Categorical Predictors and Contrasts

Turning scientific comparisons among groups into estimable linear functions

Statistics

Regression

Categorical Predictors

Contrasts

Indicator coding, planned contrasts, marginal means, uncertainty, and interpretation for categorical predictors in regression models.

Author

Jiahong Sun

Published

October 14, 2026

Modified

October 14, 2026

1 Categories Enter through a Design Matrix

A categorical predictor records membership in a finite set of levels: treatment arm, genotype, tissue, sequencing batch, or disease subtype. Regression software cannot estimate a separate unconstrained coefficient for every level together with an intercept, because the level indicators sum to the intercept. It therefore represents a factor with $K$ levels using $K-1$ independent columns.

For a one-factor conditional-mean model,

\[ Y_i=\mu_{g(i)}+\varepsilon_i, \qquad \operatorname{E}(\varepsilon_i\mid g(i))=0, \]

the scientifically meaningful parameters are the level means $\boldsymbol\mu=(\mu_1,\ldots,\mu_K)^\mathsf{T}$. A coding scheme chooses a full-rank coordinate system for the same $K$ means. Changing coordinates changes the displayed coefficients, but not fitted values or any estimable comparison computed consistently from the model.

This distinction is central:

Coding coefficients describe a parameterization.
A contrast describes a scientific comparison.
An estimate is the contrast evaluated from the fitted model.

2 Treatment Coding and Its Interpretation

With three levels—control, drug A, and drug B—and control as the reference, treatment coding gives

\[ \operatorname{E}(Y_i\mid g_i) =\beta_0+\beta_A I(g_i=A)+\beta_B I(g_i=B). \]

The coefficients are

\[ \beta_0=\mu_C, \qquad \beta_A=\mu_A-\mu_C, \qquad \beta_B=\mu_B-\mu_C. \]

Thus the intercept is the control mean and each slope is a difference from control. Changing the reference level changes all three coefficient meanings, but the fitted group means remain $(\hat\mu_C,\hat\mu_A,\hat\mu_B)$.

Reference coding is convenient when one level is a genuine scientific baseline. It is less useful when the questions concern a dose trend, an average of several controls, or A versus B. Those questions should be stated as contrasts rather than inferred indirectly from whichever coefficients software happens to print.

3 A Contrast Is the Estimand

A linear contrast of level means is

\[ \psi=\mathbf c^\mathsf{T}\boldsymbol\mu =\sum_{k=1}^{K}c_k\mu_k, \qquad \sum_{k=1}^{K}c_k=0. \]

The zero-sum condition makes the comparison invariant to adding the same constant to every level. Examples for $(C,A,B)$ include

Question	Contrast weights $\mathbf c$	Estimand
A versus control	$(-1,1,0)$	$\mu_A-\mu_C$
B versus control	$(-1,0,1)$	$\mu_B-\mu_C$
A versus B	$(0,1,-1)$	$\mu_A-\mu_B$
Average drug versus control	$(-1,1/2,1/2)$	$(\mu_A+\mu_B)/2-\mu_C$

Let $\hat{\boldsymbol\beta}$ be the fitted coefficient vector with covariance $\widehat{\mathbf V}$. If the desired comparison is $\psi=\mathbf l^\mathsf{T}\boldsymbol\beta$, then

\[ \hat\psi=\mathbf l^\mathsf{T}\hat{\boldsymbol\beta}, \qquad \widehat{\operatorname{SE}}(\hat\psi) =\sqrt{\mathbf l^\mathsf{T}\widehat{\mathbf V}\mathbf l}. \]

An interval or test follows from this estimate and standard error using a reference distribution appropriate to the design and covariance estimator. The contrast should be chosen because it answers a prespecified question, not because it produces the smallest p-value.

4 Other Coding Systems

Different full-rank codings emphasize different coordinates:

Coding	Typical coefficient interpretation	Important caution
Treatment/reference	Each nonreference level minus a baseline	Intercept and slopes depend on the chosen reference
Sum-to-zero	Deviations from an equally weighted grand mean	The grand mean is not the sample-size-weighted mean in an unbalanced design
Helmert	A level compared with an average of subsequent or previous levels	Direction and scaling conventions differ among software
Polynomial	Linear, quadratic, and higher-order patterns across ordered levels	Meaningful only when ordering and spacing are defensible
Custom	Coordinates aligned with planned scientific hypotheses	Must have adequate rank and documented weights

Orthogonal contrast vectors satisfy $\mathbf c_j^\mathsf{T}\mathbf c_k=0$ for $j\ne k$. In a balanced one-way design, their sums of squares partition the factor sum of squares. In unbalanced or covariate-adjusted data, algebraic orthogonality of the weights does not necessarily imply uncorrelated estimates.

A factor with $K$ levels contributes $K-1$ degrees of freedom. A single coefficient test answers one coordinate-specific question; an omnibus $K-1$ degree-of-freedom test asks whether all relevant level differences are zero. A nonsignificant omnibus test does not validate every equality, and a significant omnibus test does not identify which contrast matters.

5 Adjustment Changes What Is Being Compared

With additional covariates,

\[ \operatorname{E}(Y_i\mid g_i,\mathbf z_i) =\alpha+\text{factor terms}(g_i)+\mathbf z_i^\mathsf{T}\boldsymbol\gamma, \]

factor coefficients are conditional comparisons at the same $\mathbf z$. If there are no factor-by-covariate interactions, the adjusted difference is constant over $\mathbf z$. With interactions, there is no single group difference until values or a distribution of $\mathbf z$ are specified.

Estimated marginal means compute model-based predictions for each level and average them over a declared reference distribution of covariates. Equal weighting across cells, weighting by the observed sample, and weighting by a target population answer different questions. In unbalanced observational data, calling one of them “the adjusted mean” without stating the averaging rule hides part of the estimand.

Adjusted comparisons also require overlap. If a treatment occurs only in one batch, the treatment and batch columns are aliased: no coding scheme can recover their separate effects. Sparse treatment-by-tissue cells may technically permit a fit while making the requested contrast largely extrapolative.

6 Worked Example: Three Library-preparation Methods

Suppose RNA integrity is measured for independent donor samples processed by a standard method $S$ or alternatives $A$ and $B$. The fitted adjusted means are

\[ (\hat\mu_S,\hat\mu_A,\hat\mu_B)=(7.10,7.45,7.20). \]

The planned contrast comparing the average alternative with the standard method is

\[ \hat\psi =-\hat\mu_S+\tfrac12\hat\mu_A+\tfrac12\hat\mu_B =0.225. \]

That estimate does not equal the A-versus-S coefficient, the B-versus-S coefficient, or an unweighted average of all three means. Its standard error must use the covariance among the three fitted means; treating their individual standard errors as independent is generally wrong.

If the same donor material was split across methods, donor is a blocking or repeated-measures unit and the model must retain that pairing. If different donors received different methods, donors are the independent experimental units. Technical lanes, aliquots, and reads do not increase the biological replication unless the target explicitly concerns technical repeatability.

7 Diagnostics, Sensitivity, and Reporting

Before interpreting a factor or contrast:

Verify levels and reference meaning. Check spelling, unused levels, ordering, and whether a numeric code was accidentally treated as continuous.
Inspect the design matrix. Confirm rank, columns, aliases, empty cells, and the coding actually used by the software.
Inspect cell counts and overlap. Report independent units per relevant factor combination, not only the total record count.
Check the outcome model. Residual, variance, dependence, and influence diagnostics still apply; factor coding does not repair misspecification.
Define the contrast family. Prespecify the comparisons and use simultaneous intervals or multiplicity adjustment when many claims form one family.
Test coding invariance. Refit with another full-rank coding and confirm that fitted values and explicitly reconstructed contrasts agree.
State the averaging population. For marginal means, document covariate values and weights.

Report the contrast weights or plain-language comparison, estimate, uncertainty interval, scale and units, covariance method, adjustment set, and multiplicity procedure. A coefficient table alone is not a complete account of a multi-level factor.

8 Connections and Take-home Messages

Multiple linear regression defines the design-matrix model in which factor columns are fitted.
Least squares and the geometry of regression explains rank, estimability, and covariance of linear functions.
Regression uncertainty and coefficient interpretation develops conditional coefficients, contrasts, and prediction uncertainty.
Interaction terms and effect modification shows why a factor effect can depend on another predictor.
Bonferroni, Holm, and Hochberg procedures addresses multiplicity across planned or exploratory contrasts.

A coding scheme parameterizes group means; a contrast defines the scientific comparison.
Full-rank recoding changes coefficient labels, not fitted values or consistently computed estimable contrasts.
In unbalanced data, adjusted or marginal means depend on an explicit covariate distribution and weighting rule.
Empty cells, aliasing, and poor overlap are design problems that recoding cannot solve.
Biological inference follows the number and structure of independent samples, not the number of technical observations.

9 References

Fisher, R. A. (1935). The Design of Experiments. Oliver and Boyd.
Searle, S. R., Speed, F. M., and Milliken, G. A. (1980). Population marginal means in the linear model: an alternative to least squares means. The American Statistician, 34(4), 216–221. https://doi.org/10.1080/00031305.1980.10483031
Schad, D. J., Vasishth, S., Hohenstein, S., and Kliegl, R. (2020). How to capitalize on a priori contrasts in linear (mixed) models: a tutorial. Journal of Memory and Language, 110, 104038. https://doi.org/10.1016/j.jml.2019.104038
Gelman, A., Hill, J., and Vehtari, A. (2020). Regression and Other Stories. Cambridge University Press. https://doi.org/10.1017/9781139161879
R Core Team. contrasts: get and set contrast matrices, and model.matrix: construct design matrices. R statistical software documentation. https://stat.ethz.ch/R-manual/R-devel/library/stats/html/contrasts.html
Lenth, R. V. emmeans: estimated marginal means. R package documentation. https://rvlenth.github.io/emmeans/

--- title: "Categorical Predictors and Contrasts" subtitle: "Turning scientific comparisons among groups into estimable linear functions" description: "Indicator coding, planned contrasts, marginal means, uncertainty, and interpretation for categorical predictors in regression models." author: "Jiahong Sun" date: "2026-10-14" date-modified: "2026-10-14" categories: [Statistics, Regression, Categorical Predictors, Contrasts] toc: true number-sections: true code-fold: show draft: false --- # Categories Enter through a Design Matrix A categorical predictor records membership in a finite set of levels: treatment arm, genotype, tissue, sequencing batch, or disease subtype. Regression software cannot estimate a separate unconstrained coefficient for every level together with an intercept, because the level indicators sum to the intercept. It therefore represents a factor with $K$ levels using $K-1$ independent columns. For a one-factor conditional-mean model, $$ Y_i=\mu_{g(i)}+\varepsilon_i, \qquad \operatorname{E}(\varepsilon_i\mid g(i))=0, $$ the scientifically meaningful parameters are the level means $\boldsymbol\mu=(\mu_1,\ldots,\mu_K)^\mathsf{T}$. A coding scheme chooses a full-rank coordinate system for the same $K$ means. Changing coordinates changes the displayed coefficients, but not fitted values or any estimable comparison computed consistently from the model. This distinction is central: - **Coding coefficients** describe a parameterization. - **A contrast** describes a scientific comparison. - **An estimate** is the contrast evaluated from the fitted model. # Treatment Coding and Its Interpretation With three levels—control, drug A, and drug B—and control as the reference, treatment coding gives $$ \operatorname{E}(Y_i\mid g_i) =\beta_0+\beta_A I(g_i=A)+\beta_B I(g_i=B). $$ The coefficients are $$ \beta_0=\mu_C, \qquad \beta_A=\mu_A-\mu_C, \qquad \beta_B=\mu_B-\mu_C. $$ Thus the intercept is the control mean and each slope is a difference from control. Changing the reference level changes all three coefficient meanings, but the fitted group means remain $(\hat\mu_C,\hat\mu_A,\hat\mu_B)$. Reference coding is convenient when one level is a genuine scientific baseline. It is less useful when the questions concern a dose trend, an average of several controls, or A versus B. Those questions should be stated as contrasts rather than inferred indirectly from whichever coefficients software happens to print. # A Contrast Is the Estimand A linear contrast of level means is $$ \psi=\mathbf c^\mathsf{T}\boldsymbol\mu =\sum_{k=1}^{K}c_k\mu_k, \qquad \sum_{k=1}^{K}c_k=0. $$ The zero-sum condition makes the comparison invariant to adding the same constant to every level. Examples for $(C,A,B)$ include | Question | Contrast weights $\mathbf c$ | Estimand | |---|---:|---| | A versus control | $(-1,1,0)$ | $\mu_A-\mu_C$ | | B versus control | $(-1,0,1)$ | $\mu_B-\mu_C$ | | A versus B | $(0,1,-1)$ | $\mu_A-\mu_B$ | | Average drug versus control | $(-1,1/2,1/2)$ | $(\mu_A+\mu_B)/2-\mu_C$ | Let $\hat{\boldsymbol\beta}$ be the fitted coefficient vector with covariance $\widehat{\mathbf V}$. If the desired comparison is $\psi=\mathbf l^\mathsf{T}\boldsymbol\beta$, then $$ \hat\psi=\mathbf l^\mathsf{T}\hat{\boldsymbol\beta}, \qquad \widehat{\operatorname{SE}}(\hat\psi) =\sqrt{\mathbf l^\mathsf{T}\widehat{\mathbf V}\mathbf l}. $$ An interval or test follows from this estimate and standard error using a reference distribution appropriate to the design and covariance estimator. The contrast should be chosen because it answers a prespecified question, not because it produces the smallest p-value. # Other Coding Systems Different full-rank codings emphasize different coordinates: | Coding | Typical coefficient interpretation | Important caution | |---|---|---| | Treatment/reference | Each nonreference level minus a baseline | Intercept and slopes depend on the chosen reference | | Sum-to-zero | Deviations from an equally weighted grand mean | The grand mean is not the sample-size-weighted mean in an unbalanced design | | Helmert | A level compared with an average of subsequent or previous levels | Direction and scaling conventions differ among software | | Polynomial | Linear, quadratic, and higher-order patterns across ordered levels | Meaningful only when ordering and spacing are defensible | | Custom | Coordinates aligned with planned scientific hypotheses | Must have adequate rank and documented weights | Orthogonal contrast vectors satisfy $\mathbf c_j^\mathsf{T}\mathbf c_k=0$ for $j\ne k$. In a balanced one-way design, their sums of squares partition the factor sum of squares. In unbalanced or covariate-adjusted data, algebraic orthogonality of the weights does not necessarily imply uncorrelated estimates. A factor with $K$ levels contributes $K-1$ degrees of freedom. A single coefficient test answers one coordinate-specific question; an omnibus $K-1$ degree-of-freedom test asks whether all relevant level differences are zero. A nonsignificant omnibus test does not validate every equality, and a significant omnibus test does not identify which contrast matters. # Adjustment Changes What Is Being Compared With additional covariates, $$ \operatorname{E}(Y_i\mid g_i,\mathbf z_i) =\alpha+\text{factor terms}(g_i)+\mathbf z_i^\mathsf{T}\boldsymbol\gamma, $$ factor coefficients are conditional comparisons at the same $\mathbf z$. If there are no factor-by-covariate interactions, the adjusted difference is constant over $\mathbf z$. With interactions, there is no single group difference until values or a distribution of $\mathbf z$ are specified. Estimated marginal means compute model-based predictions for each level and average them over a declared reference distribution of covariates. Equal weighting across cells, weighting by the observed sample, and weighting by a target population answer different questions. In unbalanced observational data, calling one of them “the adjusted mean” without stating the averaging rule hides part of the estimand. Adjusted comparisons also require overlap. If a treatment occurs only in one batch, the treatment and batch columns are aliased: no coding scheme can recover their separate effects. Sparse treatment-by-tissue cells may technically permit a fit while making the requested contrast largely extrapolative. # Worked Example: Three Library-preparation Methods Suppose RNA integrity is measured for independent donor samples processed by a standard method $S$ or alternatives $A$ and $B$. The fitted adjusted means are $$ (\hat\mu_S,\hat\mu_A,\hat\mu_B)=(7.10,7.45,7.20). $$ The planned contrast comparing the average alternative with the standard method is $$ \hat\psi =-\hat\mu_S+\tfrac12\hat\mu_A+\tfrac12\hat\mu_B =0.225. $$ That estimate does not equal the A-versus-S coefficient, the B-versus-S coefficient, or an unweighted average of all three means. Its standard error must use the covariance among the three fitted means; treating their individual standard errors as independent is generally wrong. If the same donor material was split across methods, donor is a blocking or repeated-measures unit and the model must retain that pairing. If different donors received different methods, donors are the independent experimental units. Technical lanes, aliquots, and reads do not increase the biological replication unless the target explicitly concerns technical repeatability. # Diagnostics, Sensitivity, and Reporting Before interpreting a factor or contrast: 1. **Verify levels and reference meaning.** Check spelling, unused levels, ordering, and whether a numeric code was accidentally treated as continuous. 2. **Inspect the design matrix.** Confirm rank, columns, aliases, empty cells, and the coding actually used by the software. 3. **Inspect cell counts and overlap.** Report independent units per relevant factor combination, not only the total record count. 4. **Check the outcome model.** Residual, variance, dependence, and influence diagnostics still apply; factor coding does not repair misspecification. 5. **Define the contrast family.** Prespecify the comparisons and use simultaneous intervals or multiplicity adjustment when many claims form one family. 6. **Test coding invariance.** Refit with another full-rank coding and confirm that fitted values and explicitly reconstructed contrasts agree. 7. **State the averaging population.** For marginal means, document covariate values and weights. Report the contrast weights or plain-language comparison, estimate, uncertainty interval, scale and units, covariance method, adjustment set, and multiplicity procedure. A coefficient table alone is not a complete account of a multi-level factor. # Connections and Take-home Messages - [Multiple linear regression](multiple-linear-regression.qmd) defines the design-matrix model in which factor columns are fitted. - [Least squares and the geometry of regression](least-squares-and-regression-geometry.qmd) explains rank, estimability, and covariance of linear functions. - [Regression uncertainty and coefficient interpretation](regression-uncertainty-and-coefficient-interpretation.qmd) develops conditional coefficients, contrasts, and prediction uncertainty. - [Interaction terms and effect modification](interaction-terms-and-effect-modification.qmd) shows why a factor effect can depend on another predictor. - [Bonferroni, Holm, and Hochberg procedures](../statistical-inference/bonferroni-holm-and-hochberg-procedures.qmd) addresses multiplicity across planned or exploratory contrasts. 1. A coding scheme parameterizes group means; a contrast defines the scientific comparison. 2. Full-rank recoding changes coefficient labels, not fitted values or consistently computed estimable contrasts. 3. In unbalanced data, adjusted or marginal means depend on an explicit covariate distribution and weighting rule. 4. Empty cells, aliasing, and poor overlap are design problems that recoding cannot solve. 5. Biological inference follows the number and structure of independent samples, not the number of technical observations. # References - Fisher, R. A. (1935). *The Design of Experiments*. Oliver and Boyd. - Searle, S. R., Speed, F. M., and Milliken, G. A. (1980). Population marginal means in the linear model: an alternative to least squares means. *The American Statistician*, 34(4), 216--221. <https://doi.org/10.1080/00031305.1980.10483031> - Schad, D. J., Vasishth, S., Hohenstein, S., and Kliegl, R. (2020). How to capitalize on a priori contrasts in linear (mixed) models: a tutorial. *Journal of Memory and Language*, 110, 104038. <https://doi.org/10.1016/j.jml.2019.104038> - Gelman, A., Hill, J., and Vehtari, A. (2020). *Regression and Other Stories*. Cambridge University Press. <https://doi.org/10.1017/9781139161879> - R Core Team. `contrasts`: get and set contrast matrices, and `model.matrix`: construct design matrices. *R statistical software documentation*. <https://stat.ethz.ch/R-manual/R-devel/library/stats/html/contrasts.html> - Lenth, R. V. `emmeans`: estimated marginal means. *R package documentation*. <https://rvlenth.github.io/emmeans/>

Question	Contrast weights \(\mathbf c\)	Estimand
A versus control	\((-1,1,0)\)	\(\mu_A-\mu_C\)
B versus control	\((-1,0,1)\)	\(\mu_B-\mu_C\)
A versus B	\((0,1,-1)\)	\(\mu_A-\mu_B\)
Average drug versus control	\((-1,1/2,1/2)\)	\((\mu_A+\mu_B)/2-\mu_C\)