Abstract:
|
Count data frequently exhibit overdispersion due to an excess of zeros, unexplained heterogeneity, and/or temporal dependency. Zero-inflated (ZI) models have gained in popularity, especially during the past decade, as the modeling preference for count data with excess zeros and these models have become standard in most statistical software. Conceptually, zeros in ZI models are assumed to be a result of two types: structural and sampling. For example, in response to the question "How often did you drink alcohol during the last 30 days?" there will be individuals who drink alcohol but chose not to drink during the last 30 days (sampling zeros) and individuals who never drink alcohol (structural zeros). Here, using simulation, we address how adequately ZI models estimate the structural and sampling zeros and the resulting impact on inference. We simulate ZI data using several scenarios for the proportion of structural and sampling zeros and non-zeros. Our simulations demonstrate that the estimated structural and sampling zeros may be biased and we discuss the impact of ZI model bias on inference.
|