Robust Negative Binomial Regression
Also known as: robust NB regression, negative binomial regression with robust standard errors, sandwich-corrected negative binomial regression, NB2 robust regression
Robust Negative Binomial Regression models overdispersed count outcomes using the negative binomial distribution while protecting coefficient inference against misspecification of the variance function. It pairs maximum-likelihood estimation of the mean and dispersion parameters with sandwich (Huber-White) standard errors, yielding valid tests even when the assumed variance structure is only approximately correct.
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 Robust Negative Binomial Regression when the outcome is a non-negative integer count that shows overdispersion (variance substantially exceeds the mean) and you have reason to suspect the variance function may be only approximately correct — for example, residual clustering, mild zero-inflation, or unknown data-generating heterogeneity. It is particularly appropriate in health, epidemiology, and social-science applications where count outcomes are common and strict distributional assumptions are hard to verify. Do not use it as a substitute for properly handling zero-inflation (use a ZIP or ZINB model instead), and do not rely on it when the mean model itself is badly misspecified, because sandwich corrections address variance misspecification only — biased coefficients cannot be rescued by robust standard errors.
Strengths & limitations
- Accommodates overdispersed count data by explicitly modelling excess variance through the dispersion parameter α.
- Sandwich standard errors remain valid under moderate violations of the assumed variance function, reducing the risk of anti-conservative tests.
- Point estimates (coefficients) are identical to standard negative binomial MLE, so effect-size interpretation is unchanged.
- Widely available in major statistical software (R sandwich/MASS, Stata vce(robust), Python statsmodels).
- Preserves the interpretability of incidence-rate ratios (IRRs) from the log-link parameterisation.
- Sandwich standard errors are consistent only asymptotically; in small samples (n < 100) they can be poorly calibrated and bias-corrected versions (HC3) are preferred.
- The method corrects variance misspecification but cannot fix a misspecified mean function — if the log-linear mean model is wrong, coefficients will be biased.
- Estimating α alongside β increases the number of parameters and requires an adequately large sample relative to the number of predictors.
- Does not model structural zeros separately; severe zero-inflation still requires a zero-inflated negative binomial (ZINB) model.
- Robust standard errors can be less efficient than model-based ones when the negative binomial variance function is actually correct.
Frequently asked
How is Robust Negative Binomial Regression different from standard Negative Binomial Regression?
The coefficient estimates are identical. The only difference is in standard errors: the robust version uses sandwich (Huber-White) standard errors that remain valid when the assumed variance function is misspecified, whereas standard NB regression relies on model-based standard errors that are correct only if the full distributional assumptions hold.
When should I prefer Robust Poisson over Robust Negative Binomial Regression?
Robust Poisson regression (with sandwich SEs) is a popular choice for relative-risk estimation because the log-Poisson model with sandwich SEs is consistent for the mean even under overdispersion. Prefer the Robust NB model when overdispersion is severe — it will be more efficient and give better-calibrated confidence intervals in large samples, while the robust SEs protect against residual variance misspecification.
What sample size is needed for the sandwich standard errors to be reliable?
As a rough guide, n ≥ 100 observations is often cited for the ordinary sandwich estimator. For cluster-robust standard errors, at least 30–50 clusters are typically needed. In smaller samples, bias-corrected sandwich estimators (HC3 for independent data, CR2 for clustered data) are preferable.
Does the dispersion parameter α have a direct interpretation?
Yes. In the NB2 parameterisation, Var(Y) = μ + α·μ². An α close to zero means the data are nearly Poisson; larger α indicates greater overdispersion. A likelihood-ratio test comparing NB to Poisson (H₀: α = 0) formally checks whether the negative binomial variance structure is needed.
Can I use Robust Negative Binomial Regression with panel or clustered data?
Yes, but you should use cluster-robust standard errors rather than the ordinary (HC) sandwich estimator. Specify the clustering variable (e.g., subject or group id) so that the sandwich accounts for within-cluster correlation. If the clustering is very strong or the number of clusters is small, consider a mixed-effects negative binomial model instead.
Sources
- Hilbe, J. M. (2011). Negative Binomial Regression (2nd ed.). Cambridge University Press. ISBN: 978-0521198158
- Zeileis, A., Kleiber, C., & Jackman, S. (2008). Regression Models for Count Data in R. Journal of Statistical Software, 27(8), 1–25. DOI: 10.18637/jss.v027.i08 ↗
How to cite this page
ScholarGate. (2026, June 3). Robust Negative Binomial Regression. ScholarGate. https://scholargate.app/en/statistics/robust-negative-binomial-regression
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.
- Generalized Linear ModelStatistics↔ compare
- Negative Binomial RegressionEconometrics↔ compare
- Poisson RegressionEconometrics↔ compare
- Robust Poisson RegressionStatistics↔ compare
- Robust RegressionStatistics↔ compare
- Zero-inflated modelStatistics↔ compare