Categorical Predictors in Regression: Dummy Coding, Reference Groups, and Coefficient Interpretation

The problem: category numbers are not automatically quantities

Suppose a researcher wants to predict a continuous engagement score from participants' program route, which has four categories:

• Standard
• Referral
• Outreach
• Partnership

A data file might store these categories as:

• Standard = 1
• Referral = 2
• Outreach = 3
• Partnership = 4

Those numbers can function perfectly well as labels in a dataset. But entering that single 1–4 variable directly into an ordinary regression would tell the model to treat the numbers as quantitative values. Hayes (2022) explicitly warns that a multicategorical variable representing kind rather than quantity cannot simply be entered in this way because regression interprets its numerical values as quantities. Tabachnick and Fidell (2013) make the same distinction: arbitrary numerical assignments to qualitatively different categories do not place those categories on a quantitative continuum, and changing the numerical assignments can arbitrarily change the apparent relationship with another variable (Hayes, 2022; Tabachnick & Fidell, 2013).

For the program-route example, treating 1, 2, 3, and 4 as a single quantitative predictor would therefore impose a structure that the category labels do not justify. It would imply, among other things, that moving from Standard to Referral represents the same numerical increment as moving from Outreach to Partnership.

Dummy coding converts categories into explicit comparisons

A common solution is indicator coding, also called dummy coding. Instead of entering one arbitrary numerical category code, the researcher creates binary variables that represent specified group memberships. Hayes (2022) describes indicator coding as a system in which (g − 1) variables coded 0 or 1 represent membership in g groups. Meier et al. (2014) likewise demonstrate dummy variable regression for a nominal predictor with more than two categories (Hayes, 2022; Meier et al., 2014).

For the four program routes, suppose Standard is selected as the reference category. The coding can be:

Each explicit indicator now answers a specific comparison.

Standard does not need its own indicator because it is represented by having zero on all three indicators (Hayes, 2022).

Why four categories produce three indicators

Under this common reference-group indicator coding, a categorical predictor with k categories is represented by (k − 1) explicit indicator variables. The remaining category is represented by the all-zero pattern and becomes the reference or baseline group (Hayes, 2022).

Meier et al. (2014) illustrate the same principle with three geographic categories. They construct indicators for two categories and explicitly advise against creating a third dummy variable: when both included indicators equal zero, the observation belongs to the remaining category. Their regression intercept then represents that omitted category's mean (Meier et al., 2014).

The fitted model can therefore be written as:

Ŷ = b0 + b1(Referral) + b2(Outreach) + b3(Partnership)

What does the intercept mean?

The intercept is the predicted outcome when every predictor in the equation equals zero.

With the dummy coding above, all three route indicators equal zero for the Standard group. Therefore:

b₀ = predicted engagement for the Standard group.

When the model contains only the categorical predictor, this corresponds to the Standard group's mean engagement score. Both Hayes (2022) and Meier et al. (2014) demonstrate this reference-group interpretation of the regression constant (Hayes, 2022; Meier et al., 2014).

This is one reason reference categories matter. They determine the baseline represented by the intercept and the comparisons represented by the dummy-variable coefficients.

What does each dummy coefficient mean?

Under indicator coding, each coefficient compares its category with the reference category (Hayes, 2022).

Suppose the synthetic fitted equation is:

Predicted Engagement = 62.0 + 4.5(Referral) − 2.0(Outreach) + 7.0(Partnership)

The intercept is 62.0, so the predicted engagement score for Standard-route participants is 62.0.

For Referral participants:

62.0 + 4.5 = 66.5

For Outreach participants:

62.0 − 2.0 = 60.0

For Partnership participants:

62.0 + 7.0 = 69.0

The coefficients therefore have direct group-comparison interpretations.

Referral coefficient: b₁ = 4.5

Referral participants are predicted to score 4.5 points higher on engagement than Standard participants.

Outreach coefficient: b₂ = −2.0

Outreach participants are predicted to score 2.0 points lower on engagement than Standard participants.

Partnership coefficient: b₃ = 7.0

Partnership participants are predicted to score 7.0 points higher on engagement than Standard participants.

Translation from output to English

Suppose a software-neutral coefficient table reports:

The translation should proceed from the coding, not from the coefficient table alone.

Output: Intercept = 62.0

“The predicted engagement score for participants in the Standard reference group is 62.0.”

Output: Referral = 4.5

“Participants entering through Referral are predicted to score 4.5 points higher on engagement than participants entering through the Standard route.”

Output: Outreach = −2.0

“Participants entering through Outreach are predicted to score 2.0 points lower on engagement than participants entering through the Standard route.”

Output: Partnership = 7.0

“Participants entering through Partnership are predicted to score 7.0 points higher on engagement than participants entering through the Standard route.”

This translation follows directly from reference-group indicator coding: the intercept identifies the reference group's predicted value, while each dummy coefficient gives the difference between its group and that reference (Hayes, 2022; Meier et al., 2014).

If additional predictors are present, add the appropriate qualification:

“Holding the other predictors in the model constant, participants in Partnership are predicted to score 7.0 points higher than participants in the Standard reference group.”

Hayes (2022) notes that indicator-coded multicategorical variables can be included alongside additional antecedent variables, allowing group differences to be estimated while the other modeled antecedents are held constant.

Do not read comparisons that the coefficients do not provide

With Standard as the reference category, the coefficient of 4.5 directly compares Referral with Standard. The coefficient of 7.0 directly compares Partnership with Standard.

Neither coefficient, by itself, is the Referral-versus-Partnership comparison.

From the fitted values, the estimated difference between Partnership and Referral is:

69.0 − 66.5 = 2.5

Equivalently:

7.0 − 4.5 = 2.5

Changing the reference category

The reference group is a feature of the coding system, not an inherently privileged group.

Suppose the research question is primarily about how every route compares with Referral rather than Standard. The researcher can recode the indicators so that Referral receives zeros on every indicator:

Using the same synthetic fitted group values, the equation would now have an intercept of 66.5, because Referral is the new reference.

The coefficients become:

• Standard versus Referral: 62.0 − 66.5 = −4.5
• Outreach versus Referral: 60.0 − 66.5 = −6.5
• Partnership versus Referral: 69.0 − 66.5 = 2.5

The group predictions themselves have not changed. What changed is the parameterization: the intercept now identifies Referral, and the coefficients express comparisons with Referral. Hayes (2022) explicitly demonstrates changing the reference group and rerunning a model so that coefficients and tests provide the comparisons associated with the newly selected reference group.

Choosing a reference group should therefore follow the research question. A control, standard-practice, untreated, conventional, or otherwise substantively meaningful group may make interpretation especially clear, but a different reference can be selected when a different set of comparisons is more useful.

Why not include all four dummy variables with the usual intercept?

The (k − 1) construction is not an arbitrary convention. With an intercept and mutually exclusive indicators for every category, the category indicators contain redundant information: for every observation, exactly one of them equals 1 and their sum equals the intercept column of ones.

Reference-group coding avoids this redundancy by representing one group through the all-zero pattern. Hayes (2022) describes multicategorical predictors as requiring (g − 1) variables under these coding systems, while Meier et al. (2014) demonstrate the practical version of the same rule by using two indicators for three groups.

The result is an interpretable baseline-and-differences parameterization:

reference prediction + category difference

Categorical predictors alongside other regression predictors

Dummy coding does not restrict the model to group comparisons alone. Indicator variables can appear alongside quantitative and other categorical predictors in a multiple regression model (Hayes, 2022).

For example:

Ŷ = b0 + b1DReferral + b2DOutreach + b3DPartnership + b4X

where X might be a quantitative predictor.

Now the dummy coefficients describe estimated differences from the Standard reference category while X is held constant. Conversely, the coefficient for X describes its modeled association with the outcome while program route is held fixed (Hayes, 2022).

Multicategorical predictors in interaction and moderation models

A multicategorical predictor can also participate in an interaction or moderation model. Hayes (2022) extends ordinary moderation analysis directly to multicategorical focal predictors and moderators.

Suppose program route has four categories and W is a continuous moderator. With Standard as the reference category, the model can include three route indicators, W, and three corresponding products:

Y = b0 + b1D1 + b2D2 + b3D3 + b4W + b5D1W + b6D2W + b7D3W + e

For a four-category focal predictor, the three D variables represent the groups, and the three D × W products allow the relationship involving W to differ across those group comparisons. Hayes (2022) describes this as the extension of moderation to a multicategorical antecedent: a g-group predictor is represented by (g − 1) coding variables, and moderation requires corresponding products between those coding variables and the moderator.

The reference category remains central to interpretation. Under indicator coding, the lower-order and interaction coefficients are expressed relative to that baseline coding. Hayes (2022) also shows that changing the reference group can be useful for obtaining conditional effects and comparisons for another group.

When the moderator itself is multicategorical, the same coding logic applies: the multicategorical moderator is represented through (g − 1) coding variables, and the required interaction products are constructed accordingly (Hayes, 2022).

Checks before interpreting dummy-variable regression

Correct dummy coding does not make the rest of regression diagnostics unnecessary. It solves the problem of representing the categorical predictor; it does not establish that the overall regression model is adequate.

Before interpretation, researchers should at minimum verify the coding against the original categories and confirm which category is the reference. Because arbitrary codes for qualitatively different categories can impose artificial numerical relationships, the coding scheme should be documented rather than inferred later from variable names or coefficient signs (Hayes, 2022; Tabachnick & Fidell, 2013).

The broader assumptions and diagnostics appropriate to the regression model also remain relevant. Meier et al. (2014) discuss model specification and multicollinearity as regression concerns, while Hayes (2022) emphasizes that regression interpretation and inference depend on the assumptions of the fitted model (Hayes, 2022; Meier et al., 2014).

A compact interpretation workflow

When encountering categorical variables in regression, use the following sequence:

1. Identify what the variable represents. If its categories are qualitative rather than meaningful numerical quantities, do not interpret arbitrary category codes as a quantitative scale (Hayes, 2022; Tabachnick & Fidell, 2013).

2. Choose the comparisons you need. Under reference-group dummy coding, select a substantively useful reference category.

3. Create (k − 1) indicators for k categories. The reference category receives zero on every indicator (Hayes, 2022; Meier et al., 2014).

4. Interpret the intercept first. It identifies the predicted outcome for the reference group at zero on any other predictors.

5. Interpret each dummy coefficient as a difference from the reference group. Do not describe it as a generic “effect of category.”

6. Recode the reference when another comparison is needed. Changing the baseline changes the coefficient comparisons, not the underlying fitted group predictions.

7. For moderation, preserve the multicategorical structure. Represent the g groups with (g − 1) coding variables and construct the corresponding interaction terms rather than treating arbitrary group numbers as a continuous moderator (Hayes, 2022).

8. Check the complete regression model. Dummy coding addresses representation of the categorical predictor, not every assumption or specification issue in regression.

Common pitfalls

Limitations and interpretation boundaries

Reference coding is also only one possible coding system. Indicator coding is popular because its coefficients have straightforward comparisons with a baseline group, but alternative contrast systems can encode different research questions. Coefficients must always be interpreted according to the coding actually used (Hayes, 2022).

Finally, the numerical example in this Resource is entirely synthetic. Its purpose is to demonstrate how to interpret dummy variables, not to provide empirical evidence about real programs or participants.

Conclusion

The essential rule for categorical variables in regression is simple: numbers used to label categories should not silently become quantities.

With dummy coding, a k-category predictor is commonly represented by (k − 1) explicit indicators and one reference category. The reference group's coding determines the intercept, while each indicator coefficient describes a comparison between its category and that reference group (Hayes, 2022; Meier et al., 2014).

Once that structure is clear, coefficient interpretation becomes much easier. Start with the reference category, translate the intercept, identify exactly which comparison each coefficient represents, and change the reference group when the research question requires another comparison. For interactions and moderation, retain the multicategorical structure rather than collapsing arbitrary group labels into a quantitative scale (Hayes, 2022).

That discipline turns dummy variable regression from a coding exercise into what it actually is: a way of specifying the group comparisons the regression model is being asked to estimate.

References

Hayes, A. F. (2022). Introduction to mediation, moderation, and conditional process analysis: A regression-based approach (3rd ed.). The Guilford Press.

Lovric, M. (Ed.). (2011). International encyclopedia of statistical science. Springer. https://doi.org/10.1007/978-3-642-04898-2

Meier, K. J., Brudney, J. L., & Bohte, J. (2014). Applied statistics for public and nonprofit administration (9th ed.). Cengage Learning.

Tabachnick, B. G., & Fidell, L. S. (2013). Using multivariate statistics (6th ed.). Pearson.