Count data modelling application.

Abstract

The rapid increase of total children ever born without a proportionate growth in the Nigerian economy has been a concern and making prediction with count data requires applying appropriate regression model.. As count data assumes discrete, non-negative values, a Poisson distribution is the ideal distribution to describe this data, but it is deficient due to equality of variance and mean. This deficiency results in under/over-dispersion and the estimation of the standard errors will be biased rendering the test statistics incorrect. This study aimed to model count data with the application of total children ever born using a Negative Binomial and Generalized Poisson regression The Nigeria Demographic and Health Survey 2013 data of women within the age of 15-49 years were used and three models applied to investigate the factors affecting the number of children ever born. A predictive count modelling was also carried out based on the performance evaluation metrics (root mean square error, mean absolute error, R-squared and mean square error). In the inferential modeling, Generalized Poisson Model was found to be superior with age of household head (𝑃<.0001), age of respondent at the time of first birth (𝑃<.0001), urban-rural status (𝑃<.0001), and religion (𝑃<.0001) being significantly associated with total children ever born. In the predictive modeling, all the three models showed almost identical performance evaluation metrics but Poisson regression was chosen as the best because it is the simplest model. In conclusion, early marriage, religious belief and unawareness of women who dwell in rural areas should be checked to control total children ever born in Nigeria.