Sampling-Based versus Design-Based Uncertainty in Regression Analysis

Causal Question / Estimand

What does a regression standard error actually quantify? The paper distinguishes uncertainty about a descriptive or causal estimand arising from sampling (the sample is a draw from a larger population) versus from assignment/design (which units are treated), and shows the two can differ.

Identification Strategy

Develops a framework where the population may be (nearly) fully observed, so the usual sampling-based view is the wrong source of randomness. Under design-based uncertainty, potential outcomes are fixed and randomness comes from treatment assignment; the appropriate variance — and hence the standard error — can be smaller than the conventional robust (sampling-based) one, because there is no uncertainty about units already in hand. Provides estimators that are valid for the design-based estimand.

Key Assumptions

A well-defined assignment mechanism; clarity about the target estimand (descriptive vs. causal) and the population (sample vs. super-population). Potential-Outcomes, SUTVA.

Threats to Validity

Mismatch between the estimand/population the researcher intends and the one the standard error implicitly assumes; using sampling-based robust SEs when the relevant uncertainty is design-based (or vice versa) gives misleading inference.

Setting / Data

n/a — econometric theory, with regression as the running example.

Key Claims

  • Conventional robust standard errors implicitly assume sampling-based uncertainty; when the population is fully observed, they overstate uncertainty about causal/ descriptive estimands.
  • The correct variance depends on whether the estimand is descriptive or causal and on the source of randomness.
  • Provides standard errors valid under design-based uncertainty.

Connections

Citation

Abadie, A., Athey, S., Imbens, G. W., & Wooldridge, J. M. (2020). Sampling-Based versus Design-Based Uncertainty in Regression Analysis. Econometrica, 88(1), 265–296.