A model averaging approach for the ordered probit and nested logit models with applications

Abstract

This document provides the results of four experimental designs in addition to the two designs considered in the paper.Design 3: The data are generated based on the ordered probit model, with J = 3,and Z i4 each distributed as i.i.d N (0, 1), Z i3 distributed as i.i.d Bernoulli(0.4), and γ set to one of the following scenarios: S1: γ = (0.15, 0.35, 0.075, −0.02) S2: γ = (0.15, 0.35, 0, 0) S3: γ = (0.15, 0, 0, 0) Design 4: The data are generated based on the nested logit model. We let τ = 0.225, and set all other parameters to the same values as in Design 3. The purpose of these two designs is to ascertain the differences in results under an increased number of explanatory variables.

Optimal designs for model averaging in non-nested models

“…In particular, model averaging with (fixed) uniform weights yields very reasonable results. These observations coincide with the findings of Schorning et al (2016) and Buatois et al (2018) who compared model averaging and model selection in the context of dose finding studies (see also Chen et al (2018) for similar results for the AIC in the context of ordered probit and nested logit models).…”

Optimal Designs for Model Averaging in non-nested Models

“…Summarizing, for small sample sizes model averaging performs better than estimation after model selection. These observations coincide with the findings of Schorning et al (2016) and Buatois et al (2018) who compared model averaging and model selection in the context of dose finding studies (see also Chen et al (2018) for similar results for the AIC in the context of ordered probit and nested logit models). In particular, model averaging with (fixed) uniform weights yields very reasonable results in our case.…”