The Causal Interpretation of the AKM Estimand

Abstract

The Abowd–Kramarz–Margolis (AKM) regression is widely used to analyze firm wage premia. We derive a general potential outcome decomposition of the AKM estimand and show that it is a worker-weighted average of causal effects of moving between pairs of firms. Under strong assumptions, this decomposition collapses so that the AKM estimand for a specific firm equals the causal effect of moving from the left-out firm to that firm. We give necessary and sufficient conditions for the AKM estimand to have this causal interpretation and propose tests for the sufficient conditions. In an application to Italian administrative data we document that the standard AKM estimator can be severely biased when these conditions fail. We propose a matched-AKM estimator that consistently estimates causal firm effects under weaker and more realistic assumptions. Joint paper with X. Huang, T. Lamadon, M. Mogstad, and A. Shaikh.