Probit Coefficients on the Latent Index

Why Probit Coefficients Need Careful Interpretation

A probit model describes the conditional probability of a binary outcome, but its coefficients are not expressed directly in probability units. Each coefficient shifts a latent index, and the model maps that index to a probability through the standard normal cumulative distribution function. This distinction matters: a coefficient of 0.4 does not mean that the probability of the outcome rises by 40 percentage points.

The interpretation is still useful. Holding other regressors fixed, a probit slope tells us how much the normalized latent index changes when a regressor changes by one unit. Since the normal cumulative distribution function is increasing, the coefficient’s sign also tells us the direction in which the conditional outcome probability moves. The size of that probability change, however, depends on where the index starts.

The Normalized Latent-Index Model

Let \(Y_i\) be a binary outcome for observation \(i\), and let \(x_i\) be a vector of observed regressors, including an intercept if the model has one. The probit model can be represented using a latent variable \(Y_i^*\):

The latent variable is not directly observed. We observe only whether it crosses the threshold zero. The disturbance variance is fixed at one: this is the probit model’s normalization, which sets the scale used to express the coefficients. Define the normalized latent index as \(\eta_i=x_i'\beta\). The model’s conditional probability of outcome one is

where \(\Phi\) is the standard normal cumulative distribution function. The probability of outcome zero is \(1-\Phi(\eta_i)\). These expressions follow from the threshold rule: \(Y_i=1\) exactly when \(\varepsilon_i>-x_i'\beta\), and the standard normal distribution is symmetric around zero.

What a Slope Coefficient Says

Write the index as \(\eta=\beta_0+\beta_1x_1+\cdots+\beta_kx_k\), where \(\beta_0\) is the intercept and the other coefficients correspond to regressors. If \(x_j\) increases by one unit while the remaining regressors stay fixed, the index changes by \(\beta_j\). More generally, for a change of \(c\) units in \(x_j\), the index changes by \(c\beta_j\), provided the model remains linear in that regressor.

The word normalized is important. The latent disturbance variance is fixed at one, so coefficients are measured on the scale implied by that normalization. They describe movements in the model’s normalized index, not directly in the observed binary outcome or in probability points. A coefficient’s numerical value also depends on the units in which its regressor is recorded.

Why the Sign Determines the Direction

Although a coefficient is not a probability change, its sign has a clear directional interpretation in the standard probit model. The function \(\Phi\) is strictly increasing: for any \(a>b\), \(\Phi(a)>\Phi(b)\). Increasing a regressor \(x_j\) by one unit changes the index by \(\beta_j\), with other regressors fixed. If \(\beta_j>0\), the new index is higher and therefore the conditional probability of \(Y=1\) is higher. If \(\beta_j<0\), the new index is lower and the conditional probability is lower. If \(\beta_j=0\), changing that regressor alone leaves the index and fitted conditional probability unchanged in this specification.

The argument can be written directly. Let the original index be \(\eta\), and change \(x_j\) by \(c>0\). The new index is \(\eta+c\beta_j\). Therefore,

When \(\beta_j>0\), \(c\beta_j>0\), so \(\eta+c\beta_j>\eta\), and strict monotonicity gives a positive difference. When \(\beta_j<0\), the inequalities reverse and the difference is negative. This establishes the direction without treating \(\beta_j\) as a probability change.

Binary Regressors and Units of Measurement

The index interpretation also applies when a regressor is binary. A change from zero to one changes the index by its coefficient, holding the other regressors fixed. If that coefficient is positive, the model assigns the \(x_j=1\) profile a higher conditional probability of \(Y=1\) than the otherwise identical \(x_j=0\) profile. The coefficient itself is still an index difference, not the probability difference between those profiles.

For a continuous regressor, the choice of measurement units changes the reported coefficient while leaving the model’s fitted index unchanged when the coefficient is converted consistently. For example, an effect reported per hour will have a different numerical coefficient from the same relationship reported per minute. Always state the regressor’s units before interpreting the coefficient.

Scope and Common Pitfalls

A probit coefficient supports a conditional model interpretation: it describes how the index changes when one regressor changes and the others are held fixed. It does not establish that changing the regressor would cause the outcome to change. A causal interpretation requires an appropriate research design and assumptions beyond the probit link itself.

• Wrong: “The coefficient is 0.40, so the probability rises by 40 percentage points.” Correction: The index rises by 0.40 for a one-unit increase, holding other regressors fixed. The probability is \(\Phi(\eta)\), so its change depends on the starting index.
• Wrong: “A positive coefficient means the probability rises by the same amount for everyone.” Correction: A positive coefficient gives the same index shift for the same regressor change, but the probability difference can vary with the covariate profile.
• Wrong: “The coefficient is interpretable without knowing the regressor’s units.” Correction: State the units. A coefficient per minute and a coefficient per hour can differ numerically even when they encode the same relationship.
• Wrong: “A positive coefficient proves that increasing the regressor causes the outcome.” Correction: The sign describes the fitted conditional relationship; causation requires additional assumptions or design.