Skip to content
Scintria
ToolsGrant writerFunding
HomeLiteratureClinical trialsEvidence LabProtocol packData analysisLearnProjectsLibraryReferencesApplications
Start free
ScintriaStart free

Learn / Biostatistics / Logistic and Poisson regression

Poisson regression: counts, rates and person-time

14 minute read · 6-question quiz

Download this lesson as a colour Word or PDF study guide with the figures and answers. Included in paid plans; reading and quizzes stay free.Sign in

What you will learn

  • Calculate incidence rates and rate ratios from events and person-time

  • Fit and interpret a Poisson model with an offset

  • Detect overdispersion and know the alternatives (negative binomial, robust SEs)

When the outcome is a count of events observed over varying lengths of follow-up, with people entering and leaving at different times, a proportion is the wrong summary. Use a rate: events divided by the total person-time at risk.

Incidence rate = (number of new events)/(total person-time at risk)

If 24 children have malaria during 1,200 child-years of follow-up, the rate is 24/1,200 = 0.02 per child-year = 20 per 1,000 child-years. Someone followed for 6 months contributes 0.5 person-years; someone followed for 3 years contributes 3.

Worked example: a trial with person-time

In a malaria trial, intervention villages and comparison villages are followed for different totals of child-time:

ArmEpisodesChild-yearsRate per 1,000 child-years
Intervention241,20020.0
Comparison451,00045.0
  • Rate ratio (IRR) = 20/45 = 0.44: the intervention cut the rate by 56%.
  • Rate difference = 45 − 20 = 25 fewer episodes per 1,000 child-years.
  • 95% CI for the IRR: SE(ln IRR) = sqrt(1/24 + 1/45) = 0.253, so exp(ln 0.444 ± 1.96 × 0.253) = 0.27 to 0.73.
StudyEstimate (95% CI)Intervention vs comparison0.44 (0.27–0.73)0.51Incidence rate ratio (log scale); 1 = no difference
The rate ratio and its 95% CI exclude 1, so the intervention clearly reduced malaria incidence.

The Poisson regression model

To adjust for covariates, model the log of the expected count:

ln(μ) = ln(person-time) + β_0 + β_1 X_1 + …

The ln(person-time) term is an offset: a predictor whose coefficient is fixed at 1 so the model describes the rate rather than the raw count. Then e^(β) is a rate ratio (also called incidence rate ratio, IRR). With a binary arm indicator, e^(β_1) reproduces the 0.44 above. Adding age, sex and season adjusts it.

Poisson regression suits: malaria episodes per child-year, adverse events per patient-month, hospital admissions per year, deaths per 1,000 person-years, deaths by district with the population as offset.

Yes/no outcomeRatesTime to eventOdds ratio or risk ratioLogistic regressionExample: dead at 28 daysRate ratioPoisson regression +offsetExample: episodes perchild-yearHazard ratioCox regressionExample: time to relapse
Choosing a regression by the type of outcome.

Overdispersion

A true Poisson variable has variance equal to its mean. Real counts are often overdispersed: some children have many episodes, others none, because of unmeasured susceptibility or clustering. Ignoring it makes SEs too small and CIs too narrow, so the intervention appears more convincingly effective than it is.

Check: compare the model's deviance or Pearson chi-square with its degrees of freedom; a ratio well above 1 (say > 1.5) indicates overdispersion.

Remedies:

  • Negative binomial regression adds an extra parameter for the extra variation.
  • Quasi-Poisson or robust (sandwich) standard errors adjust the SEs.
  • For clustered data (villages, schools), use GEE or mixed models with cluster as a random effect.

Also watch for excess zeros: counts with many more zeros than Poisson predicts (children who never attend the clinic) may need zero-inflated or hurdle models.

Rates versus risks

A risk is a proportion over a defined period in a fixed population: unaffected by when events happen within it. A rate uses person-time and handles varying follow-up and competing exits. Do not call a rate a risk, and do not compute 'rate per person' ignoring the follow-up duration.

Person-time must be counted correctly

Count only time at risk. After a first episode of a disease that confers immunity, the patient is no longer at risk. Include time from all participants, including those lost to follow-up (up to the date last seen). Counting only completers biases the denominator.

Summary

  • A rate = events ÷ person-time; the rate ratio compares two rates.

  • Poisson regression models log counts with ln(person-time) as an offset; exp(β) is a rate ratio.

  • Overdispersion (variance > mean) narrows SEs falsely; use negative binomial, quasi-Poisson, robust SEs or cluster methods.

  • Use logistic regression for binary outcomes at fixed follow-up, Poisson for rates, Cox for time-to-event.

Check your understanding

Question 1
1

45 events occur over 1,000 person-years. The rate per 1,000 person-years is:

Question 2
2

In Poisson regression, e^β for a treatment indicator is the:

Question 3
3

The offset term ln(person-time) is included so that the model:

Question 4
4

Variance of the counts is 3 times the mean. The best description is:

Question 5
5

Which regression best suits 'time until first relapse' with censoring?

Question 6
6

A village trial analysed with Poisson regression ignoring clustering will usually have:

Draft study guide. Spotted an error or something unclear? Tell us at hello@scintria.com. This is educational material, not clinical or regulatory advice.

Prediction models: discrimination, calibration and ROC curvesTime-to-event data, censoring and the Kaplan-Meier curve