Impact of Imputation on Performance of Goodness-of-Fit Tests for the Logistic Panel Data Model

Abstract

Goodness-of-fit tests aim at discerning model misspecification and identifying a model which is poorly fitting a given data set. They are methods used to determine the suitability of the fitted model. The subject of assessment of goodness-of-fit in logistic regression model has attracted the attention of many scientists and researchers. Several methods for assessing how well observed data can fit into logistic regression models have been proposed and discussed where test statistics are functions of the observed data values and their corresponding estimated values after parameter estimation. Considering a correctly specified panel data model with balanced data set, the conditional maximum likelihood estimates of the parameters are less biased and the estimated response variable values are actually in the neighborhood of the observed values. Relative to the induced biases of the parameter estimates resulting from imputation of missing covariates, the performances of the goodness-of-fit tests may be misjudged. This study looks at the susceptibility of the goodness-of-fit tests for logistic panel data models with imputed covariates. Simulation results show that Bayesian imputation impacts less on the goodness-of-fit test statistics and therefore stands out as the better technique against other classical imputation methods. An increased proportion of missingness however appeared to reduce the confidence interval of the test statistics which in turn reduces the chances of adopting the model under study.