Probit Links and Latent-Variable Interpretation

From a Latent Response to a Binary Biomarker

A binary biomarker records whether a measured condition is present, not how far a subject lies above or below a biological boundary. A probit model gives a useful way to represent the probability of that condition: it supposes that the binary observation is obtained by thresholding an underlying continuous response. That latent response is a modeling device; it is not an additional measurement in the data.

In a longitudinal setting, let \(Y_{ij}\) be the binary biomarker for subject \(i\) at visit \(j\), coded 1 for presence and 0 for absence. Let \(\mathbf{x}_{ij}\) be the fixed-effect covariate vector and \(\mathbf{z}_{ij}\) the random-effect design vector. A probit mixed model specifies the probability of a positive response conditional on subject-specific random effects.

The conditional probability is subject-specific because it conditions on \(\mathbf{b}_i\). The model also requires a distribution for those effects, commonly \(\mathbf{b}_i\sim N(\mathbf{0},\mathbf{D})\), where \(\mathbf{D}\) is a covariance matrix. In a joint model, the same random effects may enter the time-to-event submodel as well. The interpretation here concerns the longitudinal binary response, conditional on those effects.

Deriving the Probit Link

Introduce a latent response \(L_{ij}\) and define the recorded biomarker by whether that response exceeds a threshold. For the probit model, the threshold is set to zero and the latent residual has a standard normal distribution.

Assume the latent residuals are independent across visits conditional on the random effects and covariates, and independent of the random effects. The event \(Y_{ij}=1\) is then the event that \(\varepsilon_{ij}>-\eta_{ij}\). Symmetry of the standard normal distribution turns this tail probability into the normal cumulative distribution function evaluated at \(\eta_{ij}\).

The unit residual variance is an identification convention, not a claim that a physical latent measurement has known variance. If the latent residual standard deviation were an unknown \(\sigma\), the probability would be \(\Phi(\eta_{ij}/\sigma)\). Multiplying both the latent mean and residual scale by the same positive constant would leave the thresholded observations unchanged. The binary data therefore identify coefficient-to-scale ratios, not an unrestricted latent scale and coefficients separately. Fixing the residual variance at 1 defines the scale on which probit coefficients are reported.

Interpreting Fixed Effects and Random Effects

A fixed-effect coefficient changes the conditional latent mean, measured in residual standard-deviation units. For a continuous covariate \(x\) with coefficient \(\beta_x\), a one-unit increase in \(x\), holding the other covariates and random effects fixed, changes \(\eta_{ij}\) by \(\beta_x\). It does not generally change the probability by a constant amount: the probability change depends on the starting value of \(\eta_{ij}\), because \(\Phi\) is nonlinear.

For a subject with random intercept \(b_{0i}\) and random slope \(b_{1i}\) for time \(t_{ij}\), the predictor might include \(b_{0i}+b_{1i}t_{ij}\). The random intercept shifts that subject’s latent response at every visit. The random slope changes the subject-specific latent trajectory per unit of time. These are deviations from the population fixed-effects trajectory, not direct changes in probability or evidence that one subject causes a different response.

Conditional and Population-Average Probabilities

A probability conditional on a particular random effect is not the same as the probability for a randomly selected subject. The latter averages over the random-effects distribution. Under normal random effects, that average has a convenient form. Write \(\mathbf{z}_{ij}\) for the random-effect design at a given visit, and suppose \(\mathbf{b}_i\sim N(\mathbf{0},\mathbf{D})\) independently of \(\varepsilon_{ij}\). Then \(\mathbf{z}_{ij}^{\mathsf T}\mathbf{b}_i+\varepsilon_{ij}\) is normal with variance \(\mathbf{z}_{ij}^{\mathsf T}\mathbf{D}\mathbf{z}_{ij}+1\). Consequently,

This is a marginal probability for a subject drawn from the random-effects distribution at that covariate pattern. It is not obtained by simply setting the random effects to zero: that gives the conditional probability \(\Phi(\mathbf{x}_{ij}^{\mathsf T}\boldsymbol{\beta})\). Averaging over random effects moves the probability toward 0.5 when the fixed predictor is nonzero. The amount of movement depends on the random-effect variance at that visit.

Assumptions, Diagnostics, and Alternatives

The latent derivation is useful only when its assumptions match the intended model. In this formulation, the binary response is Bernoulli conditional on the random effects and covariates; latent residuals have a standard normal distribution and are independent across visits conditional on those effects; and the random effects have mean zero and a specified covariance distribution, often multivariate normal. The fixed-effect predictor must also represent the relevant covariate and time pattern adequately. In a joint model, the longitudinal and event components may share random effects, but the probit link by itself does not specify the event process or justify that sharing.

Binary observations cannot directly reveal whether an individual latent residual was normal. Diagnostics therefore focus on implications that can be assessed from the observed data and fitted model:

• Compare observed positive-response frequencies with fitted probabilities across meaningful covariate and time groups. Persistent discrepancies can signal a poor mean trajectory, omitted predictors, or an unsuitable link.
• Check whether fitted probabilities are systematically too high or low, including near the extremes. Calibration problems can remain even when an overall fit summary looks acceptable.
• Inspect convergence, standard errors, and estimated random-effect variances and correlations. Boundary estimates, unstable estimates, or sensitivity to reasonable starting values can indicate weak information or an overly complex random-effects structure.
• Assess whether the conditional-independence structure is plausible after accounting for the modeled random effects. Remaining patterns in repeated responses may indicate that the model has not captured important within-subject structure.

A common mistake is to interpret \(\beta_x\) as a constant percentage-point change, or to use \(\Phi(\mathbf{x}^{\mathsf T}\boldsymbol{\beta})\) as the population-average probability when random effects are present. The correction is to state the conditioning explicitly: \(\beta_x\) is a latent-scale shift, a conditional probability uses specified random effects, and a population-average probability integrates over their distribution. Another pitfall is comparing probit coefficients numerically with logistic coefficients as if they shared a scale. They do not. Both links map predictors to probabilities, but their coefficient scales differ; estimates should be interpreted using their own link and model.

The logistic mixed model is a common alternative, with a logit link and conditional log-odds coefficients. The probit model instead has a direct thresholded-normal interpretation and coefficients on a latent standard-deviation scale. Neither link is universally preferable: the choice should reflect the scientific interpretation, modeling context, and fit. The latent interpretation is not proof that a real, normally distributed biological quantity exists, and estimates can be sensitive to the assumed random-effects distribution and covariance structure. The introductory material on binary biomarkers with logistic mixed models covers the related mixed-model setting; the distinction to retain here is the probit scale and its latent-normal derivation.