Generalizing the Results from Social Experiments: Theory and Evidence from India

Abstract

To illustrate the intuition behind the methodological contribution, I begin by laying out a simple example based on a stylized version of McKenzie and Woodru↵ (2008). Microentrepreneur profits can take two values: high or low. Suppose we have obtained experimental results that tell us giving cash transfers to microentrepreneurs caused a large increase in the share realizing high profits in location e, from 1 3 of all entrepreneurs to 2 3 . The program is intended to relax credit constraints, allowing small-scale entrepreneurs to take advantage of potentially high returns to capital and begin a virtuous cycle of capital accumulation and business growth. We observe only these binary outcomes and no covariates.We would like to know what the results from location e tell us about the causal e↵ect we can expect in location a, where no cash transfer (CT) program was implemented. Whereas 1 3 of microentrepreneurs had high profits without CTs in location e, 1 2 of microentrepreneurs are high-profit in location a. How will the di↵erence in the no-CT profits moderate the ATE

Temporal validity as meta-science

“…Given that a single study is necessarily restricted in its scope, the premise of the causal empiricist paradigm is to aggregate findings from across many studies such that the union of their scope covers the entirety of the covariate and treatment component spaces. This knowledge aggregation is not trivial, however, and this subject has been the focus of much recent methodological attention (Athey and Imbens, 2016;Dehejia et al, 2021;Egami and Hartman, 2018;Gechter, 2015;Green and Kern, 2012;Hartman et al, 2015;Ho et al, 2007;Imai et al, 2013;Kern et al, 2016;Slough and Tyson, 2022a;Stuart et al, 2011;Taddy et al, 2016;Wager and Athey, 2017).…”