Asymptotic and Finite Sample Bias Correction via Resampling Methods for Latent Variable Models

• Activity Number: 449
• Type: Topic Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Social Statistics Section
• Author(s): Maria-Pia Victoria-Feser*+ and Stephane Guerrier and Elise Dupuis-Lozeron
• Companies: University of Geneva and University of California, Santa Barbara and University of Geneva
• Keywords: monte carlo ; iterative bootstrap ; approximate model ; indirect inference

Abstract

The estimation of the parameters of complex statistical models, such as latent variables models, often lead to difficult optimization problems. Numerical approaches have been recently replaced by approximate models or approximate likelihoods. In this paper, we investigate the first approach which consists in using an approximate model which is likely to produce slightly biased estimates of the model parameters but that has the advantage of being easy to estimate. Then, several methodologies can be employed to reduce/remove the asymptotic and/or finite sample bias of the estimator, like is done with indirect inference. These techniques exploit Monte Carlo methods which includes, among others, the bootstrap and Markov Chain Monte Carlo methods. In this paper, we compare the properties of the different bias correction methods. We propose one that corrects asymptotic biases and carries out a finite sample bias correction. While the computational burden is drastically reduced, simulations from models with ordinal manifest variables and several latent variables also show that our proposed correction method provides unbiased estimates of the model's parameters.