Essays on Inference and Interpretation in Empirical Econometrics
Name
sarfati-sarfati-phd-economics-2026-thesis.pdf
Size
1.26 MB
Format
Adobe PDF
Checksum (MD5)
b0e9a67cae4185846805804bc7f53665
Author(s)
Sarfati, Reca
Advisor(s)
Andrews, Isaiah
Mikusheva, Anna
Abadie, Alberto
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
This thesis develops econometric tools for valid inference when researchers engage in flexible, data-dependent practices that fall outside conventional procedures. Across three chapters, I study how to preserve credible statistical conclusions when researchers deviate from pre-analysis plans, construct higher-order narratives from atomic test results, or condition their estimation on the passing of diagnostic checks.
The first chapter, joint with Vod Vilfort, studies inference when researchers deviate from their pre-analysis plans (PAPs). It is common for researchers to supplement preregistered estimates with non-prespecified findings, yet there is broad uncertainty over how to interpret such results. We consider a truth-seeking researcher who, after seeing the data, earnestly wishes to report additional estimates alongside those preregistered. We show that, even absent nefarious behavior, conventional confidence intervals and point estimators are invalid because non-preregistered estimates are only reported in a subset of potential data realizations. We propose inference procedures that account for this conditional reporting, apply them to Bessone et al. (2021), and show that adjustments can range from negligible to economically significant. We also examine robustness to misspecification.
In the second chapter, I study how to control error rates over the higher-order conclusions researchers build from collections of hypothesis tests. Economists seldom base conclusions on isolated tests; rather, they combine atomic tests in logical structures to distinguish between theories, defend causal claims, or arrange facts into coherent stories. A single set of test outcomes can support many such narratives, not all of which are true. While standard multiple testing corrections target atomic errors, they do not address errors of downstream narratives built upon them. I present a general model for the construction and testing of narratives admitting a formal definition of Type I narrative error. Partitioning narratives into monotonic and non-monotonic classes, I show that for monotonic narratives, family-wise error rate (FWER) control over atomic tests automatically delivers uniform narrative size control—a "free narrative shopping" guarantee. For non-monotonic narratives, I prove an impossibility result and provide a novel procedure linking uniform size control to joint confidence sets.
The third chapter, also joint with Vod Vilfort, studies how to incorporate the information contained in standard specification tests directly into the estimator. Empirical researchers often use diagnostic checks---covariate balance in randomized controlled trials (RCTs), pre-trends in event studies, instrument validity in IV designs---as external hurdles to estimation. We argue they should instead be integrated into estimation itself by residualizing the baseline estimator against the vector of diagnostic check statistics. This residualized estimator offers a "free lunch": it (i) eliminates inference distortions from check-based selective reporting, (ii) reduces variance without changing the estimand under correct specification, and (iii) minimizes worst-case bias under bounded local misspecification within the class of linear adjustments. Applied to the RCT in Kaur et al. (2024), residualization increases the magnitude of the baseline estimate and reduces its standard error, equivalent to roughly a 10% increase in sample size.
MIT Department
Massachusetts Institute of Technology. Department of Economics
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
