Comparing Zero-inflated Poisson, Zero-inflated Negative Binomial and Zero-inflated Geometric in Count Data with Excess Zero

Abstract

Count data often violate the assumptions of a normal distribution due to the fact that they are bounded by their lowest value which is zero. The Poison distribution is sometimes suggested but when the assumption of equal mean and variance is violated due to over-dispersion and presence of zeros we tend to look in the direction of other models. Zero-inflated data falls in this category. The zero-inflated and hurdle models have been found to fit this scenario. The proportions of zero in the data often affect the choice of the models. Our study used the Monte Carlo design to sample 1000 cases from positively skewed distribution with 1.25 as mean vector and 0.10 as zero-inflation parameter. The data was analysed using the method of the maximum likelihood estimation. The Zero-Inflated Poisson, Zero-Inflated Negative Binomial and Zero-Inflated Geometric were fitted; the standard error and Akaike Information Criterion were obtained as measures of model validation with ZIP outperformed ZINB and ZIG.