forth movement of each chain dem- we use, in which each model uses four separate MCMC
onstrates that the model is searching throughout the chains, is that comparing chain-specific distributions
parameter space, and the lack of an upward or down- demonstrates model robustness. Figures A4–A6 dem-
ward trend shows that each model does not become onstrate that the four parallel chains converge around
degenerated and trend toward a completely empty or consistent estimates for each model parameter. This
completely full network. is an important improvement upon many published
Along with mixing adequately and sufficiently ERGMs that do not use parallel processing. Although
searching the parameter space, it is also important serial processing still allows the analyst to establish
Figure A7. Distribution of simulated network statistics based upon Washington model.
Figure A8. Distribution of simulated network statistics based upon Missouri model.
Journal of Public Administration Research and Theory, 2017, Vol. 27, No. 1 181
distribution, comparing the results of several chains fit observed statistics. The unrestricted Washington
in parallel shows the extent to which model results are model (Figure A7) is the poorest fit in this regard, as
consistent and replicable. the model struggles to simulate networks that closely
Finally, in the discussion of model goodness-of-fit resemble the observed network. This is not the case
above in the main body of the manuscript, we noted for the restricted models with only sum, nonzero,
that GERGM inference is based upon comparing the and meetings attended parameters (shown in Tables
observed network to the distribution of simulated 4 and 5).
networks generated by the model. That is, if we simu- The conundrum here is that dropping model terms
late a large number of networks based upon estimated to increase the goodness-of-fit with regard to simula-
model parameters, we should expect that the simu- tion makes comparison to other network models more
lated networks generally resemble the observed net- problematic, since adding or subtracting parameters
work. To test this, we simulate 100,000 hypothetical changes each estimate (see Tables 4–6). Since our anal-
networks based upon each best fit model presented ysis focuses on comparison across the three cases, this
above and compare the distribution of network statis- issue is paramount. Moreover, given that (1) the good-
tics found in the simulated networks to the observed ness-of-fit metrics presented above (Figures A1 and
values in Washington, Georgia, and Missouri. A4) speak to adequate MCMC mixing and consistency
Figures A7–A9 plot the density of simulated val- in parameter estimation within and across MCMC
ues and show where the observed statistics fall within chains and (2) the worst fitting parameter in Figure A7
these distributions using a dashed line. In each case, is the Meetings Attending parameter, which is not of
the observed statistic is near the middle of the sim- substantive importance to the results in any case, we
ulated values and the simulated values appear to believe that it is still appropriate to make comparisons
be approximately normally distributed around the using the three models.