Dummy Variables in R: What lm() Does With Factor Predictors Under the Hood

Dummy variables in R are the 0/1 columns that lm() builds automatically whenever you pass a factor as a predictor. Every factor with k levels becomes k - 1 dummy columns, and R picks the first level alphabetically as the reference the other coefficients are measured against.

What happens when you put a factor in lm()?

Let's skip the theory for a minute and just fit a model. The iris dataset has a Species factor with three levels: setosa, versicolor, and virginica. We will regress Sepal.Length on Species and look at what R prints. The output will show two coefficients with funny names, not three, and that one missing coefficient will tell us everything we need to know about what lm() did behind the scenes.

Three species, but only two coefficients show up next to (Intercept). Where did setosa go? R did not drop it. It absorbed setosa into the intercept and now reports every other level as a difference from it. The value 5.006 is the mean Sepal.Length for setosa flowers. Add 0.930 and you get the mean for versicolor. Add 1.582 and you get the mean for virginica. Three numbers, three species, one intercept plus two dummies, and the arithmetic works out exactly.

How does R build the dummy variable matrix?

The coefficients in the previous block came from somewhere. Specifically, they came from R quietly turning the Species factor into numeric columns before handing them to the least-squares engine. That hidden set of columns is called the design matrix, and you can see it with model.matrix(). Looking at it is the single fastest way to stop treating factor handling as magic.

Read the rows one at a time. Row 1 is a setosa flower: the intercept column is 1 (always), and both dummy columns are 0. Row 51 is versicolor: intercept 1, versicolor dummy 1, virginica dummy 0. Row 101 is virginica: intercept 1, versicolor dummy 0, virginica dummy 1. Setosa has no dedicated column because "both dummies zero" is how R flags it. This is why a three-level factor produces two dummy columns, not three.

Figure 1: R routes a k-level factor through model.matrix() to produce k-1 dummy columns plus the intercept.

How do you interpret the coefficients of a factor?

The plain English reading is this: the intercept is the mean of the response variable in the reference group, and each dummy coefficient is the mean in that level minus the reference mean. You do not have to trust this interpretation; you can verify it by computing the group means and comparing them to coef(fit1) line by line.

The intercept 5.006 is exactly the setosa mean. The coefficient 0.930 on Speciesversicolor is the versicolor mean minus the setosa mean. The coefficient 1.582 on Speciesvirginica is the virginica mean minus the setosa mean. This is not an approximation. It is algebraically what least squares returns when the only predictors are dummy columns.

Figure 2: The intercept captures the reference group mean; each dummy coefficient captures how far a non-reference group sits from it.

How do you change which level is the reference?

Alphabetical order is a convenience, not a decision. In iris, setosa happens to be a natural baseline because it is visibly the smallest-flowered species. In other datasets the alphabetical first level is a random category like "AL" for Alabama, which nobody wants to interpret as a baseline. relevel() lets you pick a reference that actually means something.

The intercept is now 6.588, the virginica mean. Setosa flipped to -1.582 because its mean is 1.582 below virginica's. Versicolor is -0.652 because its mean is 0.652 below. The model's predictions, residuals, and R-squared are identical to fit1; only the labels changed. You can also use factor() with an explicit levels = argument when you want full control over level order.

Both approaches produce the same fit as fit_rl. Use relevel() when you only need to swap the baseline. Use factor(..., levels = ...) when you also want to control how other levels appear in tables and plots.

What if you want a coefficient for every level?

Sometimes the reference-plus-gaps parameterisation is awkward. You just want one coefficient per level, each equal to that level's mean. Dropping the intercept with - 1 in the formula does exactly that.

Now there are three coefficients, one per species, and each equals the species mean that you computed earlier with aggregate(). The model is mathematically the same as fit1: identical predictions, identical residuals, identical residual standard error. What changed is the parameterisation. You are just looking at the same surface from a different angle.

How do factor interactions work in lm()?

So far each coefficient was a group-mean offset. Interactions add a second dimension: they let a numeric predictor's slope change across factor levels. The way R implements this is unglamorous, it multiplies the dummy columns by the numeric column to produce new interaction columns. Same framework, more columns.

The first two coefficients describe the reference species (setosa): intercept 4.7772 and Petal.Width slope 0.9131. The two Species* coefficients shift the intercept for versicolor and virginica. The two Petal.Width:Species* coefficients shift the slope. For versicolor, the fitted equation is (4.7772 + 1.3587) + (0.9131 - 0.1989) * Petal.Width. For virginica it is (4.7772 + 1.9468) + (0.9131 - 0.5580) * Petal.Width. Each species gets its own line.

The last two columns are Petal.Width multiplied by each dummy. For a setosa row (row 1), both dummies are 0 so both product columns are 0, and the fit reduces to the reference line. For a versicolor row (row 51), only the Speciesversicolor column is 1, so only the versicolor product column carries Petal.Width. For a virginica row, only the virginica product column carries it. Nothing deeper than multiplication is happening.

What contrast schemes exist besides treatment coding?

Treatment coding is the default in R, and it is what we have been using this whole time. But it is only one of four contrast schemes that ship with base R. Switching schemes does not change how well the model fits; it changes what the coefficients mean. Pick the scheme that answers your research question directly instead of doing arithmetic on coefficients to reverse-engineer the answer.

This is the treatment (dummy) matrix you have been looking at in model.matrix() the whole time. Each non-reference level gets its own column, with a 1 in the row for that level. Now switch to sum coding, which compares each level to the grand mean instead of to a reference.

The intercept is now 5.843, the grand mean of Sepal.Length across all three species. The coefficients Species1 and Species2 are deviations of setosa and versicolor from that grand mean; virginica's deviation is implied as the negative sum of the two. The model's predictions and R² are identical to fit1. Only the lens on the coefficients changed.

Figure 3: A decision flow for picking treatment, sum, polynomial, or Helmert contrasts based on your comparison question.

Practice Exercises

These exercises combine multiple concepts from the tutorial. Use distinct variable names with the my_ prefix so your exercise code does not overwrite the tutorial variables.

Exercise 1: Verify a group mean from lm() coefficients

Using mtcars, convert cyl to a factor, fit mpg ~ cyl_f, and verify that the intercept plus the coefficient for the 6-cylinder dummy equals the mean mpg of 6-cylinder cars computed directly with aggregate(). Save the model to my_cyl_model and the check to my_cyl_check.

Exercise 2: Predict from an interaction model by hand

Using mtcars, convert cyl to a factor, fit mpg ~ cyl_f * wt, and predict mpg for a 6-cylinder car weighing 3.2 thousand pounds using only the raw coefficients (no predict() call). Store the result in my_prediction.

Complete Example: a mini-analysis of mtcars

Let's tie it all together on mtcars. We will turn cyl and am into factors with human-readable labels, set the reference levels deliberately, fit a model that controls for weight, and interpret every coefficient in plain English.

Reading the coefficients one at a time: an 8-cylinder automatic car at zero weight (not a real car, just the math baseline) is predicted to do 31.0 mpg. Switching to 6 cylinders at the same weight and transmission adds 3.0 mpg. Four cylinders adds 6.9 mpg. A manual transmission adds 1.8 mpg. Every extra 1000 pounds of weight subtracts 3.2 mpg. The coefficient on wt is the same regardless of cylinder count because this model is additive, not interactive.

You can see the factor-to-dummy mechanics at work. The Mazdas are both 6-cylinder manuals with cyl_fsix = 1 and am_fmanual = 1. The Datsun is a 4-cylinder manual with cyl_ffour = 1. Their predicted mpg values come from plugging these 0/1 columns into the coefficient formula, and that is the whole story of how R handles categorical predictors in lm().

Summary

Key takeaways:

1. A factor with k levels becomes k - 1 dummy columns plus an intercept. The omitted level is the reference.
2. The intercept equals the reference group's mean (under default treatment coding). Every dummy coefficient is "this level's mean minus the reference's mean".
3. Changing the reference with relevel() shifts coefficient labels and values but never the predictions, residuals, or R².
4. Interactions between factors and numerics are just new design-matrix columns built by multiplying dummy columns by the numeric column.
5. Treatment coding is the default; sum, Helmert, and polynomial coding exist for specific comparison questions. The fit is identical, the interpretation is not.

References

1. R Core Team. An Introduction to R, §11.1 "Defining statistical models; formulae". Link
2. Chambers, J.M. & Hastie, T.J. Statistical Models in S, 1992. Chapter 2 on design matrices and contrasts is the classic reference on how S and R build factor columns.
3. Base R documentation for contrasts and contr.treatment. Link
4. UCLA OARC. Coding for Categorical Variables in Regression Models. Link
5. STHDA. Regression with Categorical Variables: Dummy Coding Essentials in R. Link
6. Faraway, J. Linear Models with R, 2nd ed., 2014, Chapter 14 "Categorical Predictors".
7. Fox, J. & Weisberg, S. An R Companion to Applied Regression, 3rd ed., 2019, §4.3 "Factor Predictors". Link

Continue Learning

• Linear Regression Assumptions in R, diagnostic checks that still apply when your predictors are factors.
• Linear Regression, the base lm() workflow for numeric predictors, the natural prerequisite for this post.
• R Factors, how factors are stored, ordered, and manipulated outside of lm().