Common Errors In Interpreting Poisson Regression

Why Interpretation Requires More Than Reading a Coefficient

A Poisson regression coefficient is easy to misstate because it is reported on a scale different from the response mean. Other errors arise when a predicted mean is treated as an observed count, or when a comparison that holds covariates fixed is described as an overall group difference. Each mistake changes the statistical claim, not just its wording.

This tutorial focuses on those distinctions. Let \(Y_i\) be a count for observation \(i\), and let \(x_i\) be its vector of predictors, including an intercept. In the usual Poisson regression model, conditional on the predictors, \(Y_i\) has a Poisson distribution with mean \(\mu_i\), and the log link specifies that mean:

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

Here \(\beta\) is the vector of model coefficients. A fitted value \(\hat{\mu}_i=\exp(x_i^\mathsf{T}\hat{\beta})\) is the model-based fitted conditional mean for observation \(i\), given its predictors. It is not a claim that the observed count \(Y_i\) must equal that value.

Coefficient Versus Multiplicative Mean Ratio

The log link means that a coefficient describes a change in the logarithm of the conditional mean. Exponentiating the coefficient expresses the corresponding change as a multiplicative ratio of means. The ratio interpretation is often more direct, but it is not the coefficient itself.

Result: In a Poisson regression model with \(\log(\mu)=\beta_0+\beta_1 z+\sum_{j=2}^p\beta_jx_j\), increasing \(z\) by one unit while holding the other predictors fixed multiplies the conditional mean by \(\exp(\beta_1)\). The coefficient \(\beta_1\) is a change in log mean; \(\exp(\beta_1)\) is the mean ratio.

To derive the result, compare two predictor profiles that differ only in \(z\), with values \(z\) and \(z+1\). Write their conditional means as \(\mu(z,x)\) and \(\mu(z+1,x)\), where \(x\) denotes the fixed values of the other predictors. Subtract the two linear predictors:

\log\{\mu(z+1,x)\}-\log\{\mu(z,x)\}=\beta_1.

The left side is the logarithm of a ratio. Exponentiating both sides gives \(\mu(z+1,x)/\mu(z,x)=\exp(\beta_1)\), as claimed. For a change of \(d\) units in \(z\), the ratio is \(\exp(d\beta_1)\). This interpretation assumes the specified model has a linear term \(\beta_1z\) and no interaction that changes the effect across the fixed covariate values. If such terms are present, the comparison must include them. \(\square\)

Worked Example: Translating a Coefficient Correctly

Suppose a fitted model is \(\log(\mu)=1.1-0.4z+0.2x\), where \(z\) is a binary indicator and \(x\) is another predictor. Compare \(z=1\) with \(z=0\) at the same value of \(x\).

The coefficient for \(z\) is \(-0.4\). This is the change in the log conditional mean, not a decrease of \(0.4\) counts. The mean ratio is \[ \exp(-0.4)\approx 0.6703. \] Thus, at a fixed value of \(x\), the fitted conditional mean for \(z=1\) is about 0.6703 times the fitted conditional mean for \(z=0\), or about 33.0% lower because \(1-0.6703=0.3297\).

For a numerical check, take \(x=2\). When \(z=0\), the linear predictor is \(1.1+0.2(2)=1.5\), so the fitted mean is \(\exp(1.5)\approx4.482\). When \(z=1\), the predictor is \(1.1-0.4+0.4=1.1\), so the fitted mean is \(\exp(1.1)\approx3.004\). Their ratio is approximately \(3.004/4.482=0.670\), agreeing with \(\exp(-0.4)\).

Worked Example: A Coefficient Is Not a Count Difference

Suppose the coefficient of a continuous predictor \(z\) is \(0.2\). Compare predictor values \(z=3\) and \(z=5\), holding all other predictors fixed.

The change is \(d=5-3=2\) units, so the conditional mean ratio is \(\exp(2\times0.2)=\exp(0.4)\approx1.492\). The fitted conditional mean is about 1.492 times as large at \(z=5\), an increase of approximately 49.2%.

The count difference depends on the starting mean. If the fitted mean at \(z=3\) is 2, the fitted mean at \(z=5\) is \(2(1.492)\approx2.984\), a difference of about 0.984. If the starting mean is 10, the fitted mean at \(z=5\) is about 14.92, a difference of about 4.92. The same coefficient therefore does not represent a fixed increase in the expected count.

Fitted Means Versus Observed Counts

A conditional mean is an expectation over the count distribution at specified predictor values. An observed count is one realized value from that distribution. Even when a model is correctly specified, the realization may be above or below its mean; indeed, the Poisson distribution assigns probability to individual nonnegative integer counts.

Result: If \(Y\mid x\sim\operatorname{Poisson}(\mu(x))\), then \(E[Y\mid x]=\mu(x)\), but an observation \(Y=y\) need not equal \(\mu(x)\). For a fitted model, \(\hat{\mu}(x)\) estimates the conditional mean; it does not determine the realized count.

For completeness, the Poisson probability mass function is \(P(Y=k\mid x)=e^{-\mu(x)}\mu(x)^k/k!\) for \(k=0,1,2,\ldots\), with \(\mu(x)>0\). Its conditional expectation follows directly:

\begin{aligned} E[Y\mid x] &=\sum_{k=0}^{\infty} k e^{-\mu(x)} \frac{\mu(x)^k}{k!}\\ &=e^{-\mu(x)}\mu(x) \sum_{k=1}^{\infty}\frac{\mu(x)^{k-1}}{(k-1)!}\\ &=e^{-\mu(x)}\mu(x)e^{\mu(x)} =\mu(x). \end{aligned}

The sum is finite as an expectation because the displayed series equals the convergent exponential series. This establishes the mean, not a rule that each outcome equals it. Since \(Y\) has integer support, a noninteger mean cannot even be an individual possible count. \(\square\)

Worked Example: Interpreting a Fitted Mean of 2.4

For one predictor profile, suppose a fitted model gives \(\hat{\mu}=2.4\), while the observed count is \(y=5\). The fitted mean is 2.4 counts per observation under the model at that profile. The observed count is 5; it is not “wrong” merely because it differs from 2.4. The observed-minus-fitted difference is \(5-2.4=2.6\).

The fitted mean alone does not state the probability of observing 5. Under the Poisson distribution with mean 2.4, that probability is \[ P(Y=5)=e^{-2.4}\frac{2.4^5}{5!} =e^{-2.4}\frac{79.62624}{120} \approx0.0602. \] Thus, 5 is one possible count with probability about 6.02% under this particular fitted distribution. The calculation uses the fitted mean as a parameter of the predictive distribution; it does not replace the observed value with that mean.

A related mistake is to report a fitted mean as if it were the probability of a positive count. For a Poisson variable with mean \(\mu\), the probability of a positive count is \(P(Y\geq1)=1-P(Y=0)=1-e^{-\mu}\), not \(\mu\). For example, at \(\mu=2.4\), this probability is \(1-e^{-2.4}\approx0.9093\), whereas 2.4 is the expected count. These quantities have different meanings and different scales.

Conditional Effects Versus Unconditional Comparisons

A regression coefficient describes a comparison at fixed values of the other predictors when the model specifies a common coefficient effect. An overall or crude group comparison may instead compare groups with different distributions of those predictors. Such comparisons need not agree: the groups may differ both in group membership and in their covariate composition.

The distinction can be made precise. Let \(G\) be a group indicator and \(C\) a covariate. The mean in group \(g\), averaged over that group’s covariate distribution, is \[ E[Y\mid G=g]=\sum_c E[Y\mid G=g,C=c]P(C=c\mid G=g) \] when \(C\) has finitely many categories. The weights are group-specific. Consequently, even if the conditional mean ratio comparing groups is the same at every value of \(C\), the ratio of these averages can differ when \(P(C=c\mid G=g)\) differs between groups.

Worked Example: Conditional Ratio and Crude Ratio

Let \(G=0\) denote control and \(G=1\) denote treatment, and let \(C\) indicate whether a person is in an older age group. Suppose the conditional means are 1 for control-younger, 4 for control-older, 2 for treatment-younger, and 8 for treatment-older. These values are represented by the log-mean model \(\log(\mu)=\beta_0+\beta_GG+\beta_CC\), with \(\beta_0=0\), \(\beta_G=\log(2)\), and \(\beta_C=\log(4)\).

At either age category, the treatment-to-control conditional mean ratio is 2: for younger people it is \(2/1=2\), and for older people it is \(8/4=2\). This is the conditional effect described by \(\exp(\beta_G)=2\).

Now suppose 20% of the control group and 80% of the treatment group are older. The control group’s crude mean is \(0.8(1)+0.2(4)=1.6\). The treatment group’s crude mean is \(0.2(2)+0.8(8)=6.8\). The crude ratio is \(6.8/1.6=4.25\), not 2. The calculation differs because the two group means use different age weights; it does not contradict the conditional ratio of 2.

If both groups instead had the same age distribution, with proportion \(q\) older, their averaged means would be \(1(1-q)+4q\) and \(2(1-q)+8q\). The second expression is exactly twice the first, so the averaged ratio would be 2. This shows why the group-specific covariate distributions matter for an unconditional comparison.

This averaging identity is an application of the law of total expectation. It is not enough to know a regression coefficient to recover an overall group contrast: one must also specify the covariate distribution over which predictions are averaged. A conditional comparison and a crude comparison answer different questions.

Common Errors and Corrections

Wrong: “The coefficient is \(-0.4\), so the expected count falls by 0.4.” Correction: For a one-unit change, \(-0.4\) is the log-mean change; \(\exp(-0.4)\) is the conditional mean ratio. A count difference depends on the starting mean.
Wrong: “The fitted mean is 2.4, so this observation should have count 2.4.” Correction: The fitted value estimates a conditional expectation. The observed count is a random realization, and it must be a nonnegative integer.
Wrong: “The treatment coefficient is the observed overall treatment-to-control mean ratio.” Correction: In the stated model it is a conditional log-mean contrast. An overall comparison can differ if the groups have different covariate distributions.
Wrong: “A predicted mean is the probability of a positive count.” Correction: For a Poisson mean \(\mu\), the probability of a positive count is \(1-e^{-\mu}\); the mean is an expected count.

When interpreting a result, first name the quantity being reported: coefficient, exponentiated coefficient, fitted conditional mean, observed count, or group-average mean. Then state the comparison and what is held fixed or averaged over. That discipline prevents the most common changes in meaning.

Check Your Understanding

Use the distinctions in this tutorial to answer the following questions.

A Poisson regression coefficient is \(0.3\) for a predictor. What is the conditional mean ratio for a one-unit increase, and what does the coefficient itself represent?
A fitted conditional mean is \(1.7\), and the observed count is 0. Is this a contradiction? Explain.
In a model without interactions, what does exponentiating a binary-predictor coefficient compare, and which predictors are held fixed?
Two groups have the same conditional mean ratio at every covariate value but different covariate distributions. Must their crude mean ratio equal the conditional ratio? Explain.
If a Poisson fitted mean is \(\mu\), write the probability of observing at least one event and distinguish it from the mean.