Interpreting The Poisson Regression Intercept

What the Intercept Refers To

In a Poisson regression, the intercept is not a context-free summary of the response. Its meaning depends on where the predictors are set to zero. With the log link, exponentiating the intercept gives the model’s conditional mean at that covariate origin. Understanding what the origin represents is therefore essential to interpreting the intercept.

Let $Y_i$ be a count response for observation $i$, and let $x_i=(x_{i1},\ldots,x_{ip})^\top$ denote its $p$ predictors. The Poisson regression model specifies a conditional distribution given the predictors, with a log-linear conditional mean.

Definition: In the Poisson regression model $$ Y_i\mid x_i \sim \operatorname{Poisson}(\mu_i), \qquad \log(\mu_i)=\beta_0+\sum_{j=1}^{p}\beta_jx_{ij}, $$ the intercept is $\beta_0$. The covariate origin is the predictor vector $x_i=0$, meaning that every predictor is zero. The model-implied conditional mean at that origin is $\exp(\beta_0)$.

The expression “every predictor is zero” must be read using the actual definitions and coding of the predictors. For a continuous predictor measured from its natural zero, zero has that measurement’s usual meaning. For an indicator variable, zero denotes whichever category was coded as zero. For a centered predictor, zero denotes the value used as the center.

Why the Exponentiated Intercept Is the Mean

The interpretation follows directly from the conditional-mean model. A Poisson random variable with parameter $\mu_i$ has conditional expectation $\mu_i$, and the log-link equation specifies that mean.

Result: Under the stated Poisson regression model, at the covariate origin, $$ E[Y_i\mid x_i=0] =\mu_i\big|_{x_i=0} =\exp(\beta_0). $$ Thus, $\beta_0$ is the log of the conditional mean at zero-valued predictors, and $\exp(\beta_0)$ is that conditional mean.

To verify the result, substitute $x_{ij}=0$ for each $j$ into the linear predictor:

\log(\mu_i)\big|_{x_i=0} =\beta_0+\sum_{j=1}^{p}\beta_j(0) =\beta_0.

Exponentiating both sides gives $\mu_i|_{x_i=0}=\exp(\beta_0)$. Because $E[Y_i\mid x_i]=\mu_i$ under the Poisson model, this is the conditional mean of the response at the origin. If the model is fitted, the corresponding estimate is $\exp(\hat{\beta}_0)$; it is the mean specified by the fitted model at that same origin.

Worked Example: A Zero Origin With Two Predictors

Suppose the conditional mean for a count of clinic visits is modeled by $$ \log(\mu_i)=-0.8+0.04\,\text{age}_i+0.6\,D_i, $$ where age is measured in years and $D_i=1$ for a particular group and $D_i=0$ for the reference group. Find the model-implied mean at the covariate origin.

At the origin, age is zero and $D_i=0$. Substitution gives $$ \log(\mu_i)=-0.8+0.04(0)+0.6(0)=-0.8. $$ Therefore, $$ \mu_i=\exp(-0.8)\approx 0.4493. $$ The intercept is $-0.8$, and its exponential, approximately $0.449$, is the model-implied mean number of visits for the reference group at age zero. This is a precise interpretation of the fitted model; whether age zero is a useful or scientifically relevant reference for the study population is a separate question.

Centering Changes the Reference Point

A common way to make an intercept more interpretable is to center a continuous predictor. Centering subtracts a chosen constant from each value, so that zero on the transformed scale corresponds to that chosen value on the original scale. The model’s conditional means need not change: the intercept changes to match the new coordinate origin.

Theorem: Suppose the linear predictor is $\beta_0+\beta^\top x$, where $x$ is a $p$-dimensional predictor vector. Choose a fixed vector $c$, and define the centered predictor $x^*=x-c$. Then the same linear predictor can be written as $$ \beta_0^*+\beta^\top x^*, \qquad \beta_0^*=\beta_0+\beta^\top c. $$ Consequently, $\exp(\beta_0^*)$ is the conditional mean at $x^*=0$, which is the original covariate value $x=c$.

Proof. Since $x^*=x-c$, we have $x=x^*+c$. Substitute this identity into the original linear predictor:

\beta_0+\beta^\top x =\beta_0+\beta^\top(x^*+c) =(\beta_0+\beta^\top c)+\beta^\top x^*.

Thus setting $\beta_0^*=\beta_0+\beta^\top c$ gives exactly the same linear predictor for every $x$. The conditional mean is the exponential of that linear predictor, so it is also unchanged for every observation. At $x^*=0$, the original value is $x=c$, and the linear predictor is $\beta_0+\beta^\top c=\beta_0^*$. Hence the conditional mean there is $\exp(\beta_0^*)$. $\square$

This reparameterization is one-to-one: given $\beta_0^*$, the original intercept is $\beta_0=\beta_0^*-\beta^\top c$, and the slopes remain unchanged. If the model is estimated by maximizing the likelihood and a finite maximum exists, this change of coordinates leaves the likelihood at every represented model unchanged. The fitted intercept therefore changes to represent the new origin, while the fitted conditional means do not change.

Worked Example: Centering Age at 50 Years

Suppose a fitted Poisson regression for a count outcome has $$ \log(\mu_i)=-1.2+0.05\,\text{age}_i, $$ where age is measured in years. First, the intercept on the original age scale describes the mean at age zero: $$ \exp(-1.2)\approx 0.301. $$ Now define $\text{age}_i^*=\text{age}_i-50$. Since $\text{age}_i=\text{age}_i^*+50$, rewrite the predictor: $$ -1.2+0.05\,\text{age}_i =-1.2+0.05(\text{age}_i^*+50) =1.3+0.05\,\text{age}_i^*. $$ The centered intercept is $1.3$, so its exponential is $$ \exp(1.3)\approx 3.669. $$ On the centered scale, zero means age 50, so $3.669$ is the model-implied conditional mean at age 50. Checking in the original equation gives $\exp(-1.2+0.05(50))=\exp(1.3)$, the same value. The change from approximately $0.301$ to $3.669$ reflects a changed reference point, not a changed fitted model.

Origins Depend on Coding Choices

Centering is one way to choose an origin; coding choices for categorical predictors also determine what zero means. With a binary indicator $D$, where $D=0$ denotes a reference category, setting $D=0$ in the intercept interpretation selects that reference category. If a continuous predictor is centered as well, the intercept refers to the reference category at the centering value.

Worked Example: A Centered Score and a Reference Category

Consider the model $$ \log(\mu_i)=0.7-0.03(\text{score}_i-40)+0.4D_i, $$ where $D_i=1$ denotes a treatment group and $D_i=0$ denotes the reference group. The continuous score has been centered at 40. At the covariate origin, $\text{score}_i-40=0$ and $D_i=0$. Therefore, $$ \log(\mu_i)=0.7,\qquad \mu_i=\exp(0.7)\approx 2.014. $$ The exponentiated intercept is the model-implied mean at score 40 in the reference group. It does not describe the treatment group, because that group is represented by $D_i=1$.

To see the effect of using the uncentered score, expand the score term: $$ 0.7-0.03(\text{score}_i-40)+0.4D_i =1.9-0.03\,\text{score}_i+0.4D_i. $$ On this scale, the intercept is $1.9$, and its exponential is $\exp(1.9)\approx 6.686$. This is the model-implied mean at score zero in the reference group. The two intercepts differ because score zero and score 40 are different reference values, even though both equations describe the same conditional means.

When the Intercept Is Useful—and When It Is Not

An intercept is especially easy to interpret when the covariate origin represents a meaningful, well-supported setting. For example, centering a predictor at a typical value can make the intercept describe the mean at that value. In a model with indicator variables, the origin often selects the reference categories. In either case, the interpretation comes from the model specification, not from the coefficient alone.

If zero is impossible or far outside the observed range, $\exp(\hat{\beta}_0)$ is still the fitted model’s conditional mean at the zero vector, but it may be an extrapolation. The data may contain little or no information directly about that setting. Centering can give the intercept a more relevant reference point; it does not, by itself, improve the model’s fit or change which conditional means the model represents.

Common Mistake

Wrong reasoning: “The exponentiated intercept is the overall average count in the sample.” That conclusion does not follow from the model. The intercept refers to the conditional mean at the covariate origin, whereas an overall sample average combines responses from the covariate values actually represented in the sample.

Correction: State the predictor values that define zero before interpreting $\exp(\hat{\beta}_0)$. If zero is not a useful reference, recenter the relevant predictors and interpret the resulting intercept at the new origin. Do not call it a sample-wide mean unless the model and covariate structure justify that interpretation.

Key takeaway: In Poisson regression, $\exp(\beta_0)$ is the conditional mean at the point where all coded predictors equal zero. Centering or recoding changes which point is called zero and therefore changes the intercept’s interpretation, while an equivalent reparameterization preserves the model’s conditional means.

Check Your Understanding

Answer each question by identifying the predictor values at the model’s origin.

In a Poisson model with $\log(\mu_i)=0.3+0.02x_i$, what conditional mean does the exponentiated intercept describe?
A binary predictor is coded $D=0$ for usual care and $D=1$ for a new treatment. What category does the intercept refer to when all other predictors are zero?
If age is centered by subtracting 60, what original age does a centered age value of zero represent?
Why can $\exp(\hat{\beta}_0)$ fail to equal the sample average count even when the Poisson regression is correctly specified?
For a centered predictor $x^*=x-c$, how does the intercept change, and what original predictor value does its new exponentiated value describe?