Title: Monte Carlo studies when all models are wrong
Authors: Mark De Rooij - Leiden University (Netherlands) [presenting]
Abstract: In statistics we often use Monte Carlo simulation studies to investigate properties of statistical models in different circumstances. Most of the time we adopt a true data generating model and study the effect of misspecification of errors, structural form, or other disturbances on the bias and variance of parameter estimates and their standard errors. However, in data analytic settings most statisticians agree that all models are wrong. Therefore, the above mentioned Monte Carlo studies have little utility for data analysis. We will discuss this problem in detail and show alternative ways of performing simulation studies with targeted models. These models do not assume that the true model can be attained but instead define a target of analysis. We will illustrate the methodology using linear regression and latent variable models.