Ayda Abroufarakh

York University · Economics · 2026–27 job market

Fields

Econometrics

Job market paper

Large Sample Theory for Censored Mixed Poisson Regression: Sieve Estimation with Wavelets and Neural Networks

This paper develops sieve maximum likelihood estimation for the mixed Poisson regression model with right censored counts, with the heterogeneity density represented by wavelets or by a shallow neural network with adaptive softplus activation. We prove identification and consistency, and obtain the rate n−s/(2s+1) , up to logarithmic factors, in the Hellinger metric. For the neural network sieve, the rate holds under a maintained approximation condition. Recovering the mixing density is a strongly ill-posed problem. For regular mixing densities, the intercept has zero efficient information under censoring, while the efficient information for the slopes is positive definite when the censoring threshold is at least two. Under two maintained high level conditions, the slope estimator is asymptotically normal at the semiparametric efficiency bound. In simulations, the sieve estimators give valid inference for the slopes and correct the bias of the negative binomial model when the heterogeneity is not Gamma. In an application to the RAND Health Insurance Experiment, the sieves fit the censored counts better than the negative binomial model on held-out data, and the estimated heterogeneity density is more concentrated at low values and has a much longer right tail than a Gamma density.