Wednesday, February 13, 2008

Multi-level Regression Between Fixed Effects and Mixed Effects Models

Stefan Sperlich
(Georg-August Universität Göttingen)

Abstract: We introduce a semiparametric class of multi-level regression models that includes mixed and fixed effects models as its two extreme cases. In some practical cases, one could consider the fixed effects model as an over parametrized model without modeling but just plugging in dummies. In other words, it suffers from "too many parameters but too little model". The mixed effects model tries to overcome this by using just random effects and therefore has "too few parameters but too restricted model", where "too restricted model" refers to the necessary model assumptions made. We propose including a nonparametric term that allows the practitioner to position the model anywhere in between these extremes, depending on its data and underlying problem. Thereby, the smoothing parameter serves as its slider. We will show that so we can filter out possible dependency between covariates and random effects. We further provide consistent bootstrap procedures for possible inference and to analyze prediction power. The positive implications of using this model are highlighted in particular for small area statistics and econometrics. This is underlined by simulation studies and a real data problem.

Notice: Note unsual time and day of the week
Wednesday, February 13, 2008
Time: 11h00
Room: Sala CTT, Edificio Quelhas, ISEG