Cumulative Probit Models and Latent Responses

From a Latent Response to Ordered Categories

An ordinal response records an order, but not necessarily equal distances between categories. A rating of “poor,” “fair,” or “good,” for example, tells us which response is higher without telling us that the gap from poor to fair equals the gap from fair to good. A latent-response model represents the ordering by supposing that a continuous, unobserved quantity lies behind the recorded category. The observed response indicates which interval that quantity falls in.

In the cumulative probit model, the latent response has a normal distribution with a mean that depends on predictors. Ordered cutpoints divide the latent scale into the intervals corresponding to the observed categories. Because the error distribution is standard normal, the probability of being below a cutpoint is given by the standard normal cumulative distribution function.

The cutpoints are shared across covariate values; the covariates enter through the single latent mean \(x^\mathsf{T}\beta\). Thus the model does not assign a separate slope to each category boundary. Its main probability statements follow directly from the threshold rule.

Deriving Cumulative and Category Probabilities

For a cutpoint \(\gamma_j\), the event \(Y\leq j\) is the event that the latent response falls at or below that cutpoint. Standardizing the normal latent response gives the cumulative probit equation.

Proof. By the threshold definition, \(Y\leq j\) if and only if \(Y^*\leq\gamma_j\). Conditional on \(x\), \(Y^*=x^\mathsf{T}\beta+\varepsilon\), so

The category event is \(\gamma_{k-1}<Y^*\leq\gamma_k\). Its probability is the probability of being at or below the upper cutpoint minus the probability of being at or below the lower cutpoint. Applying the cumulative formula, with the limiting CDF values at the infinite endpoints, gives the stated expression. Since the cutpoints are strictly ordered and \(\Phi\) is increasing, each difference is nonnegative. Summing over categories telescopes:

Thus these are valid category probabilities, and the cumulative probability at each finite cutpoint is the sum of the category probabilities from 1 through \(j\). \(\square\)

The sign in the cumulative equation is determined by the threshold event: \(Y\leq j\) means \(Y^*\leq\gamma_j\), so the predictor is subtracted from the cutpoint. The category probability is not itself a cumulative probability; it is the difference between two adjacent cumulative probabilities.

How a Common Slope Enters the Model

For each cutpoint, the same linear predictor \(x^\mathsf{T}\beta\) appears in the cumulative probability. A change in a covariate therefore shifts the latent mean while leaving the cutpoints fixed. This gives a common-slope structure across the cumulative probabilities. The cutpoints determine where the observed categories divide the latent scale; the shared coefficient vector determines how the covariates enter that scale.

Why the Latent Scale Must Be Fixed

The latent variable is a device for specifying probabilities, not a directly observed measurement. Its units cannot be determined from the categories alone. To see the issue, suppose the latent error initially has standard deviation \(\sigma>0\), rather than 1, and write the latent response as \(x^\mathsf{T}\beta+\varepsilon\), with \(\varepsilon\sim N(0,\sigma^2)\). Standardization gives

These probabilities depend on the cutpoints and coefficients only through their ratios to \(\sigma\). The following result states this scale equivalence precisely.

Proof. Substituting the rescaled quantities into the standardized argument yields

The argument of \(\Phi\) is identical at every cutpoint and covariate value, so all cumulative probabilities—and therefore all category probabilities—are identical. A scale convention is needed to specify the model; the cumulative probit convention sets \(\sigma=1\). \(\square\)

There is a related location convention. If a separate latent intercept \(\beta_0\) were included, the cumulative argument would be \(\gamma_j-\beta_0-x^\mathsf{T}\beta\). Defining new cutpoints \(\gamma_j'=\gamma_j-\beta_0\) gives the same probabilities without a separate intercept. The usual parameterization therefore puts the location information in the cutpoints rather than adding a distinct latent intercept.

Common Pitfall: Confusing Cumulative and Category Probabilities

A common mistake is to call \(\Phi(\gamma_j-x^\mathsf{T}\beta)\) the probability of category \(j\). It is instead the probability of being in category \(j\) or any lower category: \(P(Y\leq j\mid x)\). The probability of category \(j\) is the difference between the cumulative probability at its upper cutpoint and the one at its lower cutpoint.

A second sign error is to use \(\Phi(x^\mathsf{T}\beta-\gamma_j)\) for \(P(Y\leq j\mid x)\). The threshold event is \(x^\mathsf{T}\beta+\varepsilon\leq\gamma_j\); rearranging gives \(\varepsilon\leq\gamma_j-x^\mathsf{T}\beta\), not the reverse. A reliable check is that increasing the latent mean makes it less likely to fall below a fixed cutpoint.

The cumulative probit model is useful because a small set of ordered cutpoints and a common predictor specify a full probability distribution over the observed categories. The latent construction supplies a direct derivation, while differencing the cumulative probabilities turns that construction into probabilities for the categories actually recorded.