Indicator Variables for Two-Group Comparisons

Representing Two Groups With One Indicator

A multiple regression equation can include predictors that describe categories as well as numerical measurements. When a predictor has exactly two categories, one simple representation is a zero-one indicator: assign 0 to one category and 1 to the other. This lets the fitted equation give a separate prediction for each category while retaining a single coefficient for their modeled difference.

The coefficient depends on which category is assigned 1. It does not, by itself, establish that group membership causes a change in the response. In an observational setting, for example, groups may differ in other ways. The regression coefficient describes the relationship represented by the fitted model, with the other included predictors held fixed.

The Model and Its Interpretation

Consider a response \(Y\), a two-category predictor \(D\), and one other predictor \(X\). A multiple regression model with no interaction between \(D\) and \(X\) can be written as

Here, \(i\) indexes the observations, \(\beta_0\) is the intercept, \(\beta_1\) is the coefficient for \(X\), \(\beta_2\) is the coefficient for the indicator, and \(\varepsilon_i\) represents the part of the response not captured by the model. The conditional mean model is

For \(D=0\), the modeled mean response at a particular value \(X=x\) is \(\beta_0+\beta_1x\). For \(D=1\), at that same value of \(X\), it is \(\beta_0+\beta_1x+\beta_2\). Subtracting the first expression from the second gives \(\beta_2\). Thus, the indicator coefficient is a modeled group difference at a fixed value of \(X\). With several other predictors, the same comparison holds when all of them are fixed at the same values.

This interpretation relies on the additive model: it assumes the modeled difference between the groups is constant across the values of the other predictors. It also assumes that the stated regression form is a reasonable description of the conditional mean response. Standard regression inference additionally relies on appropriate error assumptions, such as independent errors and constant error variance; normal errors are used for exact small-sample inference. Those inference assumptions are not needed merely to read the fitted coefficient as a model-based difference.

A coefficient is not automatically a comparison of the groups’ unadjusted averages. If other predictors are in the model, \(\beta_2\) compares modeled means at matched values of those predictors. Interpreting such a comparison also requires that the model can reasonably make the comparison: if the groups have no observations near the same values of the other predictors, the fitted comparison may rely heavily on extrapolation.

Why the Coefficient Is a Difference

The interpretation follows directly from the model. For a general set of other predictors \(X_1,\ldots,X_k\), write the additive model as

At fixed values \(x_1,\ldots,x_k\), the modeled mean for the group coded 0 is \(\beta_0+\sum_{j=1}^{k}\beta_jx_j\). For the group coded 1, it is \(\beta_0+\sum_{j=1}^{k}\beta_jx_j+\beta_D\). Their difference is

All terms that apply equally to both groups cancel. This is why the indicator coefficient compares the groups while holding the other modeled predictors fixed. It is an algebraic interpretation of the specified model, not a claim that every pair of individuals with the same measured predictors will have responses differing by exactly that amount.

Changing Which Group Is Coded 1

The choice of which category receives 0 is a coding choice. Reversing the coding reverses the direction of the comparison, so the indicator coefficient changes sign. The fitted values themselves do not have to change; the equation is simply expressed with a different reference category.

Common Mistakes and Limits

• Ignoring the coding direction. A coefficient of 3 does not mean simply “the group has a response 3 units higher.” State which group is coded 1, which is coded 0, and take the difference in that order.
• Calling an adjusted comparison an average difference. With other predictors in the model, the indicator coefficient is a modeled difference at the same values of those predictors. It need not equal the raw difference between the two groups’ sample averages.
• Claiming causation from the sign of the coefficient. A positive or negative coefficient describes a modeled association. Without a design and assumptions that support causal conclusions, it does not show that changing group membership would change the response.
• Assuming the group difference must be constant in reality. The additive equation imposes a constant modeled difference across values of the other predictors. If a difference varies with another predictor, this model may not represent it adequately.

When reporting a fitted indicator coefficient, give the response units, name the group coded 1 and the group coded 0, and specify that other included predictors are held fixed. This wording makes the comparison’s direction and adjusted nature explicit.