|
|
 |
| |
|
|
|
Impact of Magnitude of Zero Inflation of Covariates on Statistical Inference and Model Selection |
|
|
|
PP: 287-292 |
|
|
|
doi:10.18576/jsap/100201
|
|
|
|
Author(s) |
|
|
|
Milan Bimali,
Songthip T. Ounpraseuth,
David Keith Williams,
|
|
|
|
Abstract |
|
|
| Modeling approaches that do not consider zero-inflation are inappropriate in modelling the relationship between zero-inflated outcomes and covariates. Similarly, the association between zero-inflated covariate and outcome using the aforementioned approaches is also prone to estimation and inferential errors. The case of zero-inflated covariate despite being observed in a wide variety of scenarios has attracted little attention. While the need to develop and implement specialized approach to model the association between zero-inflated covariate and outcome is indisputable, a more fundamental question that needs to be explored is whether the magnitude of zero inflation is large enough to warrant concern and whether the degree of this concern depends on the overall size of the data and the analysis objective. The present paper employs extensive simulation-based approach to assess the effect of magnitude of zero-inflated covariate on a number of statistical metrics, such as error rates and variable selection rates across a wide spectrum of sample size in the context of two commonly used modeling approach – logistic regression, and linear regression.
|
|
|
|
|
|
|
|
