Choosing a Reference Category for Clear Interpretation

Why the Reference Category Matters

When a categorical predictor has several groups, a regression model needs a way to express their outcome differences. In reference coding, one group is the reference category; each other group’s coefficient describes its difference from that baseline. Choosing the reference does not change what the data say about group means, but it does determine which comparisons appear directly in the coefficients.

That makes reference choice a matter of interpretation, not a search for the category that makes the model “best.” If the question is whether a program improves outcomes relative to no program, the control group is a natural baseline. If the question concerns a new option relative to usual practice, the standard option may be more useful. In other settings, the highest category may be the most meaningful benchmark.

Definition: In a reference-coded model for a categorical predictor, the reference category is the category represented by all indicator variables being zero. The intercept represents the model’s fitted outcome for that category (in a model containing only the categorical predictor). Each other category’s coefficient is its fitted outcome minus the reference category’s fitted outcome.

For a categorical predictor with \(K\) categories, one category is omitted and the other \(K-1\) categories receive indicator variables. In a model with only this predictor, let \(\mu_g\) denote the fitted mean for category \(g\), and let \(r\) denote the chosen reference category. Then the model can be written as:

\widehat{Y} = \beta_0 + \sum_{g \ne r} \beta_g I(\text{category}=g)

Here \(I(\text{category}=g)\) equals 1 for an observation in category \(g\), and 0 otherwise. The fitted value for the reference category is \(\beta_0\), while the fitted value for category \(g\) is \(\beta_0+\beta_g\). Thus \(\beta_0=\mu_r\) and \(\beta_g=\mu_g-\mu_r\). A positive coefficient means the category’s fitted outcome is higher than the reference category’s; a negative coefficient means it is lower. The meaning of “higher” or “lower” depends on the outcome’s units and direction.

A Short Justification: What Changes When the Baseline Changes?

Suppose the group fitted means are already represented using reference category \(r\). The fitted mean of any group \(g\) is \(\mu_g\). If category \(s\) is selected as the new reference, the intercept must be \(\mu_s\), and the coefficient for any other group \(g\) must be \(\mu_g-\mu_s\). Substituting these values gives the same fitted mean for every group:

\mu_s + (\mu_g-\mu_s) = \mu_g

This identity explains why changing the reference category changes the coefficient labels and values but not the group fitted means. It is a re-expression of the same set of group means, not a different fit to the data. The coefficient for a particular group does change because it now answers a comparison against a different baseline.

Choose a Baseline That Matches the Question

Start by stating the comparison of interest in words. Then choose as the reference category the group that makes that comparison appear directly as a coefficient. A coefficient for category \(g\) always compares \(g\) with the selected reference—not with every other group and not with an overall average.

Common choices include:

Control or no-exposure category: useful when the main question is how each intervention or exposure compares with the absence of that intervention or exposure.
Standard or usual-practice category: useful when the question is how alternatives compare with the option people currently receive or commonly use.
Highest category: useful when the highest level is a substantively important benchmark and comparisons with that level are the focus.

The most common category is sometimes a convenient reference, but frequency alone does not determine the best choice. Consider the research question, the audience, and how the sign of each coefficient will read in context. A category should not be chosen merely because it produces positive coefficients, a small intercept, or a preferred-looking result.

Worked Example: Use the Control Group as the Reference

Hypothetical data: A study compares a control group, a standard program, and an intensive program. The outcome is a task-completion time in minutes; lower times indicate faster completion. Suppose there are 20 participants in each group, with the following total completion times:

Category	Number of participants	Total minutes	Mean minutes
Control	20	240	12
Standard program	20	180	9
Intensive program	20	140	7

The means are the totals divided by the group sizes: \(240/20=12\), \(180/20=9\), and \(140/20=7\) minutes. Because the stated question is how each program compares with no program, take control as the reference. The intercept is the control mean, and the other coefficients are differences from that mean:

\widehat{Y} = 12 - 3I(\text{standard}) - 5I(\text{intensive})

For the standard program, the coefficient is \(9-12=-3\) minutes. For the intensive program, it is \(7-12=-5\) minutes. The fitted values are 12 minutes for control, \(12-3=9\) minutes for standard, and \(12-5=7\) minutes for intensive. In this model, the coefficients directly report that the program groups have mean completion times 3 and 5 minutes lower than control, respectively.

This choice makes those control comparisons easy to read. It does not, by itself, show that either program caused the difference. A causal conclusion would depend on the study design and other evidence; the coefficient here describes a difference in fitted group means.

Worked Example: Use the Standard Category as the Reference

Use the same hypothetical group means, but suppose the central question is whether either program differs from the standard program. Choose standard as the reference. The intercept is now 9 minutes. The coefficients are differences from 9:

\beta_{\text{control}} = 12-9=3,\qquad \beta_{\text{intensive}} = 7-9=-2

The corresponding model is:

\widehat{Y} = 9 + 3I(\text{control}) - 2I(\text{intensive})

For control, the fitted value is \(9+3=12\) minutes. For standard, both indicators are zero, so the fitted value is 9 minutes. For intensive, the fitted value is \(9-2=7\) minutes. The intensive-program coefficient now directly answers the planned comparison: its fitted mean completion time is 2 minutes lower than the standard-program mean.

Notice that the control coefficient is positive here, whereas it was the reference category in the first example and had no coefficient there. This is not a change in the underlying group means. It reflects that control now takes 3 minutes longer than standard, while intensive takes 2 minutes less.

Worked Example: Use the Highest Category as the Reference

Hypothetical data: Suppose a training course is offered at low, moderate, and high intensity. The outcome is a knowledge score in points. The question is whether the lower-intensity options fall below the high-intensity benchmark. Assume 10 learners in each category, with total scores of 420, 560, and 710 points, respectively. The group means are \(420/10=42\), \(560/10=56\), and \(710/10=71\) points.

Use the highest category as the reference. Its fitted mean, 71 points, is the intercept. Each other coefficient is that category’s mean minus 71:

\beta_{\text{low}} = 42-71=-29,\qquad \beta_{\text{moderate}} = 56-71=-15

The model is:

\widehat{Y} = 71 - 29I(\text{low}) - 15I(\text{moderate})

For low intensity, the fitted score is \(71-29=42\) points. For moderate intensity, it is \(71-15=56\) points. For high intensity, both indicators are zero, giving 71 points. The negative coefficients mean that the low- and moderate-intensity group means are 29 and 15 points below the high-intensity mean. They do not mean that increasing intensity necessarily causes a score increase; the model summarizes these group means, and the interpretation of cause depends on study design.

A high-category baseline is useful here because the question explicitly treats high intensity as the benchmark. If the practical question instead concerned the difference from usual or moderate training, moderate would be a more direct reference.

What Reference Choice Does—and Does Not—Change

In a model containing only a categorical predictor, changing the reference rearranges the intercept and category coefficients so they still reproduce the same group means. In a model that also includes other predictors, the reference still determines which category comparisons are represented by the category coefficients. When the model has no interactions involving that factor, those coefficients describe category differences at the same values of the other predictors, according to the model.

The reference choice does not change the observed data, the set of category means represented by the model, or the fitted values obtained from an equivalent recoding. It does change which particular comparison a coefficient reports. Consequently, a coefficient’s numerical value and sign cannot be interpreted without knowing the reference category. If the question is about a comparison that is not represented directly, the relevant difference must be formed by subtracting the appropriate fitted means or coefficients.

Keep the category labels explicit when communicating results. “The coefficient is -2” is incomplete. “The intensive-program mean is 2 minutes lower than the standard-program mean” states the comparison and its units. If estimates from a fitted model are adjusted for other predictors, describe the comparison as adjusted and specify what is held constant where relevant.

Common Mistakes and a Practical Rule

Choosing a baseline because it makes coefficients look favorable. This changes the displayed comparisons, not the evidence in the data. Choose a baseline from the question, not from the preferred signs.
Calling a coefficient a difference from the overall mean. Under reference coding, it is a difference from the reference category. Other coding schemes can define coefficients differently.
Thinking the reference category is excluded from the model. It is included and has a fitted mean. It is represented by the intercept rather than by its own indicator coefficient.
Assuming a negative coefficient is inherently beneficial. Its substantive meaning depends on the outcome: a lower waiting time may be desirable, while a lower test score may not be.
Treating a group difference as a causal effect. The coefficient compares fitted outcomes; a causal interpretation requires appropriate design and assumptions beyond the choice of reference category.

A useful final check is to name the baseline, state what the intercept represents, and complete the sentence “the coefficient for category \(g\) is the fitted difference between \(g\) and ___.” If the blank is the category intended by the research question, the reference choice is serving its purpose.

Key takeaway: Choose the reference category to make the most important comparison direct and clear. Control, standard, and highest categories can all be appropriate baselines when they match the question. Changing the baseline changes coefficient interpretation, not the group fitted means.

Check Your Understanding

Answer each question using the meaning of reference-coded coefficients.

A model compares control, standard, and intensive programs. The main question is whether each program differs from control. Which category is a natural reference, and what does the intensive coefficient represent?
In a model with standard as the reference, the intensive coefficient is \(-2\) minutes. State the comparison in context and explain what the negative sign means.
A three-level predictor has fitted means of 42, 56, and 71 points for low, moderate, and high categories. If high is the reference, calculate the low-category coefficient.
Does switching from control to standard as the reference change the fitted mean for the control group? Explain why or why not.
Why is choosing the reference category that produces all positive coefficients not a sound reason for choosing it?