The paper
This article examines finite-sample bias in PPML estimates of structural gravity models with exporter and importer fixed effects. It combines a theoretical characterization, simulations, and an application to observed trade data to assess when bias matters and how to reduce it.
Research findings
- Bias can remain noticeable in datasets with tens of thousands of observations, especially with noisy data and sector-level estimation.
- The leading bias shrinks with the square of the number of countries. Bootstrap and split-sample jackknife corrections can reduce it, with results depending on the data.
Implications for applied research
- Compare standard and bias-corrected estimates when gravity coefficients feed into assessments of trade agreements or other policy changes.
- For counterfactual modeling, assess the stability of the point estimates used as inputs; a large observation count alone does not establish that finite-sample bias is negligible.