Zero-Inflated Negative Binomial (ZINB) Regression
Also known as: ZINB, ZINB regression, zero-inflated negative binomial model, Sıfır-Şişirilmiş Negatif Binom Regresyonu (ZINB)
Zero-Inflated Negative Binomial regression is a count model, introduced by Greene (1994), that handles count data showing both an excess of zeros and overdispersion. It combines a binary inflation process that generates structural zeros with a negative binomial count process, making it one of the most widely used distributions for real-world count data.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use ZINB when the dependent variable is a count that shows both an excess of zeros and overdispersion (variance well above the mean), with a reasonable sample size of at least about 50 observations. A Cameron-Trivedi overdispersion test should be run first to confirm overdispersion, and a Vuong test can be used to compare ZINB against a plain zero-inflated Poisson (ZIP) model. It suits cross-sectional and panel structures and accepts continuous, categorical, or binary predictors in either the count or the inflation part.
Strengths & limitations
- Handles both excess zeros and overdispersion at once, which a zero-inflated Poisson model cannot.
- The negative binomial dispersion parameter absorbs extra variance, giving more honest standard errors than Poisson-based alternatives.
- Separates structural zeros from sampling zeros, allowing different predictors for the inflation and the count processes.
- Requires a fairly large sample (at least about 50 observations) to estimate the inflation, count, and dispersion parameters stably.
- The two-part mixture is harder to specify and interpret than a single-equation count model.
- If the data are not genuinely overdispersed, the simpler zero-inflated Poisson may be preferred; model choice should be checked with a Vuong test.
Frequently asked
How is ZINB different from zero-inflated Poisson (ZIP)?
Both mix a structural-zero process with a count process, but ZINB uses a negative binomial count part that adds a dispersion parameter. This lets ZINB handle overdispersion, where the variance exceeds the mean, which ZIP cannot. A Vuong test helps decide between them.
How do I know my data are overdispersed?
Run a Cameron-Trivedi overdispersion test before fitting. If the variance of the count is significantly larger than its mean, overdispersion is present and a negative binomial component is warranted rather than a Poisson one.
How does ZINB differ from a hurdle model?
ZINB treats zeros as a mixture of structural zeros and ordinary count zeros, so zeros can come from both processes. A hurdle model treats every zero as coming from a single gate (zero versus positive) and models only positive counts in the second part.
What sample size do I need?
Because ZINB estimates an inflation equation, a count equation, and a dispersion parameter together by maximum likelihood, a reasonable sample of at least about 50 observations is recommended for stable estimates.
Sources
- Greene, W. H. (1994). Accounting for Excess Zeros and Sample Selection in Poisson and Negative Binomial Regression Models. NYU Working Paper. link ↗
How to cite this page
ScholarGate. (2026, June 1). Zero-Inflated Negative Binomial (ZINB) Regression. ScholarGate. https://scholargate.app/en/statistics/zero-inflated-negative-binomial
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Beta RegressionStatistics↔ compare
- Hurdle ModelStatistics↔ compare
- Negative Binomial RegressionEconometrics↔ compare
- Poisson RegressionEconometrics↔ compare
- Zero-Inflated Poisson RegressionStatistics↔ compare