Unobserved Heterogeneity in Observational Studies of Political Behavior
<p>This dissertation comprises three chapters dealing with unobserved heterogeneity in observational studies. In Chapters 2 and 3, I develop new estimators that deal with unobserved heterogeneity in the cases in which panel data is not available or the outcome of interest is binary, respective...
| Summary: | <p>This dissertation comprises three chapters dealing with unobserved heterogeneity in observational studies. In Chapters 2 and 3, I develop new estimators that deal with unobserved heterogeneity in the cases in which panel data is not available or the outcome of interest is binary, respectively. In Chapter 4, I analyze the effect of parties' contacting voters on the extent of tactical voting in the 2015 and 2017 United Kingdom General Elections, applying the estimator developed in Chapter 3.</p>
<p>In Chapter 2, I develop a semi-parametric two-step estimator for linear models with unobserved individual level heterogeneity that can be applied on a series of Repeated Cross-Sections, when panel data is unavailable. I show that this estimator provides consistent and asymptotically normal estimates of the parameters of interest. Identification relies on a restriction that requires the conditional expectation of the unobserved individual-level heterogeneity on observed characteristics to be continuous. Using Monte Carlo simulations, I show that this estimator typically outperforms other available alternatives. In particular, it typically has a smaller Root Mean Squared Error, and a relatively small bias that disappears for moderate sample sizes. Furthermore, it is robust to mild violations of the continuity assumption. Finally, I also show that this estimator can recover sensible estimates compared to those from an real panel.</p>
<p>In Chapter 3, I propose a method for estimating binary outcome models with panel data in the presence of unobserved heterogeneity, called the <i>Penalized Flexible Correlated Random Effects</i> (PF-CRE) estimator. I show that this estimator produces consistent and efficient estimates of the model parameters. PF-CRE also provides consistent estimates of partial effects, which cannot be calculated with existing consistent estimators. Using Monte Carlo simulations, I show that PF-CRE performs well in small samples. To demonstrate that accounting for unobserved heterogeneity has important consequences for empirical analysis, I use PF-CRE in three studies of voting behavior: tactical voting during the 2015 British Election, support for the Brexit referendum of 2016, and vote choice in the 2012 U.S. Presidential election. In all three cases, I find that ignoring the unobserved heterogeneity leads to an overestimation of the effects of interest, and that PF-CRE is a valid approach for the analyses.</p>
<p>In Chapter 4, I apply the PF-CRE estimator developed in Chapter 3 to the study of tactical voting in the United Kingdom General Elections of 2015 and 2017. In particular, I study the effect that party contacts during the electoral campaigns has on the probability that voters decide to cast a tactical vote for a less preferred party when their most preferred party is out of the race. I show that these effects are of moderate size, but substantively important. For example, during the 2017 election, contact by the most preferred party discouraged tactical voting by 7.02%, while contact by the most preferred viable party encouraged it by 13.41%. Combining counterfactual simulations with Multilevel Regression and Poststratification I estimate the effect that party contact has on the seat distribution in Westminster through tactical voting. My results show that between 9 and 18 seats change hands, depending on the election. Importantly, the Conservative party would have obtained a majority in 2017 had non-viable parties given up contacting their supporters.</p> |
|---|