Elections are a blunt tool for accountability. Can enhanced politician-voter communication in the periods between elections improve democratic outcomes? We partner with 20 provincial assembly members in Pakistan and design an experiment with Interactive Voice Response (IVR) – a technology that enables politicians to robocall a large number of voters in their own voice to ask them questions and receive feedback. We randomize whether respondents receive a call soliciting preferences about upcoming decisions the politician must make. A follow-up call randomizes how responsive the politician is to voters’ preferences. We study the effect of this communication on voters’ evaluations of the incumbent, their views on government performance, and the prospects for electoral accountability. We also study how politicians allocate effort and make decisions in response to increased information from certain areas of their constituency.
Social spillovers and female political participation in Khyber Pakhtunkhwa - With Saad Gulzar (Stanford) and Muhammad Yasir Khan (UC Berkeley) (+)
There exists a large political participation gap between men and women in Pakistan, including at the preliminary level of voter registration. There are at least 10 million fewer registered women than men in Pakistan, and this project seeks to explain this gap by studying how social networks condition the political participation of males and females in Khyber Pakhtunkhwa.
Institutional change and dynastic persistence in Pakistan: the impact of electoral reapportionment and education minimums - With Ali Cheema (LUMS) and Farooq Naseer (LUMS) (+)
The pervasiveness of families in electoral politics around the world highlights the ability of elites to self-perpetuate, even in the face of considerable institutional change. In this paper, we test how political dynasties in Pakistan responded to two institutional reforms introduced with the explicit goal of curtailing elite power. We demonstrate that electoral reapportionment and the introduction of education minimums---two key components of a reform package designed by General Musharraf in Pakistan following his 1999 coup---had different consequences for political families, a powerful elite group in Pakistan. Using a pseudo-regression discontinuity design leveraging district level reapportionment formulae, we estimate that an additional seat within a district decreases the probability a constituency is won by a dynastic candidate by 9 percentage points, diluting the power of the incumbent elite. On the other hand, using a differences-in-differences design, we estimate that the disqualification of incumbents via education minimums only causes churning within the elite; while incumbents were disqualified, their family members won at high rates in their stead and dynastic prevalence in politics was unchanged. We argue that these differential responses to institutional reforms may be the result of diminishing marginal returns to the number of family members in elected office. These results highlight the usefulness of legislature size and reapportionment to curtailing the power of incumbent elites.
Estimating NOMINATE scores over time using penalized splines - With Jeffrey B. Lewis (UCLA) (+)[working paper]
DW-NOMINATE scores are the most widely-used measure of congressional legislators' positions in an abstract "ideology" space. By constraining how individual legislators' positions can change over their careers, DW-NOMINATE produces estimates that are comparable across time, allowing DW-NOMINATE scores to serve as the basis of much of the research on political polarization (for example, Binder 2014; McCarty et al. 2006). However, recent studies have raised concerns about the plausibility of DW-NOMINATE's strong constraints on member's ideological trajectories and how those constraints affect inferences we make about polarization (for example, Bateman and Lapinski 2016). In this paper, we develop Penalized Spline DW-NOMINATE (PSDW-NOMINATE), a new, flexible, approach to estimating the trajectories of legislators' ideal points over time within the NOMINATE framework. We use penalized spline functions (Eilers and Marx 1996, 2010) to model each legislator's ideal points over her career. PSDW-NOMINATE allows us to consider a continuum of degrees of constraint and to explore how the constraint that is placed on members' movements affects inferences about political polarization.
Kernel Regularized Logistic Regression: avoiding misspecification bias while maintaining interpretability for binary outcome regressions - With Chad Hazlett (UCLA) (+)
When faced with a binary outcome, investigators face two specification challenges: a "structural component" (e.g. XB) that makes use of covariate data, and a link function connecting the probability of observing a "1" to this structural component. Investigators typically have little or no theoretical guidance on either point. Though the choice of link function will often have little substantive impact, the choice of structural component can easily lead to severe misspecification bias. Machine learning methods can more flexibly model Pr(Y=1\|X), mitigating both the structural and link function specification concerns. However, social scientists often find machine learning tools inappropriate as they rarely enable estimate the usual quantities of interest, such as marginal effects and their standard errors. We describe Kernel Regularized Logistic Regression (KRLogit), a kernel logistic ridge regression approach that seeks to avoid costly misspecification errors through a flexible modeling approach that nevertheless allows for interpretation and inference similar to traditional generalized linear models. We automatically produce estimates that are as or more intuitive than standard interpretations for conventional logit or probit model coefficients. The method is an extension of Kernel Regularized Least Squares (Hainmueller & Hazlett 2014), but performance is improved both in theory and practice for binary outcomes. We provide the KRLogit method as an option in the existing KRLS package, also adding cluster- and hetreoskedasticity-robust standard error options, weights, and a new approximation technique that greatly improves speed for both binary and continuous outcomes.