Faculty Presentations
Achen, Christopher. “Registration and Voting under Rational Expectations: The Econometric Implications.” «download»
Alone among modern democracies, the United States makes voter registration a personal responsibility rather than a governmental function. In almost all states, registration deadlines occur well before elections. Failure to register by the deadline makes the probability of voting exactly zero. This sequential feature of the registration and voting decisions has been skipped over by most researchers, who simply ignore registration. Others, notably Timpone (1998), have used the seemingly appropriate Heckman-style selection model, but have arrived at findings difficult to believe. This paper investigates the appropriate choice of a registration model under a rational expectations assumption about the desire to vote, showing that, rather surprisingly, conventional selection models will generally perform less well than ignoring the selection effect of registration entirely. However, neither is quite correct. Finally then, the paper proposes and tests a flexible model for registration as a step toward substantively appropriate joint modeling of registration and voting.
Adolph, Christopher. “A Latent Strength Model of Partially Observed Rank Data.”
Fully observed rank data, in which each observation receives a unique value from the set $1,\ldots,n$, presents challenges for conventional statistical techniques, most notably dependence across observed ranked individuals. Moreover, rank data in social settings is often only partially observed: rank is sometimes known only up to a lower and upper bound, as in the case of individuals observed to lie in the same tier of an organizational chart. The statistical literature on rank data models focuses on the analysis of multiple fully-observed rankings of the same individuals by different judges; however, in politics, we often (partially) observe only a single ranking, which may be the outcome of a multiseat election or the actual disposition of a bureaucratic hierarchy. We propose model of partially observed rank data which assumes the order of each pair of ranked individuals $\{i,j\}, i \ne j$ is a function of their relative latent strength, $\mu_i - \mu_j$. Each individuals latent strength is parameterized as a function of covariates and an individual random effect. We use Bayesian methods to estimate a cross-level mixed effects model of each possible pairwise comparison, imputing as needed any comparisons missing due to the presence in the same ranking tier. By estimating the model at this level, we respect the dependence in the ranked data and avoid discarding any information about partially ranked individuals. We discuss extensions to time series ranked data, and applications to partially observed election data.
Anderson, William, & Shane Nordyke. “A Mixed Logit Model of Presidential Choice in the U.S. House.”
We examine presidential agenda setting in the U.S. House of Representatives and, to do so, we employ a data set of 3,500 bills—half of which are presidential position bills and half a random sample of non-presidential position bills—introduced in the House of Representatives during the Eisenhower through Clinton administrations. We specify three models—mixed logit, conditional logit, and multinomial logit—and draw inferences about the most appropriate technique to use when examining presidential decisions, many of which are conditioned by past presidential behaviors. Our substantive findings suggest that presidential agenda-setting relies on three essential features: the president's legislative acumen, his timing, and the partisan and economic contexts within which he takes positions in the House. These covariate effects wax and wane across model specifications, however. We conclude with a broader discussion of the importance of granular data collection in the presidency subfield and how the variety of choice models now available to scholars provide tremendous analytical leverage in a subfield not long known for sophisticated quantitative analyses.
Andreadis, Ioannis, & Theodore Chadjipadleis. “Voter transition estimation in multiparty systems.” «download»
Recent advances in the field of ecological inference have provided researchers with new tools to estimate voter transition in two-party systems. Although some researchers have dealt with the R x C ecological inference problem, voter transition estimation remains a difficult and tedious goal. As a result scholars of multi-party systems still struggle with their electoral data. In this paper we present a new approach and we propose a new method that deals with this issue.
Although a large body of research exists investigating the degree to which "candidates of choice" elected in majority-minority districts (MMDs) provide substantive, as well as descriptive representation (e.g., Cameron, Epstein, and O'Halloran 1996; Lublin 1999; Epstein and O'Halloran 1999), questions still remain about differences in legislative behavior between representatives elected from MMDs and those not. Recent research has applied modern statistical techniques to measure polarization of majority-minority elected legislators in the South Carolina state senate after the 1992 redistricting (Epstein, Herron, O'Halloran, and Park 2007), and the cosponsorship networks of minority legislators in the House (Epstein, Fowler, and O'Halloran 2007). This paper advances the recent literature by utilizing a Bayesian item response model of legislator ideal points to test whether representatives from MMDs are more polarized ideologically than other partisans, as well as other representatives from districts with large (but not majority) minority populations. In addition, differences in majority-minority elected legislators' positions within ideological networks of agreement and influence are examined.
Researchers analyzing multilevel, panel, or TSCS data must often make a choice over whether to employ a "fixed effects" (FE) or "random effects" (RE) (aka, "random intercept") modeling approach. Both approaches account for unobserved heterogeneity in the outcome at the cluster level, i.e., across level-2 units of analysis (e.g., countries in typical TSCS data, individuals in panel data). Importantly, each approach accounts for this heterogeneity in different ways, and each approach produces different interpretations of coefficients. Pros and cons exist for both approaches. A complaint with FE is that it eats up too many degrees of freedom and is inefficient. Also, the FE approach does not allow for the inclusion of level-2 variables, that is, variables that are constant within a cluster (e.g., time-invariant variables in TSCS and panel data). For the RE approach, many highlight the controversial assumption that the level-2 random effect be uncorrelated with level-1 variables. On the upside, FE eschews the above-mentioned problem with random effects by confining all variation to within clusters (or level-2 units). And a major advantage of RE is that one can include level-2 variables in the model, which is often important for testing hypotheses of interest. While debates continue about which approach is best for certain situations, I contend that a core issue with these types of data continues to be ignored: the problem of cluster confounding. Cluster confounding occurs when a level-1 variable (e.g., a time-varying covariate in panel and TSCS data) exhibits distinct within-cluster and between-cluster effects, yet one only includes the original level-1 variable in the model without distinguishing these two types of variation in that variable. As a result of not distinguishing within- and between-cluster variation in a level-1 variable, the within- and between-cluster components of the variable are combined, or confounded, together, producing a sort of weighted average of the within and between effects. If the within and between effects of a level-1 variable are the same, which is something we can test for, then cluster confounding is not a problem. An example of cluster confounding in TSCS data, where countries are level-2 units, is where the within-country, or longitudinal, effect of a level-1 variable is distinct from that variable's between-country, or cross-sectional, effect. For hierarchical data where individuals are level-1 units and, e.g., schools are level-2 units, the within-school effect of a student-level variable is a purely individual effect, whereas the between-school effect is an aggregate, school-level effect. These effects may differ, but any differences will go unnoticed unless one directly accounts for and tests for cluster confounding. In this paper, I propose a unified and simple solution to the problem of cluster confounding, which should be of general interest to analysts employing models for multilevel, panel, and TSCS data. First, I show how to calculate within-cluster and between-cluster manifestations of a level-1 variable. This is done by subtracting the cluster mean of the variable from the original value of the variable. The effect of this new variable will be the within-cluster effect, and the effect of the cluster mean will be the between-cluster effect. I then suggest estimating a random-intercept (or, more generally, a random-coefficient) model that includes the within- and between-cluster versions of the level-1 variables, as well as any level-2 variables. I discuss the precise interpretations of each effect. Note that this procedure satisfies the controversial assumption associated with the RE approach that the level-2 random effect be uncorrelated with the level-1 variables. Since all of the variation in a level-1 variable is now solely confined to either within or between clusters, the level-1, or within-cluster variables, will be completely uncorrelated with the level-2, or between-cluster, random effect. I also highlight how the within-cluster effects will be the same as estimates from an FE model, though the standard errors will be different. Additional benefits of the proposed solution include: (1) the ability to make clearer substantive interpretations about the precise manner in which a variable exhibits effects at different levels; (2) the ability to test for cluster confounding with a Wald test of the equality of the within- and between-cluster effects; and (3) the ability to include level-2 variables in the model, something the FE approach cannot accommodate. To illustrate the proposed methodology, I present two empirical applications. First, in the TSCS context, I revisit the debate between Blaydes (2004, 2006) and Goodrich (2006) over the most appropriate modeling strategy for testing the "rewarding impatience" hypothesis that oil production by OPEC countries is tied to the manner in which they discount future gains. I suggest that the key hypothesis demands testing the between-country effect of the key level-1 variable (per capita oil reserves), which entails estimating both within- and between-country effects of this covariate. Second, in the area of more traditional multilevel data, as well as in the binary choice context, I apply the methodology to Epstein et al.'s (2006) study of Senate voting on Supreme Court confirmations (where Senate votes are level-1 units and confirmation contexts are level-2 units) to highlight differences between the within-context and between-context (or aggregate) effects of party and ideology in Senate confirmation votes. I also incorporate and discuss a random coefficient model specification, present interpretations of cross-level interactions, and present some post-estimation strategies for making substantive interpretations of the results. In sum, the methodology proposed, as well as the empirical applications of the method, highlight the importance of understanding, accounting for, and testing for cluster confounding in multilevel, panel, and TSCS data. The methodology is capable of uncovering fresh insights regarding substantive problems in political science and producing more precise statistical tests of hypotheses.
Boehmke, Frederick J., & Meissner, Christopher M. “Modeling Sample Selection for Durations with Time-Varying Covariates, With an Application to the Duration of Exchange Rate Regimes.” «download»
We extend previous estimators for duration data that suffer from non-random sample selection to allow for time-varying covariates. Rather that a continuous-time duration model, we propose a discrete-time alternative that models the (constant) effects of sample selection at the time of selection across all years of the resulting spell. Properties of the estimator are compared to those of a naive discrete duration model through Monte Carlo analysis and indicate that our estimator outperforms the naive model when selection is non-trivial. We then apply this estimator to the question of the duration of monetary regimes.
Good statistical methods ought to both be easy to use and understand for methodologists, but they should also allow nonspecialists to readily and accurately appraise the quantitative evidence those specialists produce. Matching distinguishes itself, in principle, in its simplicity and thus appeal to non-technical audiences. But whether or not a given match is in some senses "good" is not so clear. Clearly, we want any matching procedure to produce sets of "treated" and "control" units which are "similiar" or "balanced" enough. But what is the standard to which we ought to compare a given measure of balance? What balance measures are best? How should we diagnose the effectiveness of a given matching? Our project aims to answer some of these questions using the randomized experiment as a benchmark for comparison. That is, we would like to claim that if a given set of matchings produces balance that is indistinguishable from that which an ideal randomized experiment would produce, then we should call that particular set of stratifications "balanced". In addition, we propose that estimation of treatment effects from such a "balanced" matching use the properties of randomization rather than assumed models. After all, if we cannot distinguish a set of matches from an experiment, then we might as well analyze it as an experiment. Thus, we hope to enhance the transparency and simplicity of matched analyses while providing a persuasive answer to some of the thornier questions raised by matching. So far we have developed theory for this approach and applied it to political science problems in two publications (Hansen and Bowers 2008a, 2008b). And now, aided by a recent NSF grant (SES-0753164 and SES-0753168) we are turning to the creation and elaboration of software to implement these methods. For Polmeth 2008, we propose to present a poster to garner comments and criticism related to our development of an add-on package for R. As we intend for it to be useful for routine propensity diagnostics, we would like for it also to present meaningful descriptive statistics and helpful plots. We have some ideas about how to do that, but hope for discussion at the conference to sharpen these ideas. It is at present adapted to our tastes and working style: we hope that presentation as a poster at the conference will help us identify ways to cater to other tastes and working styles also. Last but not least, we are in the market for important substantive questions and interesting data sets with which the techniques might be fruitfully applied or tested. Thus, we hope that our participation in the political methodology conference will help us improve our work at the same time as it exposes others to our ideas. Hansen, Ben B. and Jake Bowers. 2008a. ``Covariate balance in simple, stratified and clustered comparative studies.'' \emph{Statistical Science}. to appear in volume 23. Hansen, Ben B. and Jake Bowers. 2008b. ``Attributing Effects to A Cluster Randomized Get-Out-The-Vote Campaign.'' \emph{Journal of the American Statistical Association}. to appear.
Bowyer, Benjamin. “The Trouble with Tobit: A District-Level Sample Selection Model of Voting for Extreme Right Parties in Europe, 1980-2004.” «download»
The growing electoral success of extreme right parties (ERPs) in many European countries has sparked academic interest in explaining variation in extreme right success. However, much of the extant research on the electoral success of extreme right parties suffers from at least two types of selection bias. The first involves the selection of cases and occurs when only those national elections that were contested by extreme right parties are included in the cross-national analysis. To address this problem, a growing number of scholars of ERP electoral support employ Tobit models to analyze national-level election results pooled across countries and election years. However, this approach conceals a second source of selection bias: ERPs are extremely selective about which election districts within a country they choose to contest. The correct specification of this process of self-selection requires the recognition of two fundamental points. First, the causal factors that determine whether an extreme right party contests an election are not identical to those that influence its share of the vote if it does appear on the ballot. Second, this decision about when and where to field candidates is one that is observable at the level of the election district. This paper argues that the appropriate way to model is as a Heckman sample selection model estimated at the level of electoral district. I present a preliminary analysis of a dataset that pools district-level election results for eighteen European countries from 1980-2004 (N=12,050), the results of which demonstrate the value of this approach.
Brandt, Patrick. “Bayesian Markov-Switching Models and Forecasting of Endogenous Conflict Data.”
We develop a Bayesian Markov-switching model for forecasting inter- and intra-state conflicts. This model allows us to identify shifts or switching processes in the parameters and generate dynamic inferences about the path of conflict over time. Most importantly, we show how these models generate superior probabilistic forecasts of the trend and timing of political conflict. This allows us to identify periods of instability that may not be captured by non-Bayesian, fixed parameter models. Finally, we show that these models generate forecasts that are superior to those generated by other existing methods.
Breding, Mary. “Field Research and Non-Government Organizations: To Engage or Not to Engage?”
What are the costs and benefits of working with non-government organizations to conduct field research? Starting a research project in a new country can be a daunting task—whether studying political elites, surveying public opinion, or engaging in other kinds of analysis. Researchers are presented with options for establishing different kinds of affiliations to set up shop for their work. This paper considers one kind of affiliation—the non-government organization route. NGOs potentially provide a great base to researchers—a well-networked group of individuals with language skills and connections that can be costly and take time to establish free of institutional support. On the other hand, NGO's are often mission-driven with normative policy agendas that may stand in the way of an empirical research agenda. I outline costs and benefits to in conducting empirical research through non-government affiliations. I further present examples of NGO-researcher affiliations in political science—including two on-going research affiliations I have maintained to collect substantial data over a three-year period in Bangalore, India.
In the comparative political economy of rich democracies there is a long tradition of classifying countries into one of a small number of categories based on their economic institutions and policies. The most recent of these is the Varieties of Capitalism project which posits two major clusters of nations: Coordinated and Liberal Market Economies. This classification, the result of rankings and expert judgements over a large number of dimensions, has generated controversy. We leverage recent advances in mixture model-based clustering to see what the data say on the matter. We find that there is considerable uncertainty around the number of clusters and, barring a few cases, which country should be placed in which cluster. Moreover, when viewed over time both the number of BIC-maximizing clusters and their membership change considerably. As a result, arguments about who has the ``right'' typology are misplaced. We conclude that these categories do not measure anything meaningful, especially in the context of time series cross section data, and should not be employed as indicator variables. We argue that the real value of both Esping-Andersen's work and the varieties of capitalism project have been largely masked by easy-to-remember typologies.
This paper uses a natural experiment in Switzerland to estimate the causal effect of political interest on party identification and ideological orientation. The key to the natural experiment is that Switzerland has three different language regions, which correspond to the languages spoken in the neighboring countries—France, Germany, and Italy. Because there is a common media market, German/French/Italian speaking Swiss are more likely to follow the news through the media of the neighboring country that corresponds to their language. During the election and the time afterwards when the new government forms the media will be filled with more political content (though not Swiss-specific content), thus exogenously increasing the political interest of the Swiss citizens who speak the language of the country with the election. Exploiting this natural experiment we find that political interest decreases the likelihood that people identify with a party but increases the likelihood that they identify their own ideological position.
Clarke, Harold, & Marianne Stewart. “The 2009-10 British Election Study.”
The purpose of the poster will be to introduce colleagues and graduate students interested in the science of electoral choice to the design features of the 2009-10 British Election Study (BES), and to discuss opportunites for collaboration and participation. The new BES has exciting innovations including massive rolling campaign and post-election panel survey, feedback-to-respondent and other experiments leveraging internet technology and monthly continuous monitoring surveys for the 2008-2012 period with substantial research opportunities (question batteries and experiments-similar to U.S. TES) submitted by electoral researchers (including graduate students) and other interested user communities. BES PI's are: Harold Clarke and Marianne Stewart, UTD, and David Sanders and Paul Whiteley, U of Essex. Clarke and Stewart will attend the methods meeting and do the poster session.
It has become common to discuss the problem of sample selection, whether traditional (Heckman) or strategic (Signorino), in terms of omitted variable bias. This paper takes a close look at the interaction between the omitted "selection" variable and other possibly omitted variables. The analysis demonstrates that correcting for selection bias can produce results that are, in a variety of senses, worse than not correcting for the selection bias. We characterize the conditions under which this result is likely to occur and discuss formal sensitivity analysis as a way of alleviating the problem.
Dunning, Thad. “Model Specification in Instrumental-Variables Regression: Assessing Homogenous Partial Effects.” «download»
In many applications of instrumental-variables regression, researchers seek to defend the plausibility of a key assumption: the instrumental variable is independent of the error term in a linear regression model. Although fulfilling this exogeneity criterion is necessary for a valid application of the instrumental variables approach, it is not sufficient. In the regression context, the identification of causal effects depends not just on the exogeneity of the instrument but also on the validity of the underlying model. In this paper, I focus on one feature of such models: the assumption that variation in the endogenous regressor that is related to the instrumental variable has the same effect as variation that is unrelated to the instrument. In many applications, this assumption may be quite strong, but relaxing it can limit our ability to estimate parameters of interest. After discussing two substantive examples, I develop analytic results (simulations are reported elsewhere). I also present a specification test that may be useful for determining the relevance of these issues in a givenapplication.
Can the Supreme Court persuade the public to agree with its rulings on controversial social issues? Or do the Court's pronouncements on these issues cause the Court to lose credibility with those who disagree with it? Both of these questions have been the topics of normative and positive theorizing and analysis of observational data. But to our knowledge, these questions have never been explored experimentally using a nationally representative sample of participants. In this paper, we use an experiment that we embedded in the 2006 Cooperative Congressional Election Study (CCES) to assess the relationship between Supreme Court rulings on three controversial topics (abortion, flag burning, and homosexual sex), public opinion on these issues, and the public's evaluation of the Court. We find that learning of the Court's ruling decriminalizing gay sex in Lawrence v. Texas leads a small, statistically significant proportion of respondents to change their attitudes to agree with the Court compared to those in a treatment group who are not informed of the decision. But we find no evidence of similar movement in attitudes on abortion (after being informed of Roe v. Wade) and flag burning (Johnson v. Texas). We do, however, find that being informed of the Johnson and Lawrence rulings has much larger effects on respondents' attitudes about the Court itself.
Elff, Martin. “A Spatial Model of Electoral Platforms.” «download»
The reconstruction of political positions of parties, candidates and governments has made considerable headway during the last decades, not the least due to the efforts of the Manifesto Research Group the and Comparative Manifestos Project, which compiled and published a data set on the electoral platforms of political parties from most major democracies for most of the post-war era. A central assumption underlying the coding of electoral platforms into quantitative data as done by the MRG/CMP is that parties take positions by selective emphases of policy objectives, which put their accomplishments in a most positive light (Budge 2001) or are representative for their current polital/ideological positions. Consequently, the MRG/CMP data consist of percentages of the respective manifesto texts that refer to various policy objectives. As a consequence both of this underlying assumption and of the structure of the CMP data, methods of classical multivariate analysis are not well suited to these data, due to the requirements to the data for an appropriate application of these methods (van der Brug 2001; Elff 2002). The paper offers an alternative method for reconstructing positions in political spaces based on latent trait modelling, which both reflects the assumptions underlying the coding of the texts and the peculiar structure of the data. Finally, the validity of the proposed method is demonstrated with respect to the average position of party families within reconstructed policy spaces. It turns out that communist, socialist, and social democrat parties differ clearly from “bourgeois” parties with regards to their positions on an economic left/right dimension, while British and Scandinavian conservative parties can be distinguished from Christian democratic parties by their respective positions on a libertarian/authoritarian and a traditionalist/modernist dimension. Similarly, the typical political positions of green (or “New Politics”) parties can be distinguished from the positions of other party families.
Fukumoto, Kentaro. “Continuous or Ordered Event History Analysis.”
Political scientists are interested in not just when an event happens but also what kind of events happen. Moreover, events themselves are related to when they happen. Scholars usually use event history analysis to examine which factors affect when events happen. If events are classified into several discrete categories, researchers may employ competing risks models. But common methods of event history analysis cannot deal with continuous or ordered events, a lacuna which the current paper addresses. For example, when will an election be called, and how many votes will the governing parties be likely to win? Smith (2004) argues that if the election is called earlier than expected, government wins even with fewer votes. Why? If the government anticipates bad news such as an economic recession, it will call elections earlier than planned. But, since voters rationally expect this, they also anticipate that something is wrong and fewer of them support the government in the voting booth. In this case, we cannot regress the vote rate on the election time because what matters is not the actual time of the election but its relative time against its expected time. The basic model of this paper assumes that the log of the duration variable (such as inter-election time) and an event variable (such as number of votes) follow the bivariate normal distribution. If duration is censored, survival probability is used as in typical event history analysis, ignoring the event variable. The quantity of the most interest, that is, the relationship between duration and the event, is captured by the covariance parameter, which is expected to be positive in this case. Another application of this approach is the duration of warfare and the terms of peace. Slantchev (2004) maintains that, the longer wars last, the worse the terms of peace are for initiators. The original article employs a probit model where an ordered war outcome is the dependent variable and the predicted duration of war is an independent variable. However, omitted variables may affect both the duration of war and the terms of peace and, therefore, estimates are likely to be biased. To avoid such kinds of problems, this paper assumes that the latent variable of the probit model (continuous war outcome) and the log of the duration variable follow bivariate normal distribution. Some readers may wonder about the cases where duration does not follow log normal distribution. The present paper proposes a generalized multivariate normal distribution where any form of baseline hazard function (other than log normal, such as Weibull or log logistic) and any form of event distribution (other than normal, for example beta) are correlated. This new model is applied to election data and war termination data mentioned above. This model enables scholars to study the relationship between time of events and contents of events, which is an important and promising topic in political science.
We are interested in this project in so-called ``dark networks:'' social networks characterized by low visibility and low interactivity between nodes along with high degrees of uncertainty about the connections. Furthermore, the edges of interest typically have multiple qualitative aspects and are likely to change over time. Although connections between individuals change character temporally, such changes typically cannot be measured directly and therefore require some form of estimation. These are fundamentally distinct types of networks since the actors are ``trading efficiency for secrecy'' (Fellman & Wright 2004). Since regular networks may contain missing information for less nefarious reasons, our methodology has broader applications. Our technique elicits prior distributions from experts as input to a Bayesian update of network information. Elicitation is based on the notion that analysts have subjective probability distributions (SPDs) about the attributes of the members of the dark network and also about the existence of links between the actors. Elicitation is a means by which analysts can provide these key pieces of information, information that can be input into link prediction models or into various forms of network analysis. Our new technology has several distinguishing features. To begin, it uses a new form of visualization. Social scientists and statisticians for many years have realized that humans ("assessors") have an easier time communicating their SPDs when they can visualize them. Therefore assessors have been asked to draw their probability distributions on graph paper or to compare their predictions with those that might come from "probability wheels." We build on these approaches but create a new form of computerized visualization that is easy for humans to understand and manipulate. Our technology is interactive and automated. Assessors work on their own at a computer screen. They adjust their visual representations of their SPDs until they are satisfied that these representations are accurate. Then they move on to the next task. The tasks they are given are a mix of what is called "indirect" and "direct" elicitation. The visualization element asks for a qualitative representation of the SPD; it emphasizes what psychologists call correspondence constancy. This element is initialized by a direct query about (numerical) centrality or the mean of a hypothetical experiment related to the attribute or link in question. This paper reports methodologies and results from experiments conducted at the University of Minnesota's new Behavioral Science Laboratory conducted during the Spring of 2008. Approximately 80 undergraduate volunteers were put through the system using the characteristics of selected populations and a British soap opera as target information. The subjects were shown partial videos from an episode and then elicited for relational information. We then repeat the process by showing an additional video clip and assess the improvement in elicitations that results from more narrative information. Our results add to the young experimental literature as well as the more mature (but still small) statistical literature on elicited priors. We expect to contribute to knowledge about attribute elicitation: how well subjects do in relation to the true population proportions under the conditions of the laboratory. More centrally, we address conditions under which accurate probabilistic statements can be made by subject matter experts without resorting to extensive training in probability theory.
Glynn, Adam N. “Nonparametric Estimation of Mechanism Specific Causal Effects.” «download»
Political scientists often cite the importance of mechanism specific causal knowledge, both for its intrinsic scientific value and as a necessity for informed policy. However, outside the framework of additive linear regression models with homogenous causal effects, mechanism specific effects are, in general, not estimated explicitly. Counterfactual causal models allow the formal definition of such concepts as direct, indirect, and mechanism specific effects, and the derivation of conditions for their identification (point or interval). In this paper, I demonstrate the use of counterfactuals to decompose causal effects into mechanism specific effects, showing that estimation and bounding can be accomplished with minor adjustments to standard techniques. Iillustrate this methodology with examples from American and Comparative Politics.
Golder, Matt. “Empirically Modeling the Government Formation Process: A Nested Mixed Logit.”
Virtually all theoretical analyses of the government formation process indicate that it occurs in stages — a formateur is initially chosen who then tries to form a government that contains her own party. While a consensus has recently emerged that the conditional logit (CL) model is the most appropriate empirical strategy for evaluating which parties make it into government, the CL model does not capture the sequential nature of the government formation process. We propose an alternative empirical model — a nested mixed logit — that does capture this aspect of how governments form, thereby narrowing the gap between theory and empirics. Our empirical strategy also allows us to avoid some of the methodological pitfalls that are, to a large extent, unavoidable when using a CL model in this setting. In addition to these methodological benefits, our model allows us to evaluate a number of additional hypotheses, particularly related to the choice of formateur, that cannot be tested in the standard CL framework. We evaluate our hypotheses linking ideological, institutional and ‘size’ variables to the choice of government using a new data set that we constructed containing information on almost 190,000 potential governments drawn from 402 government formation opportunities in 17 non-presidential democracies from 1945 to 1998.
Grant, Tobin. “Legislative and Executive Policy Production in the U.S. since 1789.”
This poster examines the causes and consequences of policy production in the legislative and executive branches since 1789. We utilize new multi-measure indexes of policy production that measure policy achievements in the legislative and executive branches during the entire history of the U.S. under the Constitution. We examine legislative and executive policy production both separately and as part of the U.S. macro political system, focusing on institutional and economic determinants of policy production. In a forthcoming article in Political Analysis, Nate Kelly and I develop an index of legislative productivity since 1789. In this paper, we build on this work by applying a similar methodology to executive branch policymaking. Using data from historical archives, measures developed by other social scientists, and expert surveys; we develop a measure of presidential policymaking. We then utilize these two measures to assess the determinants of policymaking in the legislative and executive branches. We focus especially on how policymaking in one branch responds to policy productivity in the other. We examine the relationship between the two measures using tests for fractional cointegration and error correction models. Our data, importantly, allow us to analyze policymaking through the entire history of the United States.
Models designed for limited dependent variables such as binary responses, ordered responses, counts, and so on, are becoming increasingly common in political science. Empirical researchers using limited dependent variable models often give little attention to the coefficient estimates, which cannot be interpreted as straightforwardly as OLS coefficients, and instead focus on predicted probabilities, counts, etc. Since these models are nonlinear, the effect of a change in the independent variable of interest depends on the values of the other independent variables and thus, the estimated effects are sensitive to how one generates the predictions. In presenting post-estimation predictions for a change in an independent variable of interest there are two general approaches for dealing with the other independent variables in the model: 1) specify an ‘average’ case (individual, state, country, etc.) by setting the values of the other independent variables to their respective means (or modes for dummy variables); or 2) leave the values of the other independent variables at their observed values. While textbook treatments do not favor one approach over the other, a content analysis of the American Political Science Review, American Journal of Political Science, and Journal of Politics reveals that the ‘average’ case approach is considerably more common. Moreover, standard software such as Gary King's Clarify and Scott Long's SPost make the ‘average’ case approach much more simple to implement. Using a variety of models we show that under most circumstances the observed value approach is preferable to the more common and simple ‘average’ case approach. Given the pervasiveness of limited dependent variable models it is essential that we develop a set of best practices for the interpretation of the results from these models. Toward this end we offer advice for determining which approach is best based on characteristics of the research question and data set.
When the sample is small and the number of potential confounders large, propensity scoring may seem to have little to offer. It works surprisingly well, however, when combined with complementary scoring methods, flexible matching, formal diagnostics and simple post-matching adjustments. In our motivating example, close propensity score matching was not possible; yet by matching relatively coarsely on propensity and "prognostic" scores (Hansen, 2008, Biometrika), it was possible to balance k=27 covariates with only n=67 subjects, and to reduce substantially the bias of effect estimation. Our post-matching adjustments bear a close resemblance to recommendations of Ho, Imai, King and Stuart (2007, Polit. Anal.), and more distantly resemble some of Achen's (2002, Annu. Rev. Polit. Sci.). In a thorough simulation study built around our motivating example, we evaluate the Type I error, bias, and power of our method and related ones that involve some but not all of its contributing techniques. The results show, among other things, that to ensure coverage of confidence intervals for the treatment effect: (i) overlap of treatment and control groups on the estimated propensity is *not* necessary; (ii) covariate balance in the matched sample is necessary but not sufficient; and (iii) balanced matching in combination with any of a range of post-matching adjustments does the trick.
Hays, Jude, & Aya Kachi. “Estimating Interdependent Duration Models with an Application to Government Formation and Survival.” «download»
This paper is part of a larger project in which we develop methods for estimating the causal effects of variables on (1) the duration of bargaining processes, broadly defined, and (2) the survival of bargained outcomes when both are jointly determined. There are many potential applications in political science including, but not limited to, the duration of war and survival of cease-fire agreements, coalition formation and government survival, and negotiations over and enforcement of international agreements. Our primary claim is that, in most cases, it is inappropriate to estimate the effects of variables on these two durations—the bargaining and the outcome—in isolation. Our argument is motivated by game theoretic models that show bargaining duration is correlated with the survival of bargained outcomes because players incorporate their beliefs about the survival of bargained outcomes into their decision-making at the bargaining stage. To address this problem, we develop, and examine the properties of two maximum likelihood estimators—a seemingly unrelated regresssions (SUR) estimator and a limited information maximum likelihood (LIML) estimator. We use both estimators to analyze the duration of government formation and survival in a sample of European parliamentary democracies over the period 1945 to 1998. We conclude that estimated effects based on single equation models of either government formation or survival, the predominant method of analysis in the existing literature, are likely biased because they fail to capture significant indirect effects generated by strategic and other forms of interdependence that link the two durations.
Three-Player Decision Games are by far the most difficult to model despite the wide array of rational-choice theories available that should otherwise simplify the least complexities about individuals, interests, institutions, or information. For regardless of ‘strategic ambiguity,’ spurious clarity, expected utility, or loss aversion, such terms and conditions about ‘risk’ itself assure those initial agents versus certain subsequent structures via parsimony as pragmatism. All told, the unsaid truth about tripartite interactions lie with the individualistic disconnects about actors and factors that defy yet define the one field of political science amid such multiplicatus areas. By applying then the principles of ‘extended deterrence’ which designate a tripartite security dilemma between the unsatisfied ‘challenger,’ threatened ‘protégé,’ and undecided ‘defender,’ such limitations to the IR (International Relations) field as a whole could now approximate other areas—judicial politics at supreme courts, no-confidence votes/snap elections, ‘iron triangles,’ or ‘just war’ interventions—that beget yet bestow such scope and methods from political science which hereby reflect one of four possible realities amid ‘Settlement, Separation, Standby, or Sucker.’
This paper introduces a new data set of 225 "State of the City" addresses to isolate key influences on local political agendas. To analyze the almost 25,000 sentences from big-city Mayors, the paper integrates multiple text-analytic methods. In the burgeoning field of quantitative textual analysis, some researchers opt for "supervised" approaches where the topics of interest are specified ex ante, while others use "unsupervised" approaches to identify the clustering patterns inherent in the data. This paper shows how these two approaches, while conceptually distinctive, can be integrated to improve categorization. The paper first uses a multinomial mixture model to compare the clustering patterns in the speeches with the categorization defined ex ante. This allows for adjustments in the coding scheme to consolidate hard-to-differentiate categories, provides an estimate of classification uncertainty, and permits improved instructions to coders. The paper then uses a multi-level model for compositional data to predict shifts in local agenda-setting over time. It shows that much of the observed variation in local political agendas is cross-sectional rather than longitudinal, illustrating the constraints that powerfully limit urban political agendas.
Iacus, Stefano M., Gary King, & Giuseppe Porro. “Matching for Causal Inference without Balance Checking.” «download»
We address a major discrepancy in matching methods for causal inference in observational data. Since these data are typically plentiful, the goal of matching is to reduce bias and only secondarily to keep variance low. However, most matching methods seem designed for the opposite problem, guaranteeing sample size ex ante but limiting bias by controlling for covariates through reductions in the imbalance between treated and control groups only ex post and only sometimes. (The resulting practical difficulty may explain why many published applications do not check whether imbalance was reduced and so may not even be decreasing bias.) We introduce a new class of “Monotonic Imbalance Bounding” (MIB) matching methods that enables one to choose a fixed level of maximum imbalance, or to reduce maximum imbalance for one variable without changing it for the others. We then discuss a specific MIB method called “Coarsened Exact Matching” (CEM) which, unlike most existing approaches, also explicitly bounds through ex ante user choice both the degree of model dependence and the causal effect estimation error, eliminates the need for a separate procedure to restrict data to common support, meets the congruence principle, is approximately invariant to measurement error, works well with modern methods of imputation for missing data, is computationally efficient even with massive data sets, and is easy to understand and use. This method can improve causal inferences in a wide range of applications, and may be preferred for simplicity of use even when it is possible to design superior methods for particular problems. We also make available open source software which implements all our suggestions.
Imai, Kosuke. “Causal Inference with Measurement Error: Nonparametric Identification and Sensitivity Analyses of a Field Experiment on Democratic Deliberations.” «download»
Political scientists have long been concerned about the validity of survey measurements. Indeed, given the increasing use of survey experiments, measurement error represents a threat to causal inference. Although many have studied classical measurement error in linear regression models where the error is assumed to arise completely at random, in a number of situations the error may be correlated with the outcome. Such differential measurement error often arises in retrospective studies where the treatment is measured after the outcome is realized. We analyze the impact of differential measurement error on causal estimation by deriving the sharp bounds of the average treatment effect. The proposed nonparametric identification analysis avoids arbitrary modeling decisions and formally characterizes the roles of additional assumptions. We show the serious consequences of differential misclassification and offer a new sensitivity analysis that allows researchers to evaluate the robustness of their conclusions. Our methods are motivated by a field experiment on democratic deliberations, in which one set of estimates potentially suffers from differential misclassification. We show that an analysis ignoring differential measurement error may considerably overestimate the causal effects. The finding contrasts with the case of classical measurement error which always yields attenuation bias.
Keele, Luke, Corrine McConnaughy, & Ismail White. “Adjusting Experimental Data” «download»
Randomization in experiments allows researchers to assume that the treatment and con- trol groups are balanced with respect to all characteristics except the treatment. Random- ization, however, only makes balance probable, and accidental covariate imbalance can occur for any specific randomization. As such, statistical adjustments for accidental imbalance are common with experimental data. The most common method of adjustment for accidental imbalance is to use least squares to estimate the analysis of covariance (ANCOVA) model. ANCOVA, however, is a poor choice for the adjustment of experimental data. It has a strong functional form assumption, and the least squares estimator is notably biased in sample sizes of less than 500 when applied to the analysis of treatment effects. We evaluate alternative methods of adjusting experimental data. We compare ANCOVA to two different techniques. The first technique is a modified version of ANCOVA that relaxes the strong functional form assumption of this model. The second technique is matching, and we test the differences be- tween two matching methods. For the first, we match sub jects and then randomize treatment across pairs. For the second, we randomize the treatment and match prior to the estimation of treatment effects. We use all three techniques with data from a series of experiments on racial priming. We find that matching substantially increases the efficiency of experimental designs.
In this paper, we introduce FactionSim and PMFserv to the political science community through its application to monitor, assess, and forecast intrastate crises (civil wars, coups, crackdowns, revolutions, transitions to democracy, etc.) in a way that supports decisions on how to allocate diplomatic, informational, military, and economic resources to harness them in a desirable direction. We build ten virtual Asian countries using this joint model for this purpose. The ten countries are: Bangladesh, Burma, China, Fiji, Indonesia, Pakistan, Philippines, Sri Lanka, Thailand, and Vietnam. FactionSim and PMFserv (www.seas.upenn.edu/~barryg/HBMR.html) are the results of >eight years of research at the University of Pennsylvania's Ackoff Collaboratory for the Advancement of Systems Approach. They synthesize over 100 best-of-breed theories from the literature, including a Value System Module (Goal-Standards-Preference Trees, Bayes Importance Estimators, Multi-Attribute Utility Functions, Affective Reasoning, Cognitive Appraisal), Personality Profiling Tools (Hermann Political Leader Profile Instrument, Modified Maslow-Follower Profile, Hofstede Cultural Factors Instrument, UN Globe Study Cultural Factors), Social Relationship Module (InGroup Hierarchy Designator, InGroup-OutGroup Alignment/Trust/Credibility, Automated Motivational Congruence Assessment, Identity Repertoire Theory, Eidelson “Dangerous Ideas” Model, Hirshman Exit/Voice/Loyalty Model), Decision Processes/Choice Module (Subjective Expected Utility, 5 Stress-Based Coping Styles (3 of them are algorithms of Nobel Prizes), Campaign Plans and State Transition Nets, Model of Others' Model of Me), and Socio-Cultural Game Leader-Follower Theory (Group Leader-Follower Role(s), Rival InGroup Leaders, InGroup-to-OutGroup Alignment Model, Dynamic Realignment, Insurgency Model, Tribal Credo). The agents are unscripted, but use their value systems and decision-making protocols to react to actions as they unfold and to plan out responses in the world they populate. When the agents are placed together, their affinities with each other and their proclivities to form groups are precipitated and macro-behaviors tend to emerge from the sum of individual micro-decisions. In past studies, PMFserv agents have successfully been developed and tested for simulating diverse ethno-political conflicts such as: the Palestinian Intifadah, 3rd Crusade, Somalian clans in Bakarra Market, the separatism movement of Southern Thailand, multiple factions, leaders, and followers of the current Iraqi conflict, and numerous world leaders. Many of these simulated agents exhibit 80% or better correlation with decisions by their real world counterparts in historical eras and test datasets.
A growing body of literature has established the significant role that street-level bureaucrats play in the implementation of social policy. In particular, research has highlighted the influence of professional norms on how bureaucrats do their jobs. Yet the available research presents competing explanations for where these norms originate. On the one hand, we know that professional norms are often highly correlated with individual demographics like party identification, gender and race. On the other hand, we hypothesize that these norms are shaped by the resource constraints and management styles that bureaucrats face in their day-to-day working environments. If this is the case, variation between bureaucrats should be dictated more by characteristics of the offices in which individuals work than by distinctions between individual themselves. In large part, previous research on this topic has employed simple regression analyses. This has contributed to the inability of this literature to successfully adjudicate between competing hypotheses. In this paper, I use a natural experiment and multivariate matching to assess the effects of correctional officers' professional experiences on their attitudes towards correctional policy. In California, correctional officers in their first two years of employment are randomly assigned to work with either higher or lower security inmates. I show that assignment to these different working environments leads officers to adopt significantly different criminal justice policy preferences. Original data for this research was gathered through a survey of almost 6,000 correctional officers across all 33 of California's adult state prisons.
Morton, Rebecca. “Experimental Methods in Political Science.”
This poster will be on a book manuscript on Experimental Political Science. Actually, I was supposed to do this poster at Penn State but because my flight got canceled I missed the faculty poster session. I am supportive of having faculty posters and would like to do one on this book ms. for this year. Note that in my opinion a presentation on the book manuscript works best as a poster, not a paper presentation. So if you choose not to have a faculty poster session, it shouldn't be considered for a regular presentation.
Nickerson, David. “Can Voter Turnout Contaminate Neighborhoods?”
Numerous studies have found neighborhoods to be relatively homogenous with regards to voting behavior. A prominent hypothesis is that social networks and interpersonal influence generate common forms of behavior. An alternative hypothesis is that the homogeneity is the result of particular types of people selection into a neighborhood. We examine these two competing hypotheses by employing exogenous shocks to a neighborhood's propensity to turnout in an election. Voter turnout has been found to be contagious within a household, but does the lesson apply to neighborhoods where the social ties are much weaker? Re-examining eight randomized Get Out the Vote field experiments, we find small but statistically and substantively significant effects of voter mobilization on uncontacted persons within the neighborhood. Using census data as a proxy for human and social capital, little heterogeneity is observed across types of neighborhoods.
Noel, Hans. “Interpreting Legislator Ideal Points with Help from the Political Discourse.”
I use a unique dataset of the opinions of political pundits to get leverage on the meaning of the estimated ideal points from NOMINATE or similar techniques. The interpretation of the dimensions of legislator preferences is more art than science. Most scholars (including Poole and Rosenthal) examine the angle of the cutting lines of various bills to identify the meaning of the estimated dimensions. I use an alternative approach. I assume that non-elected pundits represent "ideology" as it is defined in the public sphere, while elected representatives reflect ideology as well as party and other considerations. I estimate a common space by identifying bills that match the opinions expressed by ideological pundits. In this common space, the pundits are arrayed in essentially a straight line, defining a single dimension of ideology. In the 1950s, this dimension is orthogonal to the main cleavage between Democrats and Republicans. This confirms the interpration that the party dimension and the ideological dimension diverged in this period. I then estimate a common space with pundits from 1950 and legislators from later periods. This suggests that the new, primary dimension that comes to predict preferences in later years is the same as that defined by pundits in 1950. I argue that this means that the pundits are central in the defining of ideology, and that politicians and parties follow.
Social psychological theories have long held that extrinsic rewards have the capacity to depress intrinsic motivation. This study uses field experimental techniques to test this hypothesis within the context of voting. Voters in the November 2007 municipal elections in California were randomly assigned to receive a financial reward (of varying levels) for participating in the election. The results of the experiment suggest nominal incentives thwart intrinsic motivation and diminish turnout, while nontrivial incentives may boost motivation and stimulate participation.
Park, Won-ho. “Ecological Inference with Covariates.” «download»
The building block of ecological inference strategies is to construct a two-by-two table that describes the individual-level relationship from aggregate information. Extensions to this baseline model, whichever particular technique is employed, have been developed in the context of constructing bivariate R-by-C tables. However, another important and substantively interesting extension is a model that would let the researcher include additional covariates into the model and is yet to be fully discussed and developed. In the paper, I propose a method of moment estimator that incorporates covariates into the ecological inference process by extending Thomsen (1987)'s voter transition model. I apply the developed model to estimate the impact of demographic variables on turnout in South Korean voters over time, especially around democratization, using precinct-level electoral returns and census records.
The work of the political arithmeticians of the Enlightenment paved the way for the numerous empirical investigation techniques that were developed during the last two centuries in the fields of economics, sociology and psychology. These new approaches included such important analytical tools as household budget surveys, the measurement of income distribution and income growth and the assessment and monitoring of poverty levels (Collette 2000, 2), techniques that later were fully deployed in the studies of quality of life. While the most prominent scientists in those fields are known to be English, Scottish or French, studies from other nationalities' authors are now almost forgotten. In particular, I am referring to the Italian tradition of the XVIII and XIX century—as like Lagrange, Dellepiane and Niceforo. Lagrange is most known as mathematician, and played a major rule in the application of mathematics to the science of society, as many others in that period (Quetelet, Graunt, etc.). although Italian, he had some important functions during French Revolution, as the study on the needs of homogenization of weights and measures. With French scientists (and hommes d'Etat) Vauban, Necker, Lavoisier and Chaptal, Lagrange was an appreciated supporter of statistics as a necessary aid to the State administration. In this spirit he wrote a brief—but important—essay published in 1795-96 concerning State needs and average alimentary consumption. A further step in the historical development of the concept of quality of life came when the scientifically inclined began to insist that concrete developments in different fields of activity could be treated without the subjective intervention of the researcher. This led to the formulation of so-called indices of progress (Bossard 1931: 12). In 1912 Dellepiane made a remarkable effort to work out in rather extreme detail such an ‘objective’ analysis. Among the dimensions which he included are the amelioration of material well-being; high development of social and industrial machinery; elevation of coefficients of nuptiality and natality and many others (ibidem). In 1921 Niceforo wrote a book (Les Indices numériques de la Civilisation et du Progrès) that gained the international attention. Robert Park was looking for the criteria and the indices of progress and—partially—found them in this ‘little volume that seriously attempts to answer these questions, and this much may be said for it: If the answers it gives seem inadequate, the author has at least faced the problem and no one is likely to tell us how little we positively know about progress than the men who have tried to measure it’ (Park 1922: 671). Lazarsfeld too defined the work by Niceforo ‘the most creative effort to give structure to the ever-increasing mass of data’ (1961: 310-311). The majority of the studies on the rising science of society's measurement and her techniques have one fallacy in common: they don't consider the evolution in their social context. An adequate examination of the affirmation of concepts and practices used in assessing the welfare of States and individuals should assume that it is misleading to see it as a field of knowledge developing simply by its own internal logic and giving rise to value-free techniques. As the historical perspective by which Bulmer, Bales and Sklar (1991) analyzed the national traditions of social survey in Britain, France, Germany and United States, my desire will be to understand the origins of Italian modern empirical social investigation, an important part of the history of both social policy and of the social sciences. The focus will be on: a) to consider the historical interaction with conceptual change in other sciences, with the needs of production and with theological (Hacking 1990; 1999), political (Baker 1975) and ideological developments (Brian 1994; Bosher 1970; Piovani 2006); b) reconstruct the evolution of concepts as ‘well-being’ or ‘quality of life’ from the point of view of both individuals and States (Phillips 2006); c) discovery similarities and dissimilarities between authors and the historical, economic and political forces that influenced (in one direction or in the other) the development of studies in the assessment of the quality of life.
Peress, Michael. “Estimating Proposal and Status Quo Locations Using Voting and Cosponsorship Data.” «download»
Theories of lawmaking generate predictions for the policy outcome as a function of the status quo. These theories are difficult to test because existing ideal point estimation techniques do not recover the locations of proposals or status quos. Instead, such techniques only recover cutpoints. This limitation has meant that existing tests of theories of lawmaking have been indirect in nature. I propose a method of directly measuring ideal points, proposal locations, and status quo locations on the same multidimensional scale, by employing a combination ofvoting data, bill and amendment cosponsorship data, and the congressional record. My approach works as follows. First, we can identify the locations of legislative proposals (bills and amendments) on the same scale as voter ideal points by jointly scaling voting and cosponsorship data. Next, we can identify the location of the final form of the bill using the location of last successful amendment (which we already know). If the bill was not amended, then the final form is simply the original bill location. Finally, we can identify the status quo point by employing the cutpoint we get form scaling the final passage vote. To implement this procedure, I automatically coded data on the congressional record available fromwww.thomas.gov. I apply this approach to recent sessions of the U.S. Senate, and use it to test the implications of competing theories of lawmaking.
Rivers, Douglas, & Stephen Ansolabehere. “What is the Margin of Error?”
The reported margin of error for a survey is intended to have 95 percent coverage. In the case of election forecasts, from either pre-election telephone polls or exit polls, the actual coverage can be computed by comparing the survey estimate with the tabulated vote. Using data from the 2008 Presidential primaries, we show that the actual coverage levels of the reported margins of error are substantially below their nominal levels. The reported margins of error are misleading because of three factors: (1) an inconsistent estimate of sampling error is routinely employed for pre-election telephone polls; (2) non-response introduces a systematic bias that inflates the mean square error; and (3) differential non-response introduces an additional component of variance. These are analyzed within the context of a Bayesian hierarchical model and a "Bayesian Margin of Error" is proposed. Empirical examples from pre-election telephone polls and the National Exit Poll in the 2008 Presidential primaries are presented.
Seawright, Jason. “Approaches to Pseudo-Panel Inference: Comparing Real-World Properties.”
Scholars have proposed a number of second-best solutions to the problem of making inferences about change over time in individual survey responses for situations where panel data are unavailable and must be replaced with a sequential collection of cross-sectional data. Solutions include cohort averaging, matching estimators using time as the treatment of interest, two-stage auxiliary instrumental variables, and multiple imputation. Under specific assumptions, each of these approaches can produce consistent pseudo-panel inferences. However, finite-sample properties are not well understood; neither has the relevant literature provided much guidance about the relative usefulness of each approach under suboptimal conditions. This paper presents some guidance regarding the relative utility of each method for real-world data, drawing on Monte Carlo analysis based on pseudo-panel analysis of two panel data sets.
Randomized experiments are seen as the most rigorous methodology for testing causal explanations for phenomena in the social sciences. A fundamental assumption behind the analysis of randomized experiments is that of SUTVA (Stable Unit Treatment Value Assumption). SUTVA states that there is no inference between units; the assignment of an individual to treatment should have no affect on outcomes for individuals assigned to control. SUTVA rules out spillovers, the kind of communication that could occur over social networks, where treatment individuals communicate with control individuals. While this assumption is fundamental to analysis, field experiments are rarely attentive to the possibility of spillovers. Moreover, multilevel environments provide excellent opportunities for designing experiments which allow for explicit tests of spillovers. This paper presents a social network model that gives a theoretical structure for modeling spillovers under different assumptions about communication in social networks. We then develop a corresponding statistical model in which to analyze how spillovers may introduce bias in the estimation of treatment effects. We analyze actual data from an experiment conducted in California during the 2006 election that was designed to measure spillovers at both the precinct and household level. We also provide a recommendation for how to design experiments in a multilevel setting that test explicitly for spillovers by implementing a multiple level randomization design.
Sinclair, Betsy, & Delia Bailey. “Political Networks and the Impact of Term Limits.”
We examine roll call data from the California state assembly to investigate whether or not there is a change in legislative productivity with the implementation of term limits. We measure legislative productivity in terms of individual agreement between legislators who vote similarly on roll call bills. For each vote in the state assembly from 1975-2006, we measure the intensity of similarity between any two legislators by the number of similar votes (Yea, Nay or Abstain). We then use these measurements to evaluate legislative agreement via alpha centrality, a network classification of level of agreement between legislators. We compare our alpha centrality measurements over time to evaluate the impact of term limits on legislative productivity before and after the imposition of term limits in 1990. In addition, we are able to evaluate these measurements with respect to seniority and leadership position, as well as partisan affiliation. This allows us to draw hypotheses about the effect of term limits on bipartisan legislative productivity.
Spirling, Arthur, & Michael Peres. “Scaling the Critics: Uncovering the Latent Dimensions of Movie Criticism with an Item Response Approach.” «download»
We study the critical opinions of expert movie reviewers as an item response problem. We develop a framework that models an individual's decision to approve or disapprove of an item. Using this framework, we are able to recover the locations of movies and ideal points of critics in the same multi-dimensional space. We demonstrate that a three dimensional model captures much of the variation in critical opinions. The first dimension signifies movie ‘quality’ while the other two connote the nature and subject matter of the films. We then demonstrate that the dimensions uncovered from our ‘threshold utility model’ are statistically significant predictors of a movie's success, and are particularly useful in predicting the success of ‘independent’ films.
Cross-national empirical research on political and bureaucratic corruption relies on indexes like those compiled by Transparency International and the World Bank Institute. Such indexes summarize the perceptions of local business actors, foreign investors, and country experts. Recently, objective measures of corruption, based on the efficiency of public infrastructure provision, have been proposed, and used to assess the prevalence of corruption across regions in a single country. The efficiency in the transformation of infrastructure investment into actual infrastructure is affected by the degree of corruption. Data on actual infrastructure and public goods investment are available for a large set of countries. I propose a cross-national measure that treats corruption as a latent predictor of efficiency in public good provision, as well as a latent predictor of corruption perceptions. In order to recover the latent corruption variable, I estimate a simultaneous equation model via MCMC to ``triangulate'' the information provided by the perception indexes and the information provided by the data on public good investment and provision. Simulation with fake data shows that the model correctly recovers, on average, the ``true'' latent degree of corruption in each country. Finally, differences and similarities of the proposed measure with the available indexes are discussed.
Troeger, Vera E. “Problematic Choices: Testing for Correlated Unit Specific Effects in Panel Data.” «download»
The (generalized) Hausman specification test (Hausman 1978) is the gold-standard for political scientists using time-series cross-section data to check whether unit specific effects are correlated with right-hand-side variables. More than 300 articles (published in SSCI journals) over the last 20 years in Economics and Political Science used the Hausman test to justify the model choice, e.g. whether to employ a fixed effects or random effects/ pooled OLS specification. The asymptotic properties of the Hausman test and its variants are well known and formal power analyses have shown that the Hausman test performs reasonably well. Yet, the differences in the estimates of fixed effects and random effects models in finite samples can originate from two different sources: On the one hand, the Hausman test might rightly pick up differences that are caused by the inconsistency of the random effects estimator if unit specific effects are correlated with any of the explanatory variables and the random effects model therefore produces biased coefficients. On the other hand, differences might also stem from the inefficiency of the fixed effects estimator if explanatory variables are rarely changing and therefore only have a very small within variation. This inefficiency does not only lead to large standard errors but also to very unreliable point estimates that might be far away from the true relationship. While the Hausman test (and especially more recent variants and augmentations of the specification test) acknowledge the inefficiency of the fixed effects model and control for the differences in the asymptotic variances of the two estimators, this inefficiency still leads often to wrongly rejecting the null-hypothesis and the conclusion that the fixed effects estimator is the model of choice. In International Relations and International and Comparative Political Economy where many of our explanatory variables measure institutions which do not change much over time this result might be especially harmful since the fixed effects model in this case produces very unreliable point estimates. This paper analyses the finite sample properties and power of the Hausman specification test and its augmentations by using Monte Carlo experiments. It shows under what conditions, e.g. the size of the correlation between unit specific effects and explanatory variables, and the between-within variance ratio of right-hand-side variables, the Hausman test generates misleading results. Based on these analyses the paper suggests a variant of the Hausman test that takes these finite sample properties into account.
Vasiliu, Ana. “Networking Common Knowledge.”
What does it take for the interactions within a group to inhibit or to encourage the formation of common knowledge? The paper compares the information output from groups formed with and without the distinct purpose of sharing information to produce a persistent pool of shared knowledge. A collection of empirical cases points to the possibility that a small core of common knowledge in the form of standards of evidence can have a sizable effect on the emerging rules of engagement. Networking appears to economize on the repeated interactions needed to achieve coordination by individual accumulation of coincidental but not shared information. The result reflects on recent developments in the understanding of verisimilitude in a collaborative context.
Wittenberg, Jason. “How Similar Are They? Rethinking Electoral Congruence.” «download»
Electoral continuity and discontinuity have been a staple of voting research for decades. Most researchers have employed Pearson's r as a measure of congruence between two electoral outcomes across a set of geographic units. This paper argues that that practice should be abandoned. The correlation coefficient is almost always the wrong measure. The paper recommends other quantities that better accord with what researchers usually mean by electoral persistence. Replications of prior studies in American and comparative politics demonstrate that the consequences of using r when it is inappropriate can be stark. In some cases what we think are continuities are actually discontinuities.
