Probit Slopes on the Latent Scale

From a Probit Coefficient to a Latent-Scale Shift

A probit coefficient is easy to misread because it appears in a model for probabilities but is not measured in probability units. In a cumulative probit model, the coefficient describes how a covariate shifts a latent response relative to the model's fixed residual scale. A probability is obtained only after that shift is combined with a cutpoint and passed through the standard normal cumulative distribution function.

This distinction matters in interpretation. A coefficient of \(0.8\) does not mean that a cumulative probability rises by 0.8, or by 80 percentage points. It means that a one-unit increase in the covariate shifts the latent response by \(0.8\) residual standard deviations. The corresponding probability change depends on the cutpoint and the covariate value.

Definition: Use the cumulative probit convention \(C_j(x)=P(Y\leq j\mid x)=\Phi(\alpha_j-\beta x)\), where \(C_j(x)\) is the probability of being at or below cutpoint \(j\), \(\Phi\) is the standard normal cumulative distribution function, \(\alpha_j\) is the cutpoint, and \(\beta\) is the slope for covariate \(x\). Equivalently, write a latent response \(W=\beta x+\varepsilon\), with \(\varepsilon\sim N(0,1)\), and regard \(Y\leq j\) as the event \(W\leq\alpha_j\). Under this convention, \(\beta\) is the change in the conditional latent-response mean per one-unit increase in \(x\), in units of the latent residual standard deviation.

The sign convention is important. With \(C_j(x)=\Phi(\alpha_j-\beta x)\), a positive \(\beta\) shifts the latent response upward as \(x\) increases. It therefore makes being at or below a fixed cutpoint less probable. A negative \(\beta\) shifts the latent response downward and makes that cumulative event more probable. A model written with the opposite sign in its linear predictor will have coefficients with the opposite sign interpretation.

Why the Latent Scale Is Standardized

The latent response is not directly observed; only which side of each cutpoint it falls on is observed. Consequently, its scale cannot be determined separately from the cutpoint spacing and slope. The usual probit convention fixes the residual variance at one, so that latent shifts can be reported in residual-standard-deviation units.

Result: If instead the latent residual has distribution \(N(0,\sigma^2)\), with \(\sigma>0\), then \(P(Y\leq j\mid x)=\Phi((\alpha_j-\beta x)/\sigma)\). The cumulative probabilities therefore depend on the ratios \(\alpha_j/\sigma\) and \(\beta/\sigma\), not on the three quantities \(\alpha_j,\beta,\sigma\) separately.

To verify the result, let \(W=\beta x+\varepsilon\), where \(\varepsilon\sim N(0,\sigma^2)\), and let \(Y\leq j\) correspond to \(W\leq\alpha_j\). Then

\begin{aligned} P(Y\leq j\mid x) &=P(\beta x+\varepsilon\leq\alpha_j\mid x)\\ &=P\left(\frac{\varepsilon}{\sigma}\leq \frac{\alpha_j-\beta x}{\sigma}\,\middle|\,x\right)\\ &=\Phi\left(\frac{\alpha_j}{\sigma} -\frac{\beta}{\sigma}x\right). \end{aligned}

Here \(\varepsilon/\sigma\sim N(0,1)\), which gives the last equality. Multiplying \(\alpha_j\), \(\beta\), and \(\sigma\) by the same positive constant leaves both ratios unchanged and hence leaves every cumulative probability unchanged. Fixing \(\sigma=1\) sets the latent scale; it is what makes \(\beta\) interpretable as a shift in residual-standard-deviation units. \(\square\)

Worked Example: Interpreting a One-Unit Covariate Change

Suppose a cumulative probit model has \(\beta=0.8\), and compare covariate values \(x=0\) and \(x=1\). What is the latent-scale interpretation of this change?

In the latent representation, the conditional latent-response mean is \(\beta x\). At \(x=0\), it is \(0.8(0)=0\); at \(x=1\), it is \(0.8(1)=0.8\). Thus the one-unit increase shifts the latent response by \(0.8-0=0.8\) residual standard deviations. This conclusion concerns the latent scale, not a change of \(0.8\) in any cumulative probability.

For a two-unit increase, the shift would be \(\beta(2)=0.8(2)=1.6\) residual standard deviations. The same coefficient therefore describes a constant latent-scale shift per unit, while the associated probability changes need not be constant across covariate values or cutpoints.

Calculating a Fitted Cumulative Probability

To obtain a probability, substitute the covariate value and cutpoint into the cumulative probit equation, then evaluate the standard normal distribution function. The quantity inside \(\Phi\) is a standardized distance from the cutpoint. Its sign follows from the specified convention: the argument is \(\alpha_j-\beta x\).

Worked Example: Computing a Fitted Cumulative Probability

Suppose \(\alpha_j=0.7\), \(\beta=0.8\), and the covariate is \(x=1\). Calculate the fitted probability of being at or below cutpoint \(j\).

First compute the argument of the standard normal distribution function:

\alpha_j-\beta x=0.7-(0.8)(1)=-0.1.

Therefore the fitted cumulative probability is

C_j(1)=\Phi(-0.1)\approx 0.4601722.

The interpretation is that, at \(x=1\), the model assigns probability about \(0.4602\), or \(46.02\%\), to outcomes at or below cutpoint \(j\). The value is not the coefficient, nor is it obtained by subtracting the coefficient from a probability. It comes from applying \(\Phi\) to the model's linear predictor.

The sign can be checked against the latent interpretation. At \(x=1\), the latent mean is \(0.8\), above the cutpoint \(0.7\). Thus the event that the latent response is at or below that cutpoint has probability less than one half, consistent with \(\Phi(-0.1)\approx0.4602\).

Probability Effects Depend on Where They Are Evaluated

A fixed latent shift does not generally produce a fixed probability shift. The cumulative distribution function is nonlinear: equal changes to its argument can correspond to different changes in its value, depending on the starting argument. The following derivative makes that dependence explicit.

Result: For finite \(\alpha_j,\beta\) and real \(x\), the derivative of the fitted cumulative probability is \(\dfrac{dC_j(x)}{dx}=-\beta\phi(\alpha_j-\beta x)\), where \(\phi\) is the standard normal density. Since \(\phi(z)>0\) for every finite \(z\), the derivative has sign opposite to \(\beta\), and its magnitude depends on the cutpoint and covariate value.

To derive this, the standard normal cumulative distribution function is differentiable, with derivative \(\Phi'(z)=\phi(z)\). Apply the chain rule to \(C_j(x)=\Phi(\alpha_j-\beta x)\). The derivative of the inner argument with respect to \(x\) is \(-\beta\), so

\frac{dC_j(x)}{dx} =\Phi'(\alpha_j-\beta x)(-\beta) =-\beta\phi(\alpha_j-\beta x).

Because a normal density is strictly positive, this derivative is negative when \(\beta>0\), positive when \(\beta<0\), and zero when \(\beta=0\). Its magnitude is not just \(|\beta|\): it is \(|\beta|\) multiplied by the density at the current standardized distance. This proves both the direction and the location dependence of the local probability effect. \(\square\)

For a finite covariate change \(\Delta x\), the corresponding change is instead \(C_j(x+\Delta x)-C_j(x)\). It can be calculated by evaluating the two cumulative probabilities and subtracting, or expressed as the integral of the derivative over the interval. The derivative at one value of \(x\) is a local rate of change, not necessarily the finite change over an entire interval.

Worked Example: Comparing a Finite Probability Change with a Latent Shift

Continue with \(\alpha_j=0.7\) and \(\beta=0.8\). Compare \(x=0\) with \(x=1\), and calculate the change in the fitted cumulative probability.

At \(x=0\), the argument is \(0.7-(0.8)(0)=0.7\), so \(C_j(0)=\Phi(0.7)\approx0.7580363\). At \(x=1\), the argument is \(0.7-(0.8)(1)=-0.1\), so \(C_j(1)=\Phi(-0.1)\approx0.4601722\).

Using the unrounded standard normal probabilities, the finite change is

C_j(1)-C_j(0) =\Phi(-0.1)-\Phi(0.7) \approx -0.2978642 \approx -0.2979.

Thus the cumulative probability decreases by about 29.79 percentage points over this one-unit increase in \(x\). On the latent scale, the same covariate change shifts the response upward by exactly \(0.8\) residual standard deviations. These are different descriptions of the change: one is a latent-scale shift, and the other is a probability difference at this cutpoint and over this particular covariate interval.

Worked Example: A Negative Slope and the Direction of Change

Let \(\alpha_j=0\) and \(\beta=-0.5\). Calculate the cumulative probabilities at \(x=0\) and \(x=1\), and interpret the change.

At \(x=0\), the argument is \(0-(-0.5)(0)=0\), so \(C_j(0)=\Phi(0)=0.5000000\). At \(x=1\), the argument is \(0-(-0.5)(1)=0.5\), so \(C_j(1)=\Phi(0.5)\approx0.6914625\). The finite change is \(0.6914625-0.5000000=0.1914625\), an increase of about 19.15 percentage points.

Meanwhile, the latent mean changes from \((-0.5)(0)=0\) to \((-0.5)(1)=-0.5\), a shift of \(-0.5\) residual standard deviations. The downward latent shift makes being below the fixed cutpoint more likely, which agrees with the increase in the cumulative probability.

Common Mistake: Reading a Slope as a Probability Change

A common incorrect interpretation is: “The probit coefficient is \(0.8\), so a one-unit increase in \(x\) raises the probability by \(0.8\).” This treats a coefficient on the latent scale as if it were a constant probability difference. It is wrong both because the sign may be opposite for a lower-tail cumulative probability and because the normal cumulative distribution function is nonlinear.

The correction is to separate the two calculations. First interpret \(\beta\Delta x\) as the latent-response shift in residual-standard-deviation units. Then, if a probability effect is requested, evaluate the cumulative probability at each covariate value using \(\Phi(\alpha_j-\beta x)\), and subtract the two fitted probabilities. For a local effect, use \(-\beta\phi(\alpha_j-\beta x)\) at the specified covariate value and cutpoint.

For example, in the preceding finite-change calculation, the latent shift was \(0.8\) standard deviations, but the cumulative probability changed by approximately \(-0.2979\), not \(+0.8\). Neither number replaces the other: they answer different questions.

Key takeaway: In the cumulative probit convention \(C_j(x)=\Phi(\alpha_j-\beta x)\), the slope describes a constant shift of the standardized latent response per covariate unit. A cumulative probability is obtained by applying \(\Phi\); its local and finite changes depend on the cutpoint and covariate values.

Check Your Understanding

Answer each question using the convention and interpretation established above.

In a cumulative probit model with slope \(\beta=0.6\), what latent-scale shift corresponds to a three-unit increase in \(x\)?
If \(\alpha_j=0.4\), \(\beta=0.5\), and \(x=1\), what standard normal argument is used to calculate \(C_j(x)\)?
Under the stated sign convention, does a positive slope make a fixed lower-tail cumulative probability increase or decrease as \(x\) increases? Explain using the latent representation.
Why is the usual latent residual variance fixed at one when interpreting probit slopes?
Why does the coefficient alone not determine the finite change in a cumulative probability for a given covariate contrast?