|
Abstract:
|
Often times, we encounter the situation in which the population of interest yields count data yet without the value of zero. The zero-truncated count models provides a way to model such response variable and have been in use for long time. However, the presence of inflation (i.e., more than expected counts than underlying distributions e.g. Poisson and Negative Binomial distribution) in non-zero counts remains a problem to models as such. A newly proposed method, the multiple-inflation Poisson models (Su et al., 2013), can accommodate the multiple-inflation in counts data with zero. In this research, we propose Zero Truncated Multiple-Inflation Count (ZTMIC) models by using the mixture of a cumulative logit model and a zero truncated count model, where the mixing probabilities are formulated with a logistic regression. An EM algorithm along with numerical optimization is proposed to obtain the maximum-likelihood estimates. The performance of proposed model is compared with other count models.
|
