Modelling count data with extra observed zeros

Objective: To determine if the commonly used Poisson regression model fits count data when there are excess zero counts present, and to investigate the use of alternative models. Research design and methods: A total of 1229 children with type 1 diabetes (mean age 11.7 years and sd 4.1) were studied. The gender distribution was even with 605 (49.2%) males. Prospective assessment of severe hypoglycaemia (an event leading to loss of consciousness or seizure or resulting in a hospital admission) was made over the nine year period, 1992-2001. Patients were seen with their parents every three months. Data were analysed using the Poisson regression, negative binomial, zero-inflated Poisson (ZIP) and zero-inflated negative binomial (ZINB) models. The over dispersion, likelihood ratio statistics and goodness of fit statistics were calculated. Results: A comparison of all four models was made. The Poisson regression model did not fit the data well. The negative binomial is a popular alternative and fits better. The ZIP models fitted the data better than Poisson. Conclusions: The use of Poisson regression models may lead to biased parameter estimates. We recommend the use of either the negative binomial or zero-inflated models to examine risk factors associated with disease outcome. Statistical software is available to fit these models. (author abstract)