Categorical Predictors & Dummy Coding

Regression works with numbers: it multiplies each predictor by a coefficient. What do you do with a predictor like treatment group, country, or education level, where the values are categories, not amounts? You can't just label them 1, 2, 3 (that would falsely claim group 3 is "three times" group 1). You use dummy coding.

How dummy coding works

For a categorical variable with k groups, you create k − 1 new 0/1 indicator variables. One group is left out as the reference; each dummy variable flags membership in one of the other groups. Three groups (Control, Drug A, Drug B) with Control as reference becomes two dummies:

Reading the coefficients

Each coefficient has a simple meaning:

• The intercept is the predicted outcome for the reference group (all dummies = 0).
• Each dummy coefficient is the difference between that group's mean and the reference group's mean.
• Each coefficient's p-value tests whether that group differs significantly from the reference.

Choosing the reference

Switch the reference group above and watch the coefficients change, but notice that the predicted group means stay identical. The reference is just the baseline you measure everything against, so pick the one that makes your comparisons most meaningful: the control condition, the placebo, the "standard" category. The model is the same either way, and only the meaning of the individual coefficients changes.

The same analysis twice, on the study-methods data

Here is the same analysis done both ways, on one file. study-methods.csv holds 105 students, 35 in each of three revision conditions, with an exam score each. Section 2.2 runs it as a one-way ANOVA. Here the same file is analyzed as a regression.

Three groups means two dummies. With rereading as the reference, d₁ is 1 for a flashcards student and 0 for everyone else, and d₂ is 1 for a practice-testing student and 0 for everyone else. A rereading student is 0 on both, so the reference group needs no column of its own. Fit ŷ = b₀ + b₁d₁ + b₂d₂ and the Coefficients box shows:

Every number in that table comes from the group means. The group means are 66.143, 74.429 and 76.829. The intercept is the rereading mean exactly. 8.286 is 74.429 − 66.143 and 10.686 is 76.829 − 66.143. So each coefficient is a distance from the reference, and a group's predicted mean is the intercept plus its own coefficient. The two standard errors are identical because the groups are the same size. Both are √(MS_E(1/35 + 1/35)) = 2.331, the standard error of the pooled pairwise comparison in Section 2.3.

The model's ANOVA table, printed above the coefficients, is:

That is the one-way ANOVA table, cell for cell. The regression sum of squares is SS_G and the residual sum of squares is SS_E. The two df are I − 1 and N − I, and F(2, 102) = 11.57, p < .001 is the number Section 2.2 reports on this same file. R² = .185 is η². The residual standard error √95.116 = 9.753 is s_p, the pooled SD that every follow-up comparison uses. Nothing has been approximated: it is one analysis printed in two layouts.

Compared with the ANOVA table, the regression layout gives every difference a confidence interval. You can say that flashcards beat rereading by between 3.66 and 12.91 points, not just that the difference is significant. It also gives each dummy a t, which is the pairwise comparison of that group against the reference on N − I degrees of freedom. What it leaves out is the third comparison. No row in the table compares flashcards with practice testing, because neither of them is the reference. That gap is 2.40 points, t(102) = 1.03, p = .31, and you get it by refitting with a different reference or through a contrast. The F test does not cover it either, because the F test never tests any single comparison.

Why the third dummy has to go

Add d₃, flagging rereading, and the model no longer has a unique fit. For every student in the file d₁ + d₂ + d₃ = 1, and the intercept is itself a column of 1s, so the four columns are exactly linearly dependent. This is perfect collinearity. The estimates are not just unstable: infinitely many sets of coefficients fit these data equally well, and the arithmetic cannot choose between them. Software responds by dropping a column and fitting the model without it, sometimes with only a blank row or a footnote to show it. This is the dummy variable trap, and using k − 1 dummies avoids it.

The other coding scheme you'll meet

Dummy coding (software often calls it treatment coding, and it is R's default) is not the only way to turn groups into numbers. Effect coding uses the values −1, 0 and 1, and it changes the baseline. The intercept becomes the grand mean of all groups, not one group's mean, and each coefficient becomes that group's distance from the grand mean. Both schemes fit the same model and give the same predictions and R². They differ only in what the individual coefficients mean. The difference matters once you add an interaction. Under dummy coding the lower-order terms describe the reference group, not an average, and reading them as averages is the most common misreading of a factorial regression.

Usually you code dummies so that you can add covariates beside them in the next model. The guide One Question, an ANOVA and a Regression takes a three-group study through that sequence: ANOVA first, then the dummies alone to show that the two agree, then the controls. Assignments usually ask for that order, and it keeps the coefficients easy to read.