Inference After Model Selection in Linear Mixed Effect Models

• Activity Number: 191
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Consulting
• Author(s): Adam Sima*+
• Keywords: Inference ; Model selection ; Variable selection ; Type-I error ; Bias

Abstract

The increased computing capabilities that are available in statistical software packages have made intricate model selection strategies readily available. Often times these selected models are used for inference on the same data as used for model selection. However, the literature has shown that severe biases in the standard error of the estimates result from this common use. This results in an inflation of the Type-I error rate, resulting in hypothesis tests that are liberal when compared to the nominal level. Several authors have proposed methodology to rectify these biases for independent data. Methods are less well known for the correlated data case (e.g., mixed effect models). The current methods to perform post-model selection inference on independent data are reviewed. Additionally, the inflation of the Type-I error rates is shown to exist for mixed effect models, and a potential solution to this problem is discussed.