Bayesian and Likelihood Inference for 2 x 2 Ecological Tables: An Incomplete-Data Approach

Ying Lu Department of Sociology, University of Colorado at Boulder, Boulder, CO 80309 e-mail: ying.lu{at}colorado.edu Aaron Strauss Department of Politics, Princeton University, Princeton, NJ 08544 e-mail: abstraus{at}princeton.edu e-mail: kimai{at}princeton.edu (corresponding author) Ecological inference is a statistical problem where aggregate-leveldata are used to make inferences about individual-level behavior.In this article, we conduct a theoretical and empirical studyof Bayesian and likelihood inference for 2 x 2 ecological tablesby applying the general statistical framework of incompletedata. We first show that the ecological inference problem canbe decomposed into three factors: distributional effects, whichaddress the possible misspecification of parametric modelingassumptions about the unknown distribution of missing data;contextual effects, which represent the possible correlationbetween missing data and observed variables; and aggregationeffects, which are directly related to the loss of informationcaused by data aggregation. We then examine how these threefactors affect inference and offer new statistical methods toaddress each of them. To deal with distributional effects, wepropose a nonparametric Bayesian model based on a Dirichletprocess prior, which relaxes common parametric assumptions.We also identify the statistical adjustments necessary to accountfor contextual effects. Finally, although little can be doneto cope with aggregation effects, we offer a method to quantifythe magnitude of such effects in order to formally assess itsseverity. We use simulated and real data sets to empiricallyinvestigate the consequences of these three factors and to evaluatethe performance of our proposed methods. C code, along withan easy-to-use R interface, is publicly available for implementingour proposed methods (Imai, Lu, and Strauss, forthcoming). Authors' note: This article is in the part based on two workingpapers by Imai and Lu, "Parametric and Nonparamateric BayesianModels for Ecological Inference in 2 x 2 Tables" and "QuantifyingMissing Information in Ecological Inference." Various versionsof these papers were presented at the 2004 Joint StatisticalMeetings, the Second Cape Cod Monte Carlo Workshop, the 2004Annual Political Methodology Summer Meeting, and the 2005 AnnualMeeting of the American Political Science Association. We thankanonymous referees, Larry Bartels, Wendy Tam Cho, Jianqing Fan,Gary King, Xiao-Li Meng, Kevin Quinn, Phil Shively, David vanDyk, Jon Wakefield, and seminar participants at New York University(the Northeast Political Methodology conference), at PrincetonUniversity (Economics Department and Office of Population Research),and at the University of Virginia (Statistics Department) forhelpful comments.     

UK OC OK? Interpreting Optimal Classification Scores for the U.K. House of Commons

Iain McLean Nuffield College, Oxford University, Oxford OX1 1NF, United Kingdom. e-mail: iain.mclean{at}nuffield.ox.ac.uk e-mail: spln{at}mail.rochester.edu (corresponding author) Poole's (2000, Non-parametric unfolding of binary choice data.Political Analysis 8:211–37) nonparametric Optimal Classificationprocedure for binary data produces misleading rank orderingswhen applied to the modern House of Commons. With simulationsand qualitative evidence, we show that the problem arises fromthe government-versus-opposition nature of British (Westminster)parliamentary politics and the strategic voting that is entailedtherein. We suggest that political scientists think seriouslyabout strategic voting in legislatures when interpreting resultsfrom such techniques.     

Statistical Backwards Induction: A Simple Method for Estimating Recursive Strategic Models

Curtis S. Signorino 303 Harkness Hall, Department of Political Science, University of Rochester, Rochester, NY 14627 e-mail: curt.signorino{at}rochester.edu Robert W. Walker Department of Political Science, Center for Applied Statistics, Washington University in Saint Louis, Campus Box 1063, One Brookings Drive, St. Louis, MO 63130 e-mail: rww{at}wustl.edu e-mail: mbas{at}gov.harvard.edu (corresponding author) We present a simple method for estimating regressions basedon recursive extensive-form games. Our procedure, which canbe implemented in most standard statistical packages, involvessequentially estimating standard logits (or probits) in a manneranalogous to backwards induction. We demonstrate that the techniqueproduces consistent parameter estimates and show how to calculateconsistent standard errors. To illustrate the method, we replicateLeblang's (2003) study of speculative attacks by financial marketsand government responses to these attacks. Authors' note: Our thanks to Kevin Clarke, John Londregan, JeffRitter, Ahmer Tarar, and Kuzey Yilmaz for helpful discussionsconcerning this paper. A previous version was presented at the2002 Political Methodology Summer Meeting.     

Expert Opinion, Agency Characteristics, and Agency Preferences

David E. Lewis Woodrow Wilson School of Public and International Affairs, Princeton University, Princeton, NJ 08540 e-mail: delewis{at}princeton.edu e-mail: clinton{at}princeton.edu (corresponding author) The study of bureaucracies and their relationship to politicalactors is central to understanding the policy process in theUnited States. Studying this aspect of American politics isdifficult because theories of agency behavior, effectiveness,and control often require measures of administrative agencies'policy preferences, and appropriate measures are hard to findfor a broad spectrum of agencies. We propose a method for measuringagency preferences based upon an expert survey of agency preferencesfor 82 executive agencies in existence between 1988 and 2005.We use a multirater item response model to provide a principledstructure for combining subjective ratings based on scholarlyand journalistic expertise with objective data on agency characteristics.We compare the resulting agency preference estimates and standarderrors to existing alternative measures, discussing both theadvantages and limitations of the method. Authors' note: We thank Tom Hammond, George Krause, and JoshuaTucker for helpful comments. We are grateful to Simon Jackmanand Shawn Treier for generously providing their code and oursurvey respondents for their time and expertise.     

Comparing Legislators and Legislatures: The Dynamics of Legislative Gridlock Reconsidered

Lawrence S. Rothenberg Department of Political Science, University of Rochester, Rochester, NY 14627 e-mail: lrot{at}mail.rochester.edu (corresponding author) Although political methodologists are well aware of measurementissues and the problems that can be created, such concerns arenot always front and center when we are doing substantive research.Here, we show how choices in measuring legislative preferenceshave influenced our understanding of what determines legislativeoutputs. Specifically, we replicate and extend Binder's highlyinfluential analysis (Binder, Sarah A. 1999. The dynamics oflegislative gridlock, 1947–96. American Political ScienceReview 93:519–33; see also Binder, Sarah A. 2003. Stalemate:Causes and consequences of legislative gridlock. Washington,DC: Brookings Institution) of legislative gridlock, which emphasizeshow partisan, electoral, and institutional characteristics generatemajor legislative initiatives. Binder purports to show thatexamining the proportion, rather than the absolute number, ofkey policy proposals passed leads to the inference that thesefeatures, rather than divided government, are crucial for explaininggridlock. However, we demonstrate that this finding is underminedby flaws in preference measurement. Binder's results are a functionof using W-NOMINATE scores never designed for comparing Senateto House members or for analyzing multiple Congresses jointly.When preferences are more appropriately measured with commonspace scores (Poole, Keith T. 1998. Recovering a basic spacefrom a set of issue scales. American Journal of Political Science42:964–93), there is no evidence that the factors thatshe highlights matter. Authors' note: Thanks to Sarah Binder and Keith Poole for furnishingdata used in our analysis and to Chris Achen and Kevin Clarkefor advice. All errors remain our own. Online appendix is availableon the Political Analysis Web site.     

The Effect of Survey Mode and Sampling on Inferences about Political Attitudes and Behavior: Comparing the 2000 and 2004 ANES to Internet Surveys with Nonprobability Samples

Jon A. Krosnick Departments of Communication, Political Science, and Psychology, Stanford University, 434 McClatchy Hall, 450 Serra Mall, Stanford, CA 94305 e-mail: krosnick{at}stanford.edu e-mail: neilm{at}stanford.edu (corresponding author) Since the inception of the American National Election Study(ANES) in the 1940s, data have been collected via face-to-faceinterviewing in the homes of members of area probability samplesof American adults, the same gold-standard approach used bythe U.S. Census Bureau, other federal agencies, and some nongovernmentresearchers for many of the most high-profile surveys conductedtoday. This paper explores whether comparable findings aboutvoters and elections would be obtained by a different, considerablyless expensive method: Internet data collection from nonprobabilitysamples of volunteer respondents. Comparisons of the 2000 and2004 ANES data (collected via face-to-face interviewing withnational probability samples) with simultaneous Internet surveysof volunteer samples yielded many differences in the distributionsof variables and in the associations between variables (evencontrolling for differences between the samples in reportedinterest in politics). Accuracy was higher for the face-to-face/probabilitysample data than for the Internet/volunteer sample data in 88%of the possible comparisons. This suggests that researchersinterested in assuring the accuracy of their findings in describingpopulations should rely on face-to-face surveys of probabilitysamples rather than Internet samples of volunteer respondents. Authors' note: We thank Randy Thomas of Harris Interactive andMorris Fiorina for very helpful suggestions. Jon Krosnick isUniversity Fellow at Resources for the Future.     

How Smart Is my Dummy? Time Series Tests for the Influence of Politics

Kevin Grier Department of Economics, 335 Hester Hall, University of Oklahoma, Norman, OK 73019 e-mail: angus{at}ou.edu Of necessity, many tests for political influence on policiesor outcomes involve the use of dummy variables. However, itis often the case that the hypothesis against which the politicaldummies are tested is the null hypothesis that the interceptis otherwise constant throughout the sample. This simple nullcan cause inference problems if there are (nonpolitical) interceptshifts in the data and the political dummies are correlatedwith these unmodeled shifts. Here we present a method for morerigorously testing the significance of political dummy variablesin single equation models estimated with time series data. Ourmethod is based on recent work on detecting multiple regimeshifts by Bai and Perron. The article illustrates the potentialproblem caused by an overly simple null hypothesis, expositsthe Bai and Perron model, gives a proposed methodology for testingthe significance of political dummy variables, and illustratesthe method with two examples.

Before the curse of statisticsfell upon mankind we lived a happy, innocent life —HilaireBelloc, On Statistics

     

A Robust Transformation Procedure for Interpreting Political Text

Georg Vanberg Department of Political Science, University of North Carolina, Chapel Hill, NC 27599-3265 e-mail: gvanberg{at}unc.edu (corresponding author) In a recent article in the American Political Science Review,Laver, Benoit, and Garry (2003, "Extracting policy positionsfrom political texts using words as data," 97:311–331)propose a new method for conducting content analysis. TheirWordscores approach, by automating text-coding procedures, representsan advance in content analysis that will potentially have alarge long-term impact on research across the discipline. Toallow substantive interpretation, the scores produced by theWordscores procedure require transformation. In this note, weaddress several shortcomings in the transformation procedureintroduced in the original program. We demonstrate that theoriginal transformation distorts the metric on which contentscores are placed—hindering the ability of scholars tomake meaningful comparisons across texts—and that it isvery sensitive to the texts that are scored—opening upthe possibility that researchers may generate, inadvertentlyor not, results that depend on the texts they choose to includein their analyses. We propose a transformation procedure thatsolves these problems. Authors' note: We would like to thank Ken Benoit, Michael Laver,three anonymous referees, and the editor for comments on earlierversions of this article.     

A Comment on Diagnostic Tools for Counterfactual Inference

Alexander Michaelides London School of Economics, Department of Economics, Houghton Street, London WC2A 2AE, UK e-mail: a.michaelides{at}lse.ac.uk We evaluate two diagnostic tools used to determine if counterfactualanalysis requires extrapolation. Counterfactuals based on extrapolationare model dependent and might not support empirically validinferences. The diagnostics help researchers identify thosecounterfactual "what if" questions that are empirically plausible.We show, through simple Monte Carlo experiments, that thesediagnostics will often detect extrapolation, suggesting thatthere is a risk of biased counterfactual inference when thereis no such risk of extrapolation bias in the data. This is becausethe diagnostics are affected by what we call the n/k problem:as the number of data points relative to the number of explanatoryvariables decreases, the diagnostics are more likely to detectthe risk of extrapolation bias even when such risk does notexist. We conclude that the diagnostics provide too severe atest for many data sets used in political science. Author's note: We thank Komei Fukuda, Don Green, Alan Gerber,and Jasjeet Sekhon for their generous help, Mike Kane for assistancewith R programming, and five anonymous referees for constructivecomments.     

Improving Data Quality: Actors, Incentives, and Capabilities

Devesh Kapur Centre for Advanced Study of India, University of Pennsylvania, 3600 Market Street, Suite 560, Philadelphia, PA 19104 e-mail: dkapur{at}sas.upenn.edu e-mail: herrera{at}fas.harvard.edu (corresponding author) This paper examines the construction and use of data sets inpolitical science. We focus on three interrelated questions:How might we assess data quality? What factors shape data quality?and How can these factors be addressed to improve data quality?We first outline some problems with existing data set quality,including issues of validity, coverage, and accuracy, and wediscuss some ways of identifying problems as well as some consequencesof data quality problems. The core of the paper addresses thesecond question by analyzing the incentives and capabilitiesfacing four key actors in a data supply chain: respondents,data collection agencies (including state bureaucracies andprivate organizations), international organizations, and finally,academic scholars. We conclude by making some suggestions forimproving the use and construction of data sets. Authors' note: For generous comments at many stages in the paper,the authors would like to thank Dawn Brancati, Bear Braumoeller,Kanchan Chandra, Jorge Dominguez, Errol D'Souza, Richard Grossman,Ana Grzymala-Busse, Andrew Kydd, David Laitin, Daniel Posner,Jasjeet Sekhon, Hillel Soifer, Jessica Wallack, and Steven Wilkinsonand the Comparative Politics Research Workshop at Harvard University,and the anonymous reviewers from Political Analysis. The authorstake full responsibility for any errors. An earlier versionof this paper was presented at the American Political ScienceAssociation Annual Meetings, Boston, MA, August 2002.     

Ignoramus, Ignorabimus? On Uncertainty in Ecological Inference

Thomas Gschwend Center for Doctoral Studies in Social and Behavioral Sciences, University of Mannheim, D7, 27, 68131 Mannheim, Germany e-mail: gschwend{at}uni-mannheim.de Ron J. Johnston School of Geographical Sciences, University of Bristol, Bristol BS8 1SS, UK e-mail: r.johnston{at}bristol.ac.uk e-mail: elff{at}sowi.uni-mannheim.de (corresponding author) Models of ecological inference (EI) have to rely on crucialassumptions about the individual-level data-generating process,which cannot be tested because of the unavailability of thesedata. However, these assumptions may be violated by the unknowndata and this may lead to serious bias of estimates and predictions.The amount of bias, however, cannot be assessed without informationthat is unavailable in typical applications of EI. We thereforeconstruct a model that at least approximately accounts for theadditional, nonsampling error that may result from possiblebias incurred by an EI procedure, a model that builds on thePrinciple of Maximum Entropy. By means of a systematic simulationexperiment, we examine the performance of prediction intervalsbased on this second-stage Maximum Entropy model. The resultsof this simulation study suggest that these prediction intervalsare at least approximately correct if all possible configurationsof the unknown data are taken into account. Finally, we applyour method to a real-world example, where we actually know thetrue values and are able to assess the performance of our method:the prediction of district-level percentages of split-ticketvoting in the 1996 General Election of New Zealand. It turnsout that in 95.5% of the New Zealand voting districts, the actualpercentage of split-ticket votes lies inside the 95% predictionintervals constructed by our method. Authors' note: We thank three anonymous reviewers for helpfulcomments and suggestions on earlier versions of this paper.An appendix giving some technical background information concerningour proposed method, as well as data, R code, and C code toreplicate analyses presented in this paper are available fromthe Political Analysis Web site. Later versions of the codewill be packaged into an R library and made publicly availableon CRAN (http://cran.r-project.org) and on the correspondingauthor's Web site.     

A Bayesian Multilevel Modeling Approach to Time-Series Cross-Sectional Data

Joseph Bafumi Department of Government, Dartmouth College,6108 Silsby HallHanover, NH 03755 e-mail: joseph.bafumi{at}dartmouth.edu Luke Keele Department of Political Science, Ohio State University,2137 Derby Hall, 154 N Oval Mall, Columbus, OH 43210 e-mail: keele.4{at}polisci.osu.edu David Park Department of Political Science, George Washington University,1922 F Street, N.W. 414C, Washington, DC 20052 e-mail: dkp{at}gwu.edu e-mail: bshor{at}uchicago.edu (corresponding author) The analysis of time-series cross-sectional (TSCS) data hasbecome increasingly popular in political science. Meanwhile,political scientists are also becoming more interested in theuse of multilevel models (MLM). However, little work existsto understand the benefits of multilevel modeling when appliedto TSCS data. We employ Monte Carlo simulations to benchmarkthe performance of a Bayesian multilevel model for TSCS data.We find that the MLM performs as well or better than other commonestimators for such data. Most importantly, the MLM is moregeneral and offers researchers additional advantages. Authors' note: A previous version of this article was presentedat the 2005 Midwest Political Science Meeting. We would liketo thank the following for comments and advice in writing thispaper: Andrew Gelman, Nathaniel Beck, Greg Wawro, Sam Cooke,John Londregan, David Brandt. Any errors are our own.     

Analysis of Cluster-Randomized Experiments: A Comparison of Alternative Estimation Approaches

Lynn Vavreck UCLA, Department of Political Science, 4289 Bunche Hall Box 951472, Los Angeles, CA 90095-1472 e-mail: lvavreck{at}ucla.edu (corresponding author) Analysts of cluster-randomized field experiments have an arrayof estimation techniques to choose from. Using Monte Carlo simulation,we evaluate the properties of point estimates and standard errors(SEs) generated by ordinary least squares (OLS) as applied toboth individual-level and cluster-level data. We also compareOLS to alternative random effects estimators, such as generalizedleast squares (GLS). Our simulations assess efficiency acrossa variety of scenarios involving varying sample sizes and numbersof clusters. Our results confirm that conventional OLS SEs areseverely biased downward and that, for all estimators, gainsin efficiency come mainly from increasing the number of clusters,not increasing the number of individuals within clusters. Wefind relatively minor differences across alternative estimationapproaches, but GLS seems to enjoy a slight edge in terms ofthe efficiency of its point estimates and the accuracy of itsSEs. We illustrate the application of alternative estimationapproaches using a clustered experiment in which Rock the VoteTV advertisements were used to encourage young voters in 85cable TV markets to vote in the 2004 presidential election. Authors' note: We thank Rock the Vote for permission to usetheir public service announcements in this field experiment.The authors are grateful to Alan Gerber for suggestions throughoutthe design phase of this project. We are also grateful to DanKotin and Margaret Coblentz, who worked with cable operators,distributed the advertisements, and assembled the data. We thankTerence Leong for his programming expertise. Replication materialsare available on the Political Analysis Web site.     

Comparing Incomparable Survey Responses: Evaluating and Selecting Anchoring Vignettes

Jonathan Wand Department of Political Science, Encina Hall, Room 308 West, Stanford University, Stanford, CA 94305-6044 e-mail: wand{at}stanford.edu e-mail: king{at}harvard.edu (corresponding author) When respondents use the ordinal response categories of standardsurvey questions in different ways, the validity of analysesbased on the resulting data can be biased. Anchoring vignettesis a survey design technique, introduced by King et al. (2004,Enhancing the validity and cross-cultural comparability of measurementin survey research. American Political Science Review 94 [February]:191–205), intended to correct for some of these problems.We develop new methods both for evaluating and choosing anchoringvignettes and for analyzing the resulting data. With surveyson a diverse range of topics in a range of countries, we illustratehow our proposed methods can improve the ability of anchoringvignettes to extract information from survey data, as well assaving in survey administration costs.     

The Introduction of Voter Registration and Its Effect on Turnout

David M. Konisky Department of Political Science, Massachusetts Institute of Technology, 77 Massachusetts Avenue, Room E53-386, Cambridge, MA 02139 e-mail: sda{at}mit.edu e-mail: konisky{at}mit.edu (corresponding author) Studies of voter turnout across states find that those withmore facilitative registration laws have higher turnout rates.Eliminating registration barriers altogether is estimated toraise voter participation rates by up to 10%. This article presentspanel estimates of the effects of introducing registration thatexploits changes in registration laws and turnout within states.New York and Ohio imposed registration requirements on all oftheir counties in 1965 and 1977, respectively. We find thatthe introduction of registration to counties that did not previouslyrequire registration decreased participation over the long termby three to five percentage points. Though significant, thisis lower than estimates of the effects of registration fromcross-sectional studies and suggests that expectations aboutthe effects of registration reforms on turnout may be overstated.     

Does Mode Matter For Modeling Political Choice? Evidence From the 2005 British Election Study

Harold D. Clarke and Marianne C. Stewart School of Economic, Political and Policy Sciences, University of Texas at Dallas, Richardson, TX 75083 e-mail: hclarke{at}utdallas.edu e-mail: mstewart{at}utdallas.edu Paul Whiteley Department of Government, University of Essex, Colchester, England CO4 3SQ e-mail: whiteley{at}essex.ac.uk e-mail: sanders{at}essex.ac.uk (Corresponding author) Although political scientists have begun to investigate theproperties of Internet surveys, much remains to be learned aboutthe utility of the Internet mode for conducting major surveyresearch projects such as national election studies. This paperaddresses this topic by presenting the results of an extensivesurvey comparison experiment conducted as part of the 2005 BritishElection Study. Analyses show statistically significant, butgenerally small, differences in distributions of key explanatoryvariables in models of turnout and party choice. Estimatingmodel parameters reveals that there are few statistically significantdifferences between coefficients generated using the in-personand Internet data, and the relative explanatory power of rivalmodels is virtually identical for the two types of data. Ingeneral, the in-person and Internet data tell very similar storiesabout what matters for turnout and party preference in Britain.Determining if similar findings obtain in other countries shouldhave high priority on the research agenda for national electionstudies. Authors' note: We thank the U.K. Economics and Social ResearchCouncil (ESRC) and Gary Williams, Senior Science Manager atthe ESRC, for their generous support of, and interest in, thisproject. We also thank the editor and anonymous reviewers forhelpful comments and suggestions.     

In Defense of Comparative Statics: Specifying Empirical Tests of Models of Strategic Interaction

Christopher Zorn Department of Political Science, University of South Carolina, Columbia, SC 29208 e-mail: zorn{at}sc.edu e-mail: ccarrub{at}emory.edu (corresponding author) Beginning in 1999, Curtis Signorino challenged the use of traditionallogits and probits analysis for testing discrete-choice, strategicmodels. Signorino argues that the complex parametric relationshipsgenerated by even the simplest strategic models can lead towildly inaccurate inferences if one applies these traditionalapproaches. In their stead, Signorino proposes generating stochasticformal models, from which one can directly derive a maximumlikelihood estimator. We propose a simpler, alternative methodologyfor theoretically and empirically accounting for strategic behavior.In particular, we propose carefully and correctly deriving one'scomparative statics from one's formal model, whether it is stochasticor deterministic does not particularly matter, and using standardlogit or probit estimation techniques to test the predictions.We demonstrate that this approach performs almost identicallyto Signorino's more complex suggestion. Authors' note: We would like to thank Randy Calvert, Mark Hallerberg,Andrew Martin, Eric Reinhardt, Chris Stanton, and Craig Voldenfor their valuable feedback on this project. All remaining errorsare our own. Replication materials are available at the PoliticalAnalysis Web site.     

Random Coefficient Models for Time-Series-Cross-Section Data: Monte Carlo Experiments

Jonathan N. Katz Division of the Humanities and Social Sciences, California Institute of Technology, Pasadena, CA 91125 e-mail: jkatz{at}caltech.edu e-mail: nathaniel.beck{at}nyu.edu (corresponding author) This article considers random coefficient models (RCMs) fortime-series–cross-section data. These models allow forunit to unit variation in the model parameters. The heart ofthe article compares the finite sample properties of the fullypooled estimator, the unit by unit (unpooled) estimator, andthe (maximum likelihood) RCM estimator. The maximum likelihoodestimator RCM performs well, even where the data were generatedso that the RCM would be problematic. In an appendix, we showthat the most common feasible generalized least squares estimatorof the RCM models is always inferior to the maximum likelihoodestimator, and in smaller samples dramatically so. Authors' note: We gratefully acknowledge the financial supportof the National Science Foundation. Katz also acknowledges thesupport of the Center for Advanced Study in the Behavioral Sciences.We are thankful to Jake Bowers, Rob Franzese, Andy Gelman, SandyGordon, Bill Greene, and Luke Keele for comments; to Larry Bartelsfor always reminding us that our judgment may outperform thedata; as well as to the anonymous reviewers of Political Analysis.     

Event Dependence and Heterogeneity in Duration Models: The Conditional Frailty Model

Suzanna De Boef and Kyle A. Joyce Department of Political Science, 219 Pond Laboratory, The Pennsylvania State University, University Park, PA 16802 e-mail: sdeboef{at}psu.edu e-mail: kjoyce{at}psu.edu e-mail: jboxstef+{at}osu.edu (corresponding author) We introduce the conditional frailty model, an event historymodel that separates and accounts for both event dependenceand heterogeneity in repeated events processes. Event dependenceand heterogeneity create within-subject correlation in eventtimes thereby violating the assumptions of standard event historymodels. Simulations show the advantage of the conditional frailtymodel. Specifically they demonstrate the model's ability todisentangle the sources of within-subject correlation as wellas the gains in both efficiency and bias of the model when comparedto the widely used alternatives, which often produce conflictingconclusions. Two substantive political science problems illustratethe usefulness and interpretation of the model: state policyadoption and terrorist attacks. Authors' note: Three anonymous reviewers gave valuable advice.Replication materials and an online appendix are available onthe Political Analysis Web site. Any errors are our own responsibility.