Confidence Intervals for Indicator-Variable Coefficients

What an Indicator-Variable Interval Describes

A regression coefficient for a group indicator describes a difference between two groups. Its confidence interval expresses uncertainty about that difference. When the model includes other predictors, the coefficient describes the group difference at the same values of those predictors; this is an adjusted difference. The reference group matters because the coefficient is defined by comparing the other group with that reference.

For example, suppose a study compares an outcome score for people receiving a new program with the score for people receiving standard care. Coding standard care as the reference group lets the program coefficient represent new program minus standard care. If age is also included in the model, that coefficient compares the two groups at the same age, under the model’s assumptions. It does not, by itself, show that the program caused a difference.

The Model and the Meaning of the Coefficient

Let \(D_i=1\) for the group of interest and \(D_i=0\) for the reference group. If \(Z_i\) is another predictor, consider the model

Here, \(Y_i\) is the outcome for individual \(i\), \(\beta_0\) is the intercept, and \(\beta_D\) is the indicator-variable coefficient. The model assumes the expected outcome is described by the stated linear expression and that the errors have mean zero, are independent, have a common variance, and are normally distributed for the conventional small-sample t interval to have its exact reference distribution. The predictors must also be estimable: in particular, the indicator cannot be an exact linear combination of the other model terms. For an adjusted comparison to be meaningful, the groups must have enough overlap in the other predictors to support comparing them at comparable predictor values.

The fitted mean for the reference group at a given value \(Z_i=z\) is \(\beta_0+\beta_Zz\). The fitted mean for the group coded 1 at the same value is \(\beta_0+\beta_D+\beta_Zz\). Subtracting the reference-group mean from the other-group mean gives

Thus, in this model, \(\beta_D\) is a difference in expected outcomes, in the units of \(Y\), adjusted for \(Z\). The difference is constant across values of \(Z\) because this model has no interaction between group and \(Z\). If the model changes, the coefficient’s meaning may change too.

The conventional coefficient interval has the form

where \(\hat{\beta}_D\) is the estimated coefficient, \(SE(\hat{\beta}_D)\) is its standard error, and \(t^*\) is the t critical value for the chosen confidence level and residual degrees of freedom. The interval is expressed in the outcome’s units. Its endpoints describe plausible values for the modeled group difference, not a range containing individual outcomes.

A Two-Group Model Without Other Predictors

The simplest case has only an intercept and the group indicator: \(Y_i=\beta_0+\beta_DD_i+\varepsilon_i\). With this coding, the fitted mean for the reference group is \(\beta_0\), and the fitted mean for the group coded 1 is \(\beta_0+\beta_D\). Their difference is \(\beta_D\), so the indicator coefficient is the difference between the two sample group means in the fitted model. This is an unadjusted difference because no other predictors are held fixed.

For the ordinary pooled-variance two-group regression, the standard error of the difference is \(s_p\sqrt{1/n_0+1/n_1}\), where \(s_p\) is the estimated common within-group standard deviation and \(n_0,n_1\) are the group sample sizes. The t interval uses \(n_0+n_1-2\) residual degrees of freedom. This calculation relies on the two groups’ errors having a common variance; the standard t procedure also assumes independent, normally distributed errors for exact small-sample coverage.

Interpreting an Adjusted Difference

When other predictors are included, the coefficient interval is still in the outcome’s units, but its comparison is conditional on those predictors. For a group coefficient, “adjusted for age,” for example, means the model compares the groups at the same age. It does not mean that every possible difference between the groups has been removed, nor does it guarantee the groups are comparable in unmeasured ways.

A 95% confidence interval has a repeated-sampling interpretation: if the same sampling process and interval procedure were repeated many times under the model assumptions, about 95% of the resulting intervals would cover the true coefficient. For the particular interval calculated from one sample, it is not correct to say there is a 95% probability that the fixed coefficient lies within it.

The Reference Group Determines the Direction

The coefficient is a subtraction in a chosen direction: group coded 1 minus reference group. If the reference group changes, the comparison reverses. The estimated difference changes sign, and the endpoints of its confidence interval change sign and switch order. The underlying fitted comparisons have not gained or lost information; only their direction has changed.

Common Mistakes and Why the Interval Matters

A coefficient interval is useful because it shows both the estimated direction and the range of uncertainty in the outcome’s units. Two intervals can lead to different practical conclusions even if their point estimates are similar: a narrow interval indicates greater precision, while a wide interval leaves more uncertainty about the size of the difference. Whether the range is important in practice depends on the subject matter and what differences would matter to decision-makers.

• Ignoring the coding direction. A positive coefficient means the group coded 1 has a higher modeled mean than the reference group, not necessarily that a named group is higher unless the coding is stated.
• Calling every coefficient an adjusted difference. In a model with only the indicator and intercept, the coefficient is an unadjusted group difference. It is adjusted only for other predictors actually included in the model.
• Reading the interval as a range for individuals. The interval concerns the coefficient, or modeled group difference, not where most individual outcomes fall.
• Claiming causation from an adjusted interval. Adjustment for measured predictors alone does not establish that group membership caused the outcome difference.
• Reversing only the estimate when recoding. Reversing the comparison changes both interval endpoints’ signs and their order: \((L,U)\) becomes \((-U,-L)\).