Website CV Paper
Fields
Econometrics, Innovation
Job market paper
Semiparametric Panel Data Models with Interactive Fixed Effects
Methods for estimation and inference in semiparametric panel data models with both correlated random effects and interactive fixed effects are developed. The framework combines a Mundlak (1978) projection for the time-invariant component of unobserved heterogeneity with a local-linear filtering device, in the spirit of Huang (2013), for the time-varying component. Stacking the additive time effect and the interactive factor into a joint factor vector treats unobserved time heterogeneity uniformly. The estimator augments the partially linear regression with cross-sectional averages of the dependent variable and the regressors and recovers the parametric block by Robinson (1988) partialling, requiring no iteration, no principal components and no knowledge of the number of factors; coefficients on time-invariant regressors, which fixed effects absorb, are identified through the Mundlak projection. Asymptotic properties are characterized under joint large-N, large-T asymptotics, and Monte Carlo evidence documents finite-sample performance. The estimator is applied to the firm-level idea production function, with Bloom, Jones, Van Reenen and Webb (2020) as the parametric benchmark. Returns to research intensity are increasing, and a linear approximation to the benchmark is rejected while the constant-elasticity form it approximates is not. The returns-to-scale parameter the benchmark sets by assumption is estimated here, at a value the data cannot distinguish from the assumed one but which points to ideas getting harder to find faster than the benchmark reports.