Research

How Data Choices Shape Structural Gravity Estimates: Evidence from EU Membership  under R&R

with Harald Oberhofer

This paper quantifies how the construction of trade data shapes trade policy estimates in structural gravity models. Using 14 widely employed trade datasets and alternative sample restrictions, we generate 643 estimates of the EU membership trade effect from a standard structural gravity specification. We then use these estimates in a meta-regression framework to identify how data and sample choices contribute to variation in the estimates. In a Shapley decomposition, the choices explain 69.1% of the variation, with the selection of a trade dataset alone accounting for 39.4\% of the explained variation. We further distinguish between decisions that affect the measurement of a common estimand and those that alter the underlying economic margin. Including domestic trade flows increases the estimated EU membership effect by 21.2 percentage points on average and nearly doubles the implied expenditure-weighted welfare gain in a four-sector general equilibrium model. Finally, we relate differences in data generating processes across datasets to estimated trade effects, showing that trade data construction is a first-order determinant of structural gravity results. This implies that conclusions from gravity-based policy evaluations are sensitive to data choices and require explicit consideration of data construction as part of the empirical identification.

Bounded Attention and Attenuated Elasticities  under review

with Tingmingke Lu     

We study how bounded attention affects the structural estimation of the elasticity of substitution. In a sparse-max model, equilibrium prices and expenditure shares are observationally equivalent to those of a rational market. Its elasticity equals the structural elasticity scaled by the attention weight. The supply-side orthogonality condition of the covariance-based estimator holds at this attenuated elasticity but fails at the truth. Absent outside information, the estimator converges to the attenuated elasticity. Bounded attention leads to an identification failure that finite-sample and weak-instrument corrections cannot resolve. Because attention cannot exceed one, the estimate provides a lower bound on the structural elasticity.