Zero-Inflated Data: A Comparison of Regression Models

The content addresses the challenge of modeling zero-inflated count data, using a dataset of bicycle usage as a practical example. It first illustrates the limitations of a standard Poisson regression model, which fails to accurately fit data with an excessive number of zero values. Two alternative model families are introduced to handle this issue: zero-inflated models and hurdle models. A detailed comparison, using both visual inspection and cross-entropy metrics, demonstrates that the Zero-Inflated Poisson (ZIP) model provides a significantly better fit than the standard Poisson model for this type of data.