Penalized Poisson Regression With Exposure Offsets

Why Exposure Belongs in the Offset

When counts are observed over unequal amounts of time, area, or opportunity, the count alone is not a comparable measure of event frequency. A count of 12 events over 6,000 person-years represents a lower rate than 12 events over 1,000 person-years. Poisson regression handles this difference by placing the logarithm of known exposure in the linear predictor as a fixed offset. Penalization then acts on modeled covariate effects, not on the known exposure.

Let \(Y_i\) be the count for observation \(i\), \(E_i>0\) its known exposure, and \(x_i\) its vector of predictors. Under the Poisson model, \(Y_i\) has mean \(\mu_i\), and a log link with exposure offset gives the following specification.

The coefficient vector \(\beta\) describes log-rate effects. For a one-unit increase in predictor \(j\), holding the other predictors fixed, the modeled rate is multiplied by \(\exp(\beta_j)\). The expected count is different: it is the rate multiplied by exposure, \(\mu_i=E_i\exp(x_i^\mathsf{T}\beta)\). Thus equal covariates but unequal exposures generally produce unequal predicted counts.

Penalized Objective and Score

For observed counts \(y_1,\ldots,y_n\), omit the Poisson terms \(\log(y_i!)\), which do not depend on the coefficients. A common objective to minimize is the negative log-likelihood plus a penalty:

Here \(\lambda\geq0\) controls the overall penalty and \(0\leq\alpha\leq1\) mixes the lasso and ridge components. Ridge is obtained at \(\alpha=0\), lasso at \(\alpha=1\), and intermediate values give the elastic net. If the model includes an intercept, it is conventionally excluded from the penalty; the same is true of the fixed offset, which is not an estimated coefficient. The objective can equivalently be written using \(\mu_i=E_i\exp(x_i^\mathsf{T}\beta)\); the offset contribution to the term linear in \(y_i\) is then a constant with respect to \(\beta\).

Let \(X\) be the matrix whose \(i\)th row is \(x_i^\mathsf{T}\), and let \(\mu\) and \(y\) denote the vectors of means and observed counts. Differentiating the smooth negative log-likelihood gives \(X^\mathsf{T}(\mu-y)\). Therefore, for coordinates included in the penalty, the stationarity condition is

The subgradient is \(z_j=\operatorname{sign}(\beta_j)\) when \(\beta_j\ne0\), and any value in \([-1,1]\) when \(\beta_j=0\). An unpenalized intercept instead has its ordinary score equation, with no penalty term. The offset changes each \(\mu_i\) through the known factor \(E_i\), but it creates no additional derivative with respect to \(\beta\).

Two Properties of the Offset Model

Proof. By the Poisson mean model and the log link, \(\log E[Y_i\mid x_i,E_i]=\log\mu_i=\log E_i+x_i^\mathsf{T}\beta\). Exponentiating both sides gives \(\mu_i=E_i\exp(x_i^\mathsf{T}\beta)\). Since \(E_i>0\), dividing by exposure yields \(\mu_i/E_i=\exp(x_i^\mathsf{T}\beta)\). Increasing \(x_{ij}\) by one changes the rate from its original value to that value multiplied by \(\exp(\beta_j)\), as claimed. \(\square\)

Proof. The smooth objective, up to terms independent of \(\beta\), is \(\sum_i\{E_i\exp(x_i^\mathsf{T}\beta)-y_i x_i^\mathsf{T}\beta\}\). Its gradient is \(X^\mathsf{T}(\mu-y)\), because the derivative of \(\mu_i\) with respect to \(\beta\) is \(\mu_i x_i\). Differentiating again gives \(X^\mathsf{T}\operatorname{diag}(\mu_i)X\). For any vector \(v\), \(v^\mathsf{T}X^\mathsf{T}\operatorname{diag}(\mu_i)Xv=\sum_i\mu_i(x_i^\mathsf{T}v)^2\geq0\), since every \(\mu_i>0\). This proves convexity. If \(X\) has full column rank, every nonzero \(v\) has \(Xv\ne0\), so the displayed sum is strictly positive and the smooth objective is strictly convex. Each specified penalty is convex; adding a convex function preserves convexity, and adding it to a strictly convex function preserves strict convexity. A strictly convex objective cannot have two distinct minimizers: strict convexity at their midpoint would give a value smaller than the common minimum. This proves the uniqueness statement whenever a finite minimizer exists. \(\square\)

Worked Coefficient Comparisons

The following examples use one observation and a single coefficient, with no intercept: \(x=1\), so the linear predictor is \(\log E+\beta\). This deliberately simple setup makes each penalty’s effect visible. The examples are calculations for these specified models, not a claim that penalization must reduce every coefficient or every prediction in a general multivariable model.

To compare penalty types directly, keep the \(y=6,E=3,x=1\) observation from the first example and change only the penalty. Its unpenalized log-rate coefficient is \(\log2\approx0.69315\), and the unpenalized fitted count is 6.

How to Interpret the Shrinkage

Ridge continuously shrinks coefficients, while lasso can produce exact zeros; elastic net combines these behaviors. In a multivariable fit, these descriptions concern coefficient regularization, not a guarantee that every fitted count moves toward zero or that every prediction decreases. Correlated predictors, differing covariate patterns, and the exposure factor all affect fitted counts. For instance, even with a smaller estimated rate, a sufficiently larger exposure can yield a larger predicted count. A rate ratio \(\exp(\widehat\beta_j)\) should therefore be interpreted separately from any one exposure-specific count prediction.

Common Pitfall

A frequent error is to treat \(\log E_i\) as an ordinary predictor whose coefficient should be estimated or shrunk. That changes the model: the exposure relationship is no longer fixed to be proportional. Another error is to report \(\exp(\widehat\beta_j)\) as a predicted count. It is a multiplicative rate effect for a one-unit covariate change; the expected count for observation \(i\) is \(E_i\exp(x_i^\mathsf{T}\widehat\beta)\). Always state which quantity is being reported and include the exposure when calculating a count.