Concept Architecture
Concept
Theoretically, Negative Binomial Regression is a generalised linear modelling technique used to analyse count outcomes that exhibit overdispersion, where the variance exceeds the mean. It extends Poisson regression by introducing a dispersion parameter that captures unobserved heterogeneity among observations. The method exists to provide unbiased parameter estimates and appropriate standard errors when the assumptions of the Poisson model are violated by excessive variability in count data.
Mathematically, Negative Binomial Regression models the logarithm of the expected count as a linear function of explanatory variables while assuming that the outcome follows a Negative Binomial distribution. The model is commonly derived as a Poisson?Gamma mixture and is estimated using Maximum Likelihood Estimation. Regression coefficients are interpreted on the logarithmic scale, and exponentiation of the coefficients yields incidence rate ratios that quantify multiplicative effects on the expected event rate.
In practice, Negative Binomial Regression is widely applied to overdispersed healthcare utilisation data, including hospital admissions, outpatient visits, emergency department attendances, prescription counts and adverse events. In health economics, it is frequently used to estimate treatment effects on healthcare resource use, evaluate predictors of service utilisation and support cost-effectiveness analyses involving recurrent healthcare events.
Purpose
Used to model overdispersed count outcomes, estimate incidence rate ratios, evaluate predictors of healthcare utilisation and analyse recurrent event data in health economic research.
Mathematical Formulae
Primary Formula
Negative binomial regression model:
ln(??) = ?? + ??X?? + ??X?? + ? + ??X??
where:
- ?? = expected count for observation i
- ?? = intercept
- ????? = regression coefficients
- X??X? = explanatory variables
Expected count:
?? = exp(?? + ??X?? + ? + ??X??)
Supporting Formulae
Variance:
Var(Y?) = ?? + ??� / r
where:
- r = dispersion parameter
Incidence Rate Ratio:
IRR = e?
Log-likelihood:
?(?,r) = ? ln[P(Y? | ??,r)]
Related Mathematical Methods
- Negative Binomial Distribution
- Poisson Regression
- Generalised Linear Models
- Maximum Likelihood Estimation
- Zero-Inflated Negative Binomial Regression
- Quasi-Poisson Regression
- Incidence Rate Ratio Analysis
Example
A health economist evaluates whether a disease management programme reduces annual hospital admissions.
The fitted Negative Binomial Regression model estimates:
??Treatment = ?0.36
The corresponding Incidence Rate Ratio is:
IRR = e??�??
IRR = 0.70
This indicates that patients receiving the intervention experience approximately 30% fewer hospital admissions than patients receiving standard care after adjusting for age, comorbidity and previous healthcare utilisation.
The estimated dispersion parameter is:
r = 1.9
confirming that the observed count data exhibit overdispersion relative to the Poisson model.
Excel Implementation
| Function | Example Formula | Health Economics Application |
|---|---|---|
| EXP | =EXP(B0+B1*A2+B2*C2) | Calculate expected healthcare event counts. |
| LN | =LN(D2) | Calculate log-likelihood contributions. |
| EXP | =EXP(B1) | Convert regression coefficients into incidence rate ratios. |
| SUM | =SUM(E2:E1001) | Calculate the total log-likelihood. |
| Solver | Maximise the log-likelihood by estimating regression coefficients and the dispersion parameter. | Estimate Negative Binomial Regression models. |
VBA (Optional)
A VBA routine can automate maximum likelihood estimation of Negative Binomial Regression models, calculate incidence rate ratios and generate model diagnostics for healthcare utilisation analyses.
Sources
- Hilbe JM. Negative Binomial Regression. Cambridge University Press.
- Cameron AC, Trivedi PK. Regression Analysis of Count Data. Cambridge University Press.
- McCullagh P, Nelder JA. Generalized Linear Models. Chapman & Hall.
- Greene WH. Econometric Analysis. Pearson.
- Briggs A, Claxton K, Sculpher M. Decision Modelling for Health Economic Evaluation. Oxford University Press.
- ISPOR Good Practice Reports.
Related Concepts (2)
Library
Publications
1
Statistical Analysis of Cost-Effectiveness Data — Willan & Briggs, 1st Edition ed., 2006 (John Wiley & Sons)
A synthesis of statistical methods for analysing cost-effectiveness data, including net-benefit regression, confidence intervals for the ICER, cost-effectiveness acceptability curves, and covariate adjustment. Part of the Wiley Statistics in Practice series.
BookView source →
Frequently Asked Questions (6)
What is negative binomial regression?
A modelling technique analysing count outcome data, such as hospitalisations, that shows overdispersion relative to what a Poisson model would assume.
Source: Hilbe 2011
When is negative binomial regression chosen for count outcomes?
Negative binomial regression is chosen to model count outcomes, such as numbers of hospitalisations or seizures, when those counts vary more than a Poisson model assumes. Poisson regression ties the variance of the counts to their mean, and when the real data are more scattered than that, it understates uncertainty and can mislead. Negative binomial regression relaxes the assumption by adding a parameter for the extra spread, giving more honest standard errors while still expressing effects as rate ratios. Modelling overdispersed counts is its purpose. Kirkwood and Sterne (2003) describe this technique.
Source: Kirkwood & Sterne 2003
How does negative binomial regression work?
Negative binomial regression works by modelling the expected count as a function of predictors, typically through a log link, while using the negative binomial distribution, which includes a dispersion parameter allowing the variance to exceed the mean, to account for overdispersion. The coefficients are estimated by maximum likelihood. So negative binomial regression works like Poisson regression in relating predictors to the log of the expected count, but with the added dispersion parameter that captures the extra variability, which gives valid inference for overdispersed counts, and the exponentiated coefficients are interpreted as rate ratios, showing the multiplicative effect of predictors on the expected count.
Source: Hilbe 2011
When is negative binomial regression used?
Negative binomial regression is used when the outcome is a count that is overdispersed, showing more variability than a Poisson model assumes, as is common for counts such as hospital admissions, events, or occurrences with unobserved heterogeneity. So negative binomial regression is used for overdispersed count outcomes, where fitting a Poisson model would underestimate the standard errors and overstate significance, since the negative binomial accommodates the extra variability, giving valid inference, which is why it is preferred over Poisson regression whenever count data exhibit overdispersion, a situation checked for when modelling counts.
Source: Hilbe 2011
How does negative binomial regression differ from Poisson regression?
Negative binomial regression differs from Poisson regression in allowing the variance of the count to exceed the mean through a dispersion parameter, accommodating overdispersion, whereas Poisson regression assumes the variance equals the mean. When overdispersion is present, Poisson regression understates the standard errors, while negative binomial regression corrects for it. So the two differ in handling variability, with Poisson regression suited to counts whose variance equals their mean and negative binomial regression to overdispersed counts, which is why the negative binomial is preferred when overdispersion is detected, since it gives valid inference where the Poisson would produce overly narrow confidence intervals and inflated significance.
Source: Hilbe 2011
How are the results of negative binomial regression interpreted?
The results of negative binomial regression are interpreted through rate ratios, obtained by exponentiating the coefficients: a rate ratio above one indicates the predictor increases the expected count, below one that it decreases it, holding the other predictors constant. The dispersion parameter indicates the degree of overdispersion. So negative binomial regression results are interpreted as the multiplicative effect of each predictor on the expected count, expressed as rate ratios, analogous to Poisson regression but with valid standard errors for overdispersed data, and the estimated dispersion confirms the presence and extent of the overdispersion that motivated using the negative binomial rather than the Poisson model.
Source: Hilbe 2011
Trust Record
Verified by Dr Darrin Baines
British health economist
Professional identity: darrinbaines.org
Verification date: 19 Dec 2025
Content version: 1.0.0
Canonical Identity
- Term code
- HE-ES-SA-136
Stable URI · Machine-readable · Resolvable · CC BY 4.0