|Friday, February 19
|PS2 Poster Session 2 & Refreshments
Fri, Feb 19, 5:15 PM - 6:30 PM
A Practical Guide for Analyzing Zero-Inflated Count Data (303224)Song Wu, Stony Brook University
*Yaqi Xue, Stony Brook University
Jie Yang, Stony Brook University
Keywords: Zero-inflated count data, Poisson model, Negative Binomial model, Zero-Inflated Poisson model, Zero-Inflated Negative Binomial model, Poisson Hurdle model, Negative Binomial Hurdle model
In clinical studies, outcome variables often take the form of integers or counts with excessive zeros. For example, an interested outcome could be the number of hospital readmissions within 30 days after bariatric surgery. Although several models have been proposed for zero-inflated count data, no model always outperforms the rest ones and inferences from using different methods are often inconsistent. Therefore, it is difficult to choose the best model in practice. A practical guideline for choosing the optimal approach is presented here after a thorough comparison of six regression models popularly used for count data: Poisson, Negative Binomial, Zero-Inflated Poisson, Zero-Inflated Negative Binomial, Poisson Hurdle, and Negative Binomial Hurdle. Parameter estimates, difference between predicted and observed counts, AIC, LR statistics, Vuong statistics are compared for all these models. Both simulated data sets for different distributions of count data and several real data sets are used for illustration. How to fit these models in SAS, SPSS and R will also be introduced.