Interpreting the Logistic Intercept

What the Logistic Intercept Refers To

A logistic regression intercept is not automatically the outcome probability for a typical person, nor is it necessarily the probability for anyone in the study. Its meaning depends on the covariate values at which the model is evaluated. In the usual parameterization, the intercept describes the log odds when every modeled covariate equals zero. Interpreting it well therefore requires identifying what “zero” means for the variables in the model.

Let \(Y\) be a binary outcome, coded so that \(Y=1\) is the event of interest. For a covariate vector \(x=(x_1,\ldots,x_p)^\mathsf{T}\), write the logistic regression mean model as

Here \(\beta_0\) is the intercept, and \(\beta_j\) is the coefficient multiplying covariate \(x_j\). This tutorial concerns the meaning of \(\beta_0\) and its estimate, not the interpretation of the individual slope coefficients.

From Log Odds to Baseline Odds and Probability

The intercept is on the log-odds scale. To express it on the odds scale, exponentiate it. To express it as a probability, apply the inverse-logit function. These conversions are exact consequences of the definition of log odds.

To verify the conversion, start with \(\log\{\pi_0/(1-\pi_0)\}=\beta_0\). Exponentiating both sides gives \(\pi_0/(1-\pi_0)=\exp(\beta_0)\). Multiplying by \(1-\pi_0\) and collecting the terms involving \(\pi_0\) yields \(\pi_0\{1+\exp(\beta_0)\}=\exp(\beta_0)\), which gives the stated probability. The baseline probability is strictly between zero and one for every finite \(\beta_0\).

When a fitted model supplies an estimated intercept \(\hat{\beta}_0\), the corresponding plug-in estimates are \(\exp(\hat{\beta}_0)\) for baseline odds and \(1/\{1+\exp(-\hat{\beta}_0)\}\) for baseline probability. These are model-based quantities at the zero-covariate profile; they are not necessarily the sample event proportion.

The Meaning of Zero Depends on the Covariates

The mathematical definition is simple, but its substantive interpretation depends on how the predictors are recorded. If age is measured in years, age zero means birth. If a biomarker is recorded in its original units, zero may be outside the range observed in the study. If a binary indicator is coded zero for “not exposed” and one for “exposed,” zero refers to the unexposed category. With several predictors, the intercept refers to the profile in which all predictors simultaneously take their zero values.

That profile need not occur in the data. A model can still define a fitted probability there, provided the linear predictor is defined, but an estimate at a profile far outside the observed covariate combinations may be an extrapolation. It is important to distinguish “the model’s value at zero” from “the average outcome in the study” or “the outcome for a representative participant.”

Centering Predictors Changes the Reference Profile

Centering replaces each predictor by its deviation from a selected value. For example, if \(A\) is age and \(c\) is a chosen reference age, define \(A^*=A-c\). Then \(A^*=0\) means age \(c\), rather than age zero. The model’s fitted probabilities need not change: centering changes the coordinates used to describe the same linear predictor.

Proof. Since \(z=x-c\), we have \(x=z+c\). Substitute this identity into the original linear predictor: \(\eta(x)=\beta_0+\beta^\mathsf{T}(z+c)=\beta_0+\beta^\mathsf{T}z+\beta^\mathsf{T}c=(\beta_0+\beta^\mathsf{T}c)+\beta^\mathsf{T}z=\alpha_0+\beta^\mathsf{T}z\). This equality holds for every \(x\). Applying the inverse-logit to equal linear predictors gives equal probabilities. At \(x=c\), \(z=0\), so the centered predictor is \(\eta(c)=\alpha_0\); by the logistic model, this is the log odds at that profile. \(\square\)

The same algebra applies to estimated coefficients: if the original fitted coefficients are \(\hat{\beta}_0,\hat{\beta}\), then the corresponding centered intercept is \(\hat{\alpha}_0=\hat{\beta}_0+\hat{\beta}^\mathsf{T}c\). The centered intercept estimates log odds at the chosen reference profile. If the same likelihood model is fitted with this one-to-one reparameterization, each parameter setting in one form corresponds to a parameter setting in the other, with identical probabilities for every observation. Thus centering does not, by itself, alter the model’s fitted probabilities.

Choosing and Reporting a Reference Profile

A reference value should be selected to make the intercept interpretable for the scientific question. For a continuous predictor, a meaningful value may be a clinically relevant measurement or a typical value in the study population. For a collection of predictors, the reference profile should be described explicitly: centering age at 50 while leaving a treatment indicator at zero makes the intercept refer to age 50 in the indicator’s zero-coded category.

Centering at a sample mean can make the zero value correspond to an average predictor value, but it does not make the intercept an average predicted risk. In a model with multiple predictors, the vector of means may not describe an actual individual; moreover, the inverse-logit is nonlinear, so the predicted probability at the mean covariate vector need not equal the average of the individual predicted probabilities.

When reporting an intercept, state its scale and reference profile. For example: “The estimated intercept is \(-1.00\), corresponding to estimated event odds of \(0.368\) and estimated event probability of \(0.269\) for a 50-year-old in the zero-coded treatment category.” This avoids implying that the intercept represents the whole sample.

Common Mistakes and Why the Interpretation Matters

• Calling the intercept a probability. The intercept is a log odds. Exponentiating gives odds; applying the inverse-logit gives probability.
• Calling odds a probability. Odds of \(1/9\) correspond to probability \(1/10\), not \(1/9\). Convert using \(o/(1+o)\).
• Ignoring the zero values. In a multiple-predictor model, the intercept sets every predictor to zero at once. State what each zero represents.
• Calling the intercept the risk for a typical person. Zero values may be unusual or impossible in the study population. The intercept remains a model quantity at that profile, not automatically a sample summary.
• Assuming centering changes predictions. Centering shifts the intercept to the chosen reference profile; with the corresponding coefficient transformation, the linear predictor and fitted probabilities remain the same.

The intercept matters because it anchors the model’s predicted probabilities to a specific covariate profile. A poorly chosen zero can make the intercept scientifically unhelpful even when the rest of the model is useful. Centering offers a transparent way to select a more interpretable baseline, provided the new reference values are reported clearly.