Deriving Multiplicative Count Effects

From a Log-Mean Coefficient to a Count Ratio

In a count-response model with a log link, coefficients are expressed on the logarithmic scale. Their practical interpretation is multiplicative: a coefficient describes how the expected count changes by a factor when its predictor increases by one unit, provided the other predictors and the exposure are held fixed. The factor is obtained by exponentiating the coefficient.

The distinction between a count difference and a count ratio matters. A coefficient of $0.25$ does not say that the expected count rises by $0.25$. It says the expected count is multiplied by $\exp(0.25)$, which is about $1.284$. The derivation is a direct comparison of two log means.

Definition: Suppose a count response $Y_i$, conditional on its predictors and exposure, has mean $\mu_i=E[Y_i\mid x_i,t_i]$, with $\mu_i>0$. A log-mean model with exposure $t_i>0$ has the form $$ \log(\mu_i)=\log(t_i)+\beta_0+\sum_{k=1}^{p}\beta_k x_{ik}. $$ Here $\log(t_i)$ is an exposure offset: its coefficient is fixed at one. The $\beta_k$ are model coefficients, and $x_{ik}$ is predictor $k$ for observation $i$. If exposure is not part of the model, omit $\log(t_i)$, which is equivalent to setting $t_i=1$.

For a Poisson model, $Y_i\mid x_i,t_i\sim\operatorname{Poisson}(\mu_i)$, so the conditional mean of the count is $\mu_i$. The ratio derivation below uses the log-mean equation and the definition of a conditional mean; it does not require any further Poisson-specific calculation. In particular, the factor describes a ratio of expected counts, not a guaranteed change in an individual observed count.

Deriving the One-Unit Effect

Compare two predictor profiles that differ only in predictor $j$: the second profile increases $x_j$ by one. Keep the exposure fixed at the same positive value, and keep every other predictor fixed. The coefficient for $x_j$ contributes $\beta_j x_j$ to the first log mean and $\beta_j(x_j+1)$ to the second.

Theorem: In the stated log-mean model, if two profiles differ by exactly one unit in $x_j$ and have the same exposure and all other predictors equal, then the ratio of their conditional expected counts is $\exp(\beta_j)$.

Proof. Let $\mu_0$ denote the conditional mean at the first profile, where $x_j=x$, and $\mu_1$ the conditional mean at the second, where $x_j=x+1$. The other predictors and the exposure have the same values in both profiles. Subtracting the two log-mean equations cancels their shared terms:

\log(\mu_1)-\log(\mu_0) =\bigl[\beta_j(x+1)\bigr]-\bigl[\beta_j x\bigr] =\beta_j.

Since both means are positive, their ratio is defined. Using the logarithm identity $\log(\mu_1)-\log(\mu_0)=\log(\mu_1/\mu_0)$, then exponentiating both sides, gives

\log\left(\frac{\mu_1}{\mu_0}\right)=\beta_j \quad\Longrightarrow\quad \frac{\mu_1}{\mu_0}=\exp(\beta_j).

Thus the expected count at the higher predictor value is $\exp(\beta_j)$ times the expected count at the lower value. $\square$

This result also shows why the coefficient itself is on the log-ratio scale: $\beta_j=\log(\mu_1/\mu_0)$. Exponentiation returns the ratio on the count scale. If $\beta_j>0$, then $\exp(\beta_j)>1$ and the expected count increases. If $\beta_j<0$, the factor is between zero and one and the expected count decreases. If $\beta_j=0$, the factor is one and the expected count is unchanged by that one-unit increase, under the stated comparison.

Worked Example: A Positive Coefficient and Fixed Exposure

Suppose the number of events $Y$ follows a Poisson model with mean $$ \log(\mu)=\log(t)-2+0.25x+0.10z, $$ where $t$ is exposure, $x$ is a predictor of interest, and $z$ is another predictor. Compare $x=2$ with $x=3$, holding $t=100$ and $z=1$ fixed.

At $x=2$, the log mean and mean are $$ \log(\mu_0)=\log(100)-2+0.25(2)+0.10(1)=\log(100)-1.4, \qquad \mu_0=100\exp(-1.4)\approx24.6597. $$

At $x=3$, $$ \log(\mu_1)=\log(100)-2+0.25(3)+0.10(1)=\log(100)-1.15, \qquad \mu_1=100\exp(-1.15)\approx31.6637. $$

The ratio is $$ \frac{\mu_1}{\mu_0} =\frac{100\exp(-1.15)}{100\exp(-1.4)} =\exp(0.25) \approx1.2840. $$ The expected count is therefore about $1.284$ times as large for the one-unit increase in $x$. The two numerical means give the same check: $31.6637/24.6597\approx1.2840$. The common exposure and the other predictor's contribution cancel.

Reading the Factor as a Percentage Change

A ratio can be translated into a relative percentage change. If the ratio of the higher-profile mean to the lower-profile mean is $r$, the relative change is $r-1$, or $100(r-1)\%$. In the one-unit comparison, $r=\exp(\beta_j)$.

Result: For a one-unit increase under the fixed-exposure, fixed-other-predictor comparison, the exact percentage change in the conditional mean is $100\{\exp(\beta_j)-1\}\%$.

Derivation. By the ratio result, $\mu_1/\mu_0=\exp(\beta_j)$. Subtracting one from both sides gives $(\mu_1/\mu_0)-1=\exp(\beta_j)-1$. The relative change in the mean is also $(\mu_1-\mu_0)/\mu_0$, since $$ \frac{\mu_1-\mu_0}{\mu_0} =\frac{\mu_1}{\mu_0}-1 =\exp(\beta_j)-1. $$ Multiplication by $100$ expresses this relative change as a percentage. This is an exact transformation, not an approximation. $\square$

For a negative coefficient, the percentage change is negative; its magnitude describes the percentage decrease. The percentage is calculated relative to the lower-$x_j$ mean. It is not a fixed number of counts: the absolute difference $\mu_1-\mu_0$ also depends on the starting mean.

Worked Example: A Negative Coefficient

Suppose an emergency-department count has conditional distribution $Y\mid x,z,t\sim\operatorname{Poisson}(\mu)$, with $$ \log(\mu)=\log(t)-1+0.30z-0.40x. $$ Compare $x=0$ and $x=1$ at the same exposure $t=10$ and with $z=0$.

At $x=0$, $\log(\mu_0)=\log(10)-1$, so $\mu_0=10\exp(-1)\approx3.6788$. At $x=1$, $\log(\mu_1)=\log(10)-1.40$, so $\mu_1=10\exp(-1.40)\approx2.4660$.

The ratio is $$ \frac{\mu_1}{\mu_0} =\frac{10\exp(-1.40)}{10\exp(-1)} =\exp(-0.40) \approx0.6703. $$ Thus the mean at $x=1$ is about $0.6703$ times the mean at $x=0$. The exact percentage change is $$ 100\{\exp(-0.40)-1\}\% \approx100(0.6703-1)\% =-32.97\%. $$ The model therefore implies a decrease of about $32.97\%$ in the conditional mean for this one-unit increase, not a decrease of $0.40$ events.

Why Holding Exposure Fixed Matters

The exposure offset contributes to the log mean just like any other shared term in the subtraction: it cancels only when the exposure is the same. If two profiles differ both in $x_j$ and in exposure, the count ratio includes both changes. Specifically, if the exposures are $t_0$ and $t_1$, while the other predictors stay fixed and $x_j$ increases by one, then $$ \log(\mu_1)-\log(\mu_0) =\log(t_1)-\log(t_0)+\beta_j =\log\left(\frac{t_1}{t_0}\right)+\beta_j, \qquad \frac{\mu_1}{\mu_0} =\frac{t_1}{t_0}\exp(\beta_j). $$

For example, if exposure doubles, the expected-count ratio is $2\exp(\beta_j)$, not just $\exp(\beta_j)$. The coefficient's multiplicative factor isolates the predictor's contribution only for a comparison at equal exposure. This is why an interpretation should state what is held fixed.

Worked Example: Checking the Ratio From Two Log Means

Consider a Poisson count model $$ \log(\mu)=\log(t)-1.2+0.70x+0.15z. $$ Compare $x=3$ and $x=4$ at exposure $t=8$ and $z=2$. The coefficient for $x$ is positive, so the expected count should be multiplied by $\exp(0.70)$.

For $x=3$, $$ \log(\mu_0)=\log(8)-1.2+0.70(3)+0.15(2)=\log(8)+1.2, \qquad \mu_0=8\exp(1.2)\approx26.5609. $$ For $x=4$, $$ \log(\mu_1)=\log(8)-1.2+0.70(4)+0.15(2)=\log(8)+1.9, \qquad \mu_1=8\exp(1.9)\approx53.4872. $$

The log means differ by $1.9-1.2=0.70$, so $$ \frac{\mu_1}{\mu_0} =\frac{8\exp(1.9)}{8\exp(1.2)} =\exp(0.70) \approx2.0138. $$ The direct calculation agrees: $53.4872/26.5609\approx2.0138$. The expected count is approximately doubled, with a factor slightly greater than two.

Common Mistakes and Practical Interpretation

Reading the coefficient as an additive count change. A coefficient of $0.70$ does not add $0.70$ expected events. Exponentiate it: the factor is $\exp(0.70)\approx2.0138$.
Reporting $\exp(\beta_j)$ as a percentage increase. It is a ratio. The percentage change is $100\{\exp(\beta_j)-1\}\%$. For example, a ratio of $1.28$ corresponds to about a $28\%$ increase, not a $128\%$ increase.
Ignoring the direction of comparison. The result describes the mean at the higher predictor value divided by the mean at the lower value. Reversing the comparison gives the reciprocal factor, $\exp(-\beta_j)$.
Interpreting a mean ratio as an observed-count guarantee. The model compares conditional expected counts. Realized counts vary, even when the predictor values and exposure are identical.
Forgetting the exposure condition. If exposure changes, its ratio also contributes to the expected-count ratio. The factor $\exp(\beta_j)$ alone applies when exposure is held fixed.

In practice, a fitted model replaces the unknown coefficient $\beta_j$ with its estimate $\hat{\beta}_j$, giving the estimated multiplicative factor $\exp(\hat{\beta}_j)$. This is a point estimate of the factor; by itself it does not describe its uncertainty. The essential interpretation remains the same: state the one-unit increase, name the other quantities held fixed, and report the exponentiated coefficient as a ratio of conditional means.

Key takeaway: In a log-mean model, increasing one predictor by one unit adds its coefficient to the log mean. Exponentiating that difference gives the expected-count ratio, $\exp(\beta_j)$, when the other predictors and exposure are unchanged.

Check Your Understanding

Use the log-mean comparison and specify the direction of each ratio.

A coefficient is $0.30$. What is the expected-count ratio for a one-unit increase, and what is the corresponding percentage change?
A coefficient is $-0.20$. Is the expected count multiplied by a factor above or below one? Write the factor.
In a log-mean model with an exposure offset, why must exposure be held fixed for the coefficient's exponential alone to equal the count ratio?
If the one-unit expected-count ratio is $1.5$, what is the ratio for the reverse comparison?
Explain why a coefficient of $0.40$ does not imply an increase of $0.40$ observed events.