Negative Binomial Regression: The Overdispersion Fix, NB1 vs. NB2, and a Worked Model Comparison

When count data is overdispersed — variance meaningfully larger than the mean — a plain Poisson regression understates uncertainty: standard errors come out too small, confidence intervals too narrow, and p-values too optimistic. Negative binomial regression is the standard fix. It keeps the same log-linear structure as Poisson regression but adds a dispersion parameter that lets the variance exceed the mean instead of forcing them to be equal. This guide covers what that parameter actually does, the two ways it’s commonly built into the model (NB1 and NB2), the formal test for whether you need it at all, and a full worked comparison against a Poisson fit on the same data — the depth a brief “switch to negative binomial” mention can’t cover.

Why Poisson regression breaks down

Poisson regression assumes equidispersion: for a Poisson-distributed count with mean μ, the variance is forced to equal that same μ — Var(Y) = E(Y). That’s a property of the distribution, not something you can relax by adding covariates. Real count data routinely violates it. Citation counts, hospital readmissions, publication counts per researcher, adverse-event counts per trial site, and manuscript revision requests all tend to cluster: some units generate far more events than the model’s mean would predict, and others far fewer, because of unmeasured heterogeneity the Poisson model has no way to represent. The result is overdispersion — sample variance well above the sample mean — and it is the norm in applied count-data work, not the exception.

The practical consequence is specific: overdispersion does not usually bias the coefficient estimates (the incidence rate ratios themselves) much, but it does bias the standard errors downward. That means Wald tests and confidence intervals built from a plain Poisson fit will make effects look more precise, and more statistically significant, than the data actually supports.

The dispersion parameter, and what it actually does

Negative binomial regression is derived as a Poisson-gamma mixture: instead of assuming every observation shares exactly the same rate μ, it assumes each observation’s rate is itself a gamma-distributed random variable centered on μ, and the observed count is Poisson given that individual rate. Mixing over that unobserved heterogeneity is what produces variance in excess of the mean. The gamma distribution’s shape is controlled by a single additional parameter, usually written α (sometimes 1/θ), called the dispersion parameter:

Var(Y) = μ + αμ^k

When α = 0, the extra variance term vanishes and the model collapses exactly to Poisson — which is why Poisson is a nested special case of negative binomial, and why the two can be compared with a likelihood-ratio test (below). As α grows, the model tolerates more excess variance relative to the mean before treating an observation as unusual. α is estimated from the data along with the regression coefficients, typically by maximum likelihood, and is reported alongside them — a small but statistically significant α is itself evidence that overdispersion is real and not just sampling noise.

NB1 vs. NB2: the two parameterizations

The exponent k in the variance formula above is where the two standard parameterizations of negative binomial regression diverge. They fit the same mean structure (log(μ) = Xβ) but assume different relationships between the mean and the variance:

	NB1 (linear)	NB2 (quadratic)
Variance function	Var(Y) = μ(1 + α) = μ + αμ	Var(Y) = μ + αμ2
Excess variance grows with μ	Linearly — a constant multiple of the mean	Quadratically — overdispersion accelerates for larger counts
Default in standard software	Available (e.g. Stata’s nbreg with a linear variance option, or manual specification)	Yes — R’s MASS::glm.nb(), Stata’s default nbreg, and SAS’s default PROC GENMOD negative binomial link all fit NB2 unless told otherwise
When it’s the more natural fit	When extra variance is roughly proportional to the mean across the observed range — less common in practice	When larger-count observations show disproportionately more spread than small-count ones — the more common empirical pattern, and the reason NB2 is the near-universal default

In practice, “negative binomial regression” without qualification almost always means NB2 — it’s what every major package fits by default, and what’s reported in the applied literature unless a paper says otherwise. NB1 exists and is occasionally the better-fitting choice, but confirm which one a given software output used before comparing dispersion parameters across studies or software packages, since α is not on the same scale between the two.

Testing for overdispersion formally

Two checks are used in practice, and they answer slightly different questions:

• Informal diagnostic: compare the Poisson model’s residual deviance (or Pearson chi-square) to its degrees of freedom. A ratio meaningfully above 1 signals overdispersion. This is quick but only a rule of thumb — it doesn’t produce a p-value.
• Formal test: fit both a Poisson model and a negative binomial model on the same data, then run a likelihood-ratio test comparing the two log-likelihoods, since Poisson is nested inside negative binomial at α = 0. The test statistic is 2 × (log-likelihood_NB − log-likelihood_Poisson). Because α = 0 sits on the boundary of its allowed range (α cannot be negative), the naive reference distribution is not a standard chi-square with 1 degree of freedom — the correct reference is a 50:50 mixture of a chi-square(0) point mass and a chi-square(1), which in practice means halving the standard chi-square(1) p-value (Cameron and Trivedi’s boundary-adjusted LR test, the version most statistical packages report directly, e.g. Stata’s nbreg prints this test automatically after estimation).

A statistically significant result on that test is the formal justification for reporting negative binomial results instead of Poisson — not just visual inspection of variance versus mean.

Worked example: Poisson vs. negative binomial on the same data

Suppose a research-integrity office models the number of data-management-plan revision requests per grant application (n = 240 applications) as a function of proposal length and whether the application involves human-subjects data. A Poisson regression is fit first:

• Poisson fit: residual deviance = 612 on 237 degrees of freedom — a deviance-to-df ratio of 2.58, well above the 1.0 expected under equidispersion.
• Log-likelihood (Poisson) = −418.6

A negative binomial (NB2) model is fit on the identical specification:

• Estimated dispersion parameter: α = 0.47 (SE = 0.09) — clearly different from zero
• Log-likelihood (NB2) = −389.2

Likelihood-ratio statistic: 2 × (−389.2 − (−418.6)) = 58.8. Referred to the boundary-adjusted chi-square mixture described above, this is decisively significant (p < 0.001) — strong evidence that the negative binomial specification fits substantially better than Poisson for this data. The practical effect shows up in the standard errors: the incidence rate ratio for human-subjects involvement is nearly identical between the two models (Poisson IRR = 1.38 vs. NB2 IRR = 1.36 — the point estimates barely move), but its 95% confidence interval widens from a Poisson-based (1.19, 1.60) to a negative-binomial-based (1.09, 1.71). The Poisson model was not wrong about the direction or rough size of the effect — it was overconfident about how precisely that effect was estimated. Reporting the Poisson interval here would have overstated the strength of the evidence.

AIC and BIC tell the same story from a model-selection angle: because negative binomial estimates one additional parameter (α) than Poisson, AIC = −2(log-likelihood) + 2(k) penalizes it slightly for that extra parameter, but the log-likelihood improvement here is large enough to overwhelm the penalty — NB2’s AIC (784.4) comes in well below Poisson’s (841.2), confirming what the likelihood-ratio test already showed.

Quasi-Poisson: the lighter-weight alternative

Quasi-Poisson regression is worth naming alongside negative binomial because it solves a narrower version of the same problem. It keeps the Poisson mean structure exactly as estimated, then inflates the standard errors by an estimated dispersion multiplier (the Pearson chi-square divided by its degrees of freedom) rather than fitting a full second distribution. It corrects the standard-error problem without changing the coefficient estimates and without needing a separately estimated α. The trade-off: because quasi-Poisson isn’t a full probability model, it has no proper likelihood, so it cannot be compared with AIC/BIC or a likelihood-ratio test the way negative binomial can — that’s the main reason negative binomial is the more common default in the published literature when model comparison or prediction (not just coefficient testing) is the goal.

When overdispersion isn’t the real problem: zero-inflation

Negative binomial regression and quasi-Poisson both assume the excess variance comes from unmeasured heterogeneity spread across the whole distribution. That’s a different mechanism from zero-inflation: a distinct process generating more zero counts than either Poisson or negative binomial would predict, because some observations are structurally incapable of a nonzero count rather than just unlucky. A grant applicant who was never required to submit a data-management plan can’t generate a “revision request” count at all — that’s a structural zero, not a low realization of an underlying rate.

The diagnostic is different too: don’t just look at the deviance-to-df ratio, plot the observed proportion of zeros against what the fitted Poisson or negative binomial model predicts. If observed zeros substantially exceed the model’s predicted zero proportion even after accounting for overdispersion, that’s the signal to move to a zero-inflated model (ZIP or ZINB), which explicitly fits two processes at once: a logistic component predicting whether an observation is a structural zero, and a count component (Poisson or negative binomial) for everything else. Fitting an ordinary negative binomial model to zero-inflated data will typically still show a high α, which can be mistaken for “the overdispersion problem is solved” when the actual mechanism — a separate zero-generating process — is still misspecified.

Frequently asked questions

Is negative binomial regression always better than Poisson regression?

Not automatically. If a formal likelihood-ratio test shows α is not meaningfully different from zero, the extra parameter buys nothing and Poisson is the simpler, equally valid choice. Negative binomial is the correct default only when overdispersion is actually present and confirmed, not a blanket substitute.

Can I just always use negative binomial “to be safe”?

It’s a common practice when overdispersion is likely on substantive grounds, since NB2 collapses back toward Poisson-like behavior when α is estimated near zero. But it isn’t free: negative binomial estimates one more parameter, converges less reliably in small samples, and can behave poorly when the data is actually zero-inflated rather than generically overdispersed (see above) — running the formal test and checking the zero counts is worth the extra step rather than defaulting blindly.

How do I interpret the coefficients from a negative binomial regression?

The same way as Poisson regression: coefficients are on the log scale, and exponentiating them gives an incidence rate ratio (IRR) — the multiplicative change in the expected count for a one-unit increase in the predictor, holding other covariates constant. The dispersion parameter α doesn’t change this interpretation; it only affects the precision (standard errors) attached to those IRRs.

What sample size do I need to reliably estimate the dispersion parameter?

There’s no fixed universal threshold, but α is estimated less precisely than the regression coefficients and negative binomial models generally need more observations than an equivalent Poisson model to converge reliably, particularly with several predictors or sparse counts. If the model fails to converge or α’s standard error is very large relative to its estimate, that’s often a sample-size problem rather than evidence overdispersion is absent.

Does negative binomial regression require a link function like logistic regression?

Yes — both NB1 and NB2 use the same log link as Poisson regression (log(μ) = Xβ), which is why coefficients exponentiate to rate ratios rather than probabilities, unlike logistic regression’s logit link.

See also: Poisson distribution and Poisson regression for the equidispersion baseline this guide departs from, the chi-square test for the goodness-of-fit logic the deviance-ratio diagnostic borrows from, and the Research Methods & Statistics hub for the full quantitative-analysis series.