The paper
This article revisits influential simulation evidence that PPML performs poorly in structural gravity estimation. It shows how the size of the simulated bias depends on the amount of unexplained variation in trade data.
Research findings
- Calibrating the error variance more closely to observed trade reduces the simulated distance-coefficient bias from 27% to about 4%.
- Finite-sample bias declines with sample size and rises with data noisiness; its magnitude is not determined by the error distribution’s label alone.
Implications for applied research
- Calibrate estimator comparisons to the empirical setting before using simulation results to select a method for policy analysis.
- Treat PPML as a practical starting point for structural gravity, while checking whether the sample and variance characteristics make finite-sample bias important.