Find your next learning step
A count becomes a rate only when its exposure clock enters the model correctly. Negative-binomial regression can then represent extra-Poisson variation, but it cannot repair a wrong offset or a missing causal variable.
Open the chapter contents · Probability root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The teaching chapters remain open; these grouped routes form the index.
Route 1 · Build the rate model
Route 2 · Work the ledger
Route 3 · Diagnose and transfer
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · BUILD THE RATE MODEL
1. The exposure clock turns a count into a rate
Two groups can record the same number of events and still have different rates if one was observed for longer. For unit i, let Y_i be the event count and T_i the exposure, such as person-years, machine-hours or kilometres travelled. A log-link count model writes log(mu_i) = log(T_i) + beta_0 + beta_1 x_i, where mu_i is the expected count and x_i marks the comparison of interest.
The term log(T_i) is an offset: its coefficient is fixed at one. It converts the expected count into a rate without pretending that an extra exposure unit has an unknown effect. The rate ratio for a one-unit change in x is exp(beta_1).
This guide extends the Probability root and the exact-count route in Poisson rate-ratio inversion. The question here is different: what should happen when otherwise comparable units vary more than a Poisson model permits?
CHAPTER 2 OF 12 · BUILD THE RATE MODEL
2. Negative-binomial variance represents extra variation
A Poisson model has variance equal to its mean. Real units may differ through unrecorded frailty, repeated-event dependence or environmental heterogeneity. The negative-binomial NB2 model keeps the log mean but uses Var(Y_i) = mu_i + alpha mu_i squared, where alpha is a non-negative dispersion parameter. When alpha approaches zero, the model approaches Poisson regression.
The quadratic term matters. At mean 2 with alpha 0.2, the variance is 2.8; at mean 10 it is 30. Extra spread grows faster than the mean. A quasi-Poisson alternative instead makes variance proportional to the mean. These are distinct variance stories, not interchangeable labels.
The coefficient exp(beta_1) remains a conditional mean rate ratio under the specified covariates. Negative-binomial regression changes the likelihood and uncertainty calculation; it does not change an exposure into a confounder adjustment or prove a causal effect.
CHAPTER 3 OF 12 · WORK THE LEDGER
3. Worked example: an exposure-offset ledger
Consider six original teaching units. The comparison groups have equal total exposure here so the arithmetic is visible, but the fitted model would retain every row and its offset.
| Unit | Group x | Events Y | Exposure T | Observed rate Y/T |
|---|---|---|---|---|
| A | 0 | 4 | 100 | 0.040 |
| B | 0 | 6 | 150 | 0.040 |
| C | 0 | 2 | 50 | 0.040 |
| D | 1 | 9 | 100 | 0.090 |
| E | 1 | 18 | 150 | 0.120 |
| F | 1 | 12 | 50 | 0.240 |
Group 0 records 12 events in 300 exposure units, a rate of 0.040. Group 1 records 39 in 300, a rate of 0.130. The crude rate ratio is 0.130 / 0.040 = 3.25. Yet the comparison-unit rates range from 0.090 to 0.240, signalling variation that a two-total calculation hides.
An offset model does not divide the response before fitting. It models counts while placing log exposure in the linear predictor. That retains the count likelihood and gives longer-observed units the information their event process supports.
CHAPTER 4 OF 12 · WORK THE LEDGER
4. The offset must stay fixed
If log exposure is entered as an ordinary predictor, software estimates its coefficient. The model then says expected counts scale as T raised to a fitted power rather than directly with T. That may be a legitimate scientific hypothesis, but it is no longer the standard rate model.
Offset errors are easy to miss. Exposure measured in days for some rows and years for others changes the intercept and can distort comparisons. Zero or negative exposure cannot enter a logarithm. Exposure may also be measured with error; fixing its coefficient does not make it exact.
The clean audit is mechanical: preserve the original exposure column, state its unit, confirm all values are positive, compute log(T) once, and verify that the modelling function treats it as an offset. Recalculate a few fitted rates as fitted count divided by exposure.
CHAPTER 5 OF 12 · WORK THE LEDGER
5. A variance ledger separates Poisson and NB2 stories
The following values show how the variance assumptions diverge when alpha = 0.20. They are model expectations, not estimates from the six-row teaching dataset.
| Mean count mu | Poisson variance | NB2 variance mu + 0.20 mu squared | NB2 variance-to-mean ratio |
|---|---|---|---|
| 2 | 2.0 | 2.8 | 1.4 |
| 5 | 5.0 | 10.0 | 2.0 |
| 10 | 10.0 | 30.0 | 3.0 |
| 20 | 20.0 | 100.0 | 5.0 |
This is why a single Pearson statistic is only a starting signal. If residual spread increases roughly in proportion to the fitted mean, quasi-Poisson may describe it better. If it grows quadratically, NB2 may be plausible. Plot residuals against fitted means, time, location and exposure rather than choosing a family from one scalar alone.
CHAPTER 6 OF 12 · DIAGNOSE AND TRANSFER
6. Diagnostics ask where the variation comes from
Begin with a Poisson fit and inspect residual patterns, leverage, repeated units and the Pearson chi-square divided by residual degrees of freedom. A value well above one suggests extra variation, but it does not identify the cause. One outlier, a missing predictor, clustering and zero inflation can all inflate the statistic.
Then compare a prespecified negative-binomial fit. Check the estimated alpha and its uncertainty, fitted-versus-observed distributions, deviance or randomized quantile residuals, and sensitivity to influential units. If observations are clustered, a random-effects or correlated-count model may match the data-generating process better than an independent NB2 likelihood.
Overdispersion is a symptom. The diagnostic question is not merely “Poisson or negative binomial?” but “which omitted structure produced the extra spread, and is the chosen likelihood a defensible approximation?”
CHAPTER 7 OF 12 · DIAGNOSE AND TRANSFER
7. Changed context: the exposure is wrongly specified
Suppose hospital A contributes bed-days while hospital B contributes patient-days. Both are called “days”, yet they measure different opportunities for events. Negative-binomial dispersion can widen intervals around a misspecified mean, but it cannot recover a meaningful common rate.
A simulation study of surveillance counts found that NB2 modelling could address overdispersion under its variance structure but did not repair bias from omitted confounders or incorrect offsets. That is a crucial transfer check: improved residual spread is not evidence that the mean model is right.
Before fitting, draw the event opportunity: who or what was at risk, for how long, and under which ascertainment rule? If opportunity saturates, is interrupted or depends on the outcome, a simple log offset may not be appropriate.
CHAPTER 8 OF 12 · DIAGNOSE AND TRANSFER
8. Changed context: zeros, clusters and time dependence
Many zeros do not automatically imply a zero-inflated model. A low mean count naturally produces zeros. Compare the predicted zero frequency, examine structural reasons a unit could never experience the event, and keep the sampling design visible.
Repeated counts from one person or site violate independence even if their marginal variance resembles NB2. Time trends, outbreaks or shared shocks can create serial and spatial correlation. Cluster-robust uncertainty may protect some inferences with enough independent clusters, but it does not correct a misspecified mean or small-cluster bias.
The transfer rule is to locate dependence before choosing a distribution. A negative-binomial model is useful when its latent-heterogeneity interpretation and variance curve fit the study, not because it is a universal “overdispersion button”.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and common misconceptions
Negative binomial does not mean every unit follows the same physical process. It is a probability model for the conditional count distribution. The NB1 and NB2 parameterisations differ; software defaults must be recorded. The alpha reported by one package may be the reciprocal or a transformed version of another package’s size parameter.
The rate ratio is conditional on included covariates. It can differ from a marginal population ratio, especially with nonlinear links or random effects. A narrower information criterion does not prove the exposure or covariate set is correct.
Do not compare raw residual variance with the raw mean when fitted means vary widely. Do not delete high-count units solely because they create overdispersion. Investigate whether they reveal a separate mechanism, a data error or an expected tail.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Evidence and reproducibility route
Gardner, Mulvey and Shaw’s review, Regression analyses of counts and rates, sets out Poisson, overdispersed Poisson and negative-binomial choices. Ver Hoef and Boveng compare the linear quasi-Poisson and quadratic negative-binomial variance functions in Quasi-Poisson vs. negative binomial regression. Lindén and Mäntyniemi examine overdispersion model choice in Using the negative binomial distribution to model overdispersion in ecological count data.
For the offset warning, consult the simulation study Regression models for public health surveillance data. A reproducible record includes the event definition, exposure unit, formula with the offset marked, NB parameterisation, coefficient table, dispersion diagnostics, influential-case check, software version and a plot of residuals against fitted mean and design variables.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
Should I fit negative binomial whenever Pearson chi-square divided by degrees of freedom exceeds one?
No. Treat it as a prompt to investigate the mean structure, dependence, exposure and outliers before selecting a variance model.
Is exp(beta_1) an incidence-rate ratio?
It is a model-based conditional mean rate ratio when log exposure is a valid offset and the comparison is coded as stated.
Can the offset coefficient be tested?
You can fit a model that estimates the log-exposure coefficient as a sensitivity check, but that tests a different scaling assumption. The result does not retroactively validate a faulty exposure measure.
What if alpha is nearly zero?
The negative-binomial fit approaches Poisson. Report the boundary behaviour and avoid claiming meaningful heterogeneity from a weakly identified dispersion parameter.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Final transfer check and connected reading
Before accepting a negative-binomial rate ratio, answer six questions: What creates exposure? Is its coefficient fixed at one? Does the mean model include the important design variables? Does the residual variance curve resemble the chosen family? Where does dependence live? Does the conclusion survive removal or reclassification of the most influential unit?
Continue through the Probability root, the sparse-data guide to exact stratified Poisson tests, the broader Gold Standard of Education and the How X Works hub. The transferable skill is to keep the exposure clock, conditional mean and variance story separate enough that each can be checked.
Top · Contents · Previous · How X Works