Continuous Versus Categorical Imputation Method for Unobserved Count with Zero-Inflation

Missing data is a quite common issue in any dataset. Because the impact of missing data is substantial, several attempts have been done to produce methods to estimate the missing value. There is more than one classification for the missing imputation method; one is based on the imputation for continuous or categorical data. However, sometimes difficult to decide the best method the researcher should use due to the characteristics of the data, such as count. In this study, we evaluated two imputation methods to estimate the incomplete value in count data: Predictive Mean Matching (PMM), which is usually used for continuous data, and Multinomial Logistic Regression (MLR), a method used for categorical data. MLR showed AIC more close to the AIC of the original data; nevertheless, both methods maintained the original distribution of the data with tiny uncertainty. Further, neither did affect the selection of the best model that could fit the data better. More precisely, HZINB is the best count model fitting the original data and still outperforms the other count models after imputing the missing data using PMM and MLR.