Identity-Link Poisson Models And Feasibility

Why Feasibility Matters

A Poisson regression model describes a count using a conditional mean. With the log link, exponentiation keeps the fitted mean positive. An identity link instead makes the mean a linear function of the predictors. This gives coefficients a direct additive interpretation, but it also creates a constraint: every fitted mean must be a valid Poisson parameter.

The key task is therefore not just to calculate an expected count. It is to check whether the model produces an admissible mean across the predictor values where it will be used. A model that is feasible for the observed data may fail at a boundary or when extrapolated.

Definition: For a response $Y$ that is a count, an identity-link Poisson regression model specifies $$ Y_i\mid \mathbf{x}_i \sim \operatorname{Poisson}(\mu_i), \qquad \mu_i=E[Y_i\mid\mathbf{x}_i]=\eta_i=\beta_0+\sum_{j=1}^p\beta_jx_{ij}. $$ Conditional on the predictors, each response has a Poisson distribution and the observations are usually assumed conditionally independent. The Poisson mean must satisfy $\mu_i\ge 0$. A strictly positive mean gives a nondegenerate Poisson distribution; a mean of zero is a degenerate distribution concentrated entirely at zero.

The identity link is $g(\mu)=\mu$, so the linear predictor $\eta$ is the mean itself. In the usual Poisson model, the conditional variance is also $\operatorname{Var}(Y_i\mid\mathbf{x}_i)=\mu_i$. The identity link does not change that distributional assumption; it changes how the mean is related to the predictors.

What the Coefficients Mean

For a single quantitative predictor $x$, the model is $\mu(x)=\beta_0+\beta_1x$. The intercept is the expected count at $x=0$, if zero is meaningful and within the model's interpretation. The slope is the change in expected count for a one-unit increase in $x$. It is an additive change, not a percentage change or a rate ratio.

\mu(x+\Delta)-\mu(x) = \bigl(\beta_0+\beta_1(x+\Delta)\bigr) -\bigl(\beta_0+\beta_1x\bigr) = \beta_1\Delta

This identity justifies the additive interpretation: the intercept cancels, and the mean change depends only on the predictor change $\Delta$ and the slope. For example, a slope of $0.2$ means that a one-unit increase in $x$ adds $0.2$ to the expected count, wherever the comparison is made. It does not mean a 20% increase.

With an indicator $G$ taking values zero and one, a model such as $\mu(x,G)=\beta_0+\beta_1x+\beta_2G$ gives $\beta_2$ as the difference in expected counts between $G=1$ and $G=0$ at the same $x$. If an interaction $xG$ is included, that difference becomes dependent on $x$. In either case, feasibility must be checked for every relevant combination of predictors.

Checking the Mean Across Its Domain

A prediction is feasible only if it is nonnegative. In ordinary regression work, one generally wants the mean to be strictly positive over the range where the model is applied, so the fitted distribution is nondegenerate. A zero mean is mathematically permitted by the Poisson distribution, but it assigns probability one to a count of zero. A negative mean is never a valid Poisson parameter.

For a linear function of one predictor over a closed interval, the minimum occurs at an endpoint. Consequently, checking both endpoints is enough to establish nonnegativity throughout that interval. For multiple predictors, the same idea applies to the corners of a rectangular predictor region when the mean is linear in those predictors. If the predictor domain is not rectangular, or there are restrictions linking predictors, check the minimum over the actual domain rather than assuming all combinations are possible.

State the intended domain.
Specify the observed or planned range of each predictor and any restrictions on combinations of predictors.

Write the mean function.
For an identity link, the linear predictor is the mean; do not exponentiate it.

Find the minimum mean.
For a linear function over an interval or rectangular range, evaluate the relevant endpoints or corners.

Classify the result.
A negative value makes the model infeasible there. A zero value is a valid degenerate Poisson mean, but not a strictly positive prediction.

Worked Example: An Increasing Mean Over a Bounded Range

A hypothetical monitoring program records a count $Y$ for each inspection. Suppose that, conditional on the inspection's workload index $x$, $Y$ has a Poisson distribution with mean $\mu(x)=1.2+0.15x$. The model is intended for $0\le x\le 8$. Find the mean at $x=4$ and $x=8$, the effect of increasing $x$ by three units, and the probability of a zero count at $x=8$. Check feasibility over the stated range.

Method. Use the identity-link mean directly, then check its minimum and use the Poisson probability formula for the final part. The conditional Poisson assumption means counts are nonnegative integers with the specified mean.

At $x=4$, $\mu(4)=1.2+0.15(4)=1.2+0.60=1.80$. At $x=8$, $\mu(8)=1.2+0.15(8)=1.2+1.20=2.40$.

A three-unit increase changes the mean by $\beta_1\Delta=0.15(3)=0.45$. For instance, moving from $x=4$ to $x=7$ gives $\mu(7)=1.2+0.15(7)=2.25$, which is $2.25-1.80=0.45$ above the mean at $x=4$.

The slope is positive, so the smallest mean on $0\le x\le8$ occurs at $x=0$, where $\mu(0)=1.2$. The mean is positive throughout the domain; the model is feasible there.

For a Poisson variable with mean $\mu$, $P(Y=0)=e^{-\mu}$. Thus $P(Y=0\mid x=8)=e^{-2.40}\approx0.0907$, to four decimal places.

Check. The mean rises from $1.20$ at the lower endpoint to $2.40$ at the upper endpoint, consistent with the positive slope. The zero-count probability is less than one and positive, as expected for a positive Poisson mean. The common mistake to avoid is treating $0.15$ as a 15% increase rather than an additive increase of $0.15$ per unit.

Feasibility Boundaries and Zero Means

When a slope is negative, the mean decreases as the predictor increases. Solving $\mu(x)>0$ identifies the region of strictly positive means. Be precise about the boundary: if the mean reaches zero at a value of $x$, that value is not part of the strictly positive region, even though a Poisson distribution with mean zero can be defined there.

Worked Example: A Declining Mean That Becomes Invalid

For a hypothetical count of equipment alerts, assume $Y\mid x\sim\operatorname{Poisson}(\mu(x))$, with $\mu(x)=4.5-0.6x$. The stated predictor domain is $0\le x\le9$. Determine where the mean is strictly positive, evaluate it at $x=5$, calculate the probability of zero alerts there, and decide whether the model is feasible over the full domain.

Method. Solve the mean inequality and inspect the domain endpoint. Do not assume that a linear predictor is automatically a valid Poisson mean.

For a strictly positive mean, $$ 4.5-0.6x>0 \quad\Longleftrightarrow\quad 4.5>0.6x \quad\Longleftrightarrow\quad x<7.5. $$ Thus, within the stated domain, the mean is strictly positive for $0\le x<7.5$. At $x=7.5$, $\mu(7.5)=4.5-0.6(7.5)=4.5-4.5=0$. At $x=9$, $\mu(9)=4.5-0.6(9)=4.5-5.4=-0.9$.

At $x=5$, $\mu(5)=4.5-0.6(5)=4.5-3.0=1.5$. Therefore, $P(Y=0\mid x=5)=e^{-1.5}\approx0.2231$, to four decimal places.

The model is not feasible over the full stated domain, because it assigns a negative mean at $x=9$ and at every $x>7.5$. It is strictly positive only below $7.5$; $7.5$ is the supremum of that region, not a largest value with a strictly positive mean. The common mistake is to report $7.5$ as included in the strictly positive range. At that boundary the mean is exactly zero.

Checking Multiple Predictor Combinations

A model can be feasible for one group but not another, or feasible at the center of the data but invalid at a corner of the predictor range. The check must include all groups and the full intended range of quantitative predictors. An estimated coefficient alone does not establish feasibility; the intercept and every relevant predictor value contribute to the fitted mean.

Worked Example: A Group Effect and a Bounded Continuous Predictor

A hypothetical service team models its count of daily follow-ups using a binary indicator $G$, where $G=0$ denotes the standard workflow and $G=1$ denotes a revised workflow. Assume conditionally independent Poisson counts with mean $\mu(x,G)=1.10-0.12x-0.25G$, for $0\le x\le5$ and $G\in\{0,1\}$. Check feasibility for both workflows, interpret the coefficients, and calculate the zero-count probability for the revised workflow at $x=5$.

Method. Since the mean decreases with both $x$ and $G$, its minimum is at the largest $x$ and largest $G$. Also compare the groups at the same $x$ to interpret the indicator coefficient.

For the standard workflow, the mean ranges from $\mu(0,0)=1.10$ to $\mu(5,0)=1.10-0.12(5)=1.10-0.60=0.50$. For the revised workflow, it ranges from $\mu(0,1)=1.10-0.25=0.85$ to $\mu(5,1)=1.10-0.60-0.25=0.25$. The overall minimum is $0.25$, so all means are strictly positive on the specified predictor domain.

Holding $G$ fixed, one additional unit of $x$ changes the expected count by $-0.12$. Holding $x$ fixed, changing from $G=0$ to $G=1$ changes the expected count by $-0.25$. For example, at $x=3$, the standard-workflow mean is $1.10-0.12(3)=0.74$, and the revised-workflow mean is $0.74-0.25=0.49$.

At $x=5,G=1$, the mean is $0.25$, so $P(Y=0\mid x=5,G=1)=e^{-0.25}\approx0.7788$, to four decimal places. This probability is high because a zero count is common when the mean is small, but it is not one because the mean is positive.

Check. The minimum occurs at $x=5,G=1$ and equals $0.25>0$, confirming feasibility. The common mistake is to call the group coefficient a 25% reduction. It is an additive difference of $0.25$ in the expected count; the corresponding percentage difference would depend on the comparison mean.

Common Feasibility Traps

Checking only the observed average predictor. A positive mean near the center does not guarantee a positive mean at an endpoint or for a subgroup. Check the full intended predictor domain.
Calling a negative mean a low expected count. A Poisson distribution cannot have a negative parameter. Such a prediction is invalid, not merely unusual.
Including a zero boundary in a strict inequality. If $\mu(x)=0$ at a boundary, the mean is nonnegative there but not strictly positive there. State which condition is being checked.
Interpreting identity-link effects multiplicatively. A coefficient is an additive change in the mean. It is not a percentage change or a multiplicative factor.
Assuming observed-data feasibility guarantees safe extrapolation. Extending a linear mean beyond the range where it was assessed can eventually produce a negative prediction, especially when a slope is negative.

The identity link can be useful when additive changes in expected counts are the intended interpretation, but feasibility is a substantive limitation. If a fitted mean becomes negative in a region where predictions are required, the model as specified cannot be used there. Restricting its application to a defensible range may address extrapolation; otherwise, a different model specification may be needed. Any alternative must still be assessed against the data and the purpose of the analysis.

Key takeaway: Under an identity-link Poisson model, coefficients describe additive changes in the conditional mean. Because the mean equals a linear predictor, explicitly check its minimum over the full predictor domain: negative values are impossible, and zero values correspond only to a degenerate all-zero distribution.

Check Your Understanding

For each problem, use the stated conditional Poisson model and check the complete predictor range before interpreting its predictions.

A hypothetical count has mean $\mu(x)=0.7+0.2x$ for $0\le x\le4$. Find the mean at $x=3$, the change in mean for a two-unit increase in $x$, and whether the model is strictly positive throughout the range. (Answer: $\mu(3)=1.3$; the two-unit increase adds $0.4$; yes, the minimum is $0.7$.)
A count model has mean $\mu(t)=3-0.4t$, with $0\le t\le8$. Find the strictly positive range, the mean at $t=4$, and whether the model is feasible over the full stated range. (Answer: strictly positive for $0\le t<7.5$, zero at $7.5$, and negative at $t=8$; $\mu(4)=1.4$; not feasible over the full range.)
For a hypothetical count, $\mu(x,A)=0.6+0.3A-0.1x$, where $A\in\{0,1\}$ and $0\le x\le5$. Find the minimum mean over both groups and interpret the group difference at fixed $x$. (Answer: the minimum is $0.1$ at $A=0,x=5$, so all means are strictly positive; $A=1$ adds $0.3$ to the mean at any fixed $x$.)
A model has $\mu(x)=2.2-0.35x$ for $0\le x\le7$. Find the mean at $x=6$, determine where it is strictly positive, and state the supremum of the strictly positive range. (Answer: $\mu(6)=0.1$; strictly positive for $0\le x<44/7\approx6.2857$; the supremum is $44/7$, where the mean is zero.)
A model fitted to a hypothetical dataset has $\mu(x)=0.5+0.4x$, with observed predictors from $x=1$ through $x=4$. Evaluate the mean at the observed endpoints, then assess the proposed extrapolation to $x=-2$. (Answer: the endpoint means are $0.9$ and $2.1$, so the model is positive over the observed interval; at $x=-2$, $\mu=-0.3$, so that extrapolated Poisson prediction is invalid.)

What You'll Learn