Regency EF
Zero-Inflated Covariates: Should We Care About It? (304044)
*Milan Bimali, University of Arkansas for Medical Sciences
Songthip Ounpraseuth, University of Arkansas for Medical Sciences
David Keith Williams, University of Arkansas for Medical Sciences
Keywords: zero-inflated, modeling, variable selection
Zero-Inflated data have unusually higher proportion of zeros than expected under assumed standard distributions. There are different approaches to modeling zero-inflated outcomes such as Tobit models, hurdle model, zero-inflated Poisson models, and finite mixture model. The effect of zero-inflated covariates in the context of modeling has attracted little to no attention. This may be due to the fact that statistical modeling approaches typically do not make distributional assumptions regarding the covariates. Zero-inflated covariates are commonly seen in various disciplines including: counts of contraceptive use during sexual intercourse; consumption of specific fruits and vegetables; alcohol intake in teetotaler communities. In this project we examine the robustness of commonly used regression models such as linear regression and logistic regression in the presence of zero-inflated covariates. In particular we employ a simulation based approach to examine the relation between magnitude of zero-inflation and stability of these models in terms of error rates and variable selection.