Multiple Predictors And Conditional Mean Effects

Why Include Several Predictors?

A count outcome may vary with more than one measured characteristic. For example, the number of clinic visits might depend on age, symptom severity, and travel distance. A model with several predictors describes the conditional mean for each combination of those characteristics, rather than describing the mean as a function of only one predictor.

In a multiple-predictor model, each coefficient describes the predictor’s contribution while the other included predictors are held fixed. This is a statement about the model’s conditional mean, not automatically a causal effect. The Poisson distribution specifies the conditional response distribution; the log link and linear predictor specify how its mean depends on the covariates.

Definition: For independent count responses $Y_1,\ldots,Y_n$ and covariate vectors $x_i=(x_{i1},\ldots,x_{ip})^\mathsf{T}$, a multiple-predictor Poisson regression model is $$ Y_i\mid x_i\sim \operatorname{Poisson}(\mu_i),\qquad \log(\mu_i)=\beta_0+\sum_{j=1}^{p}\beta_jx_{ij}. $$ Here $\mu_i=E[Y_i\mid x_i]$, $\beta_0$ is the intercept, and $\beta_j$ is the coefficient of predictor $j$. Equivalently, $\mu_i=\exp(\beta_0+\sum_{j=1}^{p}\beta_jx_{ij})$, so the conditional mean is positive.

The vector $x_i$ contains the values of the predictors for observation $i$. The model uses the same coefficients across observations, but the linear predictor and conditional mean can differ because the covariate values differ. Unless stated otherwise, the model above has no interaction terms: each predictor enters additively on the log-mean scale.

What a Coefficient Holds Fixed

The phrase “holding the other predictors fixed” means comparing two covariate profiles that agree on every predictor except the one under discussion. It is not a claim that the predictors are statistically independent. It specifies the model-based comparison used to interpret a coefficient.

Theorem: In the model $\log(\mu)=\beta_0+\sum_{j=1}^{p}\beta_jx_j$, changing predictor $x_j$ by $\delta$ while keeping all other predictors fixed changes the log conditional mean by $\beta_j\delta$.

Proof. Let the original profile have predictor values $x_1,\ldots,x_p$, and let its conditional mean be $\mu$. At the changed profile, only $x_j$ is replaced by $x_j+\delta$. The two linear predictors are

\log(\mu)=\beta_0+\sum_{k=1}^{p}\beta_kx_k,\qquad \log(\mu_{\mathrm{new}})=\beta_0+\sum_{k\ne j}\beta_kx_k+\beta_j(x_j+\delta).

Subtracting the first expression from the second cancels the intercept and every unchanged predictor term, leaving

\log(\mu_{\mathrm{new}})-\log(\mu)=\beta_j\delta.

Thus the coefficient is the conditional slope of the log mean with respect to $x_j$; a one-unit increase corresponds to a change of $\beta_j$ in the log mean, with all other modeled predictors fixed. $\square$

For a continuous predictor, the corresponding effect on the mean itself depends on the current mean. Differentiating $\mu=\exp(\beta_0+\sum_j\beta_jx_j)$ with respect to $x_j$ gives

\frac{\partial \mu}{\partial x_j}=\beta_j\mu.

This follows directly from the chain rule: the derivative of the exponential is the exponential, and the derivative of the linear predictor with respect to $x_j$ is $\beta_j$. Consequently, a fixed coefficient represents a constant slope on the log-mean scale, while the change in the mean scale depends on the covariate profile.

Worked Example: Three Continuous Predictors

Suppose the number of clinic visits $Y$ is modeled using age in years, a symptom-severity score, and travel distance in kilometers:

\log\{E[Y\mid x]\} =1.20+0.015(\text{age}-50)+0.20(\text{severity})-0.03(\text{distance}).

For a patient aged 50 with severity score 3 and travel distance 10 km, the age term is $0.015(50-50)=0$. The linear predictor is

\eta=1.20+0+0.20(3)-0.03(10) =1.20+0.60-0.30=1.50.

The conditional mean is therefore $\mu=\exp(1.50)\approx 4.482$ visits. This calculation uses all three predictors to obtain the mean for the specified profile.

Now increase severity from 3 to 4 while keeping age and distance fixed. The linear predictor increases by $0.20$, from $1.50$ to $1.70$. The new conditional mean is $\exp(1.70)\approx 5.474$ visits. The severity coefficient describes this conditional change on the log-mean scale; it does not say that every patient’s expected count increases by a fixed number of visits.

At the original profile, the derivative of the mean with respect to severity is $0.20\mu$, or $0.20(4.482)\approx 0.896$ visits per severity-score unit. The derivative is evaluated at that profile, so it need not be the same at another profile with a different conditional mean.

Representing Categorical Predictors

A categorical predictor with more than two categories is usually represented by indicator variables. With $K$ categories, choose one as the reference category and use $K-1$ indicators. Each indicator equals 1 when an observation belongs to its category and 0 otherwise. The reference category has all $K-1$ indicators equal to 0. This coding lets the model compare each represented category with the reference while holding the other included predictors fixed.

Worked Example: Treatment and Sex Indicators

Suppose a count outcome is modeled using treatment status, sex, and age centered at 50 years. Let $\text{treated}=1$ for treated participants and 0 otherwise; let $\text{female}=1$ for female participants and 0 otherwise. Consider the fitted model

\log(\mu)=0.80-0.40(\text{treated})+0.25(\text{female})+0.02(\text{age}-50).

For an untreated male aged 50, all three predictor contributions are zero, so $\eta=0.80$ and $\mu=\exp(0.80)\approx 2.226$.

For a treated female aged 50, the linear predictor is

\eta=0.80-0.40(1)+0.25(1)+0.02(50-50)=0.65,

giving $\mu=\exp(0.65)\approx 1.916$. This mean combines the treatment, sex, and age contributions. To isolate the treatment coefficient, compare treated and untreated profiles with the same sex and age: the treatment indicator changes from 0 to 1, so the log mean changes by $-0.40$, with those other predictors fixed. The model does not include a treatment-by-sex interaction, so it assigns the same treatment coefficient to both sex categories.

Worked Example: A Three-Level Site Predictor

Suppose a model for a count response includes site, with categories A, B, and C, and a continuous severity score. Use A as the reference category. Define $I_B=1$ for site B and 0 otherwise, and $I_C=1$ for site C and 0 otherwise. Suppose

\log(\mu)=0.50+0.30I_B-0.20I_C+0.10(\text{severity}).

For a site A observation with severity 2, both indicators are zero. Thus $\eta=0.50+0.10(2)=0.70$, and $\mu=\exp(0.70)\approx 2.014$.

For a site B observation with the same severity, $I_B=1$ and $I_C=0$. Then $\eta=0.50+0.30+0.20=1.00$, giving $\mu=\exp(1.00)\approx 2.718$. The coefficient $0.30$ is the conditional difference in log mean for B versus A at the same severity. For site C at severity 2, $\eta=0.50-0.20+0.20=0.50$, so $\mu=\exp(0.50)\approx 1.649$. Its coefficient $-0.20$ gives the conditional log-mean difference for C versus A at the same severity.

The category indicators are not separate observations of an underlying numerical scale: site B is not “one unit more” than site A. The coefficients are comparisons with the chosen reference category, and changing the reference changes the coefficient representation.

Conditional Effects and Model Scope

The model’s conditional interpretation is useful because it makes explicit which measured predictors are included in a comparison. If age and severity both appear in the model, for example, the age coefficient describes a change in the modeled log mean at fixed severity and fixed values of any other included predictors. It is not the same as a simple unadjusted association between age and the observed counts.

Conditional interpretation also has limits. It does not establish that changing a predictor would cause the mean to change: causal conclusions require additional assumptions about study design and confounding. Nor does including several predictors guarantee that each coefficient can be estimated precisely. If predictors are redundant or nearly redundant in the observed data, it may be difficult or impossible to separate their contributions. A model-based comparison may also involve combinations of covariates that are rare or absent in the data, making that interpretation rely heavily on the assumed model form.

Finally, additivity of the linear predictor is a substantive modeling choice. In the examples above, the effect of one predictor on the log mean does not vary with the values of the others, because no interaction terms are included. If the scientific question requires that one predictor’s conditional effect vary across another predictor, the linear predictor must represent that variation explicitly. A coefficient should always be interpreted according to the terms actually present in the model.

Common Mistake

Common mistake: Reading $\beta_j$ as an unconditional association, or as the change in the mean count per one-unit increase, without checking the link and the other predictors. In a Poisson log-mean model, $\beta_j$ is the conditional change in the log mean for a one-unit predictor change, holding the other included predictors fixed. The change in the mean itself depends on the current profile, as $\partial\mu/\partial x_j=\beta_j\mu$ for a continuous predictor. This model-based conditional statement alone is not a causal claim.

Check Your Understanding

Use the model structure and conditional interpretations developed here to answer the following questions.

In a log-mean model with predictors age, severity, and distance, what does the severity coefficient describe when age and distance are held fixed?
A predictor increases by two units in a model with coefficient $\beta_j$. What is the resulting change in the linear predictor, with all other predictors fixed?
For a categorical predictor with four categories, how many indicators are needed when one category is the reference?
Why does a model without a treatment-by-sex interaction assign the same treatment coefficient across sex categories?
Does a conditional Poisson regression coefficient, by itself, establish a causal effect? Explain briefly.