Poisson Regression

A log-linear model for conditional counts and rate ratios

Statistics
Regression
Poisson Regression
Count Data
Poisson regression from its count data-generating model through likelihood estimation, coefficient interpretation, uncertainty, diagnostics, and bioinformatics design.
Author

Jiahong Sun

Published

October 18, 2026

Modified

October 18, 2026

1 The Conditional Count Model

For independent experimental units \(i=1,\ldots,n\), Poisson regression specifies

\[ Y_i\mid\mathbf x_i\sim\operatorname{Poisson}(\mu_i), \qquad \log(\mu_i)=\eta_i=\mathbf x_i^\mathsf T\boldsymbol\beta. \]

The inverse link gives

\[ \boxed{E(Y_i\mid\mathbf x_i)=\mu_i =\exp(\mathbf x_i^\mathsf T\boldsymbol\beta)}. \]

This is a generalized linear model with a Poisson response distribution and log link. It is a model for the conditional mean, not merely a rule for fitting a line to counts. The Poisson sampling distribution also implies

\[ \operatorname{Var}(Y_i\mid\mathbf x_i)=\mu_i. \]

Mean–variance equality is therefore an assumption to assess after relevant predictors and exposure have been represented. A marginal sample variance larger than the marginal mean is not enough to reject the model: different covariate patterns can have different means even when every conditional distribution is Poisson.

Quantity Meaning
\(Y_i\) Observed event count for experimental unit \(i\)
\(\mu_i\) Conditional expected count
\(\eta_i\) Linear predictor on the log-mean scale
\(\beta_j\) Conditional log mean-ratio coefficient
\(\exp(\beta_j)\) Multiplicative mean ratio for a one-unit predictor contrast
\(\hat\mu_i\) Fitted conditional mean after estimating the coefficients

The experimental unit must be identified before the likelihood is written. A count may summarize mutations in one tumor, infections in one region, or events during one observation period. Events within a count are not automatically independent experimental replicates.

2 Coefficients and Contrasts

With one predictor,

\[ \log\{\mu(x)\}=\beta_0+\beta_1x. \]

For a contrast of \(d\) units,

\[ \frac{\mu(x+d)}{\mu(x)}=\exp(d\beta_1). \]

Thus \(\exp(\beta_1)\) is a conditional expected-count ratio when exposure is common. If an exposure term is included as an offset, it becomes a conditional rate ratio. For a binary treatment \(A\) in

\[ \log(\mu_i)=\beta_0+\beta_AA_i+\boldsymbol\gamma^\mathsf T\mathbf z_i, \]

\(\exp(\beta_A)\) compares expected counts at common modeled values of \(\mathbf z\). It is an adjusted association, not automatically a causal effect. Causal interpretation additionally requires a well-defined intervention, exchangeability, positivity, consistency, and defensible handling of selection, missingness, measurement, and interference.

Interactions make the ratio depend on another variable. With a term \(\beta_{AX}A_iX_i\), the treatment mean ratio at \(X=x\) is \(\exp(\beta_A+\beta_{AX}x)\). Lower-order coefficients refer to the other variable’s reference value, just as in linear and logistic regression.

An additive count contrast is not constant under this model:

\[ \mu(x+d)-\mu(x) =\mu(x)\{\exp(d\beta_1)-1\}. \]

Absolute changes therefore depend on the baseline mean. Report fitted counts or rates alongside ratios when absolute scale matters scientifically.

3 Likelihood, Estimation, and Uncertainty

Conditional independence across experimental units gives

\[ L(\boldsymbol\beta) =\prod_{i=1}^n \frac{\exp(-\mu_i)\mu_i^{y_i}}{y_i!}, \qquad \mu_i=\exp(\mathbf x_i^\mathsf T\boldsymbol\beta). \]

Ignoring terms that do not depend on \(\boldsymbol\beta\), the log-likelihood is

\[ \ell(\boldsymbol\beta) =\sum_{i=1}^n \left\{y_i\mathbf x_i^\mathsf T\boldsymbol\beta -\exp(\mathbf x_i^\mathsf T\boldsymbol\beta)\right\}. \]

Let \(\boldsymbol\mu=(\mu_1,\ldots,\mu_n)^\mathsf T\) and \(\mathbf W=\operatorname{diag}(\mu_1,\ldots,\mu_n)\). Differentiation gives

\[ \mathbf U(\boldsymbol\beta) =\mathbf X^\mathsf T(\mathbf y-\boldsymbol\mu), \qquad -\frac{\partial^2\ell} {\partial\boldsymbol\beta\partial\boldsymbol\beta^\mathsf T} =\mathbf X^\mathsf T\mathbf W\mathbf X. \]

The maximum-likelihood estimate solves \(\mathbf U=\mathbf0\), usually by Fisher scoring or iteratively reweighted least squares. At a regular, identified fit,

\[ \widehat{\operatorname{Var}}(\hat{\boldsymbol\beta}) =(\mathbf X^\mathsf T\hat{\mathbf W}\mathbf X)^{-1}. \]

Wald intervals exponentiate coefficient limits to obtain ratio limits. Likelihood-ratio or profile-likelihood intervals can better reflect asymmetric likelihoods. Model-based uncertainty is valid only when the conditional variance and independence assumptions are adequate. A sandwich covariance can protect coefficient inference against some variance misspecification under a correct mean model and suitable independent clusters; it does not repair a wrong mean, unmeasured confounding, or too few independent clusters.

4 A Small Calculation

Suppose the fitted mean number of lesions is

\[ \log\{\mu(x)\}=1.2+0.35x, \]

where \(x\) is a centered exposure score. At \(x=0\), the expected count is \(\exp(1.2)\approx3.32\). A one-unit increase multiplies the mean by \(\exp(0.35)\approx1.42\), so at \(x=1\) the expected count is about \(4.71\). The absolute increase is about \(1.39\), not 42 lesions. At a baseline mean of 20, the same coefficient would imply an increase of about \(8.38\) counts.

The coefficient is the model parameter, \(0.35\) is its estimate in this fitted example, and \(1.42\) is the estimated conditional mean ratio. A confidence interval describes sampling uncertainty under the analysis assumptions; it does not describe unexplained heterogeneity among individual units.

5 Assumptions, Diagnostics, and Sensitivity

Assumption or issue Why it matters Assessment Response or sensitivity analysis
Count support and correct event definition Fractional, censored, duplicated, or inconsistently defined outcomes require another observation model Audit counting and preprocessing rules Redefine the outcome or model measurement and censoring explicitly
Correct log-mean structure Omitted curvature, interactions, or exposure distort fitted ratios Residual patterns against predictors and fitted means; flexible partial checks Splines, prespecified interactions, revised offsets, or stratification
Conditional Poisson variance Extra or reduced variation invalidates model-based standard errors Pearson statistic, residual deviance, replicate-level predictive checks Quasi-Poisson/sandwich inference, negative binomial, or mechanism-specific model
Independent experimental units Correlation makes naive information too large Map pairing, repeated measures, families, sites, and batches Cluster-aware covariance, GEE, mixed model, or justified aggregation
No distinct structural-zero process Inactive states or detection failure can create a separate zero mechanism Compare observed and fitted zeros conditionally; inspect assay mechanism Hurdle or zero-inflated model only when scientifically defensible
Full-rank design and adequate overlap Coefficients may be unidentified or extrapolative Rank, condition, and covariate-support checks Redesign, combine parameters, or restrict the target population
No influential experimental unit A rare covariate pattern or large count can dominate the fit Deviance/Pearson residuals, leverage, Cook-type deletion checks Verify records and report leave-one-unit-out sensitivity

For \(p\) fitted coefficients, the Pearson dispersion diagnostic is

\[ \hat\phi_P =\frac{1}{n-p}\sum_{i=1}^n \frac{(y_i-\hat\mu_i)^2}{\hat\mu_i}. \]

A value far above one flags lack of fit relative to the Poisson variance, but it does not identify the cause. Omitted predictors, unequal exposure, clustering, outliers, zero inflation, and genuine rate heterogeneity can produce similar signals and require different remedies. Residual deviance compared mechanically with its degrees of freedom can also mislead in sparse counts, small samples, or many fitted values near zero; simulation from the fitted design is often more informative.

Quasi-Poisson inference estimates a dispersion multiplier while retaining \(\operatorname{Var}(Y_i\mid\mathbf x_i)=\phi\mu_i\). Negative-binomial regression changes the variance function, commonly to \(\mu_i+\alpha\mu_i^2\). Neither choice excuses a misspecified mean or dependence structure. Compare coefficient, uncertainty, residual, and predictive conclusions across scientifically plausible models rather than selecting solely by one dispersion statistic.

6 Bioinformatics Design and Interpretation

Suppose mutation counts are measured in 120 independently sampled tumors. A model includes treatment, tumor type, sequencing center, and callable genome size. Tumors are the experimental units; mutations, reads, technical libraries, and repeated aliquots within a tumor are not 120 additional independent treatment replicates. Callable bases differ across tumors, so a log-callable-bases offset is needed if the target is a mutation rate rather than a raw count.

The Poisson variance may be too restrictive because tumors differ biologically after modeled covariates, mutation calls have sample-dependent error, and events can cluster along the genome. Diagnostics should be performed at the tumor level, with sensitivity to a negative-binomial mean–variance relation, robust or cluster-aware uncertainty where justified, influential tumors, alternative callability definitions, and batch adjustment.

For RNA-seq, library counts are commonly modeled with negative-binomial rather than Poisson variation because replicate-to-replicate variability exceeds Poisson sampling. Sample-specific normalization factors enter as offsets, but reads are technical observations and genes are parallel outcomes. Independent biological specimens support inference about a condition effect. Statistically, an exponentiated coefficient is a conditional count or rate ratio under the model; biologically, importance also depends on absolute abundance, uncertainty, technical bias, tissue composition, heterogeneity, and the experimental design.

7 Connections and Take-home Messages

  1. Poisson regression models a conditional count mean through a log-linear predictor and implies conditional mean–variance equality.
  2. Exponentiated coefficients are conditional mean ratios, or rate ratios when exposure is represented by an offset.
  3. Valid inference depends on the experimental units, mean structure, variance, exposure, identification, and dependence—not on count support alone.
  4. Overdispersion is a symptom relative to a conditional model; diagnose its mechanism before choosing quasi-Poisson, negative-binomial, mixture, or dependence-aware methods.
  5. In bioinformatics, biological samples—not reads, mutations, cells, or genes— usually provide replication for condition-level inference.

8 References

Back to top