Common Misreadings of Count Effects

Why Count Effects Are Easy to Misread

A count model often reports coefficients on a logarithmic scale, while readers usually want statements about counts: how many events are expected, how much that expectation changes, or how likely an event is. Moving between these scales requires care. A coefficient is not itself a count difference, a rate ratio is not a probability, and a fitted mean is not the count that must be observed.

The safest interpretation starts with the quantity the model describes. Let $Y$ be a count response and let $\mu(x)=E[Y\mid X=x]$ denote its conditional mean at predictor value $x$. In a log-link model, the conditional mean has the form

\log \mu(x)=\eta(x), \qquad \mu(x)=\exp\{\eta(x)\},

where $\eta(x)$ is the linear predictor. Thus, model coefficients directly describe changes in the log of the conditional mean. To interpret them on the count scale, exponentiate the relevant contrast.

Definition: For two covariate profiles $x_0$ and $x_1$, the expected-count difference is $\mu(x_1)-\mu(x_0)$, and the expected-count ratio is $\mu(x_1)/\mu(x_0)$. A log-mean contrast is $\log\{\mu(x_1)\}-\log\{\mu(x_0)\}$; exponentiating it gives the expected-count ratio.

What a Log-Scale Coefficient Says

Suppose a predictor $x$ enters the linear predictor as $\beta x$, with all other predictors held fixed. Increasing $x$ by $\Delta$ changes the linear predictor by $\beta\Delta$. The corresponding expected-count ratio is $\exp(\beta\Delta)$, not $\beta\Delta$. The expected-count difference depends additionally on the starting mean.

Result: In a log-link model, if changing a predictor from $x_0$ to $x_1$ changes the linear predictor by $\delta$, then $$ \frac{\mu(x_1)}{\mu(x_0)}=\exp(\delta), \qquad \mu(x_1)-\mu(x_0)=\mu(x_0)\{\exp(\delta)-1\}. $$ The ratio depends on the linear-predictor contrast; the difference also depends on the baseline conditional mean.

Derivation. By definition, $\log\mu(x_1)-\log\mu(x_0)=\delta$. Exponentiating both sides gives $\mu(x_1)/\mu(x_0)=\exp(\delta)$. Multiplying by $\mu(x_0)$ gives $\mu(x_1)=\mu(x_0)\exp(\delta)$; subtracting $\mu(x_0)$ yields the difference formula. $\square$

For a one-unit increase in $x$, $\delta=\beta$, so $\exp(\beta)$ is the expected-count ratio. The associated relative change is $\exp(\beta)-1$, often expressed as a percentage. Neither quantity is an additive change in expected counts. For a change of $\Delta$ units, use $\exp(\beta\Delta)$. This distinction matters especially when a predictor changes by several units or when the baseline mean varies across profiles.

Worked Example: A Coefficient Is Not a Count Difference

Suppose a log-link model has coefficient $\beta=0.40$ for a continuous predictor. At a starting profile, the fitted conditional mean is $3.2$. Calculate the effect of increasing the predictor by two units.

The linear-predictor contrast is $\delta=0.40(2)=0.80$. Therefore, the expected-count ratio is

\exp(0.80)\approx 2.2255.

The new fitted mean is $3.2(2.2255)\approx 7.1216$. The expected-count difference is $7.1216-3.2=3.9216$. Thus, the model implies an expected count about 2.23 times as large, or an increase of about 122.6% relative to the starting mean. It does not imply an increase of $0.80$ counts, $0.40$ counts, or two counts.

A Rate Ratio Is Not a Probability Change

A rate ratio or expected-count ratio compares two conditional means. It does not, by itself, compare probabilities. Counts can be greater than one, whereas the probability of an event is between zero and one. Even when the count model specifies a distribution that lets us calculate probabilities, a mean ratio does not become a probability ratio or a probability difference.

For example, if $Y\mid X=x$ has a Poisson distribution with mean $\mu(x)$, then its probability of at least one event is $1-\exp\{-\mu(x)\}$. This probability is a nonlinear function of the mean. A proportional change in the mean therefore need not produce the same proportional change in this probability.

Result: If $Y\mid X=x\sim\operatorname{Poisson}(\mu(x))$, then $$ P(Y\geq 1\mid X=x)=1-\exp\{-\mu(x)\}. $$ If two means have ratio $r$, the corresponding probabilities are $1-\exp(-\mu_0)$ and $1-\exp(-r\mu_0)$, not probabilities with ratio $r$ in general.

Derivation. The event $Y\geq1$ is the complement of $Y=0$. The Poisson probability mass function gives $P(Y=0\mid X=x)=\exp\{-\mu(x)\}\mu(x)^0/0!=\exp\{-\mu(x)\}$. Taking the complement proves the formula. Substituting $\mu_1=r\mu_0$ gives the two probabilities stated above; the formula contains an exponential, so the ratio of means alone does not imply an equal ratio of probabilities. $\square$

Worked Example: A 50% Mean Increase Is Not a 50% Probability Increase

Suppose a Poisson model gives a mean of $2$ for one profile and a rate ratio of $1.5$ for a second profile. Find the means and the probabilities of at least one event.

The second mean is $2(1.5)=3$. The probability of at least one event at the first profile is

1-\exp(-2)\approx 1-0.1353=0.8647.

At the second profile it is

1-\exp(-3)\approx 1-0.0498=0.9502.

The mean increased by 50%, but the probability increased by approximately $0.9502-0.8647=0.0855$, or 8.55 percentage points. The probability ratio is approximately $0.9502/0.8647=1.0989$, not $1.5$. These probability calculations use the Poisson distributional assumption; the mean ratio alone does not determine event probabilities for an unspecified count distribution.

Expected Counts Are Not Observed Counts

A fitted conditional mean is an expectation across the model’s outcome distribution for a given covariate profile. It need not be an integer, and it is not a promise about the particular count that will be observed. A future observation may be below or above the mean. This distinction is also why a statement about a fitted mean should not be rewritten as a statement that an individual will experience that exact count.

For a Poisson response with mean $2.4$, for instance, the observed count can be $0,1,2,\ldots$, even though the mean is $2.4$. The model assigns positive probability to counts other than $2.4$; the mean summarizes the distribution rather than naming its realized outcome.

Worked Example: A Fitted Mean and an Observed Count

At a particular profile, a Poisson model gives fitted mean $\mu=2.4$. Suppose the observed count is $Y=4$. Does the observation contradict the fitted mean?

No. The fitted mean is $E[Y\mid X=x]=2.4$, while the observed count is one realized value, $4$. Under the stated Poisson model, the probability of observing exactly four is

P(Y=4\mid X=x) =\exp(-2.4)\frac{2.4^4}{4!} \approx 0.09072\left(\frac{33.1776}{24}\right) \approx 0.1254.

Thus, $4$ is a possible observation with probability about 0.125 under this model. The calculation does not say that four is the mean; it shows how an observation is assessed relative to a distribution centered, in the expectation sense, at $2.4$.

Units, Contrasts, and Baselines

Coefficient interpretation depends on the predictor’s units. If $x$ is measured in years, its coefficient describes a one-year change; if it is recorded in months, the numerical coefficient changes, even though the model can represent the same relationship. The contrast of interest must therefore be stated explicitly before exponentiating.

To see the rescaling, write the predictor in new units as $z=x/c$, where $c>0$. Then $x=cz$, and a term $\beta x$ becomes $(c\beta)z$. The coefficient in the new units is $c\beta$, so a one-unit change in $z$ corresponds to a $c$-unit change in $x$, with ratio $\exp(c\beta)$. The modeled effect for a fixed real-world change is unchanged when the contrast is expressed consistently.

Worked Example: Interpreting a Binary Predictor on Both Scales

A binary predictor has coefficient $\beta=-0.30$, with category 0 as the reference. The fitted mean for category 0 is $4.5$. Compare category 1 with category 0.

The expected-count ratio is $\exp(-0.30)\approx0.7408$. Thus, the fitted mean for category 1 is about 74.08% of the reference mean, a relative decrease of about $1-0.7408=0.2592$, or 25.92%.

The category 1 fitted mean is $4.5(0.7408)\approx3.3336$. The difference is $3.3336-4.5=-1.1664$, or about 1.17 fewer expected counts. The additive difference is not $-0.30$; it depends on the reference mean. Also, this is a comparison of fitted conditional means, not a claim that each category 1 observation will be about 1.17 counts lower.

Common Misreadings and Their Corrections

Wrong: “The coefficient is $0.40$, so the expected count rises by 0.40.” Correction: For a one-unit increase, exponentiate the coefficient to obtain the mean ratio. To get an additive difference, multiply the relative change by the starting mean.
Wrong: “A rate ratio of 1.5 means the probability of any event is 50% higher.” Correction: A rate ratio compares means. Calculate a probability from the specified outcome distribution, if that distribution is part of the model.
Wrong: “The fitted mean is 2.4, so the observed count should be 2 or 3.” Correction: A mean is an expectation, not a restriction on possible observations. Use the outcome distribution to describe the range and probabilities of future counts.
Wrong: “The coefficient describes the effect of any change in the predictor.” Correction: State the change in predictor units. For a change of $\Delta$, use the contrast $\beta\Delta$, not $\beta$ alone.
Wrong: “A 25% lower fitted mean means every observation is 25% lower.” Correction: The model comparison concerns conditional means at specified profiles; observed counts vary around those means.

Key takeaway: Keep four quantities separate: the log-mean contrast, its exponentiated mean ratio, the baseline-dependent difference in fitted means, and the observed or future count. Name which one you are interpreting before translating it into words.

Check Your Understanding

For each question, identify the scale of the quantity being interpreted and show the necessary conversion.

A log-link coefficient is $0.2$. What is the expected-count ratio for a three-unit increase in its predictor?
A mean ratio is $0.8$, and the reference fitted mean is $10$. What is the comparison mean and the additive difference?
Why does a Poisson mean ratio of 2 not generally imply that the probability of at least one event doubles?
A fitted mean is $1.7$, but the observed count is $0$. Explain why these values are not contradictory.
If a predictor is rescaled from years to months, what must be stated before interpreting its coefficient?