- This event has passed.
Martin MUGNIER (CREST) – “Identification and (Fast) Estimation of Nonlinear Panel Models with Additively Separable Two-Way Fixed Effect “
Time: 3:00 pm – 4:15 pm
Date: 28th of September 2021
Room :3001 et visio
Martin MUGNIER (CREST) – “Identification and (Fast) Estimation of Nonlinear Panel Models with Additively Separable Two-Way Fixed Effect ”
Abstract: In this paper, we study the identification and fast estimation of a class of nonlinear panel models with additively separable two-way fixed effects widely used in empirical research. We propose a novel identification strategy and show that all structural parameters of the model (common slopes, fixed effects, and link function) can be nonparametrically identified when T is large. We propose a novel iterative Gauss-Seidel procedure to implement the routinely used MLE. During each iteration, we sequentially update the estimates of individual fixed effects, time fixed effects, and common slope parameters. Importantly, the updating of the estimated fixed effects can be fully parallelized across individuals/time periods, largely alleviating computational time and memory requirement when the dataset is large. We prove that the proposed method is numerically equivalent to the MLE under standard conditions. Extensive Monte Carlo simulations show that this procedure delivers a good approximation to the MLE using only fractional running time even in the presence of a large number of fixed effects. Joint work with Ao Wang (University of Warwick).
Organizers:
Benoît SCHMUTZ (Pôle d’économie du CREST)
Anthony STRITTMATTER (Pôle d’économie du CREST)
Sponsors:
CREST
Time: 3:00 pm – 4:15 pm
Date: 28th of September 2021
Room :3001 et visio
Martin MUGNIER (CREST) – “Identification and (Fast) Estimation of Nonlinear Panel Models with Additively Separable Two-Way Fixed Effect ”
Abstract: In this paper, we study the identification and fast estimation of a class of nonlinear panel models with additively separable two-way fixed effects widely used in empirical research. We propose a novel identification strategy and show that all structural parameters of the model (common slopes, fixed effects, and link function) can be nonparametrically identified when T is large. We propose a novel iterative Gauss-Seidel procedure to implement the routinely used MLE. During each iteration, we sequentially update the estimates of individual fixed effects, time fixed effects, and common slope parameters. Importantly, the updating of the estimated fixed effects can be fully parallelized across individuals/time periods, largely alleviating computational time and memory requirement when the dataset is large. We prove that the proposed method is numerically equivalent to the MLE under standard conditions. Extensive Monte Carlo simulations show that this procedure delivers a good approximation to the MLE using only fractional running time even in the presence of a large number of fixed effects. Joint work with Ao Wang (University of Warwick).
Organizers:
Benoît SCHMUTZ (Pôle d’économie du CREST)
Anthony STRITTMATTER (Pôle d’économie du CREST)
Sponsors:
CREST