Offsets and Modeling Event Rates

Separating event intensity from unequal opportunity in count regression

Statistics
Regression
Offsets
Event Rates
Count Data
Offsets, exposure denominators, rate ratios, uncertainty, diagnostics, and bioinformatics interpretation for Poisson and related count models.
Author

Jiahong Sun

Published

October 18, 2026

Modified

October 18, 2026

1 Counts Need an Opportunity Scale

Ten events in one day and ten events in one year are equal counts but very different rates. Let \(Y_i\) be a count, \(t_i>0\) a known exposure, and \(r_i\) the conditional event rate per exposure unit. A Poisson rate model is

\[ Y_i\mid\mathbf x_i,t_i \sim\operatorname{Poisson}(\mu_i), \qquad \mu_i=t_ir_i. \]

If the rate follows a log-linear regression,

\[ \log(r_i)=\mathbf x_i^\mathsf T\boldsymbol\beta, \]

then

\[ \boxed{ \log(\mu_i)=\log(t_i)+\mathbf x_i^\mathsf T\boldsymbol\beta }. \]

The known term \(\log(t_i)\) is an offset: it enters the linear predictor with coefficient fixed at one. This constraint encodes proportionality—doubling exposure doubles the expected count while leaving the modeled rate unchanged.

Quantity Meaning and units
\(Y_i\) Events observed for unit \(i\)
\(t_i\) Exposure: person-time, area, callable bases, trials, or another opportunity measure
\(\mu_i=t_ir_i\) Expected count; dimensionless
\(r_i\) Expected events per unit exposure
\(\beta_j\) Conditional log rate-ratio coefficient
\(\exp(\beta_j)\) Conditional rate ratio for a one-unit predictor contrast

The denominator is part of the estimand. Events per person-year, per callable megabase, and per sequenced molecule answer different questions even when the numerator is the same.

2 Why the Offset Coefficient Is One

Suppose every unit has a common rate \(r\). The Poisson log-likelihood is, up to constants,

\[ \ell(r)=\sum_i\{y_i\log(t_ir)-t_ir\}. \]

Differentiating and setting the score to zero gives

\[ \frac{\partial\ell}{\partial r} =\frac{\sum_i y_i}{r}-\sum_i t_i=0, \qquad \boxed{\hat r=\frac{\sum_i y_i}{\sum_i t_i}}. \]

The likelihood pools events over pooled exposure, not by taking an unweighted average of individual rates \(y_i/t_i\). Units with more exposure contain more information under the model.

The offset algebra follows from \(\mu_i=t_ir_i\). It is not an empirical claim that must be re-estimated every time. Replacing the offset by a free term,

\[ \log(\mu_i)=\gamma\log(t_i)+\mathbf x_i^\mathsf T\boldsymbol\beta, \]

asks a different question: whether expected counts scale as \(t_i^\gamma\). Estimating \(\gamma\) can be a useful sensitivity analysis when proportionality is doubtful, but it no longer defines the prespecified rate model. A result \(\hat\gamma\ne1\) might reflect a real exposure relationship, saturation, measurement error, omitted covariates, or an inappropriate denominator.

3 Rate Ratios, Rates, and Absolute Contrasts

For a binary treatment or predictor \(A\),

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

At common \(\mathbf z\), the modeled rate ratio is

\[ \frac{r(A=1,\mathbf z)}{r(A=0,\mathbf z)}=\exp(\beta_A). \]

The rate difference is

\[ r(1,\mathbf z)-r(0,\mathbf z) =r(0,\mathbf z)\{\exp(\beta_A)-1\}, \]

so a constant rate ratio does not imply a constant rate difference. For interpretation, report adjusted rates for meaningful covariate profiles or standardize them over a declared target distribution:

\[ \hat R_a=\sum_{i=1}^n w_i \exp\left(\hat\beta_0+\hat\beta_Aa +\hat{\boldsymbol\gamma}^\mathsf T\mathbf z_i\right), \qquad \sum_iw_i=1. \]

\(\hat R_1/\hat R_0\) and \(\hat R_1-\hat R_0\) then describe population-average rate contrasts for the population represented by \(w_i\). This standardization requires uncertainty propagation and, for causal claims, the usual identification conditions in addition to a suitable rate model.

4 Choosing and Measuring Exposure

A defensible exposure denominator must satisfy three conditions:

  1. Opportunity: it represents the scale on which events could have occurred.
  2. Comparability: one exposure unit has a comparable meaning across analyzed units, possibly after modeled stratification.
  3. Temporal ordering and measurement: it is not a consequence of the event process in a way that creates selection or informative observation.

Examples include person-time at risk, geographic area under surveillance, machine operating time, callable genomic bases, and sequencing-library scale. Population size is appropriate for an incidence rate only when it adequately represents the population at risk over the observation period. Total reads may act as a library-scale adjustment, but composition bias can make raw library size a poor measure of effective sampling opportunity for a gene.

An offset is commonly treated as known. If \(t_i\) is estimated, noisy, or systematically biased, ordinary model standard errors do not include that uncertainty. Sensitivity analyses can perturb the denominator, use validation data, jointly model exposure and outcome, or propagate uncertainty by bootstrap or Bayesian computation when the estimation process is reproducible. Multiplying all exposures by a constant changes the intercept and the numerical rate unit, but not slope estimates or fitted counts.

5 Grouped Counts and Individual Events

Aggregation is coherent under a homogeneous rate within each group. If independent subunits \(j\) have common modeled rate \(r_g\) and exposures \(t_{gj}\), then

\[ \sum_jY_{gj}\sim\operatorname{Poisson} \left(r_g\sum_jt_{gj}\right). \]

The grouped analysis uses total count and total exposure. If rates vary with covariates inside a group, aggregation can discard confounding information, conceal interactions, or change the estimand. It can also hide dependence and overdispersion. Fine temporal splitting does not create independent experimental units: twelve monthly records from one person remain repeated observations from one person.

When event times and risk sets are available, survival or recurrent-event methods may represent censoring, time-varying risk, and within-unit dependence more faithfully. Poisson regression on split person-time can approximate certain piecewise-exponential hazard models, but the construction and dependence must be explicit.

6 Worked Example

Suppose two groups are followed for unequal times:

Group Events Person-years Observed rate per 100 person-years
A 18 900 2.0
B 15 300 5.0

Comparing counts alone suggests fewer events in B. Comparing exposure-adjusted rates gives

\[ \widehat{\operatorname{RR}} =\frac{15/300}{18/900}=2.5. \]

An intercept-plus-group Poisson model with \(\log(t_i)\) as the offset reproduces this rate ratio for the two aggregated rows. In individual data, additional covariates can adjust the conditional rate comparison, provided aggregation has not erased them and the likelihood reflects the actual independent units.

The estimate 2.5 is a statistical comparison. Its scientific importance depends on the event definition, absolute excess rate, follow-up mechanism, target population, uncertainty, confounding, and consequences of the event.

7 Implementation Without Treating Exposure as a Weight

In R, a basic Poisson rate model is written as ordinary, non-executed example code:

fit <- glm(
  events ~ treatment + age + batch + offset(log(person_time)),
  family = poisson(link = "log"),
  data = dat
)

exp(coef(fit)["treatment"])

offset(log(person_time)) changes the modeled mean. It is not a frequency, precision, survey, or inverse-probability weight. A zero exposure makes \(\log(t_i)\) undefined and usually indicates that the unit contributes no at-risk time; it requires a design decision, not an arbitrary small replacement. Missing exposure should be handled as missing data with its mechanism considered.

Predicted counts and rates are different outputs. A fitted count for specified exposure \(t\) is \(t\exp(\mathbf x^\mathsf T\hat{\boldsymbol\beta})\); the fitted rate sets \(t=1\) on the chosen scale. Software prediction calls must therefore provide or deliberately neutralize the offset according to the requested target.

8 Assumptions, Diagnostics, and Sensitivity

Assumption or issue Failure mechanism Diagnostic question Response or sensitivity analysis
Exposure measures opportunity Denominator does not represent time or units truly at risk Could an event occur throughout every recorded exposure unit? Refine risk time, eligibility, callable region, or observation model
Expected count is proportional to exposure Saturation, depletion, detection limits, or duration effects violate unit slope Do residuals vary with exposure; does a free \(\log(t)\) coefficient differ materially from one? Stratify time, add duration structure, or compare constrained and free-slope models
Exposure is measured accurately Error in an offset biases rates and understates uncertainty How was exposure estimated and validated? Perturb offsets, use validation data, or jointly model measurement error
Conditional mean is log-linear Missing curvature, interactions, or time trends distort rate ratios Residual and partial-pattern checks Splines, interactions, time strata, or revised covariates
Variance and dependence are represented Heterogeneity or repeated units makes Poisson uncertainty too small Conditional residuals, dispersion, cluster map, predictive checks Negative binomial, cluster-aware covariance, GEE, mixed or recurrent-event model
Observation is non-informative Follow-up or callability can depend on prognosis or outcome Why does exposure end or become unobservable? Censoring/observation model, inverse-probability methods, sensitivity analysis

Always inspect counts, rates, exposures, and residuals together. A large raw count may be expected under large exposure, while a modest count under tiny exposure can be influential. Check conclusions under scientifically credible alternative denominators, especially when preprocessing or detection rules determine the offset.

9 Bioinformatics Design and Interpretation

Consider somatic mutation counts from independently sampled tumors. The numerator is the number of validated mutations and the exposure might be callable megabases, not total genome length. A model can use \(\log(\text{callable Mb})\) as an offset and estimate treatment, tumor-type, and batch rate ratios. Tumors are the experimental units. Mutations and reads within a tumor are event-level or technical observations, not independent tumor replicates.

Callability is estimated from depth, mapping quality, filtering, copy number, and other sample-dependent processes. If disease state affects tissue quality and therefore callable sequence, the offset can participate in selection or measurement bias rather than merely correct opportunity. Report the callability definition and assess plausible alternatives. A statistically precise mutation rate ratio does not by itself establish a biological mechanism.

For RNA-seq, an effective library-size or normalization factor is usually an offset in a negative-binomial log-linear model. It makes expected counts proportional to estimated sampling scale, while coefficients compare expression rates on that scale. Composition bias means raw total reads may not be the right offset. Biological specimens provide condition-level replication; technical lanes or reads chiefly improve measurement. Sensitivity should examine normalization choice, batch allocation, dispersion, influential specimens, and the intended expression estimand.

10 Connections and Take-home Messages

  1. An offset separates the event rate from unequal opportunity by fixing the log-exposure coefficient at one.
  2. The exposure denominator defines the scientific rate estimand; it is not a generic weight or harmless preprocessing constant.
  3. Exponentiated slope coefficients are conditional rate ratios, while absolute rate differences depend on the baseline rate.
  4. Estimated, outcome-dependent, or poorly measured exposure requires uncertainty propagation and sensitivity analysis.
  5. Aggregation and record splitting do not create biological replication; the likelihood and uncertainty must follow the true experimental units.

11 References

Back to top