Pitfall: Treating Conditional Coefficients As Marginal Effects

Why a Fixed-Effect Coefficient Can Be Mislabelled

A generalized linear mixed model (GLMM) describes responses conditional on random effects. Its fixed-effect coefficients therefore have a conditional interpretation, even when the scientific question concerns an average response across the random-effects distribution. The two interpretations can coincide in special cases, but they are not interchangeable by default.

The distinction matters especially for nonlinear links. In a logistic GLMM, for example, a treatment coefficient describes a log-odds difference at fixed random effects. It is not automatically a difference in population-averaged response probabilities, nor is its exponent automatically a marginal odds ratio. To answer a response-scale question, calculate response means after integrating over the random effects, then contrast those means.

Definition: Suppose \(Y_{ij}\) is the response for unit \(j\) in cluster \(i\), \(b_i\) is the cluster's random-effect vector, and \(x_{ij}\) is its covariate vector. A GLMM specifies a conditional mean \(\mu_{ij}(b_i)=E(Y_{ij}\mid b_i,x_{ij})\) through \(g\{\mu_{ij}(b_i)\}=x_{ij}^{T}\beta+z_{ij}^{T}b_i\). The fixed-effect coefficient \(\beta_k\) describes the change in the conditional linear predictor for a one-unit increase in covariate \(x_k\), holding the random effects and other covariates fixed. A population-averaged response mean at covariates \(x\) is \(m(x)=E_b\{\mu(x,b)\}=\int\mu(x,b)\,dF(b)\), where \(F\) is the specified random-effects distribution. A response-scale contrast compares such integrated means, for example \(m(x_1)-m(x_0)\).

This definition presumes the model's conditional mean and random-effects distribution are specified, and that the integral exists. In the examples below, the distribution is stated explicitly, so the averaging operation is fully determined. For a binary response, the response mean is a probability; for a count response, it is an expected count or rate, according to the model. The contrast's units follow the response scale.

Conditional Coefficients and Integrated Contrasts

For a binary outcome with a logit link, the conditional model can be written \(P(Y=1\mid x,b)=\operatorname{logit}^{-1}(x^T\beta+z^Tb)\). If \(x_k\) is a binary treatment indicator with no treatment interaction, then \(\beta_k\) is the conditional log-odds difference between treatment and control at the same \(b\) and covariate values. The conditional odds ratio is \(\exp(\beta_k)\). The marginal probabilities instead are integrals of the two conditional probabilities:

m(a)=\int \operatorname{logit}^{-1}\{\eta(a,b)\}\,dF(b), \qquad \Delta=m(1)-m(0).

Here \(a\) denotes treatment status, and \(\eta(a,b)\) is the model's linear predictor at that status and random effect. In general, \(\Delta\) is not \(\beta_k\), and it is not obtained by applying the inverse link to an average linear predictor. The response is transformed first at each random-effect value; those conditional responses are then averaged.

Worked Example: A Logistic Random-Intercept Model

Consider a hypothetical binary-outcome model with treatment coefficient \(\beta_T=1\), no other covariates, and random intercept \(b\) taking values \(-1\) and \(1\) with probability \(0.5\) each. The conditional probability is \(\operatorname{logit}^{-1}(\beta_T T+b)\). Find the conditional odds ratio and the population-averaged probability difference.

Conditional interpretation. At any fixed \(b\), treatment increases the log odds by 1. Thus the conditional odds ratio is \(\exp(1)=2.718\). This is an odds ratio comparing treatment and control at the same random-intercept value.

Integrate the control probability. For \(T=0\), the two conditional probabilities are \(\operatorname{logit}^{-1}(-1)=0.2689\) and \(\operatorname{logit}^{-1}(1)=0.7311\). Therefore \(m(0)=0.5(0.2689+0.7311)=0.5000\).

Integrate the treatment probability. For \(T=1\), the two linear predictors are 0 and 2. The probabilities are \(\operatorname{logit}^{-1}(0)=0.5000\) and \(\operatorname{logit}^{-1}(2)=0.8808\). Therefore \(m(1)=0.5(0.5000+0.8808)=0.6904\).

The marginal probability difference is \(0.6904-0.5000=0.1904\), or 19.04 percentage points. It is not the coefficient 1, which is on the log-odds scale, and it is not the conditional odds ratio 2.718. The integrated probabilities imply a marginal odds ratio of \(\{0.6904/(1-0.6904)\}/\{0.5000/(1-0.5000)\}=2.230\), also different from 2.718.

Why the Scales Differ

The key mathematical issue is that averaging and applying a nonlinear inverse link do not generally commute. Even in a model with only a random intercept, the conditional effect is evaluated at each \(b\), while the population-averaged effect compares two averages over \(b\). With random slopes, the treatment or covariate effect itself varies across clusters, so integration reflects both the response curve's nonlinearity and that variation.

For a differentiable inverse link \(h=g^{-1}\), consider a continuous covariate \(x_k\) whose only contribution to the linear predictor is \(\beta_k x_k\). Then \(m(x)=E_b[h(\eta(x,b))]\). If differentiation can be interchanged with integration, the marginal response derivative is \(E_b[h'(\eta(x,b))\beta_k]\). This expression shows why \(\beta_k\) alone is not the marginal response derivative: the inverse-link slope \(h'(\eta)\) varies with the random effect and covariate values.

For the logit link, \(h'(\eta)=h(\eta)\{1-h(\eta)\}\), so the derivative is \(\beta_k E_b[p(x,b)\{1-p(x,b)\}]\), where \(p(x,b)=h(\eta(x,b))\). This derivation assumes the random-effects distribution does not itself change with \(x_k\), the predictor has the stated form, and the derivative is integrable. If the model includes a random slope for \(x_k\), the conditional derivative also includes that random slope; the marginal derivative is the expectation of the full conditional derivative.

Worked Example: A Random Slope Changes the Average Probability Difference

Suppose a hypothetical logistic model has no random intercept, fixed treatment coefficient \(0.8\), and a random treatment slope \(b_T\) equal to \(-0.8\) or \(0.8\) with equal probabilities. The linear predictor is \((0.8+b_T)T\). Calculate the conditional probability difference at \(b_T=0\) and the integrated probability difference.

At \(b_T=0\), the control probability is \(\operatorname{logit}^{-1}(0)=0.5000\). The treatment probability is \(\operatorname{logit}^{-1}(0.8)=0.6900\). The conditional probability difference at this random-slope value is \(0.6900-0.5000=0.1900\).

For the integrated contrast, the control probability is 0.5000 for either random slope. Under treatment, the linear predictors are \(0.8-0.8=0\) and \(0.8+0.8=1.6\), giving probabilities 0.5000 and 0.8320. Thus \(m(1)=0.5(0.5000+0.8320)=0.6660\), \(m(0)=0.5000\), and \(m(1)-m(0)=0.1660\), or 16.60 percentage points.

The fixed coefficient 0.8 is the treatment effect on the conditional log-odds scale when the random slope is zero. The integrated probability contrast averages over the stated slope distribution and is a different estimand. Neither the conditional probability difference at \(b_T=0\) nor \(\beta_T\) can replace that integration.

When a Coefficient Does Equal a Marginal Effect

Avoid the opposite overstatement: conditional and marginal effects do not always differ. Equality depends on the model, the scale, and the contrast. For example, consider a Poisson log-link model with a random intercept \(b\), a treatment coefficient \(\beta_T\), and a common random-intercept distribution under both treatment conditions: \(E(Y\mid T,b)=\exp(\beta_0+\beta_T T+b)\). Then \(m(T)=\exp(\beta_0+\beta_T T)E\{\exp(b)\}\). Provided \(E\{\exp(b)\}\) is finite and the same distribution applies in both conditions, the marginal rate ratio is \(m(1)/m(0)=\exp(\beta_T)\). Thus \(\beta_T\) is the log of the marginal rate ratio in this special case. It still is not the marginal rate difference \(m(1)-m(0)\).

Worked Example: A Special Case for a Log-Link Model

Suppose a hypothetical Poisson random-intercept model has conditional mean \(E(Y\mid T,b)=\exp(\beta_0+\beta_T T+b)\). Let \(\exp(\beta_0)=2\), \(\beta_T=\log(1.5)=0.4055\), and let \(b\) equal \(\log(2)\) or \(-\log(2)\) with equal probabilities. Find both integrated means and compare the rate ratio with the rate difference.

The random-intercept multiplier has expectation \(E\{\exp(b)\}=0.5(2+0.5)=1.25\). Thus \(m(0)=2(1.25)=2.500\). Under treatment, \(\exp(\beta_0+\beta_T)=2(1.5)=3\), so \(m(1)=3(1.25)=3.750\).

The marginal rate ratio is \(3.750/2.500=1.500=\exp(\beta_T)\). The marginal rate difference is \(3.750-2.500=1.250\) expected events per unit of exposure, if the model's response is a rate per unit exposure. The coefficient 0.4055 is therefore interpretable as a log rate ratio here, but not as a rate difference. This equality relies on the log link, random-intercept structure, and common random-effects distribution across the two conditions.

A Reliable Interpretation Workflow

Before reporting a coefficient as an effect, identify its scale and conditioning. Then decide whether the scientific question asks for a conditional comparison at fixed random effects or an average response contrast over a random-effects distribution. The two are legitimate but distinct estimands. For a population-averaged response question, compute the integrated response means at the specified covariate settings and contrast them on the requested scale.

Name the coefficient's scale.
For a logit link, a fixed treatment coefficient is a conditional log-odds difference, holding random effects fixed. For a log link, it is a conditional log-rate or log-mean difference.

Specify the response-scale question.
For example, a probability difference between treatment and control, or a ratio of expected rates. State the covariate settings used for the comparison.

Integrate each prediction.
For each condition, evaluate the conditional mean across the random-effects distribution and average those response-scale values.

Contrast and report in context.
Subtract or divide the integrated means as required, then state the response scale and units. Do not relabel a conditional coefficient as that contrast.

Common Mistakes and Diagnostics

Calling \(\beta_T\) a probability difference. A logistic coefficient is on the log-odds scale. Calculate integrated probabilities and subtract them to report a population-averaged probability difference.
Exponentiating and calling the result a marginal odds ratio. In a logistic GLMM, \(\exp(\beta_T)\) is the conditional odds ratio at fixed random effects. A marginal odds ratio must be formed from the integrated probabilities if that is the target.
Assuming all conditional effects differ from marginal effects. Some models preserve a contrast on a particular scale. Verify the algebra and model assumptions, as in the log-link random-intercept example, rather than assuming either equality or inequality.
Ignoring random slopes or interactions. If the treatment effect varies with covariates or random slopes, a single fixed coefficient may describe only a conditional component. Integrate predictions under the full fitted linear predictor for each comparison condition.
Reporting a marginal contrast without checking the computation. Compare integrated predictions with conditional predictions at representative random-effect values, and verify that the reported contrast uses the response scale and units requested. Large discrepancies can be real; also check the integration method, model specification, and numerical accuracy.

The conditional GLMM and a population-averaged analysis answer different questions; neither is universally preferable. A conditional coefficient is useful when the target is a comparison within clusters at fixed random effects. An integrated contrast is appropriate when the target is an average response over the stated random-effects distribution. The main limitation is that integrated estimates depend on the fitted model and its distributional assumptions, so they should not be reported as assumption-free summaries.

Check Your Understanding

For each problem, identify the relevant scale and determine whether the fixed coefficient is itself the requested population-averaged response effect.

A logistic GLMM has treatment coefficient \(\log(2)\). What does this coefficient imply conditionally, and does it by itself give the marginal probability difference?
A log-link random-intercept model has treatment coefficient \(\log(1.2)\), with the same random-intercept distribution under both conditions and a finite exponential moment. What marginal contrast does exponentiating the coefficient give?
For a continuous covariate in a logit random-intercept model, explain why the marginal probability derivative is generally not \(\beta_k\).
A researcher reports an exponentiated logistic fixed effect as the population-averaged odds ratio. State the correction.

Solutions. (1) The conditional odds ratio is \(\exp\{\log(2)\}=2\) at fixed random effects; the coefficient alone does not give the marginal probability difference, which requires integrating the two response probabilities and subtracting them. (2) The marginal rate ratio is \(\exp\{\log(1.2)\}=1.2\) under the stated model conditions; the marginal rate difference is not given by this coefficient. (3) The marginal derivative is \(\beta_k E_b[p(x,b)\{1-p(x,b)\}]\), assuming the random-effects distribution is unchanged with the covariate. The expectation factor generally is not 1, so the derivative is not generally \(\beta_k\). (4) Describe the exponentiated coefficient as the conditional odds ratio. To report a population-averaged odds ratio, first integrate the predicted probabilities under each condition and form the odds ratio from those integrated probabilities.