Poisson Coefficients As Multiplicative Mean Ratios

Why Exponentiate a Poisson Coefficient?

A Poisson regression coefficient is on the scale of the log conditional mean. That makes the coefficient additive in the model equation, but it does not directly describe the change in the expected count. Exponentiating the coefficient translates it into a ratio of conditional means, which is often easier to interpret: it tells us how the expected count is multiplied when a predictor changes and the other predictors stay fixed.

The key is to compare two predictor settings that differ only in the predictor of interest. This is a conditional comparison: the model holds the other included predictors constant. The result is a ratio of expected responses, not a prediction of how much an individual count will change.

The One-Unit Mean Ratio

Let $Y$ be a count response and let $x=(x_1,\ldots,x_p)$ be a vector of predictor values. Write $\mu(x)=E[Y\mid x]$ for the conditional mean. In the standard Poisson regression model with a log link,

\log \mu(x)=\beta_0+\sum_{k=1}^{p}\beta_k x_k.

Here $\beta_0,\ldots,\beta_p$ are the model coefficients. Since the logarithm is defined only for positive values, this model specifies a positive conditional mean. To interpret $\beta_j$, compare the mean at $x$ with the mean when just $x_j$ is increased by one unit. Let $x+e_j$ denote the latter predictor setting, where $e_j$ is one in position $j$ and zero elsewhere.

Theorem: Under the log-linear conditional mean model above, if both $x$ and $x+e_j$ are valid predictor settings, then $$ \frac{\mu(x+e_j)}{\mu(x)}=\exp(\beta_j). $$ Thus $\exp(\beta_j)$ is the conditional mean ratio for a one-unit increase in $x_j$, holding every other predictor fixed.

Proof. At the original setting,

\log\mu(x)=\beta_0+\sum_{k=1}^{p}\beta_kx_k.

At the changed setting, the predictors other than $x_j$ are unchanged, while $x_j$ is replaced by $x_j+1$. Therefore,

\log\mu(x+e_j) =\beta_0+\sum_{k\ne j}\beta_kx_k+\beta_j(x_j+1) =\log\mu(x)+\beta_j.

Exponentiating both sides gives $\mu(x+e_j)=\mu(x)\exp(\beta_j)$. The conditional mean $\mu(x)$ is positive, so division by it gives the stated ratio. $\square$

The value of the ratio determines the direction of the change. If $\exp(\beta_j)>1$, the conditional mean is larger after the one-unit increase. If it is between zero and one, the conditional mean is smaller. If it equals one, the conditional mean is unchanged. Because the exponential of any finite coefficient is positive, the ratio cannot be negative.

Worked Example: A Positive Coefficient

Suppose a fitted Poisson regression includes a continuous predictor $x_1$, measured in years, and has coefficient $\hat{\beta}_1=0.08$. Interpret a one-year increase while holding the other predictors fixed.

The estimated conditional mean ratio is

\exp(\hat{\beta}_1)=\exp(0.08)\approx 1.0833.

Thus, for two predictor settings that differ by one year and agree on all the other predictors, the first setting's estimated conditional mean is about $1.0833$ times the second's. Expressed as a percentage change, this is

100(1.0833-1)\%\approx 8.33\%.

So the estimated conditional mean is about 8.33% higher for a one-year increase in $x_1$, with the other predictors fixed. This is a relative change in the mean, not an increase of 0.08 expected counts.

Ratios and Percentage Changes

A ratio and a percentage change describe the same comparison in different forms. If the new conditional mean is $R$ times the old one, then the relative change is $R-1$. For a one-unit change in $x_j$, the ratio $R$ is $\exp(\beta_j)$, so the exact percentage change is $100\{\exp(\beta_j)-1\}\%$. This formula applies to negative coefficients as well as positive ones.

Result: In the log-linear mean model, the percentage change in the conditional mean for a one-unit increase in $x_j$, holding the other predictors fixed, is $$ 100\{\exp(\beta_j)-1\}\%. $$

Derivation. The one-unit ratio is $\mu(x+e_j)/\mu(x)=\exp(\beta_j)$. By the definition of percentage change, the change relative to the original mean is the new mean minus the original mean, divided by the original mean, and multiplied by 100%. Substituting the ratio gives

100\frac{\mu(x+e_j)-\mu(x)}{\mu(x)}\% =100\left\{\frac{\mu(x+e_j)}{\mu(x)}-1\right\}\% =100\{\exp(\beta_j)-1\}\%.

The percentage is negative when the ratio is below one. Its magnitude is not generally equal to the percentage increase for the reverse comparison: reversing the comparison uses the reciprocal ratio. $\square$

Worked Example: A Negative Coefficient

Suppose the estimated coefficient for a predictor measured in ten-unit increments is $\hat{\beta}_2=-0.25$. Find the conditional mean ratio and percentage change for a ten-unit increase, holding the other predictors fixed.

First exponentiate the coefficient:

\exp(-0.25)\approx 0.7788.

The conditional mean after the increase is therefore about 0.7788 times its previous value. The percentage change is

100(0.7788-1)\%\approx -22.12\%.

The negative sign indicates a decrease: the estimated conditional mean is about 22.12% lower for a ten-unit increase in this predictor. It would be incorrect to say that it is 77.88% lower; 0.7788 is the fraction remaining, not the fraction lost.

Changes of More Than One Unit

The same model gives a simple extension when $x_j$ changes by an amount other than one. For a change of $d$ units, with the other predictors fixed, the change in the linear predictor is $d\beta_j$. The ratio consequently depends on both the coefficient and the size of the change.

Result: If $x$ and $x+d e_j$ are valid predictor settings in the log-linear mean model, then $$ \frac{\mu(x+d e_j)}{\mu(x)}=\exp(d\beta_j). $$

Proof. The terms in the linear predictor for $x$ and $x+d e_j$ differ only in the $j$th term. Subtracting the first linear predictor from the second gives

\log\mu(x+d e_j)-\log\mu(x) =\beta_j(x_j+d)-\beta_jx_j=d\beta_j.

The left-hand side is $\log\{\mu(x+d e_j)/\mu(x)\}$. Exponentiating both sides yields the claimed ratio. $\square$

For example, a two-unit increase has ratio $\exp(2\beta_j)$, not $2\exp(\beta_j)$. Multiplicative changes compound: applying the one-unit ratio twice gives $\exp(\beta_j)\exp(\beta_j)=\exp(2\beta_j)$.

Worked Example: A Binary Indicator

Suppose $z$ is a binary indicator, coded 0 for the reference category and 1 for the other category. In a Poisson regression with coefficient estimate $\hat{\beta}_z=0.60$, interpret the conditional mean ratio for category 1 relative to category 0, holding the other predictors fixed.

Changing from 0 to 1 is a one-unit increase, so the estimated ratio is

\exp(0.60)\approx 1.8221.

The model therefore estimates a conditional mean in category 1 about 1.8221 times the mean in category 0 for the same values of the other predictors. Equivalently, the estimated conditional mean is

100(1.8221-1)\%\approx 82.21\%

higher in category 1. The interpretation depends on the coding: category 0 is the reference, and the ratio describes category 1 relative to it. This model-based association alone does not establish that changing category membership would cause the mean response to change.

Conditions and Common Pitfalls

The constant ratio follows from a linear predictor on the log-mean scale with a main-effect term $\beta_jx_j$. It applies only to predictor settings for which the comparison is meaningful and allowed by the model. It is not automatically the interpretation of a coefficient in a model that includes interactions or nonlinear terms involving $x_j$: in those models, the change in the linear predictor can depend on the other predictors or on the starting value of $x_j$.

Keep three quantities distinct: $\beta_j$, the ratio $\exp(\beta_j)$, and the percentage change $100\{\exp(\beta_j)-1\}\%$. The coefficient is a change in log conditional mean; the exponentiated coefficient is a multiplier; and subtracting one from that multiplier gives the relative change. For small coefficients, $\beta_j$ can be a rough approximation to the relative change on a decimal scale, but it is not the exact percentage change.

Wrong: “A coefficient of 0.08 means the expected count increases by 0.08.” Correction: The one-unit mean ratio is $\exp(0.08)$; the corresponding percentage change is $100\{\exp(0.08)-1\}\%$.
Wrong: “A ratio of 0.78 means the mean decreases by 78%.” Correction: It means about 78% of the original mean remains, or a decrease of about 22%.
Wrong: “The ratio compares observations with different values of all predictors.” Correction: This coefficient interpretation compares predictor settings that differ in $x_j$ and hold the other included predictors fixed.
Wrong: “A conditional mean ratio is a causal effect.” Correction: The ratio is a model-based conditional association; a causal interpretation requires additional design and assumptions.

Key takeaway: In a Poisson regression with a log-linear conditional mean, exponentiating $\beta_j$ gives the factor by which the conditional mean changes for a one-unit increase in $x_j$, holding other predictors fixed. Subtract one from this factor to obtain the exact relative change.

Check Your Understanding

Use the log-linear conditional mean model to answer each question.

A predictor has coefficient $0.20$. What is the one-unit conditional mean ratio, and what is the corresponding percentage change?
A predictor has coefficient $-0.40$. Is the conditional mean ratio above or below one? What does that imply about the direction of change?
Why is a ratio of $0.65$ described as a 35% decrease rather than a 65% decrease?
For a change of three units in $x_j$, what expression gives the conditional mean ratio when the model contains the term $\beta_jx_j$?
What does the coefficient of a binary indicator compare, and why does that comparison by itself not prove a causal effect?