Exact Probit Marginalization With Gaussian Random Intercepts

Why Probit Marginalization Has an Exact Form

A probit generalized linear mixed model describes an outcome conditional on a cluster-specific random effect. The resulting probability answers a conditional question: what is the probability of the outcome for a cluster with a specified random-intercept value? A population-averaged probability instead averages over the distribution of random intercepts. For a Gaussian random intercept and a probit link, this averaging can be done exactly: it changes the scale of the linear predictor by a known factor.

This identity is useful when a fitted conditional model is used to report population-averaged predictions or to compare conditional and marginal coefficients. It does not make the two estimands interchangeable. The conversion depends on the model assumptions, and it does not remove the dependence among observations in the same cluster.

The usual binary GLMM formulation also assumes that outcomes are conditionally independent given the random effects and covariates. That assumption describes the joint distribution within a cluster; the marginalization identity below concerns a single observation's mean. Even when conditional independence holds, observations sharing \(b_i\) are generally dependent after \(b_i\) is integrated out.

Deriving the Marginal Mean

Write \(\eta_{ij}=\mathbf{x}_{ij}^{T}\boldsymbol{\beta}\) for the fixed part of the linear predictor. The marginal mean is the expectation of the conditional probability over \(b_i\). To evaluate it, introduce an independent standard normal variable \(Z\). Since \(\Phi(a)=P(Z\leq a)\), the conditional probability can be written as \(P(Z\leq \eta_{ij}+b_i\mid b_i)\). Averaging over \(b_i\) gives:

The final step follows because \(Z\) and \(b_i\) are independent normal variables, so \(Z-b_i\sim N(0,1+\tau^2)\). Standardizing this normal variable yields the stated probability. This argument justifies the identity directly; it is not an approximation.

Let \(D=\sqrt{1+\tau^2}\). Then the marginal mean has a probit form with linear predictor \(\eta_{ij}/D\). If the conditional model includes an intercept, its marginal counterpart is divided by \(D\) as well as every slope. Thus, when the same covariates and a constant random-intercept variance apply throughout the population, the marginal coefficient vector on the probit scale is \(\boldsymbol{\beta}/D\). These are coefficients for the marginal mean, not new conditional coefficients.

This is a statement about the probit scale. A coefficient is not a change in probability by a fixed amount: the probability change depends on the starting linear predictor. For a continuous covariate \(x\) with conditional slope \(\beta_x\), the marginal probability derivative is \((\beta_x/D)\phi(\eta/D)\), where \(\phi\) is the standard normal density. It varies with the covariate values through \(\eta\).

Conditional Predictions and Population-Averaged Predictions

A conditional prediction for a specified cluster uses its random-intercept value: \(\Phi(\eta+b_i)\). For a new cluster drawn from the population, the appropriate prediction averages over the random-intercept distribution and equals \(\Phi(\eta/D)\). Setting \(b_i=0\) instead gives \(\Phi(\eta)\), the prediction for a cluster at the mean of the random-intercept distribution. Because \(\Phi\) is nonlinear, this is generally not equal to the population average.

Assumptions, Diagnostics, and Boundaries

The exact rescaling requires the specified conditional probit mean, a normally distributed random intercept with mean zero and variance \(\tau^2\), and independence between that random intercept and the covariates for the stated population averaging. The probit latent-error variance is fixed at one; that convention is what produces \(1+\tau^2\) in the denominator. The transformation also presumes the same random-intercept variance for the observations being compared.

In practice, the identity is exact for the model, but the model and its estimated parameters may be imperfect. Check whether the binary-outcome and probit mean structure are reasonable, whether there are important omitted predictors or cluster-level differences, and whether the estimated random-intercept variance is plausible. Compare predicted and observed outcome frequencies across relevant covariate ranges and clusters, and examine whether a few clusters dominate the estimated heterogeneity. Such checks assess model adequacy; they do not prove the normality assumption.

If uncertainty for a marginal coefficient or prediction is required, uncertainty in both \(\boldsymbol{\beta}\) and \(\tau^2\) must be reflected. Simply transforming a point estimate gives a point estimate, not a confidence interval. A delta-method calculation or a bootstrap that refits the model can propagate this uncertainty.

The result is specific to a Gaussian random intercept and a probit response. With a zero random-effect variance, \(D=1\) and the conditional and marginal means coincide. With a non-Gaussian random effect, a different response link, or a variance that changes across observations, this simple constant rescaling need not hold. In particular, random slopes require a different treatment; no random-slope formula is implied here.

A marginal model fitted directly, such as a probit GEE, targets a population-averaged mean through its own estimating approach. The identity provides a way to obtain that mean from the specified probit random-intercept model. It does not guarantee that a separately fitted marginal model has coefficients equal to the transformed GLMM estimates in finite samples, or when the marginal mean model and covariate specification differ. Nor does the marginal mean identity specify the full joint distribution of outcomes within clusters.