Statistics Glossary → Zero-inflated and hurdle models
Zero-inflated and hurdle modelsconcept
Two families of models for count data with far more zeros than a standard Poisson or negative binomial model predicts, which explain the zeros with a separate process.
Zero-inflated models (ZIP, ZINB) assume two kinds of zeros. 'Structural' zeros come from units for which the event can never occur, and 'sampling' zeros from units for which it could occur but did not this time. The model estimates the probability of belonging to the structural-zero group with a logistic part, and the counts of the other units with a Poisson or negative binomial part. A hurdle model handles the problem in two stages: first a binary model separates zero from positive, then a zero-truncated count model explains the positive counts.
The choice between them should be theoretical. If some zeros are counts that 'could have been positive but were not', the zero-inflated model fits better; if all zeros come from the same 'not crossing the hurdle' process, the hurdle model is more suitable. Part of the excess zeros may stem from overdispersion, so a negative binomial model should be included in the comparison. The Vuong test is used to compare non-nested models, but its validity for comparing a zero-inflated model with a standard one is debated; AIC, BIC and a comparison of expected and observed zero counts should also be considered.
An everyday example
Visitors to a national park are asked how many fish they caught that day. Seventy per cent answer zero, because most visitors never intended to fish, and some anglers simply had bad luck. A zero-inflated model handles 'did the person intend to fish?' and 'if so, how many fish?' with separate parts.
In MerQur
In Poisson / negative binomial regression, the 'Zero model' option offers (none), Zero-inflated (ZIP / ZINB) and Hurdle models and reports the raw, AIC- and BIC-corrected Vuong tests; it is intended for cases where there are clearly more zeros than the standard model expects.
Analyses where this concept appears
Poisson / Negative Binomial Regression (Count Regression)
Related concepts
Overdispersion and the scale parameterDistribution family (GLM)