Modified Poisson regression
Modified Poisson regression estimates risk ratios for a binary outcome by pairing a log-link Poisson fit with robust standard errors.
Fit
When to use it, and when not to
Use it when
- You need a risk ratio for a binary outcome at a defined time.
- You need adjustment for recorded baseline factors.
Not when
- Your outcome counts repeat events. Use Poisson regression for event counts or rates. Read: Poisson regression
- You sample cases and controls by outcome status. Use logistic regression for an odds ratio. Read: Logistic regression
Before you start
What you need
- Binary outcome and follow-up windowoutcomeTypeCode the event as one and its absence as zero. Record whether everyone has complete follow-up.
- Exposure contrast and baseline covariatessummaryMeasureName the risk ratio of interest and choose adjustments before seeing the result.
- Independent people or defined clustersclusteredState the sampling unit. Missing outcomes need an assumption and a plan.
- The log risk follows the specified predictor model.How to check: Check continuous terms and interactions. Robust errors do not correct a wrong mean model.
- Binary outcomes need sandwich, or robust, standard errors.How to check: Verify the variance option explicitly. Ordinary Poisson errors use the wrong variance for a binary outcome.
- People are independent for an individual sandwich variance.How to check: For shared sites or repeated outcomes, account for clusters. Few clusters require a justified small-sample method.
- A causal claim needs adequate control of baseline confounders.How to check: Choose factors using subject knowledge. Adjustment alone cannot remove bias from unrecorded causes.
- The analysis matches the stated missing-data assumption.How to check: Missing at random means gaps depend on recorded factors. Account for those factors and assess plausible departures.
Plain language
How it works
Fit a Poisson model with a log link to zero-or-one outcomes, without a person-time offset. The model describes risk, not a count rate. Exponentiate the exposure coefficient to obtain a risk ratio at fixed covariate values. State any interactions, since these can change that ratio.
A binary outcome does not have Poisson variance. Use the sandwich variance: the observed residual spread corrects the uncertainty without changing the fitted coefficients. Zou develops this approach for binary prospective data. McNutt and colleagues explain why an adjusted odds ratio cannot simply be relabeled a risk ratio.
Worked example
A worked example
- Setting
- A complete cohort has two baseline severity strata. Each stratum has 100 exposed and 100 unexposed people, with one binary outcome per person.
- Question
- What is the exposure risk ratio after accounting for severity?
- Approach
- Low severity: exposed 20/100, unexposed 30/100. High severity: exposed 40/100, unexposed 70/100. Fit intercept, exposure and severity, without interaction. Use Poisson fitting with an individual HC0 sandwich variance, without a small-sample multiplier.
- Result
- The adjusted risk ratio is 0.600, with 95% robust Wald interval 0.471 to 0.765. The log coefficient is -0.511; its robust standard error is 0.124.
- What it means
- The fitted exposed risk is 0.188 at low severity and 0.413 at high severity. The estimated risk is 40.0% lower at either severity level.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Robust standard errors are essential for a binary outcome fitted with a Poisson model. They do not repair a wrong risk model, confounding or dependence. Clustered data need a suitable cluster variance method; fitted risks can exceed one.
- Reporting ordinary Poisson errors for a binary outcome.Instead: Use the sandwich variance and a log-scale Wald interval; exponentiate both limits.
- Interpreting a fitted value greater than one as a probability.Instead: Check all fitted risks. An impossible risk warns that the mean model needs attention.
- Calling a cross-sectional result a future risk ratio.Instead: With a cross-sectional binary outcome, name the result a prevalence ratio. Define the design and time window.
Inside Mallard
How Mallard uses this
Mallard offers this lesson in Learn search and the Atlas. The method registry links this method to its published analysis article and code.
modified-poisson-adjusted-risk-ratio).analysis.modified_poisson, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- A modified poisson regression approach to prospective studies with binary data (opens in a new tab) · Am J Epidemiol · doi 10.1093/aje/kwh090
- Estimating the relative risk in cohort studies and clinical trials of common outcomes (opens in a new tab) · Am J Epidemiol · doi 10.1093/aje/kwg074