Joyce	Kyle	The Pennsylvania State University	Applying Computational Methods to a Theory of War Expansion	The war between Austria-Hungary and Serbia that began in July 1914 quickly expanded to include additional states. Only six months into the war, two additional states had joined on the side of Austria and five on the side of Serbia. By the time the war ended in November 1918, thirteen additional states had joined; three on the side of Austria-Hungary and ten on the side of Serbia. What began as a war between two states grew into what we now know as World War I, a war that alone is responsible for an astonishing 27% of the fatalities across all interstate wars between 1816 and 2001. Most interstate wars (80%) do not follow this pattern; they end without expanding to include states beyond the original belligerents. But those wars that do expand (20%) account for an exorbitant 88% of the fatalities across all interstate wars during this time period. This paper examines three interrelated puzzles: 1) Why do some wars expand while others do not? 2) How does war expansion alter the dynamics of war? 3) How does the participation by third parties influence the likelihood of participation by other third parties? I investigate these puzzles using a combination of formal and empirical techniques. I begin by employing agent-based modeling to run computer-based simulation experiments, from which I derive propositions based on the emergent behavioral patterns of third parties during an ongoing war and on the effects of third party intervention on the dynamics of war. I then subject the propositions derived from these computer experiments to empirical evaluation using statistical analysis. To model war expansion I created an artificial international system populated by agents (i.e., states) who interact in an anarchic environment. The agents are heterogeneous in their power (i.e., military capability) and the utility they derive from each of the initial participants winning the war (i.e., affinity). Following the onset of a war between two of these agents, each agent not involved in the war (i.e., third party) decides whether or not to participate. Each third party’s decision is a mathematically prescribed function of that state’s expected utility, calculated as a product of its power and its affinity for the initial participants. The decision of whether or not to join is endogenous to the dynamics of the war; as the war persists, each agent re-evaluates its decision to participate and adapts its behavior to the current conditions of the war (i.e., who now is winning the war and which other third parties now are participating) and decides anew whether or not to participate. This artificial world is analogous to a laboratory in which I “grow” war expansion to examine how the micro-level behavior of many locally-interacting agents generates the emergent behavioral patterns of war expansion. These patterns serve as theoretical predictions that we should see emerge in the real world under parallel conditions. After conducting experiments in this artificial world (in the form of computer simulations; n.b., these simulations are already completed) under varying initial conditions, I derive hypotheses based on the emergent behavioral patterns. The first hypothesis predicts the effect of third party participation on the probability of a war ending at a specific point. The second hypothesis relates the timing of third party participation and the side on which that participation occurs. Both of these hypotheses are evaluated using competing risks hazard model on a sample of 79 interstate wars between 1816-2001 complied by the Correlates of War project. The combination of formal and statistical methods will provide additional insight into why wars expand than either method used in isolation and thus provides an improvement on previous studies of war expansion.
Cranmer	Skyler	University of California, Davis	Multiple Imputation for Discrete Variables	This poster will present a technique for imputing missing observations in discrete data. The technique used is fractional hot deck imputation; because the imputed value is a draw from the conditional distribution of the variable with the missing observation, the discrete nature of the variable is maintained as its missing values are imputed. Multiple imputation, as it is typically practiced, should not be run on discrete data because it assumes continuity and replaces missing discrete values with nonsensical non-discrete values; a binary variable for gender might be imputed with a value of .6 with traditional multiple imputation. The problem of nonsensical imputations for discrete data worsens the less continuous the data are. The primary weakness of traditional hot decking as compared with multiple imputation, an inability to make confidence statements about the imputed values, is overcome by weighting imputed values by the fraction of the original weight of the missing element assigned to the value of the donor observation based on its degree of affinity with the incomplete observation. The proposed poster will detail the mathematics of fractional hot deck imputation as well a R package to accomplish it which is currently under development.
Crow	David	University of Texas at Austin	Turnout Overestimation in Mexico	Now that Mexican elections have progressed from window dressing for dominant party authoritarianism to a genuine mechanism for choosing political leadership, scholars—and political strategists—want to know who votes and why. As in other democratic countries, however, self-reported voting in surveys is overstated. In 2000, for example, the National Poll on Political Culture and Citizen Practices showed 81% turnout, but the official election returns, only 64%. As elsewhere, the sources of this phenomenon are two: respondents may falsely report having voted, and the sample may overrepresent those who did vote (a matter of selection bias). The overestimation matters not only for description and cross-national comparison, but because the nonrandom errors involved may bias individual-level estimations of the influences on voting. If, for example, either claiming falsely to have voted or turning up in the sample is positively related to income (which is positively related to voting), income's effects on the probability of voting will be overestimated. In the U.S., the National Election Study sometimes validates claims of voting by comparing them to official voting records, a costly but simple expedient, but in Mexico this solution is unavailable. Ballot materials are destroyed six months after the election and federal law prohibits release of individual voting information. Validated vote data are thus unobtainable. That leaves several other possibilities. Using data from the Comparative Study of Electoral Systems (CSES) 2000 Mexican survey, this paper tries and examines the implications of three corrective methods. The first is to use survey “filter questions” to reclassify people who claim to have voted but could not have done so. For example, I recategorize as non-voters those who claimed to have voted but whose voter credentials were not marked by polling officials and then rerun the analysis on the reclassified sample. The second approach is to down-weight respondents claiming to have voted, on the expectation that people who say they didn’t vote are telling the truth more often than those who say they did. The probability of actually voting, conditional on sociodemographic characteristics like age, sex, and income, can be estimated by combining census data with official turnout figures and by exit polls. I use these estimates to form sample weights. The third approach, following Brehme (2001), is to estimate Heckman-type selection models with auxiliary information to correct for "stochastic truncation", or unit non-response, in which we have no information on non-respondents. I use aggregate data to estimate a “dose-probit” selection model. (Since the inverse Mills ratio is based on aggregate, rather than individual, data I call the method a "Heckman-type" method.) Then I incorporate these estimates into the outcome model, using post-election surveys, to examine the causes of voting. The results suggest that prior studies have generally overemphasized the impact of sociodemographic variables on voting. With the biases from turnout overestimation are corrected for, political attitudes—especially trust that the elections are clean and fair—appear to predict voting better than income, education, or age.
Butler	Daniel	Stanford	Making the Call: The Effect of Media Projections on Turnout and Election Outcomes	In the aftermath of the 2000 presidential election, some conservatives raised concerns that the media’s early projection of Florida to Gore had cost Bush votes by decreasing turnout in the Republican-dominated western part of the state. Similar concerns were raised by Democrats following the landslide election of 1984. Are these concerns correct? Do early media projections of the presidential winner affect turnout or election outcomes? Contrary to much of the popular wisdom and previous academic research, I present evidence suggesting that the early media projections of the winner do not affect turnout, the vote share in the presidential race, or the vote share in the congressional races. The advantage of this study over previous research is that I leverage the natural experiment created in Kentucky where there are two time zones and the networks have projected the election winner after polls in the eastern part of the state have closed but while polls in the western half are still open. Comparing precincts within a state, confounding factors such as state laws, other statewide races, etc., are controlled for. Further, I match precincts in the treatment group to similar precincts in the control group based on registration data to improve the quality of the counterfactuals used in estimating the average treatment effects.
Corstange	Daniel	University of Michigan, Ann Arbor	Sensitive Questions, Truthful Answers? Modeling the List Experiment with LISTIT	Many of the social phenomena we are most interested in are also extremely difficult to study via surveys when people lie about or otherwise hide socially unacceptable or incriminating behaviors. To measure sensitive behaviors and opinions, I present an unobtrusive measure called a list experiment (originally Kuklinski et al.), which I have improved and extended to enable multivariate analysis via the new LISTIT statistical estimator. The procedure, briefly, splits the survey sample into control and treatment groups. The former receives a series of non-sensitive yes/no questions to be answered individually. The latter receives the same set of non-sensitive questions plus one sensitive question in a list format, and are asked how many of the list items they do/believe, and specifically not which ones. Appropriate statistical analysis allows us to estimate the determinants of the sensitive opinion in the population without ever knowing which individual survey respondents actually hold the sensitive opinion. I first detail the derivation of the new statistical estimator, and then demonstrate the procedure with original survey data collected in Lebanon in 2005, using items on restrictions on the suffrage, willingness to live in mixed neighborhoods, and participation in the pro-/anti-Syrian demonstrations in the spring of 2005.
Lauderdale	Benjamin	Princeton University	Pass the Pork: A Model for Appropriations Shares in Bicameral Legislatures	While OLS models are commonly used to analyze congressional appropriations, they fail to capture the structure of Congressional representation. A model for final spending outcomes ought to reflect the overlapping representation of the House and Senate by describing district spending as the result of multiple legislators working on behalf of the district. I propose a flexible statistical model for legislator shares within each chamber that allows direct comparison with share predictions from formal modeling approaches to bicameral bargaining. This model allows disaggregation of House and Senate influence on appropriations and statistical estimation of the relative power of the two chambers using variation in funding by state size for either state- or district-level spending data. Where district level data is available, it also allows estimation of the advantages that accrue from holding certain positions in the structure of the legislature, providing a platform for empirical tests of recently developed formal models of agenda-setting, veto, and committee powers. I consider formula and earmark spending in the recently passed 2005 transportation legislation as a test of the model and find that the model accurately predicts the shape of the dependence of spending on state size for both formula and earmark spending, but that proposal power is distributed differently for those two types of appropriations.
Kellermann	Michael	Harvard University	Winning the Lottery? Randomization and committee seniority in the U.S. House of Representatives	This project examines the effects of committee seniority rankings on legislative careers in the House of Representatives. While formal models suggest that seniority systems represent an inter-generational equilibrium, rewarding senior members to provide incentives for junior members, there is relatively little evidence about the value of the seniority system to legislators themselves. In the legislative context, one might expect that members ranked higher on a committee would have a higher probability of re-election, or that lower-ranked members would be more likely to transfer to other committees. Obtaining estimates of such effects can be problematic, however, since many legislator characteristics (age, experience, etc.) co-vary with committee seniority. We take advantage of the committee assignment process used by the House Democrats in order to obtain unbiased estimates of these effects. After committee assignments are made, lots are drawn to establish rankings when two or more new committee members have equal chamber seniority. This randomization ensures that, within these groups of committee members, the committee ranks are uncorrelated in expectation with background characteristics that could introduce bias. In general, we focus on the effects of a one-rank difference in committee seniority on various career outcomes. We use both permutation and Bayesian model-based approaches as appropriate to make valid inferences about the consequences of this aspect of the congressional seniority system.
Smidt	Corwin	Ohio State University	Evaluating the Adequacy of Vector Autoregression to Model Candidate, Voter, and Media Issue Influences Across the 2000 Campaign	How successful are candidates when attempting to influence how voters evaluate them? Past works demonstrate candidates have the incentives and the abilities to influence the issue focus of voter evaluations (e.g. Johnston, Blais, Brady, and Crete 1992; Jacobs and Shapiro 1994). However, when empirically evaluating such questions these works often fail to separate and control for the potentially endogenous influences of voters and the media. That is, candidate actions within campaigns are equally a reaction to voter and media behavior as vice-versa. Consequently, the identification restrictions within structural equation techniques are inadequate and a modeling strategy that accounts for the possibilities of multiple endogenous relationships is preferred. Importantly, new data resources, such as the National Annenberg Election Study and the Wisconsin Advertising Project, allow for a vector autoregression approach which partially avoids such problems (Freeman, Williams, and Lin 1989). This poster examines the adequacy of classical vector autoregression and error correction estimates when evaluating voter, media, and candidate issue agenda series across the 2000 campaign and whether different inferences are produced within a Bayesian time series approach which has recently been argued to produce more powerful and robust results (Brandt and Freeman 2006). In particular, the stationary behavior of each series is evaluated to ascertain the accuracy of standard test distributions for each series (Sims, Stock, and Watson 1990; Freeman, Houser, Kellstedt and Williams 1998). Granger F-tests and impulse response functions are then presented to test the contrasting influences of candidate, media, and voter dynamics on each other. Finally, beneficial post-estimation techniques, that are rarely used within political science, are implemented such as eigenvalue decomposition of the coefficient matrix (Hamilton 1994), residual autocorrelation (Johansen 1995), and residual normality tests (Jarque and Bera 1987) to ascertain the validity of vector autoregression estimates. In general, the poster makes three important methodological contributions. First, it provides an appropriate test of reciprocal influences within candidate, media, and voter behavior. Second, it evaluates the stationary behavior of time series that are gaining use from new and popular data resources (e.g. Box-Steffensmeier, Darmofal, and Farrell 2005) and whether their dynamics produce different substantive inferences within a Bayesian VAR approach. Finally, it demonstrates rarely used post-estimation tests that evaluate the validity of vector autoregression and error correction estimates.
Sinclair	Betsy	Caltech	Cosponsorship Networks: Using Centrality to Understand Legislative Polarization	We examine cosponsorship data from the 93rd to 108th U.S. Congress for evidence of legislative polarization. Using two notions of centrality from the social network literature -- eigenvector centrality and degree -- we produce a centrality measure for each legislator using the cosponsorship data from that particular Congress. We then use this value as a variable to understand the relationship between centrality and party membership, Congress, committee membership, and ideology.
Bassi	Anna	New York University	Voting Systems and Strategic Manipulation	Any preference aggregation method used for electing legislatures or other representative bodies is vulnerable to strategic manipulation by voters. In particular, voters whose first preference has only minority support may vote strategically for a candidate they prefer less but with greater support among other voters. Approval voting allows these minority voters to vote for their first preference as well as less preferred candidates who they believe may garner more support. Thus, we expect approval voting to encourage more sincere voting choices. This paper presents experiments testing this prediction. In the experiments approval voting is compared to two ranked procedures (Borda count and Hare systems) and one unranked rules (Plurality voting). In the experiments there are four candidates and five different types of voters (that is, preference orderings). The results support predictions.
Bailey	Delia	Caltech	Assessing the Accuracy of Voting Technologies in California, 1962-2004	This paper utilizes data from Presidential, Senatorial and Gubernatorial elections in California, 1962-2004, to assess the accuracy of voting technologies. Unlike many recent studies of voting technology, this data allows comparison of specific voting equipment types, such as central and precinct count optically scanned ballots, and pre-scored and non-scored punch cards. A difference-in-differences estimator and fixed effects regressions are employed to measure causal effects.
Jeong	Gyung-Ho	Washington University in St. Louis	An empirical test of the multidimensional spatial voting model with roll call data	Many important concepts of the spatial voting model like the core, the uncovered set, and the Pareto set are about the locations of final winning alternative or bill. Thus, in order to test empirical validity of the spatial voting model, we need to estimate not just ideal points of legislators but also the locations of bills. However, estimating bill locations in multidimensional spaces has been tricky. Although Clinton and Meirowitz's agenda constrained ideal point estimation is an improvement on this, the limitation of this approach in a multidimensional setting is that it requires each bill to be related to only one dimension. The bills that are related to more than one dimension still cannot be estimated even after the agenda constraints. Here, I suggest that this problem can be solved if the legislature has an amendment procedure that allows votes on perfecting amendments to have precedence over votes on substitutes. This is because a perfecting amendment tends to be short and one-dimensional and because a substitute, which tends to be long and multidimensional, often has same components as the amendments that were voted on earlier. If this is the case, we can use this information to constrain the location of the substitute in one dimension to be same as that of the earlier amendment. In this paper, I report results from my simulation analysis. Then, I apply this approach to the roll call data from the U.S. Senate, where the principle of ``precedence'' accommodates estimation of bill locations. By comparing the locations of winning bills and the uncovered set calculated from the estimated ideal points, I show how we can test the predictive power of the spatial voting model with roll call data. Since empirical tests of the spatial voting model have been limited to data from experiments, this is the first empirical test of the spatial voting theory with real world data.
Whang	Taehee	University of Rochester	Symbolic uses of economic sanctions: domestic politics and international signaling	A central anomaly in debates about economic sanctions is the discrepancy between the increasing use of sanctions and the increasing pessimism regarding their effectiveness. Many scholars tend to explain this anomaly in terms of ‘the symbolic use of sanctions’ separately from ‘the instrumental use of sanctions’. The instrumental use of sanctions focuses on the extent to which the goals of the sender are accomplished as a result of sanctions. While the symbolic uses of sanctions are widely thought as alternative sources of initiating sanctions to the instrumental uses, scholars tend to use conceptually different kinds of symbolic sanctions. More importantly, little attention has been paid to their empirical evaluations. In this paper, I differentiate between two different symbolic effects of sanctions, one on international and the other on domestic audience. Then, I empirically test each symbolic impact of sanctions. First, for the international signaling impacts, I provide a fully structural model that employs a game theoretic model as a statistical model. The theoretical model formalizes a simple crisis bargaining logic with two-sided incomplete information. Not only does the empirical model take strategic interactions seriously, but also it allows me to estimate the amount of belief updating and signaling in sanctions’ episodes. Second, for the domestic political impacts, I examine the existence of audience benefits to the incumbent of sanctioning state, independent of the sanctions’ outcomes. I assess the effect of sanctions’ imposition on the US presidential approval from 1948 to 1993. Preliminary results suggest that sanctions often fail to work as a costly signal to represent the sender state's resolve and a target state does not learn much from observing sanctions. This cheapness of sanctions prevents signaling mechanism from determining the sanctions’ outcomes. Thus, one of the Fearon's rationalist accounts of war cannot be applied to economic sanctions. Sanctions rarely work as a credible indicator that reveals a priori unverifiable intension of the sender. On the other hand, I find that policymakers actually benefit by imposing sanctions in terms of increases in public support. Sanctions are an effective way of displaying ‘do something’ leadership to the public in international conflicts. Consequently, the domestic audience benefits to the incumbent help us understanding the recurrence of sanctions despite their instrumental ineffectiveness.
Terry	Will	UCSD-Political Science	Measuring Legislative Productivity in the 93rd-105th Congresses	Congressional scholars have long been interested in which legislators exert the greatest influence over the legislative process within the U.S. House of Representatives. Among the measures of legislative “power” considered in the literature are those that focus on legislative productivity—e.g., how many bills per session each legislator is able to push past the committee stage (and hence into serious contention for passage); or how many bills per session each is able to pass (cf. Frantzich 1979; Moore and Thomas 1991). Most scholars who investigate legislative productivity seem to view it as analogous to ordinary economic productivity: a member will be more productive if s/he can marshal more of the essential resources needed to “produce” bills (i.e., to push them past key milestones in the legislative process). Differences arise, however, regarding what those essential resources are. As Frantzich (1979, 410) noted a generation ago, an individual legislator’s “power” might increase with, among other things: (1) majority status (e.g., Gross 1953; Cox and McCubbins 1993; Kim 2006); (2) congressional seniority (Dendiner 1964, 17; Davidson, Koveneck and O’Leary 1966, 98); (3) possession of important committee or floor leadership posts (Froman 1967, 35); and (4) ideological moderation (Black 1958; Krehbiel 1997). In this paper, we exploit a large new dataset documenting the legislative activity of all members of the US House from the 93rd to the 105th Congresses in order to revisit the determinants of legislative productivity in Congress. We focus in particular on providing a better measure of the influence of majority status, by investigating how the productivity of individuals who switch from one party to another (and hence from majority to minority status, or vice versa) changes; and how the productivity of non-switching members changes when party control of the chamber changes. Such research designs help to control for factors such as congressional seniority, committee membership and seniority, and “fixed” characteristics of individuals. We also employ ANOVA and random coefficients modeling techniques.
Tahk	Alexander	Stanford University	A New Approach to Optimal Classification	I present some of the problems with Poole's Optimal Classification (OC) algorithm as well as introducing a new approach to the optimal classification problem. First, I outline some problems with the Optimal Classification algorithm. I present an alternative approach to optimal classification that has several advantages over Poole's OC algorithm. First, unlike the OC algorithm, this algorithm can be expected to produce optimally classifying ideal point estimates. Second, since an optimally classifying set of ideal points is typically not unique, the algorithm produces several sets of optimally classifying ideal points rather than a single estimate. I also present techniques which can increase the computational efficiency of this algorithm as well as what identification restrictions are appropriate for the optimal classification setting. Finally, I use real data from the Supreme Court to provide an example of the results of this algorithm as well as some of the problems with Poole's OC algorithm.
Jusko	Karen	University of Michigan	The Political Representation of the Poor	Who speaks for the poor represented in contemporary democratic societies? Do electoral rules create incentives for elected representatives to be more or less responsive to the poor? How do these electoral incentives affect distributive policy outcomes? Using the Luxembourg Income Study (LIS) data, my dissertation research investigates how electoral rules affect the poor in a broadly comparative analysis. The proposed research will contribute to this larger project, and concentrates on the methodological challenges of analyzing cross-national income survey data: How can poverty be measured in a cross-national context? How can the effects of redistributive policy be observed? While the LIS data offer an important opportunity to evaluate these questions, there are additional challenges that arise in their use. Countries are observed at different points in time, for example, and with varying frequency. The proposed research will address these issues, and present an analysis of LIS data that investigates the validity of empirical propositions derived in a formal theoretic model of electoral competition.
Lau	Olivia	Harvard University	Anchoring Vignettes in International Public Health Survey Data	This poster will examine the relationship between the institutions that provide public health and the institutions that drive economic growth using a quasi-experimental research design. In an experimental research design, units are randomly selected from a population and randomly assigned to treatment. By calculating the difference between the treated and control groups, researchers can evaluate the effect of the treatment, on average. Although there are different ways of calculating the ATE (see, e.g., Imbens (2004)), the inferential goals of these techniques remain the same. If unit selection is not random, but treatment assignment is, researcher can still perform inference on the causal effect of treatment conditional on the sub-population. If neither unit selection or treatment assignment are random, then inferences at the causal or population level are subject to potential bias. In this last case, we can achieve a quasi-experimental design, however, by matching treated and control units on the basis of pre-treatment covariates, and then estimating an average treatment effect. In the context of public health, this poster will employ the compound hierarchical ordered probit model (e.g., in Wand, King, and Lau 2006) to estimate a multinomial propensity score for the quality of health institutions from WHO survey data from 69 countries (the data employ vignettes to establish the cross-cultural comparability of survey responses). Matching on the quality of health institutions, this poster will estimate the causal effect of exogenous international health spending on domestic economic growth.
Purpura	Stephen	Harvard University	Sensitivity Analysis in Automated Content Analysis Using a Political Party Platform Example	For social science researchers, human annotated content analysis and classification of political text is time intensive and expensive. Now, technology from computational linguistics and semantic analysis is facilitating lower cost solutions. Wordscores, by Benoit, Laver, and Garry (2003), is an example of one such technique. Computational linguists use other techniques generally based on support vector machines, neural networks, nearest neighbor classification, Bayesian probability, decision trees, inductive rule learning, and regression models. When selecting an algorithm for a task, the computational linguistic profession focuses on a few metrics, such as prediction, recall, and inter-coder reliability statistics like Cohen's Kappa and AC1. But since coding schemes frequently measure multiple independent concepts of interest, these statistics are insufficient to judge whether an algorithm is suitable for a task. This poster examines the use of a sensitivity analysis method to isolate an algorithm’s performance against a single concept of interest. Such a method would be useful to a political scientist interested in justifying the selection of a specific algorithm for the study of (as an example) welfare and social issue attention within political party platforms. In an example, I examine a method which illuminates the issues being captured and separated well by a suite of algorithms under analysis. I take some party platforms which have been scored by expert humans, such as those developed by the Comparative Manifesto Project (CMP). Next I calculate the predictive performance of a handful of regression algorithms applied to issue attention in the party platforms. Finally, I conduct a sensitivity analysis of the algorithms by deleting sentences (the concepts or units under analysis) from the original platforms and then re-running the automated analysis techniques against the truncated platforms. The selection of sentences for deletion is random in a control study but based on the codes assigned by expert human annotation in the second pass. The result is a comparison of the predictive performance of the algorithms with specific classes of issues removed from the text. For discussion, I then propose that a method such as this should be a standard sensitivity analysis technique for evaluating the performance of computational algorithms against human annotated content analysis of political text when the unit of analysis in the human annotated work is considered independent.
Park	Jong Hee	Washington University in St. Louis	Discrete Time Series Model using Markov Chain Monte Carlo Methods	Analyzing time dependent discrete data have posed a serious challenge to applied researchers in political science. In this paper, I introduce two modeling strategies for time dependent discrete data: multiple changepoint models and nonlinear state space models using data augmentation method. These two models can be considered in the same family, time varying parameter models, in which observed data are assumed to depend on latent states. The estimation is based on Markov chain Monte Carlo algorithms. I apply these models to testing theories of international conflicts.
Mikhailov	Slava	Trinity College Dublin	Policy Preferences and Budgetary Structure Revisited: Hierarchical Bayesian Modelling	My research aims to investigate the role of institutions of representative democracy in public policy outcomes through the effects of policy preferences on budgetary composition. Addressing the party nature of modern representative democracies in Europe I assess whether policy preferences of political parties (in and outside a government) become realised in structural composition of budgetary spending. Much existing work in the field maps policy positions of political actors based on the preferences expressed in the electoral arena of political competition (election manifestos and expert surveys). Information contained in legislative arena of political competition, which is vital to the functioning of democracy in general and budgetary process in particular, is thus excluded from the analyses. At the same time, budgetary politics is predominantly analysed based on preferences expressed in the legislative arena by the voting behaviour of legislators (roll-call vote analysis). In order to integrate two strands of research and address inherently multidimensional nature of budgetary process I estimate a hierarchical Bayesian model. Pre-election policy positions are derived from the pledges made by political actors in electoral arena as stated in the election manifestos, while post-election policy positions are derived from the legislative arena utilising data on legislative voting behaviour in West European parliamentary democracies.
Fukumoto	Kentaro	Gakushuin University	The Modified Extended Beta Binomial Model of a Volatile Approval Rate by Heterogeneous Citizens: An Application to the Japanese Case	This paper intends to contribute to studies of approval rate methodologically and substantially. First let me explain methodological improvement. Preceding studies of approval have employed the linear normal model. This model, however, is flawed in two ways. First, the old model inadvertently does not take into account the fact that approval rate falls between 0 and 100. Second, scholars implicitly assume that respondents' propensities to approve the government are homogeneous, while it would sound more plausible to think that they are heterogeneous. For example, Democrats may well tend to approve a Democrat president rather than Republicans. Thus, I propose a new model, the logit modified extended beta binomial model. This model not only addresses boundedness of approval rate and citizen's heterogeneity, but also predicts that the heterogeneity of a citizen's attitude, particularly the presence of non partisan voters or lack of information, makes approval rate more volatile. My paper also sheds a new light on substantial aspects of approval rate. Political scientists have shown that several political and economic factors explain approval rate: party support, change of government, duration of government, and economic performance indicators - such as GDP growth, inflation, and unemployment. Though I also use them as explanatory variables, this work adds two new causes of popularity to the literature: favored international circumstances and a chief executive's youth. I apply this model to data gathered monthly within Japanese from 1960 to 2004. The results show that the rate of non partisan voters increase the variance of approval rate; when more people are in favor of the United States, the cabinet becomes more popular, and younger prime ministers win more of the citizens' support.
Parker-Stephen	Evan	University of North Carolina at Chapel Hill	A Motivation-Context Model of Political Learning	This study examines political-information dynamics in the American mass public. Its central premise is that citizens regularly pick up on information about changing political conditions and use it to revise or ``update'' their perceptions of these conditions. Because this process requires the merger of prior beliefs and new information, I model perception updating using the Bayesian learning framework. The central thrust of the model involves individuals' evaluations of information-signal credibility. And, to a great extent, these ``credibility calculations'' are generated by the relative balance of two motivational goals: the desire to reach accurate conclusions on the one hand, and to preserve prior beliefs on the other. The nature of particular informational contexts also plays a part. Empirical tests of the motivation-context model find differential learning patterns across issues. In large measure, this heterogeneity can be explained by variation in the costs associated with perceptual inaccuracy---termed the ``stakes of misperception.'' On issues where these stakes are relatively high---for example, on evaluations of the economy's standing---people's perceptions display marked responsiveness to changing conditions. Yet, where these stakes are relatively low---consider assessments of political parties' policy promises---people appear to be more resistant to information. Crucially, findings for heterogeneous learning patterns help to explain why scholars have arrived at contradictory conclusions about the quality of mass political learning (e.g., Gerber & Green 1998, 1999; Bartels 2002).
Hopkins	Daniel	Harvard University	Taxing Places: Estimating Contextual Influence on Attitudes toward Taxes	In recent years, survey researchers have made increasing use of contextual variables to explain political behavior in the U.S. Yet both measurement error and selection bias plague quantitative estimates of contextual influence. Past researchers typically use county- or even state-level demographics as proxies for a respondent’s local environment, and they rarely incorporate information about a respondent’s past environments. There is, then, a significant gap between our theories of contextual influence—which suggest that context’s influence should grow with time—and our estimates of contextual influence, which ignore the period of time spent in the current locality. Also, past work usually deals with the selection bias induced by residential mobility only by explaining it away: rarely are attempts made to estimate that bias directly. This paper makes use of 2004 American National Election Study data, including restricted geographic information, to provide more accurate estimates of the impact of racial and ethnic contexts on respondents’ attitudes about taxing and spending. Methodologically, it draws on recent advances in matching techniques to isolate the distinctive influence of local contexts. By employing tract-level and municipal-level contextual data alongside county-level data, this paper illustrates the extent to which county-level estimates of context can be misleading. It concludes that selection bias does not explain much of the observed correlation between local ethnic and racial diversity and attitudes toward taxation, and that moving to a more diverse environment does indeed dampen one’s support for taxing and for many forms of government spending.
Moore	Ryan	Harvard University	Estimating the Causal Effects of Mexico's "Seguro Popular"	Political science efforts to estimate causal effects should be both -- "political" and "scientific". "Political" experimental designs for estimating the causal effects of policy interventions must take the preferences of political actors seriously. But even in real-world deployment, these designs should still yield valid scientific inferences. The currently-ongoing evaluation of Mexico's "Seguro Popular" incorporates these principles. This project includes the input of federal and state officials, the creation of matched pairs of geographic health care units, the definition of several levels of analytic interest, and the estimation of causal effects at these levels. At the health unit level, this project combines completely randomized assignment with continuously-measureable compliance effort. At the individual level, it features an encouragement design (Hirano, et al. 2000) and a further statistical complication: the implementation of treatments in small areas forces researchers to confront the usual assumption of unit non-interference. Bayesian and maximum likelihood estimates of causal effects draw on unique data produced in the first wave of a 35,000 household survey completed in seven Mexican states in fall 2005.
Monogan III	James E.	The University of North Carolina at Chapel Hill	Heteroscedastic Models of Latent Trajectories: A Bayesian Approach	Several studies in political science have used latent growth models (LGMs) to capture individual change on a variable over time. These models fit a line to each respondents's observations across panel waves and handle cross-sectional predictors better than autoregressive models. In this paper, I argue that Bayesian hierarchical models allow researchers to estimate the stability of a variable in an LGM. More pervasive approaches to LGMs include confirmatory factor analysis and hierarchical linear modeling, both of which allow the researcher to model the systematic variance in linear latent trajectories. However, researchers interested in the predictability of an individual's trajectory would need to use a two-step estimation procedure with these approaches. First, the latent trajectory's slope and intercept need to be extracted from factor scores or Empirical Bayes estimates of the coefficients. Second, the researcher could model the slope and intercept with a multivariate normal distribution that includes a heteroscedastic variance model, and these variance model coefficients would speak to predictability. Bayesian hierarchical models, on the other hand, are more accommodating to variance models and therefore allow a one-step estimation procedure. I demonstrate the use of this model by showing that attitudes on social welfare are less predictable, and hence unstable, among attitudinally ambivalent individuals.
Hui	Iris	University of California, Berkeley	Is it Worth Going the Extra Mile to Improve Causal Inference?	How do we evaluate a voting system when it is only used in one county? Can we use matching methods to compare precincts in the “treatment” county with similar ones in another county? What kind of matching methods should we use and how should we incorporate geography into our matching? More generally, how can we use geographic information to enhance our ability to make causal inferences? This paper explores these questions by developing an evaluation of the Los Angeles County Inka-Vote system—a “Votomatic style punch card system adapted to optical scanning technology—that was used in 2004 and 2005. The first part of the paper provides an overview of the existing methods for evaluating the performance of voting systems and for detecting voting irregularities. In addition to using scatter plots, regression methods, and difference-in-difference designs, we suggest the use of matching methods and Geographic Information Science (GIS) tools. Intuitively, precincts that are geographically adjacent share many similar socio-economic and demographic characteristics. If all voting systems perform in the same manner, one would expect the residual vote rates of adjacent precincts along the border of two counties to be roughly similar. We explore whether adjacent precincts split by a county line are really similar, whether residual vote rates in border precincts do differ between Los Angeles County and its adjoining counties, and whether geographic propinquity adds inferential leverage beyond matching methods that are based upon Census information for precincts. Answering these questions requires overcoming several challenges. First, electoral returns are available at the precinct level while Census demographic characteristics are measured at the block (or block-group level). This is a common problem in social science research as political or social geography rarely follows Census geography. We demonstrate two methods, namely area calculation and spatial spline, for addressing this problem. Second, in order to identify precincts that are spatially adjacent, we will use network analysis to obtain more accurate measurement of physical distance. Third, we will take up a methodological question—does physical distance play a role in matching? In a spatial context, does physical distance serve as a good proxy for the measurement of precinct characteristics that are not adequately measured by Census data? Do we get the same or different results if we include physical distance as a matching factor in addition to Census information?
Diamond	Alexis	Harvard University	The Effects of UN Intervention after Civil War	A basic goal of political science is to understand the effects of political institutions on war and peace. Yet the impact of United Nations peacebuilding following civil war remains very much in doubt following King and Zeng (2006), which found that prior conclusions about these causal effects (Doyle and Sambanis 2000) had been based more on indefensible modeling assumptions than evidence. This paper revisits the Doyle and Sambanis (2000) causal questions and answers them using new matching-based methods that address the issues raised by King and Zeng (2006). Methods are validated for the Doyle and Sambanis (2000) data via their application to a dataset with similar features for which the correct answer is known. These new methods do not require assumptions that plagued prior work and are broadly applicable to important inferential problems in political science and beyond. When the methods are applied to the Doyle and Sambanis (2000) data, there is a preponderance of evidence to suggest that UN peacekeeping has a positive effect on peace and democracy in the aftermath of civil war.
Favretto	Katja	UCLA	Does Impartiality Matter? The Role of Bias in Major Power Mediation	In this paper, I rely on a crisis bargaining game to examine whether a close alignment of preferences between a major power and one of the parties involved in an international dispute affects the success or failure of the major power's attempt to negotiate a settlement in the dispute. I show that the degree of preference alignment, or bias, can influence the outcome of major power-organized settlement talks by revealing private information about whether or not the major power is committed to enforcing an agreement through the use of force. A non-monotonic relationship between bias and the outcome of the intervention exists: When bias is high, a peaceful outcome is more likely because the embattled parties are more certain the intervener is resolute. By contrast, when the degree of bias is middling, negotiations are more inclined to fail because the adversaries believe the intervener may be incapable of punishing the disfavored party if it fails to settle. When bias is low, a peaceful outcome is again more likely because the intervener seeks to locate agreements both adversaries find acceptable.
Lee	Han Soo	Texas A&M University	Government Type and Legislative Efficiency	Many political scientists have addressed the question of whether divided government causes legislative gridlock. Their empirical findings are mixed and even contradictory. My study focuses on how different measures of “legislative efficiency” and statistical methods lead to different conclusions about the relationship between government type and legislative efficiency. Mayhew (1991) measures legislative efficiency as “legislative production,” which is the number of important bills passed. Based on this measure, Mayhew (1991) concludes that divided government is not associated with legislative efficiency. However, my study challenges Mayhew’s findings since efficiency should also be based on input. Additionally, Mayhew used an inappropriate statistical method (OLS) to evaluate count data. In contrast, my study measures legislative efficiency as “legislative productivity,” which is the proportion of important bills passed from all that are considered. The study tests how government type affects legislative production and productivity using appropriate MLE methods. First, since legislative production as measured by Mayhew used counts of important bills passed, I employed Poisson/Negative Binomial regression to replicate his analysis. Second, I evaluate legislative productivity using my improved measure using Beta regression because legislative productivity is measured as the proportion of important bills passed from all that are considered. The Beta regression results show that divided government strongly affects legislative productivity.
Egan	Patrick	University of California, Berkeley	Issue Ownership, Representation, and Elections	In this paper (co-authored with Jasjeet Sekhon), we begin by briefly discussing some empirics indicating that politicians exploit "issue ownership"—-the degree to which the public trusts one party to better handle a particular issue than the opposing party—-to enact policy that is relatively unresponsive to changes in public opinion. A hypothesis that follows from this result is that voters discount the extreme positions taken by a politician on issues her party owns, while weighting more heavily the positions the politician takes on issues her party does not own. To test this hypothesis, we use matching methods to re-analyze the dataset on Congressional roll-call votes and election returns from 1956 through 1996 assembled by Canes-Wrone, Brady and Cogan (AJPS 2002). Our analysis indicates that their key finding--that members of Congress who take extreme positions are punished for doing so by voters--rests heavily on their model's assumed functional form, and fails to hold under the non-parametric matching approach. However, we do find support for our issue-ownership hypothesis: voters punish politicians who take extreme positions, but only on issues their parties do not own. We conclude with a discussion of the substantive and methodological issues raised by our analysis.
Esarey	Justin	Florida State University	Statistical Estimation of Bayesian Games: A Fixed Effects Quantal Response Model	Previous approaches to the statistical estimation of strategic games (Signorino 1999, 2003a, 2003b) assume a game with only one type of player. But much of the literature on Bayesian games is driven by the recognition that hard-to-observe player types can substantially affect the outcome of strategic interaction. In this paper, I show that type effects in this form can be explicitly incorporated into the error structure of the logistic QRE model, implying that estimators for Bayesian games can be derived from that structure. I develop two different error structures that unit heterogeneity can imply in the logistic QRE model (based on the distribution of information in the game), show that ignoring these effects biases the estimation of observable parameters, and develop appropriate corrections using a variant of the fixed effects estimator. I show that unit parameters can be directly estimated with reasonable reliability in cases where the objective utility payoffs are known (e.g., in laboratory experiments,) allowing researchers the opportunity to directly measure the magnitude of altruism and greed implied by behavior.
Rabinovich	Julia	Northwestern University	The Conditional Nature of Administrative Responsiveness to Public Opinion	This project is one of the first to explicitly study the theoretical and empirical relationship between public opinion and the policies of unelected administrative agencies. Public opinion scholars have usually paid little attention to the policymaking delegated to administrative agencies (e.g., Page and Shapiro 1992, Erikson, MacKuen and Stimson 2003). At the same time, public administration scholars have focused almost exclusively on the role of political elites, such as elected politicians and interest groups, and have rarely included the preferences of the general public in their analyses (e.g., Weingast and Moran 1983; Moe 1984; McCubbins, Noll, Weingast 1987). This research develops a theoretical framework that explains how legislative oversight – the main mechanism connecting legislative and executive institutions – affects both administrative policy-making and, consequently, administrative responsiveness to public opinion. The game-theoretical model suggests that, contrary to conventional wisdom, once congressional oversight mechanisms are taken into account, administrative responsiveness and legislative responsiveness do not always appear in tandem. For example, when the government is divided – i.e. when the respective preferences of the executive and legislative branches are on opposite sides of the political spectrum – a shift of legislative preferences toward the preferences of the general public will result in less congruence between the public’s preferences and the actual policies enacted by the administrative agencies. My empirical analysis focuses on policy-making by federal regulatory agencies. My theoretical predictions are tested using data on enforcement policy decisions made by the Federal Trade Commission and the Federal Communications Commission. I build a time-series dataset that includes measures of agencies’ performances, such as enforcement activities, official inspections, notices and citations, and criminal charges filed by the agencies against regulation violators from 1960 till 2004. Using these data, I estimate the impact of the public’s general ideology and specific regulatory policy preferences on the performance measures of the two commissions while controlling for agencies’ preferences and the mechanisms of congressional oversight. Since my theoretical model predicts that institutional constraints, such as the threat of congressional intervention, induce agencies to enact policies that are different from the policies they would have enacted in the absence of institutional constraints, I need to devise a method for obtaining measurements of ideological preferences that are independent of actual policy-making. I use two different methods of estimating commissioners’ preferences. First, I use each commission member’s partisan affiliation as a proxy for his/her ideological preferences in a given year. In addition, I take advantage of the fact that each commissioner’s nomination has to be confirmed by the Senate. The confirmation votes’ cutpoints are used as proxies for the locations of the commissioners’ respective ideal points. The main advantage of using this method is the ability to measure both legislative and administrative preferences by using the same scale. Furthermore, unlike models that use the nominating president’s ideal points to measure an agency’s preferences (e.g., Shipan 2004), this new measure allows for variations in the preferences of commissioners nominated by the same president.
Linzer	Drew	UCLA	A Comparative Analysis of Ideological Constraint in New versus Old Democracies using Latent Class Models	The promise of democratic institutions to promote economic growth and political stability remains largely unfulfilled in many new democracies. This study is motivated by the question: What is holding them back? Democratic theorists emphasize the importance of electoral accountability as the key mechanism for effective governance. I argue that ideological constraint, as set forth by Converse (1964), is a precondition for electoral accountability. Democracy works because exposure to democratic institutions increases opinion constraint. A lack of constraint reveals individuals' opinions to be largely random, which inhibits accountability. To the extent that constraint is slow to develop in new democratic regimes, the quality of governance will suffer. Testing this hypothesis requires a measure of ideological constraint that is appropriate and reliable for cross-national research. I develop just such a measure, based upon the statistical technique of latent class analysis, that is superior to existing bivariate correlational measures. Applying this methodology to data from a range of cross-national academic global opinion surveys, I measure the constraint of political ideologies in a diverse sample of new democracies as compared to established democracies. This reveals the extent to which new and established democracies actually differ in terms of their ideological context. I then demonstrate empirically that constraint varies in expected ways with exposure to democratic rule. The latent class model models the distribution of survey responses as a mixture of latent cross-classification tables across discrete ideological "types" of survey respondent. The model clusters respondents and estimates the class-conditional response probabilities to each question. It further estimates the effect of concomitant variables in predicting latent class membership. To estimate the latent class model, I employ a new R package I have developed with Jeff Lewis that, for the first time, enables R users to estimate latent class models with both polytomous outcome variables and covariates. For each surveyed country, estimates of respondents’ class membership and probability of giving each survey response are then used to measure levels of constraint both within and across class. My research into ideological constraint bridges a gap between survey-based behavioral studies of ideology and studies of issue "dimensionality" primarily using dichotomous dependent variables. It is also unique in its cross-national approach to the study of ideological constraint. Which cleavages--as well as how many--actually shape political conflict in new democracies, remains an open question. But it is answerable with the proper statistical analysis of existing cross-national public opinion data. Drew Linzer, UCLA
Levy	Naomi	UC Berkeley	National Identity Measurement: A Structural Equation Approach	The goal of this paper is to contribute to the conceptualization and measurement of national identity as it is realized at the individual level. Many scholars of identity who use quantitative methodology employ simple, one-dimensional measures such as census categories. However, national identity is a complex theoretical construct that is best treated as a multi-dimensional variable. Thus, I employ structural equation modeling to measure the multiple facets of subjects' national identities in Bosnia & Herzegovina (BiH) and Croatia. The data used in this paper derive from an original survey I administered in May 2004 and May 2005 to the students and history teachers of 21 secondary schools in 12 towns in BiH (N=2,989) and Croatia (N=1,468). This paper uses multiple measures for the strength of students' identification with both their ethnic group and the state. My analyses proceed in three stages. I begin by examining the opposition of state identity and ethnic group identity, and then turn to separate analyses of students' identification with each of these entities. This paper will be a chapter in my dissertation, which examines the broad questions of how individuals conceive of their national identities and how national identity is learned.
Ivanchenko	Roman	the Ohio State University	Combining Frailty and Cure Approaches to Analyze Interactions between Supreme Court and Congress.	One of the recent developments in the political methodology allows estimation of duration models, while accounting for the proportion of cases that are likely to never experience the event in question (Box-Steffensmeier and Jones, 2004). The use of a split pop model allows evaluating the probability of never experiencing the event in question, and it allows estimating the effects of covariates on duration for those cases that have a chance of experiencing the event. One of the assumptions made by a split-pop model is homogeneity of duration among the susceptible observations. A different approach to modeling event history data is by taking into account the unobserved heterogeneity across individual cases, which is accomplished by including the frailty component. Usually, the frailty term is assumed to follow the gamma or inverse Normal distribution, neither of which allows observations to have zero risks. This creates a problem should one need to estimate a duration model that accounts both for the proportion of the individuals that will never experience the event and for the unobserved heterogeneity. Thus, although the split pop framework accounts for some degree of heterogeneity, failure to properly account for unobserved variance can create distorted results. Conversely, incorrect results could be produced by a frailty model that does not allow for a non-susceptible group of observations. Two recent developments in biostatistical literature allow a combination of frailty and cure approaches. The use of the compound Poisson distribution allows statistical modeling of zero frailty (Aalen, 1992). Aalen (1992) suggests thinking about the compound Poisson frailty as a sum of random shocks each of which is randomly distributed. A frailty term of zero corresponds to a zero risk of experiencing the event. This corresponds to a zero probability of experiencing the event in the usual split-pop setting. Using interrelationships between the Laplace transforms and survivor functions allows one to account for frailty and derive a relatively straightforward likelihood function. Longinin and Halloran (1996) and Price and Manatunga (2001) offer a different approach by keeping the probability of never experiencing the event separately from the frailty term. With some probability an observations will be non-susceptible, and with the complement of this probability, an observation will be susceptible and suffer from heterogeneity. The frailty mixture approach allows separate statistical modeling of probabilities of belonging to the group that will never experience the event in question and duration until the event for susceptible observations demonstrating heterogeneity. This project extends the works of Aalen (1992) and Price and Manatunga (2001) by combining the frailty and split-pop analyses and applying them to a subfield of political science. This methodological approach allows a reexamination of the interactions between Congress and Supreme Court. Majority of federal statutes are likely to be never overturned by Supreme Court, thus, the split-pop approach seems to be well suited. Additionally, one needs to take into account the unobserved heterogeneity across the statutes that have a positive probability of being overturned. Therefore, the combination of frailty and split-pop analyses seems especially appropriate in the study of Court-Congress interactions. The goal of this project is to analyze the duration to invalidation of federal statutes by the U.S. Supreme Court while accounting for both the probability of no invalidation (non-controversial legislation) and unobserved heterogeneity across individual observations. The unit of analysis is a federal statute. The observations are statutes passed by Congress since the mid 1980’s. Similarly to Harvey and Friedman (2006), this analysis will concentrate on the influence that political predilections of the pivotal members of the Court and Congress, as well as the threat of legislative invalidation, might have on the duration until invalidation; however, this project will extend the cure model of Harvey and Friedman (2006) to include unobserved heterogeneity across federal statutes. Both the compound Poisson model and the frailty mixture model will be estimated. The flexibility of the frailty mixture approach will allow estimation of the time to invalidation as well as the probability of belonging to the group of laws that will never be invalidated.
Ramirez	Mark	Texas A&M University	The dynamics of congressional approval 1974-2004: Legislative productivity or gridlock	Legislative gridlock scholarship suggests public approval of Congress decreases when Congress fails to pass important legislation. Public opinion scholars, however, have found congressional approval decreases as legislative productivity increases partly due to increases in congressional conflict. What can explain the tides of congressional approval across time? Do increases or decreases in legislative productivity coincide with decreases in congressional approval? This paper argues inconsistencies in previous findings are the result of measurement error. New measurement strategies of both legislative productivity and procedural conflict lead to a more accurate depiction of the relationship between public opinion and congressional approval. A distributed lag time series regression model is used to estimate congressional approval quarterly from 1974 to 2004. The results have important implications for both the study of Congress and public opinion.