Reciprocal-Link Poisson Models

Why the Reciprocal Link Needs Care

A Poisson regression model must produce a positive conditional mean because that mean is the parameter of a Poisson distribution. The reciprocal link achieves positivity by making the inverse mean, rather than the mean itself, a linear function of the predictors. That change of scale has an important consequence: a positive coefficient usually corresponds to a lower predicted count, not a higher one.

This tutorial focuses on reading that relationship correctly and checking whether the linear predictor is valid. The coefficients and predictions in the examples are hypothetical. No interest-rate or mortality assumptions are needed; each response is a count with a stated Poisson distribution.

Definition: For independent count responses $Y_i$, a reciprocal-link Poisson regression specifies $Y_i\mid \mathbf{x}_i\sim\operatorname{Poisson}(\mu_i)$, where $\mu_i=E[Y_i\mid\mathbf{x}_i]>0$, and $$ g(\mu_i)=\frac{1}{\mu_i}=\eta_i=\beta_0+\beta_1x_{i1}+\cdots+\beta_px_{ip}. $$ The linear predictor $\eta_i$ must be strictly positive for every covariate pattern on which the model is used. Thus $\mu_i=1/\eta_i$ is positive and finite.

As with other Poisson regression models, this specification describes the conditional distribution and mean given the predictors. It assumes that each response is a nonnegative integer count and, under the model, has conditional variance equal to its conditional mean: $\operatorname{Var}(Y_i\mid\mathbf{x}_i)=\mu_i$. The independent-response assumption is appropriate only when the observations are conditionally independent; repeated or clustered observations may require a different dependence structure.

The coefficients are linear effects on the inverse-mean scale. They are not constant changes in the mean count. In particular, the reciprocal link is not the log link: an exponentiated coefficient is not, in general, a multiplicative change in the mean.

From the Linear Predictor to the Mean

Once the predictor has been calculated, the fitted mean is its reciprocal. The predictor must be calculated first; taking the reciprocal of individual terms and then adding them does not give the model mean.

\eta_i=\beta_0+\sum_{j=1}^{p}\beta_jx_{ij}, \qquad \mu_i=\frac{1}{\eta_i}, \qquad \eta_i>0.

For a one-unit increase in a continuous predictor $x_j$, holding the other predictors fixed, the linear predictor changes by $\beta_j$. If the original predictor is $\eta>0$, the new one is $\eta+\beta_j$, which must also be positive. The corresponding change in the mean is

\frac{1}{\eta+\beta_j}-\frac{1}{\eta} = -\frac{\beta_j}{\eta(\eta+\beta_j)}.

The denominator is positive whenever both predictor values are valid. Therefore, if $\beta_j>0$, the one-unit increase lowers the mean; if $\beta_j<0$, it raises the mean, provided the comparison remains in the valid domain. The size of the change depends on $\eta$, so it is not constant across observations.

Interpretation rule: A coefficient changes the inverse mean by its stated amount for a one-unit predictor change. If the new and old predictors are both positive, a positive coefficient lowers the mean and a negative coefficient raises it. Interpret the fitted mean by taking the reciprocal of the complete predictor.

Why the Reciprocal Formula Gives These Effects

The reciprocal link is defined by $g(\mu)=1/\mu$. Substituting the linear predictor into that definition gives $1/\mu=\eta$; solving this equation for $\mu$ gives $\mu=1/\eta$. The coefficient-direction result follows directly from the behavior of this reciprocal function.

For a continuous predictor $x_j$, the predictor is $\eta=\beta_0+\sum_k\beta_kx_k$, so its derivative with respect to $x_j$ is $\beta_j$. Differentiating $\mu=\eta^{-1}$ using the chain rule gives

\frac{\partial\mu}{\partial x_j} = -\eta^{-2}\frac{\partial\eta}{\partial x_j} = -\frac{\beta_j}{\eta^2}.

Because $\eta^2>0$, the derivative has the opposite sign from $\beta_j$. This establishes the direction of the local mean change. The derivative is not itself the exact change for a whole one-unit increase; the finite-change formula above gives that comparison. For a categorical predictor coded $z=0$ or $z=1$, calculate both predictors and take their reciprocals rather than using a derivative.

Worked Example: A Continuous Predictor and Mean Counts

Hypothetical setup. For each independent observation, $Y$ is the number of service interruptions in a fixed one-month period, with $Y\mid x\sim\operatorname{Poisson}(\mu(x))$. A fitted reciprocal-link model has $\eta(x)=0.25+0.05x$, where $x$ is a nonnegative index value. The coefficients are supplied, not estimated. Find the predicted means at $x=2$ and $x=6$, and compare them.

Method. Calculate each linear predictor, check it is positive, and then take its reciprocal.

At $x=2$,

\eta(2)=0.25+0.05(2)=0.25+0.10=0.35, \qquad \mu(2)=\frac{1}{0.35}=\frac{20}{7}\approx2.857.

At $x=6$,

\eta(6)=0.25+0.05(6)=0.25+0.30=0.55, \qquad \mu(6)=\frac{1}{0.55}=\frac{20}{11}\approx1.818.

The predictor increases by $0.20$, from $0.35$ to $0.55$, and the mean decreases by approximately $2.857-1.818=1.039$ interruptions per month. Both predictors are positive, so both Poisson means are valid. The positive slope lowers the mean, as the reciprocal-link interpretation predicts. Common mistake avoided: reading the positive coefficient as an increase in the expected count.

Categorical Predictors and Comparisons

For a binary indicator $z$, with $z=0$ as the reference group and $z=1$ as the comparison group, a model $\eta=\beta_0+\beta_1z$ gives

\mu_0=\frac{1}{\beta_0}, \qquad \mu_1=\frac{1}{\beta_0+\beta_1}.

Both denominators must be positive. The coefficient $\beta_1$ is the difference between the groups' inverse means, not the difference between their means. When both group means are valid, their ratio can also be written

\frac{\mu_1}{\mu_0} = \frac{\beta_0}{\beta_0+\beta_1}.

This ratio depends on both coefficients. It is not generally equal to $\exp(\beta_1)$, and $\beta_1$ alone is not a rate ratio.

Worked Example: Comparing Two Hypothetical Groups

Hypothetical setup. For each independent policy record, $Y$ is the count of reported device faults in one fixed quarter. Suppose $Y\mid z\sim\operatorname{Poisson}(\mu(z))$, where $z=0$ denotes the reference group and $z=1$ denotes a group receiving a hypothetical inspection protocol. The fitted model is $\eta(z)=0.25+0.10z$. Calculate the group means, their difference, and their ratio.

For the reference group, $z=0$:

\eta(0)=0.25+0.10(0)=0.25, \qquad \mu_0=\frac{1}{0.25}=4.

For the comparison group, $z=1$:

\eta(1)=0.25+0.10(1)=0.35, \qquad \mu_1=\frac{1}{0.35}=\frac{20}{7}\approx2.857.

The comparison-group mean minus the reference-group mean is $2.857-4=-1.143$ faults per quarter, approximately. The mean ratio is

\frac{\mu_1}{\mu_0} = \frac{20/7}{4} = \frac{5}{7} \approx0.7143.

Thus the fitted comparison-group mean is about $71.43\%$ of the reference-group mean. Each predictor is positive, so both distributions have valid Poisson parameters. Common mistake avoided: treating $0.10$ as the mean difference or as a multiplicative ratio. It is the increase in inverse mean.

Positivity Is a Model Constraint

A reciprocal-link model is usable only where its linear predictor is strictly positive. A zero predictor does not yield a finite mean, and a negative predictor would yield a negative value for $\mu$, which cannot parameterize a Poisson distribution. Consequently, checking the coefficients individually is not enough: check the complete predictor at the covariate patterns where predictions or interpretations are required.

For a finite set of observed covariate patterns, evaluate $\eta_i$ at each pattern. For predictors ranging over a continuous rectangle, a linear function reaches its minimum and maximum at boundary points; checking the relevant corners can establish whether it stays positive throughout that rectangle. If the predictor domain is unrestricted, a nonconstant linear predictor will generally take negative values somewhere. A model that is valid on the observed range may therefore be invalid under extrapolation.

Worked Example: Checking Feasibility Across a Predictor Range

Hypothetical setup. For each independent observation, $Y$ is the number of minor equipment faults in a fixed quarter, with $Y\mid x,z\sim\operatorname{Poisson}(\mu(x,z))$. The proposed model is $\eta(x,z)=0.40-0.06x+0.03z$, where $0\leq x\leq5$ and $z$ is either 0 or 1. Check that the model is valid over this full range, then predict the mean at $x=5,z=1$.

The predictor decreases as $x$ increases, because its $x$ coefficient is negative, and increases as $z$ changes from 0 to 1. Its minimum is therefore at $x=5,z=0$:

\eta(5,0)=0.40-0.06(5)+0.03(0)=0.40-0.30=0.10.

Its maximum is at $x=0,z=1$:

\eta(0,1)=0.40-0.06(0)+0.03(1)=0.43.

Every predictor on the stated range is between $0.10$ and $0.43$, so it is positive. At the requested pattern,

\eta(5,1)=0.40-0.30+0.03=0.13, \qquad \mu(5,1)=\frac{1}{0.13}=\frac{100}{13}\approx7.692.

The predicted mean is approximately $7.692$ faults per quarter. The positivity check covers the full specified range, not just the requested pattern. Common mistake avoided: assuming valid coefficients automatically imply valid means for all predictor values.

Common Mistakes and Practical Checks

Before interpreting a reciprocal-link Poisson result, apply these checks:

Invert the whole predictor. For $\eta=\beta_0+\beta_1x$, the mean is $1/(\beta_0+\beta_1x)$, not $1/\beta_0+\beta_1x$.
Check the direction on the correct scale. A positive coefficient raises the inverse mean. If the old and new predictors are positive, it lowers the mean.
Check feasibility at the patterns being discussed. A negative $\eta$ is not a small or unusual Poisson mean; it is an invalid mean.
Do not interpret coefficients as log-link effects. The reciprocal coefficient is neither a log rate ratio nor a percentage change.
Keep units in mind. If $Y$ is a count per specified observation period, $\mu$ is an expected count per that period and $\eta$ is its reciprocal. Predictions refer to the period used to define $Y$.

Key takeaway: In reciprocal-link Poisson regression, $\eta$ is linear in the predictors and $\mu=1/\eta$. Coefficient direction reverses when translated to the mean, and every predictor value used must be strictly positive.

Check Your Understanding

For each hypothetical model, use the reciprocal link and check the stated predictor domain before interpreting the mean. Answers are rounded to four decimal places where needed.

Independent responses satisfy $Y\mid x\sim\operatorname{Poisson}(\mu(x))$, with $\eta(x)=0.20+0.05x$. Calculate the means at $x=4$ and $x=8$, and give the change from the first mean to the second. (Answer: $\mu(4)=2.5000$, $\mu(8)=1.6667$, change $=-0.8333$.)
For independent quarterly counts $Y\mid z\sim\operatorname{Poisson}(\mu(z))$, let $\eta(z)=0.40+0.10z$, where $z=0$ is the reference group and $z=1$ the comparison group. Find both means and the comparison-to-reference mean ratio. (Answer: $\mu_0=2.5000$, $\mu_1=2.0000$, ratio $=0.8000$.)
Let $Y\mid x\sim\operatorname{Poisson}(\mu(x))$ with $\eta(x)=0.50-0.04x$, for $0\leq x\leq10$. Calculate the means at $x=0,5,10$, and state whether the model is feasible throughout the interval. (Answer: means $2.0000,\ 3.3333,\ 10.0000$, respectively; feasible because the minimum predictor is $0.10>0$.)
Independent counts have model $\eta(x,z)=0.15+0.02x-0.04z$, for $0\leq x\leq5$ and $z\in\{0,1\}$. Find the minimum and maximum predictors, then calculate the mean at $x=5,z=1$. (Answer: minimum $0.11$, maximum $0.25$; the model is feasible; $\mu(5,1)=1/0.21\approx4.7619$.)
A proposed model for independent Poisson counts has $\eta(x,z)=0.12-0.03x+0.02z$, for $x\in\{0,1,2,3\}$ and $z\in\{0,1\}$. Find the smallest predictor and its corresponding mean, and decide if the model is feasible over the stated domain. (Answer: minimum predictor $0.03$ at $x=3,z=0$, corresponding mean $33.3333$; feasible over the stated domain.)

What You'll Learn