Essays on Identification and Causality
MetadataShow full item record
CitationRambachan, Asheshananda. 2022. Essays on Identification and Causality. Doctoral dissertation, Harvard University Graduate School of Arts and Sciences.
AbstractThis dissertation contains three chapters in econometrics. A common theme is identification analysis, with a particular focus on understanding what researchers can learn from data under weak assumptions on economic behavior and dynamic causal effects.
The first chapter characterizes the behavioral and econometric assumptions under which researchers can identify whether expert decision makers, such as doctors, judges, and managers, make systematic prediction mistakes in observational empirical settings like medical testing, pretrial release, and hiring. Under these assumptions, I provide a statistical test for whether the decision maker makes systematic prediction mistakes and methods for conducting inference on the ways in which the decision maker's predictions are systematically biased. As an empirical illustration, I analyze the pretrial release decisions of judges in New York City.
The second chapter, which is coauthored with Neil Shephard, develops the direct potential outcome system as a foundational framework for analyzing dynamic causal effects in observational time series settings. We place no functional form restrictions on the potential outcome process nor restrictions on the extent to which past assignments may causally affect outcomes. We provide novel conditions under which popular time series estimands, such as the impulse response function, local projections, and local projection with an instrumental variable, have nonparametric causal interpretations in terms of dynamic causal effects.
The third chapter, which is coauthored with Iavor Bojinov and Neil Shephard, proposes a rich class of finite population dynamic causal effects in panel experiments. We provide a nonparametric estimator that is unbiased for these dynamic causal effects over the randomization distribution of assignments, derive its finite population limiting distribution, and develop two methods for conducting inference. We further show that population linear fixed effect estimators do not recover causally interpretable estimands if there are dynamic causal effects and serial correlation in the assignment mechanism of the panel experiment.
Citable link to this pagehttps://nrs.harvard.edu/URN-3:HUL.INSTREPOS:37371934
- FAS Theses and Dissertations