Binary Predictors and Indicator Coding

Why Code a Binary Predictor?

A logistic regression can include a predictor that records a category rather than a numerical measurement. For a predictor with two categories, an indicator variable gives each observation a numerical code: zero for one category and one for the other. This lets the model express a direct comparison between the two groups while retaining the logistic model’s interpretation in terms of odds.

For example, a hypothetical insurer might study whether a policyholder files a claim during a year and record whether the policyholder selected an optional roadside feature. The feature is categorical, but it can be represented by an indicator. The resulting coefficient compares the conditional log-odds of a claim for policyholders with the feature against those without it. The comparison is conditional on any other predictors included in the model.

The zero and one labels are coding choices, not measurements of quantity. If \(Z=0\) means “feature not selected” and \(Z=1\) means “feature selected,” then the value 1 does not mean one unit more of the feature. It identifies a category. The interpretation of the fitted coefficient must follow the category definitions.

Indicator Coding in a Logistic Model

Let \(Y\) be a binary response, and let \(p=P(Y=1\mid Z,x)\), where \(x\) represents any other predictors held in the model. A logistic regression with one binary indicator and no interaction between \(Z\) and \(x\) can be written as

Here, \(\beta_0\) is the intercept, \(\beta_1\) is the indicator coefficient, and \(\beta_2\) is the coefficient of \(x\). If there are no other predictors, the \(\beta_2x\) term is omitted. The model assumes that, conditional on the included predictors, the response is Bernoulli with success probability \(p\), and that the log-odds are represented by the stated linear predictor. When observations are modeled as independent, their conditional likelihood factors into their Bernoulli likelihood contributions.

The key interpretation is obtained by calculating the fitted log-odds at each indicator value. For \(Z=0\), the indicator contributes zero. For \(Z=1\), it contributes \(\beta_1\). Thus, at the same value of \(x\), the category coded 1 has log-odds \(\beta_1\) higher than the category coded 0. Exponentiating this contrast gives the corresponding conditional odds ratio.

Why the Coefficient Is a Log-Odds Contrast

The result follows directly from the model, rather than from a special rule about indicator variables. Write the model’s linear predictor as \(\eta(Z,x)=\beta_0+\beta_1Z+\beta_2x\). Because the logit link equates the log-odds to the linear predictor, substituting the two possible values of \(Z\) gives

Here \(p_1=P(Y=1\mid Z=1,x)\) and \(p_0=P(Y=1\mid Z=0,x)\). Subtract the second equation from the first. The intercept and the contribution from \(x\) cancel, leaving

The left side is the difference between the two conditional log-odds. Exponentiating both sides shows that the ratio of the conditional odds is \(\exp(\beta_1)\). This derivation also shows why holding other predictors fixed matters: only then do their contributions cancel. If the model includes an interaction between \(Z\) and another predictor, the contrast can depend on that predictor; the simple \(\beta_1\) interpretation no longer applies on its own.

Adding Other Predictors

When the model also includes predictors, the indicator coefficient compares the two categories at the same values of those predictors. For example, in \(\operatorname{logit}(p)=\beta_0+\beta_1Z+\beta_2x\), the conditional odds ratio is \(\exp(\beta_1)\) at every fixed value of \(x\), provided there is no interaction term involving \(Z\). This is an odds comparison, not a claim that the groups have the same baseline probability at every value of \(x\).

The fitted probabilities still depend on the other predictors. Even when the odds ratio is constant, the probability difference between categories can change with \(x\), because the logit transformation maps odds to probabilities nonlinearly.

Obtaining the Contrast from Grouped Outcomes

For a model with an intercept and one binary predictor only, the two indicator values define the two groups. If both groups contain observed successes and failures, the fitted group probabilities equal the observed proportions in those groups. The indicator coefficient can then be found as the difference between the groups’ observed logits. This calculation is a special case of fitting a logistic model; it does not replace the general model-fitting procedure when other predictors are included.

What Happens If the Coding Is Reversed?

The model’s fitted probabilities do not depend on which category is assigned zero, as long as the intercept and indicator coefficient are changed consistently. If the original coding is \(Z=0\) for category A and \(Z=1\) for category B, reversing the coding makes the new indicator \(W=1-Z\). In a model with no interaction, substituting \(Z=1-W\) gives

Therefore, the reversed coding has intercept \(\beta_0+\beta_1\) and indicator coefficient \(-\beta_1\). The odds ratio for category A relative to category B is the reciprocal of the original odds ratio. Reversing the coding changes the direction of the comparison, not the underlying fitted comparison between the two groups.

Common Exam Traps

• Reading the coefficient as a probability change. A logistic coefficient is on the log-odds scale. Exponentiate it to obtain an odds ratio; calculate probabilities separately.
• Omitting the category direction. “The odds ratio is 2” is incomplete unless it is clear which category’s odds are compared with which other category’s odds.
• Comparing observations with different covariate values. The indicator coefficient is a conditional contrast. Hold other predictors fixed when interpreting it.
• Reversing the coding but keeping the coefficient unchanged. Reversing the categories changes the coefficient’s sign and reciprocates the odds ratio.
• Assuming a constant odds ratio means a constant probability difference. The odds ratio can be constant across values of another predictor while the probability difference varies.