Graduate Student Presentations
Albright, Jeremy. “Bayesian Estimates of Party Left-Right Scores.” «download»
As a consequence of the ubiquity of spatial imagery in political science, comparative scholars have proposed a number of means for locating political parties on various policy dimensions. Yet little work has recognized that scholars of the American Congress are also actively engaged in identifying the latent preferences of political actors. This paper argues that comparative researchers can draw inspiration from advances in the estimation of left-right scores of legislators. In particular, it is possible to adapt recent applications of Markov Chain Monte Carlo simulation on Congressional roll-call data to data from content-coded election manifestos. The paper shows that MCMC produces results that correlate highly with other published scores, provides measures of uncertainty around those scores, and can uniquely incorporate prior knowledge—such as from expert surveys or previous elections—into the estimates.
Comparative analysis of elections suffers from the relatively little use of district-level aggregate electoral returns for empirical analyses compared to election research in American politics. Although a family of statistical models of multiparty vote shares has been proposed to model the compositional nature of multiparty vote shares, its application is still limited and does not generate empirical studies of district-level vote shares across various countries in the magnitude comparable to the same type of research in American Politics. The first problem is that these models cannot be used to estimate multiparty vote shares across different party systems—across districts with different number and composition of parties—in a single model. It is not uncommon to observe that the number and composition of parties running in elections vary across countries as well as across districts and time within the same country. A statistical model applied to comparative analyses of elections must accommodate this issue of diverse party systems, but the standard models of multiparty elections can be applied only to the districts with the same party systems. Another problem is that the existing models for multiparty vote shares are not readily applicable to a comparative analysis of elections across different electoral rules. The problem is twofold. First, many electoral rules allow intra-party competition—competition between candidates from the same party within a district. The existing models of multiparty elections, however, do not take intra-party competition into its perspective. Second, many countries use electoral rules with multimember districts (MMD) rather than single-member districts (SMD), and their district magnitude—the number of seats in each district—varies widely across countries and across electoral districts even within the same country. The same proportion of vote share has a quite different meaning according to district magnitude since it is translated into different numbers of seats. It is, therefore, misleading to compare across different electoral rules the impacts of certain political and socioeconomic factors in terms of their effects on vote shares. We need to gauge these impacts in terms of a comparable measurement such as winning probabilities of seats or candidates. To overcome these problems, this project proposes an extension of a statistical model of multiparty vote shares that allows: 1) a single estimation of a model across different number and composition of parties, 2) an estimation of intra-party vote shares, and 3) an estimation of winning probabilities of seats and candidates. For the first issue of different party systems, I develop a Bayesian hierarchical model of multiparty vote shares which makes possible a single estimation of the model across different party systems by partially pooling coefficients within and across districts. For the second, I also develop a statistical model of intra-party vote shares of candidates from the same party—i.e., a dependent variable is district-level vote shares of multiple candidates from the same party—which is estimated for each party separately, and propose a two-step estimation strategy of inter- and intra-party vote shares. For the third issue, I use a simulation based on the models of inter- and intra-party vote shares to estimate the winning probabilities of individual seats and candidates. With this proposed methodology, we can conduct a truly comparative analysis of elections across different electoral rules and party systems. The project uses the methodology to analyze competitiveness, incumbency advantage, and electoral cohesion of parties across developed democracies.
Bassi, Anna. “A Model of Endogenous Government Formation.”
Political parties bargain over the allocation of cabinet portfolios when forming coalition governments. Baron-Ferejohn (1989) non-cooperative theory of bargaining predicts that the proposer party enjoys a disproportionate share of government ministries. However, empirical evidence indicates that parties receive share of portfolios proportional to the amount of seats they control in the assembly, supporting Gamson's Law. This paper examines the bargaining over government formation as a process in which the role of formateur is determined endogenously inside the coalition and parties have different preferences among cabinets positions. In equilibrium, if parties have similar preferences over cabinet portfolios, the share of seats they are allocated to will be proportional to the parties' size. This paper offers an empirical analysis of the model and four alternative theories of coalition formation using data of Italian governments over the period from 1955 to 1998.
Political scientists frequently employ ratio variables such as income per capita, vote share, crime rates, or female labor force participation in their statistical analyses. Using ratios has two key advantages, in particular if a model is specified in ratio form with common components: First, akin to weighted least squares, ratio models can help reduce heteroskedasticity if the common component or scale variable is correlated with the disturbance term in a non-ratio specification. Second, ratio specifications usually require fewer degrees of freedom, which translates into additional efficiency gains. There are drawbacks, however, if a ratio specification is not appropriate, and for at least twenty years political scientists have had advice on properly specifying a model in ratio form: Do not use correlations (which do not model an intercept) with ratio variables that include common components; be wary of ratio variables if the scale variable has severe measurement error; avoid proportions (e.g. urban divided by total population) and use ratios of non-nested components instead (e.g. urban divided by non-urban population); include the inverse of the scale variable as a regressor if it is statistically significant; and verify that the ratio form specification does not in fact induce heteroskedasticity. This poster makes three contributions to this literature. First, it documents the extent to which these pieces of advice have been neglected in articles published in APSR within the last ten years. I reanalyze data from a recent APSR article to illustrate the problems caused by improperly specified ratio variables. Second, much of the literature on ratio variables is preoccupied with linear models, while my analysis extends to some non-linear specifications (chief among them fractional logit). Third, virtually all of the research in this area assumes that we know the true data-generating model that is to be estimated, but I argue that we often cannot know ex ante whether a specification with ratio variables is appropriate because we are uncertain about how the scale variable enters the data-generating process. The solution I suggest is to first estimate an unrestricted model that is agnostic as to whether the scale variable should enter the model directly as a predictor or by way of a ratio, and then examine the residuals over the scale variable as well as the joint significance of particular groups of coefficients in order to determine whether the model can safely be estimated in a more efficient ratio form. I use Monte Carlo simulations to show that this approach yields superior estimates in the face of specification uncertainty. Convenient R code to implement this procedure is provided.
Bell, Curtis. “A Multi-Dimensional Measure of Regime Similarity.”
International relations scholars often use dyads to examine the relationships among states, and these analyses require researchers to transform state-level attributes into dyadic variables. In this respect, regime type can be especially problematic. Categorical or indexed measures of regime type, such as Polity and Freedom House, obscure nuanced regime characteristics, and this problem is exacerbated when these values are transformed into measures of dyadic similarity. To capture these nuances, I measure a number of specific regime characteristics on multiple dimensions. Rather than index or collapse this variance, these values are used to place each country-year at a point in multi-dimensional space. Then, the regime similarity of each dyad is derived by calculating the distance between these points with a Euclidean formula. This method maintains more variance and specificity than conventional “joint regime/ non-joint regime” indicators, allowing for more elegant research in areas such as peace studies, interstate trade, alliance formation, and conflict intervention.
Blackwell, Matthew. “Of Doctors and Choices: Causal Effects in Time-Series Data.”
In political science, our data are often time-series data and our inference goals are often causal. Nevertheless, the connections between our data and our goals remains stagnant. The state of the art is still focused on Granger causality, a statement about correlations, not causations. We present the model of Robins (1997) which allows analysts to test for the presence of causal effects when treatments are updated over time in response to changes in covariates. We analyze the assumptions necessary to identify these causal effects and show that those treatments with simpler assignment mechanisms are fertile ground for applied use of the model.
Bohlken, Anjali Thomas. “Democracy, Autocracy and the Individual Influence of Rulers.”
Natural experiments, and instruments based on natural experiments, have become the cornerstone of dealing with many problems involving causal inference. A well-known problem with these methods is that sources of identification that are convincingly exogenous are difficult to find in practice. Another problem which has received less attention is that employing an instrument or a natural experiment could allow the estimation of the Local Average Treatment Effect, but may not allow the estimation of the Average Treatment Effect (Imbens and Angrist 1994), when in fact the latter may be the true quantity of interest. I demonstrate the perniciousness of this problem using the example of a recent study—Jones and Olken (2005)—in the "Quarterly Journal of Economics". The authors attempt to estimate the average causal impact of rulers on economic growth by focusing on a sample of leaders who died in office of natural causes. I provide evidence showing that the natural experiment employed here leads to a biased estimate of the average impact of leaders and thus leads to substantively misleading conclusions about the role of regimes in constraining individual rulers. The results from my analysis demonstrate that longer-lasting autocrats are more likely to die in office of natural causes than any other type of leader. Using a difference-in-difference approach, I also find that changes of leadership associated with longer-lasting autocrats are associated with significant changes in growth, while those associated with shorter-lasting autocrats and democrats are not. The paper brings to light a potential problem that may plague many studies that employ instruments and natural experiments as tools for causal inference.
In this paper, I will introduce a new way to understand economic voting. I argue that there is an interactive relationship between how the economy and the political environment are recognized among voters when making a vote choice. The framework for determining vote choice can be explained in the following manner: (1) During economic downturns, economic perceptions are the impetus for voters' decision making; because the economy is performing poorly, voters punish the incumbent government. (2) During economic prosperity, voters focus less on the economy and more on politics; incumbent presidents are rewarded for economic prosperity to a lesser extent because voters focus primarily on political matters. (3) During periods of mixed economic performance, voters focus on the economy; however, this focus is tinged by partisan filters. Therefore voters are concerned about the performance of the economy and use their partisanship as a short-hand device to view its performance, which affects vote choice (Fiorina 1981; Popkin 1994). Nonetheless, it is the perceptions of the state of the economy that dictates to what degree voters' economic perceptions and partisanship influence vote choice. This paper is situated within the economic voting literature to assess under what conditions vote choice follows economic trends and under what conditions it is determined by political orientation. Namely the paper will introduce a new wrinkle for understanding economic-voting by analyzing the interrelationship between voters' party affiliation and their perceptions of the economy's performance. This paper will provide a means to reassess the reasons why and understand why voters do not reward incumbents for economic prosperity to the same extent they punish incumbents for economic downturns (Bloom and Price 1975; Mercer 2005; Soroka 2006). The model used in this paper will test and analyze how voters' economic assessments influences their political decisions for U.S. presidential elections from 1956-2004. The model's framework will also be tested by an analysis of British Elections from 1974 to 2005. The common factors of economic voting will be applied to both countries to assess how economic perceptions and partisanship explain electoral outcomes. This project is significant because it clarifies the relationship between economic perceptions and political partisanship and how these factors shape vote choice. This paper will provide a better understanding of how voters make political decisions as a result of economic shifts that goes beyond the reward and punishment thesis of economic voting.
Buttorff, Gail. “Electoral Fraud in America's Gilded Age.”
The occurrence of election fraud during America's Gilded Age has been widely suspected. There is a long list of studies suggesting that election fraud was a problem during this period in American politics. Until now, most of these studies have been largely historical accounts, interpreting historical evidence that might prove the existence of electoral fraud. However, recent developments in the area of electoral studies have seen the advent of new techniques aimed at detecting instances of electoral fraud. The second-digit Benford's Law test is one such technique, which relies on identifying discrepancies between the predicted and realized distribution of vote counts. This paper extends on recent developments in election forensics. It uses the second-digit Benford's Law test in an effort to identify possible instances of election fraud during the Gilded Age. The study is largely historical, in that it focuses on US presidential elections between 1872 and 1896. The results do in fact suggest incidences of electoral fraud during this period.
Chen, Jowei. “Buying Votes with Public Funds in the US Presidential Election: Are Swing or Core Voters Easier to Buy Off?” «download»
In the aftermath of the summer 2004 Florida hurricane season, the Federal Emergency Management Agency (FEMA) distributed $1.2 billion in disaster aid among 2.6 million individual applications for assistance. This research measures the relative costs and benefits of using FEMA aid to buy votes from swing voters and core voters. First, I compare precinct-level vote counts and individual voter turnout records from the post-hurricane (November 2004) and pre-hurricane (2000 and 2002) elections to measure the effect of FEMA aid on Bush's vote share. Using a two-stage least squares estimator, with hurricane severity measures as instruments for FEMA aid, this analysis reveals that core Republican voters are most electorally responsive to FEMA aid — $7,000 buys one additional vote for Bush. By contrast, in moderate precincts, each additional Bush vote costs $21,000, while voters in Democratic neighborhoods are unresponsive to receiving FEMA aid. Additionally, by tracking the geographic location of each aid recipient, the data reveal that FEMA favored applicants from Republican neighborhoods over those from Democratic or moderate neighborhoods, even conditioning on hurricane severity, average home values, and demographics. Collectively, these results demonstrate the Bush administration's disproportionate distribution of FEMA disaster aid toward core Republican areas was the optimal strategy for maximizing votes in the Presidential election.
Chiba, Daina. “Estimating Strategic Models of Choice under Asymmetric Information.”
I construct a stochastic formal model of interstate bargaining under asymmetric information, building upon the previous works by Lewis & Schultz (2003) and Wand (2006). The perfect Bayesian equilibrium probabilities are utilized to derive the statistical model to be estimated. In a strategic environment, actors make choices in anticipation of the opponents' likely reactions, and the choice itself will in turn convey some information to the others. A major obstacle to valid causal inference in such circumstances is that, due to the strategic dependence of observations, we cannot simply relate observed outcomes to (unobserved) preferences. To explicitly model how the observable variables affect outcomes, a formal theory is particularly useful in the construction of the empirical model. The model advanced here allows the variance terms to vary so that it can capture a greater amount of Bayesian updating than the original Lewis & Schultz's (2003) model. Along with some preliminary Monte Carlo analyses, I present results from WTO trade disputes.
Christenson, Dino. “Information and Campaigns: Multivariate Matching with Exposure.”
The potential outcomes framework of causality has helped scholars question anew many findings in the social and behavioral sciences. However various limitations with the traditional matching methods may have done more to obfuscate than illuminate. In particular, complications associated with optimal non-bipartite algorithms have pushed the discipline to predominantly consider dichotomous treatment variables—or worse yet: force a dichotomous distribution onto a categorical or ordered treatment. In cases where degrees of treatment are considered, differences in the degree of treatment are often improperly handled. To that end, I further a multivariate matching method that takes into account degrees of treatment (a la Lu et al 2001). This paper is the first to demonstrate the many benefits of multivariate matching with degrees of exposure to social science questions and data. Multivariate matching with exposure takes into account the particular nature of ordered treatment variables by ensuring that pairs balance on their covariates while simultaneously diverging on their treatment exposure. In this case, the matching is non-bipartite, allowing any two observations to be paired. Following Imbens (2000) I enlist a single scalar propensity score for exposure. The algorithm required for optimal matching is derived from Derigs (1988); fortunately, Lu et al has adapted the original FORTRAN code for use in R. Given such functional simplicity and a host of appropriate questions in the social and behavioral sciences, multivariate matching with exposure should prove quite popular. I have been working with Bo Lu, a professor in the Biostatistics Department at Ohio State University, on the application of his recent methodological developments to social science data. I apply this groundbreaking methodology to study the extent to which degrees of campaign exposure affect the amount and kinds of information individuals possess about politics. The application is a natural fit for the method, as peculiar limitations of data and research design have consistently plagued this scholarship (Brians and Wattenberg 1996, Huber and Arceneaux 2007). In particular, exposure to modern campaigns is often operationalized as a simple dichotomy. Because presidential campaigns are national and pervasive, there is no clear unexposed control group, as with experimental studies; rather individuals are likely to receive varying degrees of the treatment. Furthermore, few models take into account the myriad of potential confounders to political information. Thus understanding the nexus between campaign exposure and political information requires a method that can match individuals on a host of potential confounders as well as their degree of difference on the treatment, such as multivariate matching with exposure. I utilize individual level data from the 2000 American National Election Studies supplemented by campaign exposure data from the Wisconsin Advertising Project. A comparison of average treatment effects from both models suggests that the traditional bipartite treatment severely underestimates the impact of campaign exposure on particular dimensions of political information. The result is contrary to the most recent findings in the literature, and it bolsters the sentiments of campaign advisors and democracy enthusiasts. Foremost, it suggests the danger of misspecification in various social science treatment variables.
Item response theory (IRT) models have become an increasingly popular way to estimate latent traits from political ideology to democracy. In addition to generating estimates of a latent trait for each survey respondent, IRT models can offer a picture of how each survey item behaves in relation to the latent trait. In this paper we use survey data from the American National Election Study to construct policy-based measures of ideology. We extend this analysis by estimating the degree to which each policy item discriminates between liberal and conservative respondents, the location of each item along the scale of ideology, and the consistency of each item as a measure of unobserved ideology, both across time and population subgroups. To evaluate consistency, we complete a differential item functioning (DIF) analysis, which allows us to explore ideological response differences among population sub-groups. Finally, we present the trends in item functioning in the time period between 1972 and 2004, showing the changing relationships between issues and political ideology over time. In doing so, we shed light on the changes and continuities within mass political ideology, the extent of political polarization within the electorate, and how the relationships between issues and ideology vary among social groups and between voters and non-voters.
Desmarais, Bruce A. “Discrete Measurement and Bias in Time-to-Event Models.”
Event history models have come to play a prominent role in empirical political research. A substantial proportion of these models, both parametric and non-parametric, assume that the time being analyzed follows some continuous distribution. This assumption may hold in theory, but in practice even truly continuous variables are operationalized discretely through limited measurement tools. This discrete character of measurement causes non-trivial bias in standard time-to-event models. This implies that the standard implementation of continuous-time duration models produces biased parameter estimates. In this paper I develop this bias both analytically and computationally. A simple correction based upon an existing model for censored survival times is proposed. Lastly, to demonstrate the practical utility of this correction, I re-analyze the data from a published study on the duration of conflict.
Centralized presidential review of agency rulemakings is one of the most important institutional developments of the administrative state. Since Ronald Reagan issued his famous executive order in 1981, every president has required agencies to send their pending regulatory actions to the Office of Management and Budget (OMB) for review. The OMB reviews each agency regulation to ensure that it is justified by a cost-benefit analysis and that it conforms to the president's own regulatory agenda. However, scholarship lacks both theoretical explanations and quantitative empirical tests for how different political variables impact the outcome of OMB review of agency rulemaking. The dearth of scholarship in this area raises questions. For instance, is the OMB review of agency rulemakings more vigorous during Republican rather than Democratic presidential administrations? How does the outcome of the OMB review vary with the political context in which the rule is drafted? To answer these questions, I construct an original data set consisting of all economically significant regulations reviewed by the OMB between 1981 and 2007. I perform a statistical analysis by modeling the OMB's decision to change or not change a rule under review. The model treats OMB's decision as a latent variable which is a function of the presidential administration in power plus other political and economic control variables. The empirical analysis shows that the OMB review is not a partisan but a neutral instrument that both Republican and Democratic presidents use to advance their agenda. During the Clinton administration the probability of OMB changing a rule reviewed increases by 25% in comparison to the Reagan administration and by 10% in comparison to the George W.H. Bush administration. The analysis also shows that the OMB, regardless of the president partisanship, is more likely to change the rules of liberal agencies. Moreover, the OMB is more likely to change agency regulations if the incumbent president is in a re-election year, if presidential approval increases, and later rather than earlier in the tenure of a president.
Eggers, Andrew. “MPs for Sale? Estimating Returns to Office in Postwar British Politics.”
While the role of money in policymaking is a central question in political economy research, surprisingly little attention has been given to the rents politicians actually make from politics. Using an original dataset on the size of British politicians' estates, we find that gaining a seat in the House of Commons had a large effect on personal wealth: Conservative Party MPs died with almost twice as much money, on average, as very similar Parliamentary candidates who were defeated. We find no financial benefits for candidates from the Labour party. We argue that Conservative MPs profited from office in a lax regulatory environment by using their political positions to obtain outside work as directors, consultants, and lobbyists, both while in office and after retirement. Our results are consistent with anecdotal evidence on MPs' outside financial dealings but suggest that the magnitude of influence peddling was larger than has been appreciated.
Esarey, Justin. “Bureaucratic Control: An Experiment in Strategic Budgeting.”
How closely can an executive control an administrative agency using budgetary incentives? Consonant with the literature, I assume that the executive cannot fire bureaucrats, nor directly observe their output; instead, the executive can only (a) ordinally rank bureaucrat performance, with error, and (b) offer a budget bonus to the highest-producing bureaucrats and/or threaten to cut the budget of the lowest producing bureaucrats. I assume that bureaucrats respond to financial incentives and prefer to obey elected officials, but dislike exerting effort to benefit groups with which they ideologically disagree. These assumptions imply a tournament model of employee behavior, from which I derive behavioral predictions using Quantal Response equilibrium. I test the theory's predictions using experimental data from the laboratory. The experiment divides subjects into two ideological groups, then lets them choose between working (generating income for one group), shirking (generating income for oneself only), and sabotage (generating income for the other group). The experiment manipulates whether low workers are punished with sabotage, or no punishments are offered.
While estimating ideal points in recent congresses is non-problematic (Poole 2005), determining support along specific issue dimensions can be difficult when only a limited number of votes are available. Luckily, cosponsorship data provides another manifestation of support in a policy area. Early support for legislation implies greater support along a policy dimension. I merge the date a member of Congress cosponsors legislation into a standard Bayesian MCMC item response model (Clinton, Jackman, and Rivers 2004) to obtain improved policy support scores. Finally, I use my model to gauge changes in support for gay rights policy in contemporary Congresses, an area where public support has changed dramatically in the last two decades.
This paper offers a unique framework in understanding when and why government responds to citizens' political behavior on racial issues. I argue that citizens' unconventional political actions on race changes overall public perceptions to race in a way that forces government to address a racially conscientious state of the world. Consequently, the sporadic nature of race coming on and off the public agenda influences the effectiveness of citizen's unconventional political behavior. To capture this dynamic I implement a Markov regime-switching regression model, where being racial conscientious is a latent state variable that is not directly observed by government officials, but rather is inferred through some information set that is unbeknown to the econometrician. While most Markov regime-switching models assume that regime shifts are exogenous, the model I implement relaxes this assumption and allows the Markov-switching state variable to be endogenous. The Markov regime-switching model is based on a probit-specification for the realization of the latent state. Focusing on racial and ethnic minority issues, I test my theory on executive orders, State of the Union addresses, Supreme Court cases, and protest data that range from 1954-1993. The results of the model strongly support the claim that the President and the Supreme Court is responsive to unconventional political behavior on race, however, this responsiveness is largely dictated by the oscillation of when the nation is racially conscientious and when it is not.
Goodrich, Ben. “Semi-Exploratory Factor Analysis.”
The debate between exploratory and confirmatory factor analysis (EFA and CFA) has largely been a stalemate for about 40 years now. I present a new estimator of the same population model, called semi-exploratory factor analysis (SEFA) that is part of my dissertation, is implemented in my R package (called FAiR) on CRAN, and is especially useful for the study of public opinion. SEFA almost-but-not-quite includes EFA and CFA as special cases but is more flexible and certainly captures the advantages of each while avoiding some of their disadvantages. SEFA is simply an exercise in restricted optimization where one necessary restriction is that rotational indeterminacy must be broken by allocating a user-specified number of exclusion restrictions in the coefficient matrix wherever they produce the best fit to the data. Hence, SEFA is a very difficult optimization problem that requires the use of—and many tweaks to—a genetic algorithm (namely GENOUD; see Mebane and Sekhon 2008). My use of GENOUD for optimization subject to non-linear restrictions (sometimes inequalities) is potentially applicable beyond factor analysis problems.
Goodrich, Melanie. “A Coding Methodology for Open-Ended Survey Questions.”
Open-ended survey questions promise a wealth of information to the researcher who can properly process the responses. Unlike close-ended survey questions in which a respondent is asked to select the best answer from a finite list of options, open-ended survey questions do not restrict the respondent's ability to fully express herself. This limitation of close-ended survey responses is not particularly worrisome when asking a respondent her age because the range of potential responses is known at the time of the survey's design. Thus it is possible to account for all of these potential responses when creating the set of options from which the respondent will eventually choose her response. However if the surveyor wishes to query the respondent regarding the political issues that are most important to the respondent, the limitations associated with close-ended survey responses begin to impose a structure on the resulting data that may be undesirable. By asking an open-ended question to elicit this information, the surveyor eliminates the possibility of the respondent not being able to properly answer the question because the list of options did not include the issue(s) that are most important to the respondent. The very features that make open-ended questions desirable also make the responses challenging to analyze. In this paper, I work with a set of over 25,000 survey responses. Analyzing these responses requires that I collapse what are essentially over 25,000 unique categories to a set of considerably fewer categories which are analytically useful. However, it is not immediately evident from the data in its raw form how many categories the responses can or should be broken into, let alone what these categories are. In addition, unlike other documents that political scientists have developed textual analysis methodologies for, such as party manifestos and newspaper articles, open-ended survey responses are very short and often lack proper grammatical structure. The central empirical analysis of my dissertation involves the use of the open-ended responses to the following NES question: ``What do you think are the most important problems facing this country?'' In this paper I will describe the computer assisted methodology that I have developed to identify the relevant issues of interest to respondents and to subsequently categorize each response as addressing one or more of these issues. In designing the algorithm, I had to contend with the fact that I initially knew neither what the issue categories were nor how many of them I would need. The primary innovation of the coding methodology presented in this paper is that the algorithm allows the categories to originate from the responses themselves. It would defeat the purpose of asking an open-ended survey question if the coding algorithm for the resulting responses required the analyst to first choose a set of the most politically salient issues and then to code all of the responses as addressing one or more of those issues chosen by the analyst. Once I have identified the issue categories and I have coded each response for the issue categories it addresses, I am able to show the various issues that have become increasingly important (or unimportant) to the American electorate throughout the last two decades of the twentieth century, and how the level of importance has fluctuated over time. Ultimately I intend to measure the amount of resources each political party dedicates to the problems that members of the electorate care the most about, and this algorithm allows me to identify these problems. The methodology that I present in this paper is not only useful to other researchers as a result of the data generated from using it to code the aforementioned NES question, it also provides guidance to other researchers who wish to unlock the information within open-ended survey responses in the future.
Recently developed normative theories of representation have called for the evaluation of the issues representatives discuss with constituents. But existing methods for evaluating the content of messages from representatives to constituents are prohibitively labor intensive, require strong assumptions, and are unable to provide uncertainty estimates. To overcome these challenges, I develop a Bayesian statistical model to measure a representative's expressed agenda: the issues and priorities emphasized when communicating with constituents. The expressed agenda model is capable of analyzing thousands of messages—simultaneously uncovering the topics in the texts, assigning documents to appropriate topics, and measuring the attention each representative in a legislature dedicates to the estimated topics. To estimate the posterior distribution on representatives' expressed agendas, I use a recently developed estimation technique, variational inference. Variational methods are an important alternative to MCMC for posterior estimation, particularly for posterior distributions that are poorly explored using computationally intensive methods and have yet to be employed in the political science literature. I apply the model to an original data set of over 24,000 press releases from Senate offices from the first session of the 110th Congress. The model is used to evaluate the quality of deliberation in American democracy using a set of normative standards from a deliberative model of representation.
Guillermo, Rosas, & Shomer, Yael. “Non-Ignorable Abstentions in Roll-Call Data Analysis.” «download»
How should we deal with abstentions in roll-call data analysis? Abstentions are very common in decision-making bodies around the world, and very often obey to a strategic rationale. Methods to recover ideal points from roll-call datasets — such as Nominate and MCMC IRT — are based on assumptions about the ignorability of the abstention-generating mechanism. However, the strategic character of abstentions makes the assumption of ignorability difficult to meet in practice. We discuss different abstention-generating mechanisms to understand the conditions under which they may be deemed ignorable, and extend the MCMC IRT model so as to incorporate information from abstention patterns into inference about legislators' ideal points.
In observational studies researchers often use matching to create covariate balance between a treatment and a control group in order to estimate causal effects under the assumption of unconfoundedness. In this paper we propose synthetic matching as a new method that tackles this adjustment problem from the reverse. Instead of producing matches and then checking whether balance is achieved, synthetic matching is based on a reweighting scheme that allows researchers to calibrate a set of unit specific weights so that the re-weighted groups satisfy a potentially large set of pre-specified balance constraints. Synthetic matching thereby perfectly adjusts inequalities in representation between the two groups with respect to the first, second, and possibly even higher moments of the covariate distributions. It also prevents the reduction of balance on some covariates while increasing it on others. This allows estimating treatment effects with minimal approximation error using standard inverse probability weighting estimators that reach the semi-parametric efficiency bound for the estimation of average causal effects. The calibrated weights can also be used for double robust estimation. We demonstrate the desirable properties of synthetic matching using Monte Carlo simulations. We also provide an empirical application using the Lalonde (1986) dataset. Reweighting almost perfectly adjusts the means and variances of 52 covariate combinations in this dataset, by far exceeding the level of balance obtained in the numerous previous studies on this dataset. The resulting estimate falls close to the experimental benchmark answer. Companion software is provided in the form of an R package.
Hangartner, Dominik. “A Bayesian Choice Set Model of Class Voting in the UK.”
There has been an ongoing debate about the nature of class voting in Britain. On one hand, some scholars have concluded that class is of declining importance in elections (Franklin 1987). On the other, there is also evidence contradicting this conclusion (Evans 2000; see also Elff 2007). In this paper, we (Dominik Hangartner and Marco Steenbergen) revisit the debate but from a different perspective. Using choice set models, we distinguish between class as a primary and a secondary determinant of vote choice. Class is a primary determinant if it influences the composition of the choice set, i.e. if for example working class voters systematically exclude the conservatives from the choice set and middle class voters systematically include it. Class is a secondary determinant if it does not determine the choice set but does influence which party is finally chosen. The model that we estimate is a variant on Manski's (1977) choice set model. Choice set models have been under-exploited in political science, yet offer a flexible framework for studying choice behavior, especially voting in multi-party elections. We adopt a Bayesian approach and Metropolis-Hastings sampling for the numerical computations. In addition, we propose the use of weakly-informed priors to stabilize parameters in not strictly concave regions of the likelihood. We apply the model to British electoral data starting in the early 1960s.
Bargaining games such as the Ultimatum and Principal-Agent games are commonly applied in describing political interactions such as negotiations between congress and the president or congressional oversight of the bureaucracy. Analysts typically assume that actors know perfectly each other's payoffs or that there is some constant private information about payoffs about which one can learn over the course of play. However, real bargaining involves uncertainty about players' own payoffs and often this uncertainty cannot be reduced during the game through learning. Accounting for this uncertainty provides a richer theoretical model, more reasonable behavioral predictions, and solves the statistical "zero-likelihood" problem. There are two treatments of the ultimatum game as a QRE. Yi (2005) develops a mixed QRE model which assumes that, regardless of how much uncertainty the proposer faces, the she has no interest in proposing a division of the resource they are splitting that exceeds the bounds [0,1] of the resource, either by being negative or greater in size than the resource. The result of this assumption is that it is relatively easy to calculate the likelihood of a given offer via backward induction. Ramsay & Signorino (2005) on the other hand take the observational QRE tack, using X\beta in place of \lambda u(x) to include individual characteristics of subjects in the statistical analysis. Note that \lambda is not identified because it is wrapped into the vector \beta. This allows the payoffs to take on any real values, which means that the proposer might actually most prefer to make an offer outside the [0,1] bounds of the resource. Restriction to [0,1] leads to a Tobit-like structure where the model assigns positive probability to observing offers on the boundary. Ramsay's approach offers some advantages, but it takes some work for them to show that the latent most preferred offer is always finite and unique. This paper makes several important contributions. First, it compares the implications of the different approaches taken by Yi and Ramsay to determine the difference in their substantive interpretations: Yi's approach corresponds to uncertainty about actions, and Ramsay's corresponds to uncertainty about payoffs. Next, it develops a combined model which allows for inclusion of subject-level covariates and the estimation of \lambda. Finally, it describes how to use simulation to determine the number of subjects needed to observe how covariates affect the bargaining process.
Hartman, Erin. “Appealing to the People: Debate Effects and Public Opinion.”
Debates serve a unique purpose in election in that they allow politicians to communicate directly with the electorate. A wide variety of people watch the debate, and they are exposed to all of the party leaders positions. However, the literature on the effects of these debates is mixed. Mass communication and campaign effects are inherently difficult to measure given the numerous competing messages voters are exposed to and the general stability of public opinion. On October 25, 1988, John Turner, a candidate for prime minister of Canada, took a bold stance during a debate and accused Prime Minister Brian Mulroney of having sold the country out; many people agreed that Turner won the debate. This clear change in message acts as a good natural experiment for determining if debates can have a clear effect on public opinion. Johnston, Blais, Brady and Crete (1992) showed, empirically, that those who viewed the debate had a marked increase in competence ratings, and a possible increase in vote intentions, in the period following the debate. Those that did not see the debate still showed a lagged increase, by a few days, in terms of competence ratings and vote intention. These results were derived from a rolling cross section survey conducted during the election that contained a second wave of interviews of the same voters following the election. The authors acknowledge, however, that there may be characteristics inherently different about those that watched the debate that are driving these results, and an analysis of differences in distributions of many demographic covariates supports this concern. In order to correct for the selection driving the differences among the covariates, I employ a genetic matching method in order to create a data set in which the distributions of observable characteristics is balanced across treatment and control, which can be used to estimate the average treatment effect of seeing the debate. Using this matched data set, I will conduct an analysis of the effect of the debate on voter's competence ratings of Turner in the second round of interviews and the effect on vote proportion for Turner. I will also conduct a placebo test on voter turnout rates, which arguably should not be affected by whether or not a voter observed the debate. I will also attempt to create matched data sets constrained by small windows of interview date in order to determine the effect of viewing the debate across time. I expect to find a significant difference in competence ratings among those who saw the debate and those who didn't for the first few days following the debate, after which time the information about the debate disseminates through the public to those who did not see the debate. These findings will provide insight in to how long it takes for information to travel through the public's social networks and what effect a debate in which there is a clear winner can have on public opinion.
In analyzing longitudinal data, political scientists have become well aware that correlation among observations over time is an issue that must be accounted for in order to make appropriate inferences. Specifically, when units of observation are measured for quantities of interest at several points in time, assuming that these repeated measures represent independent observations is often dubious. What has been less recognized, however, is that the independence assumptions required for regression modeling approaches are also violated in the presence of spatially correlated units. Geographic units measured at fixed time intervals represent an exceedingly common data structure in political science. But thus far, an overwhelming amount of effort has been devoted to correcting and controlling for temporal dependence, while researchers have given scant attention to the possibility of spatial dependence among observations. Though relatively less developed than that for time dependence, a toolbox does exist for the purpose of accounting for spatial dependence in statistical models for data aggregated by geographic unit. Using county-level data, this paper examines the structural correlates of the spatial clustering of voting rates in the southern United States, before and after passage of the 1960s civil rights legislation. Prior to passage, Congressional, state and local politics were one-party Democratic, and blacks were largely excluded from the process. After passage, blacks were gradually folded into the process, and two-party politics developed. Examining overall rates of participation and rates of voting for Democratic and Republican candidates as a function of demographic, economic, institutional and legal characteristics, I model voting rates as spatial autoregressive processes. In doing so, I address the question of whether a particular set of structural factors can account for the spatial dependence among observations across elections. In particular, does the spatial process change after passage of the civil rights legislation? Additionally, are voting rates consistent with a diffusion process? Model comparisons are offered between linear models incorporating spatial lag and spatial error processes, generalized linear mixed models with covariance structures that account for particular forms of spatial dependence, and Bayesian models using prior distributions to account for the covariance among observations. Results indicate that a particular specification of covariates and covariance structures may be adequate to capture the spatial dependence in the data for some years, but may represent a misspecified model in others. In some elections, the spatial structure of voting rates is consistent with a contagion process, while in others it is consistent with local heterogeneity. In either case, accounting for non-independence is critical to appropriate inference.
Hidalgo, F. Daniel. “Do Legislator Gender Quotas Change Public Policy? Evidence from Argentina.”
Using panel data from the Argentine national congress, as well as Argentine provincial legislatures, we estimate the causal effect of women political representation on public good provision. To identify the causal effect of increased women representation, we use two complementary strategies: First, we exploit a feature of the national gender quota rules that creates large discontinuous increases in the number of women legislators with small shifts in the distribution of votes across parties in each district. This discontinuity is used as an instrument for the gender share of each districts' delegation. We find that increased women representation in a given district sharply increases discretionary education spending, but decreases public investment in infrastructure. A second strategy is to exploit the adoption of gender quotas among the Argentine provinces over time. Using nonparametric matching, we estimate the impact of gender quota adoption on public good provision at the provincial level. Like at the national level, we find a large impact of increased gender equity on education spending.
Hosek, Adrienne. “Do Observational Methods Produce Reliable Results? The Use of Matching in Estimating the Treatment Effect of Class Size Reduction.” «download»
Several studies have tested the accuracy and validity of observational research methods to evaluate what estimation techniques . Randomized experiments are the gold standard of research design. When conducted correctly, such studies produce an unbiased estimate of the treatment effect for the experimental sample. Unfortunately, randomized experiments are rarely performed in the social sciences, largely due to insufficient resources. When a randomized experiment is not an option, social scientists turn to observational research methods to study the effects of a given treatment. Several previous studies have looked at the validity of using observational techniques to determine whether the reliably provide an accurate and consistent measure of a known treatment effect. In this paper, we re-eximine the work of Hollister and Wilde (2007), which did not systematically recover the experimental benchmark through propensity score analysis using data from an experimental study on class size reduction. They concluded that observational methods performed poorly based on these results. We find that they did not develop an appropriate test and thus the inability to achieve the experimental benchmark should not reflect flaws in the methodological approach, but rather stem from problems in test design.
One of the major challenges faced by researchers is that data are often available at different geographic level. For example, many researchers have collected electoral returns at the precinct level to study vote choice or voting irregularity. As precinct and Census geography are imperfectly align, there is no easy way to append Census characteristics to precincts. In the counterfactual framework, since the precinct data come with a limited set of covariates, there is a strong motivation to bring in Census characteristics to account for heterogeneity in the composition of voters and surrounding context. The conventional approach is to perform overlay function and areal interpolation using Geographic Information System (GIS). There are some major drawbacks with this approach. First, characteristics are assumed to be uniformly and homogenously distributed within each Census unit (such as tract). Second, this method conceives the environment as discrete geographic unit instead of a continuous surface. Third, there are hundreds of Census characteristics. Analysts often have to arbitrarily select a few to use for matching. Lastly, the interpolation process is labor intensive. As precinct boundary changes every election, analysts have to repeat the overlay process if the study involves more than one election. I develop an elegant solution to overcome the incompatibility in geography and to use the additional Census information to reduce unobserved heterogeneity in matching. My empirical example will take up the question—does it matter where we vote? Elections in California are held in a wide array of settings that range from churches to public facilities to private garages. Does the polling place environment affect vote choice? Do voters in religious setting vote more conservatively on proposition that invokes moral sentiments, such as abortion? Similarly, do voters tend to support propositions on public spending if they happen to vote in a public library or school? Electoral returns are collected at the precinct level for the 2005 November election. Unlike the conventional approach that arbitrarily select a few Census covariates and discard the rest of the information, I will employ the full set of Census characteristics and use principal curve to extract components that best capture the variation in context. Then using kriging, I will model the space as a continuous environment with a series of pseudo propensity scores. With spatial overlay, these pseudo propensity scores will be appended to precincts. Genetic matching will be used to account for differences in observed characteristics in the precincts that have different polling place assignment. Using sensitivity analysis, I will report the extent of unobserved bias for the following setup: a) matching only on precinct level covariates; b) matching on precinct level covariates and a few selected Census characteristics; c) matching on precinct level covariates and a full set of Census characteristics captured by my pseudo propensity scores. The last setup has the least remaining unobserved heterogeneity. In addition, the time needed to perform genetic matching significantly reduces as the full set of Census characteristics are condensed into pseudo propensity scores.
Iwanami, Yukari. “Assessing Costly Signaling Using Dyadic Time-Series-Cross-Section Data.”
Dyad-year data on international conflict have been widely used in international politics. Dyadic data (especially directed dyadic data) are useful in order to estimate the effects of interactions between players. For example, they allow us to differentiate the initiator of a crisis from the defender. By differentiating these players, we can identify the factors that affect each player's decision as well as estimate the effects of strategic interactions between players. Although empirical studies focusing on one-shot strategic interactions tend to appreciate the structure of dyadic data (Signorino, 1999), many empirical studies, which take time dependency into account, underestimate the usefulness of such data. The problem arising from this underestimation is that we cannot estimate the effects of each player's previous behavior on the current bargaining outcome. If we do not distinguish the factors that cause peace due to the defender's capitulation from those causing peace due to the challenger's inaction, we will not be able to estimate, for example, the effects of the defender's previous decision on the attacker's action today. Hence, in order to estimate the effects of interactions between players and time dependency, we use the Markov transition model. This model was introduced to political science some time ago (Duncan and Siverson, 1975). Despite its usefulness, few studies have shed light on the importance of this model until recently. Although Przeworski and Vreeland published their work in 2002, they did not intend to examine the interaction between players. In our paper, we consider a three-state Markov model with exogenous variables which accounts for the hegerogeneity and nonstationarity of the data. We use the first-order and other high-order Markov models which has not been done in previous works. The maximum likelihood method is used to obtain estimates of the transition probabilities over various time periods. Using the estimates of the transition probabilities, we obtain restricted maximum likelihood estimators. Finally, we conduct a test of the first-order assumption against high-order Markov chains. As an illustration, we apply this model to data on international conflict with directed dyad-year. We want to estimate the effects of previous interactions on decision making today as well as identify the factors that affect decisions of each player in international conflict. By comparing different models, we will show the lingering effects of previous costly signaling on current decision making. This work forms part of a doctoral dissertation on the use and effects of signaling in international conflict, which seeks to contribute to the discipline by introducing new approaches to estimate the time-dependent interactions between players.
Katz, Gabriel. “A Compositional-Hierarchical Model of Abstention under Compulsory Voting” «download»
Invalid voting and electoral absenteeism are two important sources of abstention in compulsory voting systems. Previous studies in this area have not considered the correlation between both variables and ignored the compositional nature of the data, potentially leading to unfeasible results and discarding helpful information from an inferential standpoint. In order to overcome these problems, this paper develops a statistical model that accounts for the compositional and hierarchical structure of the data and addresses robustness concerns raised by the use of small samples that are typical in the literature. The model is applied to analyze invalid voting and electoral absenteeism in Brazilian legislative elections between 1945 and 2006 via MCMC simulations. The results show considerable differences in the determinants of both forms of non-voting; while invalid voting was strongly positively related both to political protest and to the existence of important informational barriers to voting, the influence of these variables on absenteeism is less evident. Comparisons based on posterior simulations indicate that the model developed in this paper fits the dataset better than several alternative modeling approaches and leads to different substantive conclusions regarding the effect of different predictors on the both sources of abstention
Kedziora, Jeremy. “Measuring Resolve.”
Central to theories of crisis bargaining prior to the outbreak of war in internationalization are unobservable concepts such as resolve, audience costs, probability of victory in war, and others. The operationalization of such commonly encountered concepts in international relations has traditionally proven problematic for the researcher. Here we consider the measurement and causes of state resolve, and conceptualize it as a latent utility, which we define as the total amount of resources one side is willing to expend during war. We derive a structural estimation procedure to measure it, which is straightforward to implement and estimate. Doing so has three principle benefits: the computed measure is consistent with commonly used theoretical models in international relations; modeling resolve as a latent variable allows the researcher to easily quantify the uncertainty in the measure; finally, it is straightforward to estimate the causes of resolve. Since the costs of war are revealed during wartime, and these costs are crucial in determining whether or not to go to war in the first place, we apply the procedure to measure ‘revealed’ resolve of the participants in conflict during the war years of the past two centuries. The intuition is that states willing to tolerate high costs of war during the fighting must have been relatively more resolved prior to war. Our results indicate that resolve is a function of institutional parameters that indicate (1) regime type, (2) state repression, (3) the fates of leaders, and (4) material inequality prior war. The overall ranking suggests that ‘stronger’ states with less material inequality are more resolved during war, which leads to the finding that developed democratic countries are more resolved during war, due to relatively egalitarian distributions of wealth. At the same time, and somewhat paradoxically, given that democracies win more wars at less cost than autocracies, we find that non-repressive societies in which the executive is constrained are less resolved in war than repressive societies with unconstrained executives. This work forms part of a doctoral dissertation that will attempt to link the institutional structures of states more closely with its actions in the international arena. The methodological component of this work aims to treat other unobservable concepts in international relations, for example power and political relevance, as latent variables and apply new approaches and estimation techniques to determine their causes.
Kellermann, Michael. “Balancing or Signaling? Electoral Punishment in Sub-National Elections.”
In many federal countries, political parties in government at the federal level often suffer losses in state or provincial elections. Several authors (Alesina and Rosenthal, 1996; Erikson and Filippov, 2001; Kedar, 2006) argue that these losses are the result of balancing by moderate voters, who attempt to average over the policies enacted at different levels of government. An alternative explanation, however, is that electoral punishment in sub-national elections is exactly that - voters firing a shot across the bow of parties in government at the federal level. In the spirit of recent papers on voting as signaling (Piketty 2000; Razin 2003; Meirowitz and Tucker 2007), I develop a simple formal model to demonstrate that rational voters may vote counter to their policy preferences at the sub-national level in order to influence the behavior of parties at the federal level. This signaling model implies different cutpoints between the parties than those suggested by balancing models, making it possible to distinguish them empirically. The main methodological contribution of this project is to develop a coherent statistical framework in which to compare the predictions of the balancing and signaling approaches. I use a Bayesian dynamic linear model to allow the baseline level of support for parties (the "normal vote") to evolve over time. Differences in partisan preferences across sub-national units are modeled as time-invariant offsets from the time-varying normal vote, along the lines of the house effects in Beck, Jackman, and Rosenthal (2006). This approach provides a principled way to estimate the degree to which parties at the federal or sub-national level suffer when their co-partisans are in power at the other level of government. I extend the baseline model to incorporate the predictions of the balancing and signaling approaches and evaluate the empirical support for each. While the empirical examples are drawn from post-war elections in Germany and Canada, it should be useful for any country in which parties compete at both federal and sub-national levels of government.
Assessing the Effect of Terrorism on Political Attitudes Using Observational Data: A Synthetic Panel Approach Cross-sectional surveys usually offer little opportunity for assessing the causal effect of a particular event. However, when a significant political event occurs while a survey is in the field, researchers potentially have the opportunity to assess the impact of that event on political attitudes. Is such an inference possible? This project exploits a random shock (the July 2005 London Transit bombings) that occurred while the 2005 UK Home Office Citizenship Survey was in the field to explore this question. The sample is large (14,000 respondents), contains an over-sample of ethnic minorities, and has a roughly equal number of treated and control cases. The analysis first investigates whether treatment assignment (ie: whether the individual was interviewed before or after July 7) is indeed ‘random’ by assessing the balance on important covariates in the pre and post-attack samples. The project then moves on to estimate causal effects, comparing a naive estimate using the entire sample (under the assumption that treatment was random) to one produced by a synthetic panel approach. In the synthetic panel, pre-attack respondents are matched to post-attack respondents on the basis of stable observed characteristics. This set of matched pairs is then used to estimate the causal effect on outcomes of interest and on item non-response. Finally, the analysis conducts a series of placebo tests to ascertain whether either method can accurately capture a true effect. For instance, if pre-attack respondents are matched to other pre-attack cases on the same set of covariates, do they show any significant differences on the outcomes of interest? The primary goal of the paper is to explore whether such an approach to causal inference is sound and how one might test such an assumption. With respect to the substantive findings, the sample contains ample numbers of whites and ethnic minorities, respondents of high and low socioeconomic status, and a diverse range of ages, allowing the analysis to search for potentially interesting heterogeneous treatment effects. In addition, because the dataset contains information on what month the respondents were interviewed, we can also assess whether the impact of the event may have ‘decayed’ over time. Of particular interest is whether the effects fade uniformly for different demographic groups.
Research on the dynamics of collective action has paid scant attention to the potential role of foreign media in the spatial diffusion of protest behavior in authoritarian regimes. In this case study of the impact of West German television on protest diffusion during the East German revolution, we make use of a rich set of county-level covariates and detailed archival police records on protest activities. Identification is achieved by exploiting a natural experiment: West German television broadcasts could be received in most but not all parts of East Germany. We show that West German television's coverage of political dissatisfaction and protest activities in East Germany helped East Germans to coordinate their protest behavior. In the early phases of the revolution, fewer protest activities took place in counties without access to West German television than in a matched sample of counties with access to West German television. We also show that this relationship no longer existed after the fall of the Berlin Wall. As one would expect, once protests had diffused and information about protests had become readily available, the level of protest activities was no longer affected by access to West German television. A hierarchical overdispersed Poisson regression is used to multiply impute missing data on protest sizes, and genetic matching is used to find a control group of counties with access to West German television.
Kernell, Georgia. “Giving Order to Districts: Estimating Voter Distributions with National Election Returns.” «download»
Correctly measuring district preferences is crucial for empirical research on legislative responsiveness and voting behavior. This article argues that the common practice of using presidential vote shares to measure congressional district ideology systematically produces incorrect estimates. I propose an alternative method that employs multiple election returns to estimate voters' ideological distributions within districts. I develop two estimation procedures—least squared error model and a Bayesian model—and test each with simulations and empirical applications. The models are shown to outperform vote shares, and they are validated with direct measures of voter ideology and out of sample election predictions. Beyond estimating district ideology, these models provide valuable information on constituency heterogeneity, an important but understudied quality for understanding representatives' strategic behavior.
Kim, Yong Kyun. “Regime Type and Heavy Indebtedness: Quantile Regression Analysis.”
The regression curve summarizes the averages of the distributions of an outcome variable conditional on a set of predictors. While researchers are often interested in seeing if there are any significant average differences between subgroups, focusing on average behavior might be misleading in some cases. The relationship between regime type and the level of indebtedness (total external debt scaled by GDP) is such a case. While a theory suggests that democracy accumulates foreign debt less heavily than autocracy, an ordinary regression analysis fails to uncover significant differences between the two groups of countries due to the fact that the means (or the medians) of the two groups' indebtedness level do not differ substantially. However, one can find significant differences between them by looking at the entire distributions of the indebtedness level. In particular, the locations of the upper tails tend to differ starkly, and this difference in heavy indebtedness, rather than, average indebtedness, across regime types is what the theory actually suggests. Quantile regression analysis is used to capture the group differences along the entire distributions of the indebtedness level. The results show that at about 70th percentile or upper tails democratic countries accumulated significantly less foreign debts than autocratic ones by about 30-50% points even after controlling for the previous level of indebtedness. The OLS regression results, however, do not reveal such differences.
Konstantinidis, Nikitas. “The Size and Scope of International Unions: A Coalition-Theoretic Approach” «download»
This paper examines the endogenous strategic considerations in simultaneously creating, enlarging, and deepening an international union of countries within a framework of variable geometry. It essentially amounts to an examination of the equilibrium relationship between union size and scope. This an exciting research puzzle given that current game-theoretic predictions have been at odds with the empirical reality of European integration (especially throughout the nineties). Given the broad, non-issue specific nature of political unions, it seems more than plausible to assume more than one policy dimensions in the union negotiation process, thus giving rise to opportunities for issue-trading, log-rolling, as well as enhanced cooperation in the form of policy-specific subunions. What is the equilibrium (stable) size and scope of an international union and how do these variables interact? When should we expect countries to take advantage of more flexible modes of integration and how does that possibility affect the pace and depth of integration? In tackling these questions I characterize the various policy areas of cooperation only with respect to their degree of separability, their efficiency scales, the heterogeneity of preferences, and the ensuing externalities. I show both theoretically and empirically (through the examination of successive EU enlargements) that the enlargement of a union and the widening of its policy scope are too symbiotic and mutually reinforcing dynamic processes under certain conditions. From a normative point of view, a welfare analysis of policy uniformity vis-à-vis more flexible policy centralization arrangements, e.g. enhanced cooperation, may prove germane to the ongoing debate on the institutional structure of the enlarged European Union.
Kropko, Jonathan. “Using Item Response Theory to Estimate Ideology in Congress.” «download»
I use item response theory (IRT) to estimate latent ideology from selected roll-call votes in the first session of the 110th House of Representatives. Votes are selected if they are divisive, unique, but not wholly explained by party loyalties. The method is similar to the one employed by Clinton et al (2004), but does not assume a spatial structure of voting. The results demonstrate that (1) although Democrats hold a majority of the seats in the 110th House, a majority of the members have conservative ideologies, (2) the Republican party leadership is much more conservative than the Democratic party leadership is liberal, and (3) that the House is far less ideologically polarized than DW-Nominate scores would indicate.
The current matching literature offers many options for obtaining balance on a number of covariates to improve causal inference. To deal with the curse of dimensionality, one option is to collapse multiple dimensions into a scalar summary of the covariates via the propensity score, which we can then match on. However, the properties of matching via the propensity score are less well understood. For example, what is the relationship between the number of dimensions to match on and the number of observations for the prospects of achieving good balance? Which estimators of the propensity score perform the best and under which circumstances should we choose one estimator over another? This study attempts to examine the properties of matching with propensity scores using Monte Carlo simulations. The goal is to offer practical advice and rules of thumb for researchers interested in making causal inferences via matching and propensity scores.
Since the publication of Bowling Alone: America's declining social capital (Putnam 1995), political scientists as well as policy makers have started a quest for instruments to increase social capital in society. Within the tradition of the Chicago School, a number of studies focussed on how characteristics of neighbourhoods influence the social connections among inhabitants (e.g. Coulthard et. al. 2000, Wickrama & Bryant 2003). Just as many other studies with a geographical dimension, this research field is confronted with major difficulties regarding the operationalisation of aggregated research units. This general problem will be given more insight by treating the specific case of the definition of neighbourhoods in the research tradition described above. Whereas defining the frontiers of a neighbourhood excludes from the model the facilities located just outside these artificial boundaries, this particular information might have an important impact on the results. Therefore, the technique of spatial multilevel regression analysis makes it possible to take this information into account. The main focus of this paper will be to examine whether using this technique can influence the conclusions about the relation between neighbourhood facilities and social connectedness. We will do this by comparing the results of a classic multilevel and a spatial multilevel model. The data used in the research are from the General Life Quality-survey of Ghent (Belgium) during the year 2006 (n=1756).
Lauderdale, Benjamin. “Locating Supreme Court Opinions in Doctrine Space”
The development of powerful theoretical models of intra-court bargaining and judicial hierarchy has created a need to measure the doctrinal content of Supreme Court opinions. Such theories yield empirically-testable predictions, but the analysis of these predictions has been limited by the ability of scholars to measure judicial policy. This paper develops an original scaling model to estimate opinion locations along a continuous dimension using the citations between opinions as data. We assume that each opinion has a fixed location in a unidimensional doctrine space and that the probability of a citation that affirms rather than disputes the doctrine of the precedent decreases as the doctrinal distance between them increases. This proximity citation model is applied to original data based on freedom of religion opinions written by the Warren, Burger and Rehnquist Courts. We use the resulting estimates of opinion content to evaluate median and non-median voter theories of Supreme Court bargaining and opinion writing. We find striking empirical support for theoretical models that predict the majority opinion will fall at the ideal point of the median member of the majority coalition as opposed to the court's median or the opinion author's ideal point.
It is widely believed that proportional rather than majoritarian electoral systems should be more effective at bringing minority and underrepresented groups into politics. There is, indeed, weak evidence that more proportional electoral rules do a somewhat better job of descriptive representation in some realms. For instance, there are more elected women in legislatures whose members are returned by PR, ceteris paribus. Survey evidence that proportionality predicts more positive affect for politics or more attachment to particular parties, or the political system, however, is decidedly mixed. PR systems have a better record of descriptive representation of women in parliament, but higher proportions of female legislators have not necessarily strengthened the relationships between women and government. Cross-nationally, women remain less interested in politics and less attached to parties than men, even in places with comparatively strong historical records of women's representation. One difficulty in interpreting this apparent discrepancy is that electoral laws rarely change, and so a variety of contextual factors are probably collinear with proportionality of the (legislative) electoral system. One manner of mitigating issues of comparability is to eschew cross-national comparison and focus, instead, on a (rare) case where we observe a change in electoral rules, from first-past-the-post to PR or vice versa. We examine survey data from the multiple waves of the New Zealand Election Study, before and after the country's switch to PR. Substantively, our focus will be on women's attachment to the political system, broadly construed. The study will also have a distinctive methodological purpose. To manage the before-after comparison, we will use propensity-score matching for causal inference. Matching in longitudinal survey data is already somewhat novel, but there is an especially attractive feature to these data. Because the NZES includes a panel component (1573 respondents participated in both the 1990 and 1993 surveys), We will have a natural baseline with which to compare the results from the matched samples, drawn from the remaining respondents, who were randomly chosen from the electoral roll. Matching has primarily been used in quasi-experimental studies, often without a longitudinal component, and so these New Zealand data present a unique opportunity to assess how well matched samples do at approximating an observable longitudinal change in the panel of individual actually observed at the two different times.
The purpose of this project is to estimate the causal effects of transitional justice measures on democratic stability using graphical causal modeling (Bayesian networks) and (probabilistic) nonparametric structural equation approach. Transitional justice refers to the ways in which emerging democratic regimes deal with the legacies of their authoritarian predecessors, ranging from amnesty to prosecution. Two of the commonly used strategies to estimate the causal effects when back-door path causal mechanism (indirect causal effects) is present is either by balancing conditioning via stratification or weighting (for example, "matching")or adjustment-conditioning (commonly regression"). I introduce alternative strategies of conditioning when simple conditioning like matching or regression is ineffective or infeasible and when inductive causation is only possible strategy, using Pearl's front door criterion, discuss the conditions under which those strategies are effective, and, then, apply to the estimation of transitional justice implementation on democracy. Recently, some scholars argue that there is elective affinity between the preconditions of democracy such as “mode of transition” or “level of economic development‘ and successful transitional justice and estimate the effects of transitional justice either by conditioning on determinants of transitional justice implementation or by conditioning on the determinants of the stable democracy (and putting the transitional justice measures as one of the regressors). However, neither of the conditioning strategies consider the effects of unobservables (such as culture, rule of law, the anticipated effects of transitional justice) that affect both transitional justice measures and democratic stability simultaneously. I will employ the alternative conditioning strategies on unobservables (political culture and social structure) and estimate the causal effects of transitional justice on democracy (understood as a latent variable) with using weaker assumptions than the ones for traditional (deterministic) structural equation modeling. I use a cross-national dataset on transitional justice measures (about 5000 country-year) and other indicators of democracy that I have pulled together from Polity IV, the World Value Survey, various barometers (the Asia Barometer, the Latino barometro, Afrobarometer), Alvarez et al.(ALCP, 2000), Gibson (2001) and Backer (2004).
Lee, Han Soo. “Three Way Information Flow between the President, News Media, and the Public.”
This study aims at solving two puzzles: Do presidents respond to the public? Do or can presidents lead the public? Unlike prior research, this study tries to solve the puzzles by focusing on the news media as an actor that intervenes in the relationship between the public and the president. Despite the fact that the questions this study addresses are theoretically and empirically important to understanding democratic representation, no empirical study systematically examines how the three actors simultaneously interact with each other. This study argues that the relationships between the president, the news media, and the public are potentially reciprocal. To estimate the possible multi-directional relationships between variables, this study uses Vector Autoregression (VAR) methods. The VAR approach is used to examine the direction of the relationship between the three variables. Granger tests are used to examine Granger causality between the three variables. However, the absence of Granger causality does not necessary mean no causal relationship between variables in the system. As an alternative of VAR, this study uses the Moving Average Response models, which are simulations based on the estimated system. For this study I construct a data set of the news media's bias on issues (Media Liberalism) through time from 1958 to 2004. Media Liberalism is constructed based on how many liberal and conservative news stories the news media carry. If the news media express a liberal bias, there will be more news stories favoring liberal policies than conservative policies. This relative measure for the news media issue stance is comparable to the policy preference measures for the public and the president, which is available from prior research. Analyzing the dynamic relationships between the president, the news media, and the public from 1958 to 2004 in the U.S., this study uncovers that the news media influence both the president and the public. However, the public conditionally affect the president, vice versa. The president responds to the public only if the public moves together with the new media. The president affects the public only if the president speaks harmoniously with the news media. It implies that the news media intervene the relationship between the president and the public.
Levy, Naomi. “Identifying Identity: A Latent Variable Model with Multilevel Data.”
This paper contributes to the conceptualization and measurement of national identity as it is realized at the individual level, arguing that individual-level identity is best measured using a latent variable approach. I develop a latent variable model of students' state and ethno-national identities using data from twenty-one schools in twelve towns in Bosnia-Herzegovina (B-H) and Croatia collected in May 2004 and May 2005 with an original survey instrument (N = 4,457). However, since the data have a multilevel structure, with students nested within schools and towns, I must account for the hierarchical nature of the data while modeling the effects of school organization, curriculum and town-level group dynamics on the identities of these students.
Llewellyn, Morgan. “Experimental Evidence of Information Aggregation in Committee Decisions.”
One source of information within committee decisions is testimony by self-interested, outside parties. This paper investigates the ability of uninformed committee members to incorporate information into the committee decision through the recommendations of informed outside parties who possess incentives to manipulate the committee decision. Investigating this question through experimental design supplements traditional theoretical methods. The theory of such complex phenomena is incomplete. The process through which information aggregation in committees is facilitated by various institutions is almost unexplored. The experimental method helps to fill both gaps. The laboratory research design is as follows, a state dependent equilibrium exists where the underlying state of nature is chosen at random from a two dimensional set of possible states and all preferences depend upon the state drawn. The experiments consist of seven participants, five voting committee persons and two non-voting participants who have the ability to make proposals and recommendations but cannot vote. The five voting committee persons do not know the state and thus have no direct information about their own preferences. Deliberations follow procedures similar to Roberts Rules of Order and end with a choice chosen from the Cartesian plane. Two additional participants know the true state with certainty but their preferences are in conflict with each other and also with the voting committee members. The two additional members cannot vote but can make proposals. The key result of this paper is the committee process, under Roberts Rules, aggregates information regarding the true state and this information is reflected in the final decision. Furthermore, institutions which promote the development of reputations among the informed parties improve the accuracy of the final decision point relative to the fully informed equilibrium.
Malecki, Michael. “MCMCirtHier1d: Subject-Specific Covariates Implemented in MCMCpack.”
I describe and illustrate a one-dimensional hierarchical item-response model with subject-specific covariates on latent ideal point (also called "ability") parameters. The model is implemented in Martin and Quinn's MCMCpack framework (compiled C++ called from R). With this new MCMCpack module, researchers can take immediate advantage of a fast, stable sampler using tools with which they are already familiar. Because we may now be interested not only in the values of latent parameters but their determinants as well, I provide a separate function, suitable for parallel computation, to sample the marginal likelilhood of the model, enabling formal comparison of non-nested models via the Bayes factor. I include illustrations using (1) simulated data; (2) the Supreme Court 2000 term data, used as an example for MCMCpack's basic IRT1d; and (3) a more involved application to the European Parliament, where predictors for ideal points include Party Group, Member State, and a survey of legislators.
For ordinal and/or multinomial response variables, Anderson's (1984) stereotype model (or reduced-rank multinomial logit model) gives a single set of regression parameters and a series of ancillary parameters (often noted as phi) for each scale category. Anderson proposed that these parameters could be used to evaluate the ordinality assumption of the response variable as well as distinguishability of adjacent scale scores. Unfortunately, Anderson died prior to the publication of this work and applications of the model, especially in social sciences, have been sparse. Yee and Hastie (2003) renewed interest in the stereotype model, and since then, the Anderson model has been implemented in a variety of software packages readily available to social scientists, including R and Stata. Yet despite the accessibility of the model, few have applied the stereotype model to evaluate either the ordinality condition or distinguishability of scale scores. Researchers have considered the issue of collapsability of scale scores on parameter estimates (c.f. Holtbrugge and Schumacher 1991), but the sensitivity of the phi parameters to model specification remains an open question. In this paper, I use Monte Carlo simulations to examine the sensitivity of the phi parameters to a variety of model specifications. By determining the stability of the phi parameters over various model specifications, social scientists may have more confidence in the appropriateness of the stereotype model in drawing conclusions about both ordinality and distinguishability of response categories.
Spatial models of voting predict citizens' feelings of alienation and indifference strongly affect their decisions to vote, or not, in elections. Despite strong theoretical expectations about the relationship between these concepts and turnout, the magnitude and significance of the effect alienation and indifference have on turnout varies greatly between studies. One explanation for this variation is that different measurement techniques have been used to operationalize alienation and indifference between studies. Every study seems to use a slightly different method to operationalize these concepts and the validity of the resulting array of measures is questionable, at best. The purpose of this paper is to test the criterion-related validity of several common measures of alienation and indifference using a laboratory experiment to determine the most valid measure. The experimental findings suggest graded measures of alienation and indifference are more valid than dichotomous measures of these concepts and, in multiparty contexts, measures of indifference that include information about the spatial location of all parties are more valid than those that only include information about two of the parties.
Is the conventional Generalized Least Squares (GLS) estimation or currently prevailing Panel Corrected Standard Errors (PCSEs) method appropriate to analyze a dynamic Time-Series Cross-Section (TSCS) model? Many political studies (the majority of which are in comparative/ international political economy) attempting to investigate TSCS data by applying frequentist approach, which is relying on point estimates under asymptotical assumption, are deficient, especially when the studied samples and time-periods are limited. Although researchers have been aware of the common unit-specific serially correlated and jointly endogenous characteristics among TSCS data (Arellano and Bover 1995; Beck and Katz 1996, 2004; Wawro 2002; Wilson and Butler 2007), their methodological prescriptions somehow suffer difficulties to conduct data analysis in practice. The Bayesian approach, which is not required to attain large-sample size and to report precise one point estimates, but rarely used in political research, is a better alternative method to an analysis of TSCS data if the model would be considered dynamic and the analytic data is limited due to which consists of small sample sizes, missing data or unobservable variables. In this paper, I will specifically apply the Bayesian model to investigate how and why a catch-up developmental state chooses its exchange rate regime and manipulates its currency level, conditioned to certain domestic political institutions (i.e. veto players and regime types) and international economic structures (i.e. intra-regional trade integration and monetary relations). My major study cases will be ten East and Southeast Asian countries during the periods before and after the shock of Asian Financial Crisis in 1997.
Middleton, Joel A. “Is Matching Really ‘Essential’?” «download»
Imai et. al. (2007) argue that matching is ’essential’ for the analysis of cluster randomized experiments. But is matching really ‘essential’ The authors do correctly note that regression for the analysis of cluster randomized experiments can be biased (see also Middleton 2008a, forthcoming Stat. and Prob. Let.). However, this presentation argues that, in fact, prematching can be seen in a broader context of restricted randomized design—which can obviate the need for covariate adjustment and parametric assumptions more generally—while postmatching does little to resolve parametric biases. In this study several methods are compared including: (1) matching, (2) regression, (3) stratified design with unequal probability of assignment between strata (a restricted randomized design) and (4) nonparametric covariate adjustment. Methods 3 and 4 are known to sampling theorists but new to experimental science. The discussion covers an important motivation for the use of these nonparametric methods that differs from the usual motivation given for matching, namely, that regression assumptions are not justified by randomization (Freedman 2007, Middleton 2008a, 2008b). Also discussed are common misunderstandings about the relationship between randomization and regression assumptions that perpetuate the use of methods such as logistic regression that are not generally consistent for randomized experiments (Middleton 2008c).
Min, Brian. “Should We Trust Government Data? Detecting Measurement Error by Satellite.”
Measurement error is a widely acknowledged but underestimated source of bias in the analysis of many cross-national datasets. Compendiums of country-level data like the World Development Indicators underlie many claims about the impact of democracy on the provision of basic public services to their citizens. Yet when measures of service provision are reported with error, coefficient estimates will be unbiased only if the measurement error is not systematically related to democracy or other explanatory variables. I present evidence showing this assumption to be markedly false: many governments misreport measures of public service provision to international agencies and the measurement errors are highly correlated with political factors. To illustrate the potential magnitude and effects of measurement error, I compare official government estimates of the share of a country's population that lacks electricity against new satellite-derived estimates that detect unelectrified populations through an objective and unbiased procedure. These satellite estimates are based on remotely sensed images of the earth at night, providing a unique indicator of the presence or absence of outdoor lights and basic electrical infrastructure across the entire globe at a resolution of 2.7 km. While the official and satellite estimates are highly correlated, the deviations between the two measures are substantial and not random. Many governments misreport the share of their populations lacking electricity, and the magnitude of the inaccuracies are easily predicted by measures of bureaucratic capacity, corruption, and political freedoms. The results are strongly consistent with the expectation that governments respond to different incentives in the collection of accurate statistics on state performance. The findings call into question the quality of inference that can be drawn from the analysis of cross-national datasets compiled from self-reported government data and may help explain why a generation of research on the effects of democracy have proven so inconclusive.
My dissertation asks the question, why do politicians take seemingly bad long-term strategies on the issue of immigration? I aim to explain what leads politicians to raise the immigration issue and what the long-term consequences of doing so are. To do this, I formally model how parties will place themselves ideologically if their placement has to hold for two elections while the electorate's mean ideology shifts in a known way. Yet, I do not use a deterministic game because the only solution is a mixed-strategy Nash equilibrium. So rather than make parties place themselves probabilistically, it seems more realistic to have voters vote probabilistically by adding a stochastic disturbance, representing private utility, to their utility functions. The disturbance term raises a methodological puzzle, however. The utility function to parties across two elections becomes a sum of distribution functions, either logistic or normal, which are transcendental. In short, solving for the best response function from the derivative of this utility function requires solving for the party's behavior parameter, located within summed exponentials. That is algebraically impossible. To resolve this, I use simulation techniques to determine the comparative statics of this model. This poster demonstrates how I impute values into each party's expected utility functions to determine best responses across a range of parametric scenarios. This procedure distinguishes parameters that have no theoretical consequence on behavior from those that do, revealing the direction and intensity of the parameters' effects. Besides explaining the simulations, my poster also demonstrates how I apply these expectations to data when I model immigration roll call votes in the U.S. Senate as well as state-level passage of laws regarding immigration. As an application of creating empirical tests of formal models, this project speaks to an important development in game theoretic research. Formal theory in political science has successfully served to explain complex social situations in an elegant way. In doing so, nearly all formal models are reduced to simple specifications that can be solved in closed-form Nash equilibria. Increasingly, though, social scientists are finding it useful to add complexity to their models' assumptions. The downside to this, apparent in my project, is that finding best response functions from which equilibria and comparative statics are derived can be difficult or even impossible. The benefit, though, is that the assumptions are often more realistic. Thus, when the results of simpler models fail to hold up under added complexity, it seems quite likely that the simple model was insufficient to capture dynamics of the real world. My project offers a working example of how to thoroughly evaluate more realistic models.
Despite the importance of the concept of representation for democratic theory, its exact definition and meaning have been a point of continual debate and discussion. In my poster, I specify a model testing the theory that Senators are in fact responsive to multiple constituencies (Fenno 1978). In previous empirical studies Congressional behavior has often been framed as a dichotomy. Senators are said to either be representing their constituencies or else "shirking" in favor of partisanship or personal ideology (e.g. Rothenberg and Sanders 2000). I take the viewpoint that Senators may be simultaneously responsive to multiple constituencies that they need in order to achieve re-election. In particular I argue that Senators are responsive to both the general electorate constituency and the activist community upon which they depend for needed electoral resources. My poster, based on my thesis to be submitted for an M.S. in Statistics, will demonstrate some of the empirical implications of this more complex description of the linkage between public opinion and roll call votes in the U.S. Senate. In addition, it will make use a statistical framework that will allow me to deal more flexibly with some of the methodological difficulties surrounding the study of representation. This paper offers two primary contributions to the extensive literature on representation. First, I will use data from the 2000 and 2004 National Annenberg Election Study (NAES) that allows me to make more certain inferences about public opinion within each constituency due to its comparatively large sample size. Previous studies of representation have often been forced to make use of within-state sample sizes as low as 19 (Miller and Stokes, 1958). In addition, from within the Bayesian paradigm it is relatively straightforward to include our uncertainty about district-level opinions (which may vary considerably from state to state) in our final inferences. I will also use generalized linear mixed model (GLMM) techniques to demonstrate that Senators tend to be differentially responsive to these two communities depending on state-level characteristics. There are two main advantages for conducting an analysis using this techniques. First,previous studies of representation have primarily focused on dependent variables which scale roll-calls for each Senator into a single indicator of policy preferences. Using the NAES data and the GLMM framework I will be able to dispense with these practices by using relevant state public opinion information to predict issue specific roll-calls. A second major advantage of the GLMM approach is that I will be able to examine the conditions under which responsiveness to roll-calls (represented by random senator-specific coefficients) is moderated by constituency-specific traits. Works Cited: Fenno, Richard. Home Style: House Members in Their Districts. Boston: Little, Brown and Company, 1978. Miller, Warren E., and Donald E. Stokes. "The American Representation Study, 1958: Candidate and Constituent." Ann Arbor: ICPSR, 1958. Rothenberg, Lawrence S., and Mitchell S. Sanders. "Severing the Electoral Connection: Shirking in the Contemporary Congress." American Journal of Political Science 44, no. 2 (2000): 310-19.
Nyhan, Brendan. “The Network of Presidential Scandal Allegations in Congress.”
While political scientists have devoted a great deal of effort to studying the legislative aspects of interbranch conflict in American politics, quantitative research has neglected another prominent domain in which the branches frequently clash—scandal. In particular, we do not understand the ways in which members of Congress attempt to achieve their political objectives by making accusations of scandal against the president. This paper presents a network analysis of a unique dataset of presidential scandal allegations in the Congressional Record for 1985-2006 (the period for which the full text is searchable online). I focus specifically on references to specific actions taken by the administration or past actions by members of the administration as "a scandal," "-gate," or violations of the law. This simple and well-defined coding procedure should reduce error and enhance replicability. In total, I coded 141 specific allegations of scandal that were made in Congress between 1985 and 2006. These were repeated 1738 times by members of Congress (the maximum number of repetitions for one scandal allegation was 267). This dataset can be characterized as a bipartite social network with edges connecting members of Congress to scandal allegations. Building on recent studies of committee and cosponsorship networks in Congress (Porter et al 2005, 2007; Fowler 2006a, 2006b; Zhang et al 2007), I analyze the network structure of scandal allegations for each two-year term of Congress from 1985-2006 and characterize the positions of members of Congress within those networks. Specifically, I examine popular notions about central figures in Congressional scandalmongering using measures of network centrality, consider changes in network-level measures of connectedness and density over time, and characterize the partisanship and importance of different allegations. Methodologically, its focus on bipartite social networks differs from previous studies of social networks published in political science journals, which focus on unipartite networks of legislative cosponsorship and Supreme Court precedent (Fowler 2006a, Fowler et al 2007).
Ono, Yoshikuni. “Coalition Management and the Distribution of Coalition Payoffs.”
The formation of multi-party coalition government is very common in many parliamentary democracies in Western Europe. When a group of parties decides to govern collectively, they have to reach an agreement about how to allocate power among them. They typically make this decision through allocating portfolios of cabinet ministries. Empirical evidence from Western European countries shows that the party to which the prime minister belongs (PM's party) seldom receives more portfolios of cabinet ministries than its seat share in the coalition government. This empirical evidence contradicts conventional wisdom, which argues that the PM's party should receive a proportional or more than proportional share of portfolios. Why do many prime ministers in multi-party governments over-reward their junior coalition partners in distributing coalition payoffs? When does the PM's party receive a lower share of cabinet portfolios relative to its seat share? Despite the large amount of literature on portfolio allocation, the disproportional outcome and much of variation in the gap between portfolio shares and seat shares have yet to be accounted for. I will answer these questions by considering prime ministers' management activities. My paper brings new perspectives to our understanding of portfolio allocation in parliamentary democracies. In my paper, I provide a theoretical explanation for understanding the underlying mechanisms of portfolio allocation and the conditions under which the PM's party takes a lower share of cabinet portfolios relative to its seat share. In particular, I argue that the prime minister, who has a goal to maintain power, uses portfolio allocation to manage the coalition government successfully, because how portfolios are allocated affects not only the coalition formation, but also affects how much the coalition government can accomplish. To explain mechanisms of portfolio allocation, I construct a game-theoretic model, in which I show that the allocation is determined by the need to balance two factors: the threat that the junior coalition partners might leave the coalition and the need for the prime minister to obtain policy concessions from the junior coalition partners. The model allows me to explain the disproportional pattern and the variation in portfolio allocations over time that the current literature does not explain. My empirical analysis evaluates the implications of the model using data drawn from portfolio allocations in thirteen Western European countries throughout the post World War II period. The empirical results demonstrate that the distribution of coalition payoffs is not determined solely by the share of seats that parties contribute to the coalition; instead, the prime minister changes the allocation depending on the credibility of the exit threat that the junior coalition partners have and the necessity of policy concessions from junior coalition partners in managing the government. My study shows that there is more politics involved in the bargaining of cabinet portfolios than the existing literature suggests.
Owen, Erica. “A Spatial Econometric Approach to the Political Economy of Contagions.”
Currency crises spread across borders as a result of a contagion and can have significant political and welfare consequences as evidenced by the Asian crisis. There are many theories about the mechanism through which this contagion occurs, however there has been little empirical testing of this mechanism in political science. Previous quantitative work has demonstrated that political factors can affect the onset of a speculative attack, but has treated the possibility of a contagion as a nuisance to be controlled for. Spatial dependence resulting from financial integration is substantively interesting because it can help us understand the channels through which a crisis starting in one county can spread to neighboring countries that may have sound macroeconomic fundamentals. Haggard (2000) argues that politics, in particular transparency and political uncertainty, were a critical factor in determining vulnerability to a speculative attack during the Asian crisis. Investor decisions determine whether a country will be subject to a speculative attack. In finance and economics, there are three important theories that can be used to explain this decision-making and thus the spread of crises: portfolio management theory, rational contagion and herding behavior. Substantively, I seek to determine which mechanism best explains the pattern of contagion and how political variables factor into this process. There have been three main approaches to the study of spatial dependence in political science. The most common method has been to control for the occurence of the event of interest, in this case, a currency crisis, in other countries. Another approach has been to include a spatially lagged dependent variable. In a previous paper, I incorporated financial contagion into a Cox conditional gap time model of speculative attacks using three specifications of a spatially lagged dependent variable to test the mechanisms above. However, using this approach or ignoring spatial dependence completely is problematic for several reasons, including overemphasis on the importance of unit-level variables (political or economic), as well as possible bias and inconsistency of estimates. In this paper I would like to do the statistical part right and therefore will look at the advantages of a third approach which utilizes more sophisticated estimators from spatial econometrics, like the spatial probit, as illustrated in political science by the work of Hays and Franzese. I plan to look at patterns of contagion using spatial multipliers. Beyond this, I hope to incorporate the temporal aspects of the issue, possibly using a Bayesian spatial survival model. I would appreciate the opportunity to discuss this with top methodologists and get feedback from experts in spatial econometrics.
Pang, Xun. “Binary and Ordinal Time Series with AR(p) Errors: Bayesian Model Determination for Latent High-Order Markovian Processes.” «download»
In models for time series discrete responses, serial correlated errors cause serious problems of biasedness andinconsistency as well as inefficiency. To directly and adequately correct serial correlation in binary and ordinal response data, this paper proposes a probit model with errorsfollowing a $p$th-order autoregressive process, and develops simulation-based methods in the Bayesian context to handle computational challenges of posterior estimation, model comparison, and lag order determination. This time series modeldoes not depend on initial values by applying a mixed sampler ofthe Gibbs and Metropolis-Hastings algorithm. As for Baysian model determination, the auxiliary particle filter, complemented by the fixed-lag smoothing, is extended to approximate Bayes Factors formodels with latent high-order Markovian processes. Computationalmethods are tested with empirical data. Effectiveness of energy cooperation policies of the International Energy Agency onglobal oil-supply security are analyzed. Multiple models with different lag orders, together with other competitive models,are estimated and compared.
Pemstein, Daniel. “Predicting Strategic Roll Calls with Legislative Text.”
Roll call vote analysis is an increasingly popular tool in comparative legislative studies. Yet, unlike in the US Congress, where roll call votes are the norm, many legislatures studied by comparativists use roll calls for only a subset of votes. In many cases, political parties may strategically call roll to discipline members, embarrass opposing political parties, or signal their policy positions to a variety of audiences. While researchers are well aware of the potential for bias that this strategic selection introduces into roll call vote analyses, we lack predictive models of the roll call generating process. I apply statistical natural language processing techniques to legislative text to predict the occurrence of roll call votes in the European Parliament (EP), use the resulting model to examine theoretical accounts of parties' incentives to request roll calls, and quantify aspects of the selection bias characterizing EP roll call data.
Powell, Eleanor Neff. “Partisan Entrepreneurship and Career Advancement in the U.S. Congress.”
In recent decades there has been tremendous evolution in the long-standing relationship between partisan entrepreneurship and career advancement in Congress. Fund-raising for the party and congressional colleagues is a form of partisan entrepreneurship which was pioneered by a few individuals. Overtime, this practice has expanded and been formalized into cartel-like behavior by the parties with party leaders providing explicit incentives (both positive and negative) with the goal of maximizing the party's electoral success. To analyze this transforming relationship, I have compiled a new dataset composed of partisan entrepreneurship activity, legislative entrepreneurship activity, seniority, party promotions and committee promotions from 1980 to 2006. I use Kalman filters and Gibbs sampling using data augmentation to address the temporal variation in these relationships. Preliminary results suggest that partisan entrepreneurship in the form of member to member contributions has evolved into a primary determinant of career advancement. The increasing importance which members place on this partisan entrepreneurship activity has potentially important implications for policy-making and representation in Congress as members better able to contribute to others accrue more power in Congress.
Ragan, Robi. “The Complex Adaptive Congress: Inherited Status Quos.”
The use of complexity science and agent based models has been growing within social science in general and specifically within political science. The tools of complexity have been used to examine a wide range of social science issues such as cooperation (Schelling, W.W. Norton 1978), elections (Kollman, Miller and Page APSR 1992), political economy (eds. Kollman, Miller and Page MIT Press 2003), macroeconomics (Sargent, Oxford University Press 1994), and international relations (Cederman APSR 2003). As of yet, there has been little application of the tools of complexity science to issues associated with American Political Institutions. This paper advocates taking a complex adaptive systems approach to modeling the United States Congress. Traditionally, formal models of Congress have used a static, single dimensional and game theoretic approach to modeling policy formation. However, I believe that there are analytical gains to be made by using an agent based, complex adaptive approach. Systems with dynamics, heterogeneous agents, non-linear interactions, adaptation, positive and negative feedback, and externalities are prime candidates for an application of the complex adaptive systems approach. In such systems, analyzing the affect of any one variable on another, or even several variables on another, without considering the feedback in the system and underlying heterogeneity of preferences among agents will likely lead to theoretical predictions which are biased, overly simplistic, or only applicable to a highly restricted set of cases. The Congressional policy making system has many of the previously mentioned characteristics of a complex system. In order to illustrate the analytical leverage of using the complex systems approach, I first build a set of agent based generative models that computationally replicate the three most prominent models of congressional policy making: the distributive/committee model (Shepsle and Weingast APSR 1987 and Weingast and Marshall JOP 1988), the party cartel model (Cox and McCubbins, University of California Press 1993 and Cambridge 2005), and the pivotal politics model (Krehbiel Michigan Press 1991 and Chicago Press 1998, Brady and Volden Westview 1998 and 2006). Each of these models exogenously assumes which agents in the policy making system are pivotal to the policy outcomes. In the distributive model, the congressional committee with jurisdiction over the policy area has considerable influence over the policy outcome. In the party model, the majority party has considerable influence over the policy outcome. In the informational model, there are pivotal actors (based on institutional rules) who have a strong influence on the policy outcome. With the dominant models replicated, I begin to add elements of a complex system to the models. This paper is part of a larger project whose end goal is to formulate a complete generative and agent based model of the congressional policy making system. For this part of the project, the aim is to look at the implications of making the existing models fully dynamic. Each of the dominant congressional models achieves its equilibrium predictions for each policy at time t from a static game. When the existing models are made dynamic, and status quos are allowed to be inherited across time periods, radically different equilibrium predictions are made for each of the models. For example, the party cartel model (which normally predicts that outcomes will be at or near the median member of the majority party) begins to predict median outcomes. Conversely, the pivotal politics model (which in the static case tends to be more majoritarian) predicts outcomes that are significantly different from the location of the median member of the chamber. Next, rules for learning and adaptation are added to the decision making process for each member of Congress. For example, in the party cartel model, the party leaders take into account the centripetal nature of the dynamics of Congress and adopt a different set of strategies depending on the political environment in each time period. This decision rule is forward looking and adaptive with respect to its expectation formation. The results from these adaptive models produce outcomes that further deviate from the predictions of the static game theoretic models.
Ramey, Adam. “Missing in Action: A Bayesian Hierarchical Model for NA/DK Responses in Survey.”
The problem of missing data is virtually endemic to political science research. From surveys to census data to the large datasets on war in international relations, the presence of large numbers of NA's presents a great stumbling block for researchers. Though this problem is a constant worry to all researchers, how to deal with it not entirely straightforward. Listwise and pairwise deletion, both older and much simpler ways to treat missing observations, are now understood to produce large mean-squared error when data are anything except missing completely at random (MCAR). The popular imputation techniques (e.g. Rubin 1987, King et al. 2001) have become widespread and are incredibly useful. However, there are many instances when modeling the mechanism of missingness directly will improve over both of these extant techniques in terms of model fit and, more importantly, theoretical microfoundations. This paper seeks to fill this gap in the methodological literature, with a particular emphasis on survey data. Specifically, I propose a new approach to modeling NA/DK responses in ordinal survey questions that, rather than imputing hypothetical values, models the NA/DK responses as the products of choice on the part of respondents. Drawing key insights from the marketing literature (Bradlow and Zaslavsky 1999), I present a Bayesian hierarchical model that treats responses as the product of multiple latent variables: saliency, opinion, and decisiveness. It is through either a lack of saliency or a lack of decisiveness that NA/DK's are produced. Estimation is performed by MCMC methods, employing the data augmentation technique of Tanner and Wong (1987) due to the lack of closed-form conditional posterior distributions. An application to citizens' perception of candidate ideology is presented. The results provide both new and different insights that are missed by extant methods (e.g., listwise deletion and multiple imputation). In particular, the results of the model lead to different inferences with respect to the impact of race, education, and partisanship on ideological perception than either listwise deletion or multiple imputation. This paper was first prepared for the University of Rochester's advanced methods seminar (Psc 506) and was subsequently revised, submitted, and orally defended at the methods comprehensive examination at the advanced level. The paper is the first in a series of works in progress on ordinal data and dealing with problems like missingness and scale usage heterogeneity. My broader methods interests center on latent variable modeling, Bayesian techniques, hierarchical modeling, ideal point estimation, and structural estimation. I am currently working on my dissertation prospectus which seeks to reconcile the theoretical and methodological problems in the understanding of party effects in Congress.
Ramirez, Mark. “Public attitudes toward climate change: A latent variable approach.”
Scholars know very little about the nature and variability of public attitudes toward government policies aimed at mitigating the risks of global climate change. This research examines the role of scientific information and competing beliefs about the environment, the economy, and anthropocentrism in public opinion regarding policies to limit global warming using data from a national telephone survey. Latent variable modeling is used to estimate each individuals policy preference and the covariates of those preferences, in addition to the properties of each policy preference indicator.
Reeskens, Tim. “Issues in Comparative Values Research.”
Since Verba & Almond their seminal work on the “Civic Culture” (1963), comparative values research has gained widespread access in political science. More recently, by comparing Italian regions to investigate what underlying processes are “Making Democracies Work‘ (1993) —without doubt one of the main core issue in political science—Robert Putnam's emphasis on social capital has led to an increase in the political analysis of cross-national values surveys. In the last decades, various papers have been published which trying to disentangle the causal social capital chain. Next to these theoretical and empirical interests, methodological advancements have boosted this research strategy as well. Hierarchical linear modeling and structural equation models have become common analysis techniques in political science. However, even though this domain on comparative values research has contributed significantly to political sciences, numerous articles seem to forego to methodological issues that are inherent in quantitative comparative values research. Focusing on a general model for explaining cross-national differences in generalized trust (Delhey & Newton, 2005), we will be critically reflect on four issues. First of all, we will focus on differential non-response bias. It is undeniable that various countries encounter different types of unit and item non-response. Since we do know that non-response affects variable distributions, it is important to have an insight on the extent of the differential non-response bias. Second, before one can analyze values, it needs to be investigated whether the values have the same underlying structure across various cultures or countries. It can be expected that generalized trust is conceptualized differently among the Swedes than among the Italians. Third, not only because of the number of countries are limited—the United Nations consist of 192 member states—but also because of budgetary restrictions only a small sample of countries take part in comparative surveys, with about a maximum of about 60 in the World Values Survey. Yet, one needs to question whether this small number of countries is sufficient for statistical inference. Moreover, since the recent methodological innovations like complex hierarchical modeling, this problem is even more prominent. Fourth, some analyses seem to forgo to the basic assumptions in regression analysis. In this paper I would like to address the issue of how influential elements— particularly leverage points—can affect outcomes. To address these four methodological issues, we will use the European Social Survey (ESS). From the release of the first wave in 2003, the ESS has gained widespread access among scholars. This survey project is praised particularly for its methodological rigorous. The Central Steering Committee has put extensive efforts in controlling all aspects of the survey process: from the development of the survey, with strict translation procedures, to the harmonious fieldwork in the various country, an extensive procedure on gathering information on non-response like contact forms, and so on. Therefore, it can be expected that a considerable amount of error in comparative research is already under control because of the extensive centralization, which makes the survey a reliable instrument for detecting other issues in comparative values research.
Sagarzazu, Inaki. “Look who's talking: Analyzing the Dynamics of Political Discourse.”
Studies of political dialogue use as a unit of analysis either legislative or campaign discourse. The dynamics of daily political discourse in the media, however, have received very little attention. This research aims to explain the dynamics of routine political dialogue, to explore how issue salience and issue ownership are claimed by political actors through regular media outlets. To this purpose, I use a very large original database of editorial and op-ed articles published by prominent political figures in the most important Venezuelan newspapers. Including 100,000 articles from 1997 to 2007, this dataset provides a privileged view of political dialogue during Chavez tenure as chief executive. To analyze the data, I use content analysis and clustering algorithm s that extract the ideological position of actors and map social structure on issues. A multi-level analysis combines individual and contextual information to explore political dialogue.
Sekiya, Yoji. “A Structural Model of International Bargaining.”
While the importance of dynamic analysis has been acknowledged, little has been done about deriving and testing implications of dynamic games in political science. In particular, countries endlessly interact with each other, yet most analyses of international relations terminate as soon as countries reach an agreement or initiate a military conflict. To study the performance of countries in international conflict as well as the outcome of bargaining between countries in a dynamic environment, I structurally estimate parameters in a dynamic model of international bargaining in which two countries bargain over benefits (e.g., territory), they have an option of going to war, and the interaction between countries continues unless one of the countries collapses. To estimate such a model, I apply methods for estimating parameters in dynamic games developed by Bajari, Benkard and Levin (2007, forthcoming at Econometrica). The BBL estimator is a two-step algorithm. In the first step, policy functions (or strategies) and transition probabilities are estimated. Given the estimates in the first step, the second step is a simple simulated minimum distance estimator, which estimates the value functions containing information about the players' payoffs. An important assumption for the estimation is that the data are generated by Markov perfect equilibrium strategies. As a result, we can estimate equilibrium strategies, and given the estimates, we can analyze how countries would behave in different situations. That is, we can conduct policy experiments (e.g., see Merlo 1997; Diermeier, Eraslan and Merlo 2003; Diermeier, Keane and Merlo 2005). Important questions for this project are how various factors affect the performance of a country in international conflict, and how power is reflected in the outcome of bargaining between countries. The main source of the data is Diehl (1998), who studies dynamics of enduring rivalries. An important part of the estimation is that of transition probabilities, which dictate the trajectory of battle outcomes when countries fight. That is, the estimates of the transition probabilities are also estimates of countries' performance on the battlefield. This project is an important component of my doctoral dissertation for testing findings in my theory chapters. My dissertation is on bargaining behavior in international relations, and I introduce and apply new techniques for the analysis of bargaining in international relations. In another chapter of my dissertation, I apply Bayesian methods to estimate "ideologies" of countries in order to analyze how idealogies affect countries' bargaining behavior.
Slosar, Mary C. “Understanding Electoral Utilities in Candidate-Centered Electoral Contexts.”
Electoral politics in both new and established democracies have become more candidate-centered. In line with this trend, voting behavior has become more candidate-centered; voters give considerable weight to candidates' personalities in making their vote choice decisions. The relative weight assigned to such considerations compared to policy and ideological ones, however, is not constant across all voters. This paper examines the sources of this heterogeneity using data from a unique Brazilian election study. Echoing recent developments in modeling electoral choice (e.g., van der Eijk et al. 2005), I treat voters' electoral utilities as the dependent variable (rather than the more common revealed vote choice) and model the influence of voter-specific and choice-specific characteristics on these utilities. I depart from—and hope to contribute to—recent studies in one important respect: the conceptualization and operationalization of electoral utilities. Most empirical analyses of electoral utilities have focused on party choice as the relevant outcome in developing measures of electoral utilities and subsequent models. This is reasonable given that most of these studies focus on voters in Western European contexts where political parties remain important actors in electoral competition. In most new democracies, however, politics are candidate-centered rather than party-centered. Thus, we must develop strategies that allow for a wider applicability of the electoral-utility-as-dependent-variable approach. In this study, I conceptualize electoral utilities in terms of candidate choice and operationalize these utilities using candidate feeling thermometers. In addition to defending this measure on both theoretical and practical grounds, I examine the degree to which it meets the necessary criteria to represent electoral utilities. I employ this measure as the dependent variable in a model of voters' electoral decisions in the 2002 Brazilian presidential election, in which the units of analysis are voter-candidate pairings. I estimate this model using OLS with weighted error structures and conclude with a substantive discussion of the results.
Generations of scholarship have demonstrated that social interaction plays an important role in opinion formation and vote choice. However, although a sizable body of scholarship has examined strategic voting in American primary elections, the connection between interpersonal networks and this type of electoral behavior remains poorly specified. Using two original network panel studies—one following voters through the 2006 Ohio gubernatorial primary and general election, and an ongoing one examining voters in the wake of the 2008 Ohio presidential primary—I address this oversight, examining how social networks carry information about electability and structure individuals' perceptions of candidates' electoral prospects. In addition to contributing to work on strategic voting, I make both substantive and methodological contributions to the literature on social influence. Although it is established that networks influence political behavior, research in this area has been prone to problems of selection bias and reciprocal causation; virtually all studies have been static, relying on cross-sectional data. I consider the dynamics of individuals' engagement with their networks, introducing 1) new survey items designed to capture the content and range of political discussion in networks, and 2) the use of a longitudinal approach – these studies are among the first in political science to track a random sample's social networks over time. In the paper, I first model selection bias, considering the factors that drive individuals to choose/report networks. I then examine the influence of political talk in these interpersonal circles across multiple elections, looking at how the same social channels that inform voters' expectations in a primary contest evolve, operate and influence subsequent behavior. This wedding of innovative measurement and design provides an unprecedented look at information flow and everyday political talk, allows the testing of a range of substantively important hypotheses, and provides leverage on some of the classic methodological criticisms of the social influence literature.
This research examines how and to what extent mediation and its styles influence duration of international crises. The study assumes that the longer duration of international crisis increases the probability of war and subsequently imposes a greater amount of human, economic, social, and political costs on the parties and international system. Since international crisis can be ended with different outcomes depending on contents of outcome—stalemate and formal agreement—or level of tension—tension escalation and tension reduction, Cox competing risks model is employed with the system-level dataset of the International Crisis Behavior (ICB) Project. The study reports that mediated crises are more likely to have longer duration. In contrast to the expectation, mediated crisis which involves ethnicity and violence lasts much longer than unmediated ethnic, violent crisis. However it is very interesting that while mediation is associated with the longer crisis duration, mediated crises are more likely to contribute to stabilization after crisis; mediated crises are less likely to experience tension escalation. While the effect of mediation styles on crises duration is hardly discernable, formulating mediation styles are most likely to be associated with the longer duration while facilitating mediation styles are most likely to end crisis quickly. I argue that formulating mediation styles are most likely to be required in tougher crisis and need more time and resources; on the other hand, manipulative mediation styles are associated with the shorter span of international crisis because of mediators' willingness to provide resources to resolve crisis. Through this POLMETH conference, I would like to improve the causal inference of using the Cox competing risk model over Weibull regression and Cox model with baseline hazard. For providing a general theory of mediation and crisis, the study need to devise more developed conceptualizations and operationalizions to consider the effect of mediation on outcome and duration of international crisis. This great opportunity, I believe, will help me compare the strengths and deficiencies of alternative models and develop a generalizable theoretical argument.
Steinwand, Martin C. “The Strategic Interdependence of Foreign Aid: A Theoretically Informed Application of the Spatial Autoregressive Model.” «download»
Spatial statistical methods in political science provide a tool to deal with spatial and other forms of interdependence in observational data. However, political scientist have been slow to use theory in conceptualizing how political units interconnect other than through geography. In this paper, I use a game theoretic impure public good model to derive the connectivity matrix for a spatial autoregressive (SAR) statistical model. I estimate two SAR models with pure respectively impure public good weights and compare their performance in summarizing data on international aid commitments from 1974 to 2006. I find some evidence for impure public good characteristics of aid during the cold war, and strong evidence for pure public good characteristics after the end of the cold war.
Strauss, Aaron. “Cue-Taking and Campaign Microtargeting: A Bayesian Approach.”
Recent scholarship has situated party identification within the framework of Bayesian updating. This paper applies a similar Bayesian model to an electoral setting, examining the voters' choices of candidates, rather than parties. The model's theoretical implications for campaign decisions are detailed; specifically, it is asserted that campaigns should microtarget voters on issues with which the voters have direct, personal experience. A multi-wave panel survey of undergraduates tests the theoretical model on a race between two hypothetical candidates, which involves three issues. The respondents' evaluations of the candidates demonstrate the properties of a Bayesian running tally, while their opinions on the issues display too much randomness to support the model.
Su, Yu-Sung. “Causal Inference of Repeated Observations: A Synthesis of Matching Method and Multilevel Modeling.” «download»
The fundamental problem of causal inference is that an individual cannot be simultane- ously observed in both the treatment and control states (Holland 1986). The propensity score methods that compare the treatment and control groups by discarding the un- matched units are now widely used to deal with this problem. In some situations, how- ever, it is possible to observe the same individual or unit of observation in the treatment and control states at different points in time. The data has the structure that is often refer to as time-series-cross-sectional (TSCS) data. While multilevel modeling is often ap- plied to analyze TSCS data, this paper proposes that synthesizing the propensity score methods and multilevel modeling is preferable. The paper conducts a Monte Carlo simulation with 36 different scenarios to test the performance of the two combined methods. The result shows that synthesizing the propensity score matching with multilevel modeling performs better in that such method yields less biased and more efficient estimates. An empirical case study that reexamine the model of Przeworski et al. (2000) on democratization and development also shows the advantage of this synthesis.
Previous work indicates that majoritarian electoral rules tend to create greater gender imbalances within parliaments than do PR electoral rules, which suggests that majoritarian electoral rules advantage candidates (e.g., men) who belong to relatively more powerful social groups. My project tests this hypothesis by investigating whether candidates over 60, who are likely to benefit from stronger social networks and more financial resources than younger competitors and also from popular perceptions that associate age with wisdom, occupy a greater proportion of parliamentary seats in majoritarian than in PR systems. Matching is judged to be a more appropriate testing technique than regression given the imbalance between treatment and control groups in regard to the values of control variables and the existence of a relatively large reservoir of control cases in the sample. Using a sample of 207 elections from 57 countries, I examine the utility of a variety of matching procedures—including greedy pair matching, optimal pair matching, and optimal full matching —and measures of distance between treatment and control cases—including propensity scores and Mahalanobis distances—for achieving valid causal inference. All matching procedures produced relatively consistent estimates of causal effects: majoritarian electoral rules were found to increase the proportion of legislators over 60 by between three and five percent. However, particular weight is placed on the results obtained using optimal full matching with propensity scores given the difficulty of achieving balance between treatment and control groups with the pair matching techniques and particularly with Mahalanobis distances.
Theodoridis, Alexander George. “The Returns to Civic Education: Moving Beyond Selection Bias.”
Measuring the ability of civic education programs to convey values, attitudes and knowledge consistent with increased and more effective political participation is a central concern to scholars and educators, as well as the many foundations, academic units and public interest organizations deeply committed to advancing the cause of civic education. This study uses examination of the pedagogical effectiveness of one particular civic education program to provide generalizable insights regarding such programs and the civic education endeavor generally. In 2002, the University of Virginia Center for Politics pursued a study to assess the effectiveness of its civic education program, the Youth Leadership Initiative, which had recently been developed and made available to high schools and teachers in all 50 states. A national survey, conducted by the UVa Center for Survey Research, was administered to participating students and teachers as a pre- and post-test in a quasi-experimental study using a Solomon's four-group design. The resulting data set includes responses from 2,028 high school students, with measures of political knowledge, past and intended future political participation, political attitudes, civic activism, levels and types of participation in the program, external attitudinal inputs (such as from parents and teachers), and basic demographics. Treatment assignment in the initial study suffered from selection bias. This paper employs a genetic optimization algorithm developed by Jasjeet S. Sekhon (2007) to match individuals across and improve balance between experimental groups, thus using selection on observables to move toward applicability of the Rubin Causal Model (Holland 1986, Rubin 1974) and perhaps closer to the experimental benchmark (Diamond & Sekhon 2005). This improves upon previous assessments of the effectiveness of civic education programs and the extent to which political attitudes and knowledge can be transmitted through education. Where some others (e.g. Niemi and Junn) find significant effects, my analysis shows that these effects dwindle under controls for confounders and disappear almost entirely under selection on observables. One robust effect appears to be an increase in political knowledge when a participatory activity (in this case a mock election) is combined with the use of lesson plans. My findings suggest that we should look somewhat skeptically upon positive overall treatment effect estimates for civic education found in much of the recent literature on this subject, as it is possible that these effects arise from poor balance between treatment and control groups in observational and quasi-experimental studies.
An explosion of relational data has made network analysis methods of great interest to scientists of all stripes. Currently available methods, however, are traditionally limited to analyzing zero-one relationships, requiring investigators to establish a “threshold” argument for the existence of an underlying connection. We describe methods for inference on random graphs with weighted links, using this inference to generate a null distribution of possible graphs for comparison. We demonstrate our method on relational data between United States senators as revealed by their joint press releases. This work is in collaboration with Prof. Joseph Blitzstein of the Harvard Statistics Department.
Titiunik, Rocio. “Drawing Your Senator from a Jar: Term Length and Legislator Behavior.”
The tenure by which representatives are to hold their offices was carefully designed by the Framers of the American Constitution to ensure two basic conditions: that representatives be sufficiently dependent on the people, and that they face the incentives and develop the ability to pursue complex, long-term and unpopular policies when required. The adoption of different term lengths in the two legislative chambers reflected a compromise between these opposing principles. Whether term lengths operate in the way the Framers intended is ultimately an empirical question, but one that poses extraordinary methodological challenges. When legislators are elected to serve a given period of time for a given office they exhibit certain behavior, but in order to learn whether the observed behavior is being affected by the duration of the term for which they were elected, one would have to know how the legislator would have behaved if his term length had been shorter or longer. This is of course extremely difficult, since the latter is never observed. In this paper, I use the gold standard solution to this fundamental problem of causal inference, namely, the random assignment of term lengths. Exploiting a unique randomized natural experiment that takes place in the state senates of Arkansas, Illinois and Texas, I provide the first experimental estimate of the causal effect of the duration of terms on the behavior of legislators. In these states, after the elections immediately following the decennial reapportionments of 1992 and 2002, senators were randomly assigned to different groups or classes according to whether they would serve four years or two years in the legislative period immediately following the enactment of the new districts. Between censuses, state senators serve for a period of four years and are staggered so that one half (one third in Illinois) of senate seats are up for election every two years; but in the election following reapportionment, all senate seats are up for election to ensure that senators be representative of their new constituencies. The staggered structure of terms is thus broken, and the procedure by which it is reestablished is purely random. All statistical inferences are based on a randomization inference approach, directly incorporating the random assignment mechanism in the estimation of treatment effects and testing of hypotheses. Under the null hypothesis that two-year senators and four-year senators behave identically, the only random variable is the assignment of term lengths, which implies that the distribution of the chosen test statistic under the null hypothesis is completely determined by the randomization distribution of the treatment assignment. Since I know the exact randomization distribution for every state, I am able to directly incorporate the randomization process in the estimation of treatment effects and testing of hypotheses. Moreover, I formally derive the assumptions implicit in previous observational work, and show that the causal interpretation of these results is not warranted by the research designs used in these studies. I compiled a rich dataset of demographic, socioeconomic and electoral characteristics at the individual-senator level and the state-senate-district level, and use it to verify that the randomization was successfully implemented. I find that senators whose reelection is closer in time take position less often: the proportion of roll-call votes in which they vote neither yea nor nay is significantly higher when compared to senators whose reelection is two years later. The effects of term length on party discipline and passage of legislation are also considered.
Survey methodologists have shown considerable interest in recent years concerning how to deal with inaccurate answers to sensitive survey questions. (Buchman and Tracy 1982, Berinsky 2004, Corstange 2008). This paper contributes to this literature by examining respondents' self-reported voting behavior in pre- and post- 2004 presidential elections survey in Russia. In particular, we are interested in the type of behavior that Timur Kuran (1991) characterized as ‘preference falsification’—public behavior at variance with private (‘real’) beliefs to create an impression of compliance with regime expectations. Although Kuran coined this concept to compare citizens' behavior in authoritarian regimes with that of citizens in pluralist systems, we argue this concept is also useful for analyzing response patterns to sensitive survey questions across regions within countries such as Russia where there is substantial cross-regional variance in levels of political pluralism. As our unit of analysis we take respondent's self-reported choice of presidential candidate in 2004 elections and develop an empirical test for assessing the accuracy of answers. Then we examine how preference falsification affects the consistency of estimated coefficients when we fail to account for this phenomenon. The paper is organized in three sections. We start by contrasting official election results with survey data. Then we lay out our empirical strategy used to test if respondents falsify their preferences. We conclude by using Monte Carlo simulation to assess the magnitude and the direction of the bias caused by preference falsification.
The purpose of the study is to verify the hypothesis that special institutional arrangements in the Japanese electoral system of the House of Representatives, which enables political parties to provide proportional representation (PR) seats for best losers in single member districts (SMD) significantly reduces the incumbency advantage that incumbent congresspersons otherwise enjoy, and consequently contributes to more competitive and responsive elections in Japan. I use Regression Discontinuity design (Hahn, Todd, and Van der Klaauw (2001)) to estimate the incumbency advantage under the two different campaign conditions. In the 1994 electoral reforms, Japan adopted the combination of single member districts (300 districts) and regional proportional representation (200, later 180 seats from 11 districts). As a peculiar characteristic, the electoral system allows party candidates to run both in SMD and PR tiers, and attain seats from the latter tier if lost in the former. In addition, parties are allowed to rank candidates on the same rank in PR list, to provide seats to the ‘best losers’ among those listed on the same rank. The scheme has been criticized even during the legislative process. For example, McKean and Scheiner (2000) criticized the scheme because the “arrangement transforms PR representations into locally-based politicians who will rely on personalistic rather than party-based or programmatic campaigning” However, in this study I prove the positive aspect of the institutional arrangement; it reduced incumbency advantage in each district to have more competitive and responsive elections, which were otherwise stable strongholds of the incumbent, especially those of conservatives. The logic behind my assertion is as follows: the arrangement provides to the best losers in the SMD tier all the resources as an incumbent congressperson from the PR tier (such as public finance for political activities and media attention), and consequently enables them to challenge previous winners with a similar level of resources. If the logic is right, I will find large incumbency advantage when the incumbent congresspersons in SMD contest with challengers without seats, but not so much advantage when the former compete with Zombie challengers, who lost in the last election but attained seat from the PR tier, even if the other conditions are equal. To verify the hypothesis, I used a Regression Discontinuity (RD) design for the election results in 2003 and 2005, the last two elections contested almost the same parties after a decade of partisan realignment in Japan. I used the districts where ruling coalitions (mostly Liberal Democratic Party, LDP) and Democratic Party Japan (DPJ), the largest opposition party contested, and incumbents from either of the camps were running in 2005. For RD design, Hahn, Todd, and Van der Klaauw (2001) provided formal theoretical framework, in addition Lee (2008) applied the design to estimate the incumbency advantage in the US Congresspersons. I found there was only negligible size of incumbency advantage for the incumbent against Zombie incumbents. The Wald estimate of the advantage was around negative 3% but far from statistically significant. On the other hand, the estimate of the incumbency advantage against challengers without seats was almost 20% and statistically significant. The results strongly supported my hypothesis. Because some regions had more the second type of contests around the cutting point than the other regions did, I also run the analysis only with the districts in the former regions to test the robustness of the result. The advantage was roughly the same (20%), although it did not reach conventional statistical significance probably due to smaller number of observations. The institutional arrangement of double candidacy and the same rank PR listing seems to contribute more responsive elections in Japan, probably beyond the expectation of the drafters judging from the discussion during legislating process and electoral strategies adopted in the campaign just after the reform. Currently, in addition to DPJ, LDP also shifts the electoral strategy to use the arrangement to fight the opponent's incumbent more efficiently, probably out of their own research and/or political instincts. Their electoral strategies also support the hypothesis I argued in this study. Reference Hahn, Jinyong, Petra Todd and Wilbert Van der Klaauw (2001). “Identification and Estimation of Treatment Effects with a Regression-Discontinuity Design,” Econometrica, vol. 69, no. 1, pp. 201-209 Lee, David S. (2008), “Randomized Experiments from Non-random Selection in U.S. House Elections,” Journal of Econometrics, vol. 142, no. 2, pp. 675-697.
Weidmann, Nils B. “Predicting Conflict via Machine Learning.”
Machine learning deals with the development of algorithms for classification and prediction. However, machine learning has rarely been used in political science. This poster demonstrates the application of state-of-the-art machine learning techniques to the prediction of conflict in contemporary societies. A new dataset spanning the period from 1998 to 2004 containing measurements on domestic and international crises, ethnic and religious violence, insurgency, and rebellion in 29 Asian societies is divided into a training and test set. Three prominent machine learning algorithms are applied in order to create out-of-sample predictions for these event classes: A decision tree, a rule learner and a support vector machine. These algorithms achieve high accuracy in out-of-sample predictions, making them powerful tools for researchers and practitioners interested in forecasting conflicts. In general, the machine learning approach substantially outperforms conventional statistical models based on in-sample classical inference. In addition to providing examples, the poster illustrates how to utilize these algorithms in both R as well as Weka, a widely used Java-based machine learning workbench.
Weiksner, G. Michael. “Measurement Error as a Threat to Causal Inference: Acquiescence Bias and Deliberative Polling” «download»
Experiments, unlike observational studies, are rarely criticized for yielding invalid causal inferences. However, I identify measurement error as a threat to causal inference of an experiment. In particular, acquiescence bias, a common and substantial source of measurement error within surveys, may be correlated with experimental manipulations. Using data from a survey experiment embedded in a Deliberative Poll, I find that acquiescence bias causes significant measurement error and that the bias differs before and after deliberation. I conclude that even experimental researchers should heed the recommendation by questionnaire design researchers to refrain from asking agree/disagree questions completely and instead ask only construct-specific questions to avoid this threat to validity.
Wright, Dominick'. “Network Structure & Dynamics: An Empirical Survey of Terror Network Panel Data.”
Currently, there is scant knowledge about the factors contributing to the micro-level characteristics of individual terrorists and the overall structure of terror networks. While a growing circle of academic and policy minds note that terror networks are dynamic and that culprit cliques of terrorists are actually social structures embedded within larger, meta structures of social relations, there has yet to be a systematic empirical analysis in academia that moves the discussion beyond myopic conjecture. One of the most disturbing shortcomings in the literature is a near constant reference to ‘network dynamism’ without any clear indication about what exactly is changing and with respect to what in these networks. For instance, if the only thing changing over the course of time is growth in membership at a constant rate and a constant distribution of relevant characteristics, then it is plausible to argue that very little in the process has actually changed. However, if membership volume significantly varies in correspondence with the type of individuals incorporated as well as the means for incorporation, then the reference to dynamism is less vacuous. As important as answers to these types of questions are when forming micro-level theories about terror/counter-terror processes (e.g., evaluating the time-differenced effect between an act at time t and observed changes in the other at time t + i), very few studies have attempted to address the issue with real data and analysis. Using information compiled in the Global Transnational Terrorism (GTT) database (a database on the social relations between convicted terrorists, suspected terrorists, and a sample of non-terrorists), I propose to begin exploration of these critical concepts. There will be three foci in the foregoing analysis: micro-level attributes, micro-level relational characteristics, meta structure characteristics. Micro-level attributes range from social background (such as place of birth and ethnicity) to individual skill set proxies (e.g., education) and the skill sets themselves (i.e., tactical, logistical, and operational capabilities). Classification of network members according to their attributes will provide an opportunity to evaluate the relevance of alternative individual attributes within the overall percolation process relative to other relational characteristics (e.g., the number of relations at time t − 1). Lastly, aggregating observations of individuals within the network (and parallel networks under some circumstances) allows me to assess the overall structural properties of the large component (network with the most members) and the smaller components (all others). Since there are very few terror network specific theories that exist for the evaluation of these dynamic data, I will use variant percolation theories (e.g., preferred attachment, random attachment, and so forth) to evaluate the growth characteristics within the observed data.
This study introduces split population binary choice models to address irrelevant dyads in the dyadic analysis of conflict with binary dependent variables. The advantage of employing a statistical model instead of directly identifying relevant dyads manifests itself in the selection of relevant dyads: rather than researchers making take-it-or-leave-it decisions, covariates are used to estimate the latent variable relevance. An application of the model to the trade conflict debate shows that probability of conflict, the quantity of interest for the traditional binary choice model based on all dyads, is non-monotonic against trade generated by the split population model. This finding provides one explanation for why trade is found to either increase or decrease the probability of conflict in existing research as monotonicity by chance is imposed on an underlying non-monotonic relationship.
Yohai, Ian. “Not So Identical Twins: Party Positions on Immigration and Trade in the United States.”
While economic theory suggests that the distributional consequences of immigration and trade are likely to be similar, the political parties in the United States have recently taken opposing positions on the issues. Although the Republicans' business constituency and commitment to the free market makes them naturally more supportive of trade, they have grown increasingly resistant to more open immigration. Likewise, the Democrats' concern for low-skilled workers has led to a more protectionist stance of late, yet they continue to support freer immigration, even though immigration would seem to have the same impact on low-skilled workers as trade. Moreover, the current party alignments are exactly opposite of those in the late 19th century, when Republicans were more open to immigration but were the party of protectionism, and Democrats were committed to free trade but skeptical of immigration. Building on recent advances in Bayesian ideal point estimation, especially with a limited number of votes available for analysis, this poster traces the positions of the parties on both issues over time. Finally, it makes some comparisons with mass-level opinion data, which shows that while the parties have taken seemingly contradictory policy positions, the public seems to be more constrained in its attitudes.
