The 'ABCs' of `lmabc`: linear regression with categorical covariates

Overview

Regression analysis commonly features categorical (or nominal) covariates, such as race, sex, group/experimental assignments, and many other examples. These variables may appear alongside other covariates, as in the main-only (or ANCOVA) model

y ~ x + race

or interacted with other variables, as in the group-modified model

y ~ x + race + x:race

(here we use race as our default categorical variable for clarity).

The group-modified model estimates group-specific effects of x on the response y (e.g., does exposure to a pollutant x more adversely impact the health y of certain subpopulations?). Specifically, the linear model expresses the expectation of the response variable y at given x and race values: $$E(Y | X = x, race = r) = \mu(x, r) = \alpha_0 + \alpha_1 x + \beta_r + \gamma_r x$$ (here x is continuous). As a result, we obtain group-specific intercepts, $$\mu(0, r) = \alpha_0 + \beta_r$$ and group-specific slopes, $$\mu(x+1, r) - \mu(x,r) = \alpha_1 + \gamma_r.$$ The main-only model is recovered when $\gamma_r = 0$ for all race groups $r$. Then the slope is global and does not depend on race: $\mu(x+1, r) - \mu(x,r) = \alpha_1$.

However, both the main-only and the group-modified models have too many parameters. This is known as the “dummy variable trap”: to numerically encode $L$ levels of a categorical variable, only $L-1$ dummy variables are needed. Here, both models have group-specific intercepts, yet these are parametrized by group-specific coefficients $\beta_r$and a global coefficient $\alpha_0$. Clearly, this uses one more parameter than necessary. A similar problem occurs for group-specific slopes. Without modification, these parameters are not estimable or interpretable.

The central question is then, how best to parametrize (or constrain) these group-specific intercepts and slopes? The choice has significant implications for statistical efficiency, equitability, and interpretability.

Example Dataset

We generate a simulated dataset with a continuous response variable y, a continuous covariate x, and two categorical variables, race and sex.

To mimic the challenges of real data analysis, the simulated covariates are dependent: x depends on race, both in mean and distribution,

while race and sex are highly dependent categorical variables:

Simulated data proportions: sex (rows) and race (columns)

	A	B	C	D
uu	0.292	0.126	0.058	0.078
vv	0.056	0.028	0.104	0.258

We will consider estimation and inference with various combinations of the predictors x, race, and sex (and their interactions). The race and sex variables use arbitrary labeling to avoid any misleading race- or sex-specific effects in our example regression output.

Default strategies: the problem

Suppose we call lm to fit the group-modified model:

fit_lm = lm(y ~ x + race + x:race)

The summary() output appears as follows:

Default output: lm(y ~ x + race + x:race)

	Estimate	Std. Error	t value	Pr(>|t|)
(Intercept)	1.95	0.35	5.62	0.00
x	1.51	0.07	22.40	0.00
raceB	-1.27	0.41	-3.14	0.00
raceC	-2.36	0.60	-3.94	0.00
raceD	-0.96	0.36	-2.64	0.01
x:raceB	-0.86	0.12	-7.17	0.00
x:raceC	-0.59	0.12	-4.93	0.00
x:raceD	-0.56	0.09	-6.02	0.00

Immediately, we notice the absence of raceA and x:raceA. This occurs by design: lm uses reference group encoding (RGE), which parametrizes the model by deleting a reference group, here $\beta_A = 0$ and $\gamma_A = 0$. There are several limitations of this approach.

First, the x effect is misleading: because $\gamma_A = 0$, the “global” slope parameter is $$\alpha_1 = \alpha_1 + \gamma_A = \mu(x+1, A) - \mu(x,A),$$ i.e., the group-specific x effect for the reference group (race = A). However, this is not clear from the model output, which instead appears to present a “global” x effect. This presentation invites mistaken conclusions about the x effect for the broader population.

Second, this output is inequitable: it elevates one group above all others. The reference group is often selected to be White for race, Male for sex, etc., and thus induces a bias in the presentation of results. Similarly, the x:race effects are presented as deviations from the reference group x effect: for example x:raceB refers to the difference between the x effect for group B and the x effect for the reference group A. This implicitly treats one group as “normal” and the others as “deviations from normal.”

Third, RGE is not well-designed to include interactions like x:race. In addition to the difficulties with interpretations and equitability, RGE is statistically inefficient for the main effects. To see this, consider the estimates and inference for the x effect under the main-only model y ~ x + race:

Estimated x effect: lm(y ~ x + race)

	Estimate	Std. Error	t value	Pr(>|t|)
x	1.08	0.04	26.39	0

The estimates and standard errors of the x effect are considerably different. This occurs because here, the x effect refers to a global slope, $\alpha_1 = \mu(x+1, r) - \mu(x,r)$, rather than a (reference) group-specific slope—even though the output appears exactly the same. Usually, the standard errors will increase when x:race is included, since the slope is restricted to a subset of the data.

In aggregate, default lm output under RGE is at best difficult to interpret and at worst outright misleading. It suffers from alarming inequities and sacrifices statistical efficiency in the presence of (x:race) interactions. These issues also occur for the intercept parameters and compound for multiple categorical variables and interactions.

Abundance-Based Constraints (ABCs): the solution

Instead, suppose we call lmabc to fit the group-modified model:

library(lmabc)
fit_lmabc = lmabc(y ~ x + race + x:race)

The summary() output appears as follows:

ABC output: lmabc(y ~ x + race + x:race)

	Estimate	Std. Error	t value	Pr(>|t|)
(Intercept)	2.85	0.14	20.71	0.00
x	1.09	0.04	28.23	0.00
raceA	1.58	0.19	8.28	0.00
raceB	-1.10	0.16	-6.71	0.00
raceC	-1.74	0.55	-3.18	0.00
raceD	-0.29	0.14	-2.03	0.04
x:raceA	0.42	0.05	7.74	0.00
x:raceB	-0.44	0.09	-4.86	0.00
x:raceC	-0.18	0.09	-1.94	0.05
x:raceD	-0.14	0.05	-2.70	0.01

First, every race group is represented: the results do not elevate any single race group above the others. This eliminates the presentation bias and provides more equitable output.

Second, the x effect estimates and standard errors are nearly identical to those in the main-only model (${\hat\alpha}_1 =$ 1.08, $SE({\hat\alpha}_1) =$ 0.04). This illustrates two remarkable invariance properties of ABCs:

1. The estimated x effects under y ~ x + race and y ~ x + race + x:race are nearly identical; and
   
2. The standard errors of the x effects under y ~ x + race and y ~ x + race + x:race are
   1. Nearly identical when the x:race effect is small or
   2. Smaller for the group-modified model when the x:race effect is large.

In effect, ABCs allow the inclusion of (x:race) interactions “for free”: they have (almost) no impact on estimation and inference for the main x effect. With ABCs, the analyst can estimate group-specific x effects without worrying that the addition of x:race will sacrifice power for the main x effect (which occurs for RGE). And, when the interaction effect x:race is substantial, the analyst gains more power for the main x effect.

We emphasize several features of these invariance properties:

• They are unique to ABCs and do not occur for alternative approaches (default RGE, sum-to-zero constraints, etc.);
• They make no requirements about the true data-generating process; and
• They allow for dependencies between x and race (as in this simulated dataset).

The only condition is that, for continuous x, the scale of x must be approximately the same within each race group (no conditions are needed when x is categorical; see below). This is reasonable: if a “one-unit change in x” is not comparable for different race groups, then only the group-modified model that includes race-specific slopes is meaningful. In that case, there is no reason to consider the x effect under the main-only model. Empirically, these (near) invariance results are quite robust to this condition.

Finally, these results improve interpretability: all group-specific coefficients $\gamma_r$ (e.g., x:raceB) now represent the difference between the group-specific slope, $\mu(x+1, r) - \mu(x,r)$, and the properly global x effect $\alpha_1$. Similar interpretations apply to the intercept parameters.

ABCs: some details

ABCs identify the group-specific parameters by constraining the group averages to be zero, $$E_\pi(\gamma_R) = \sum_r \pi_r \gamma_r = 0$$ where $R$ denotes a categorical random variable (e.g., race) with probabilities $Pr(R = r) = \pi_r$. A similar constraint is then used for the group-specific intercept parameters: $E_\pi(\beta_R) = \sum_r \pi_r \beta_r = 0$. These linear constraints are constructed using getConstraints() and enforced during estimation, for example using ordinary least squares lmabc(), maximum likelihood glmabc(), and penalized least squares cv.penlmabc(). Modifications are available for categorical-categorical interactions.

The constraints above are general and include many special cases: RGE uses $\pi_A = 1$ (reference) and $\pi_r = 0$ otherwise, while sum-to-zero constraints use equal weights $\pi_r = 1$ for all $r$. However, the choice of $\{\pi_r\}$ is critical for equitability, statistical efficiency, and interpretability.

ABCs use the empirically-observed proportions by group, $\pi_r=$mean(race == r). As such, the parameters have a genuine “group-average” interpretation. For instance, the global slope parameter in the group-modified model is $$\alpha_1 = E_{\pi}\{\mu(x+1, R) - \mu(x, R)\} = \sum_r \pi_r \{\mu(x+1, r) - \mu(x, r)\}$$ i.e., the group average of the group-specific x effects. Similar interpretations are available for the global intercept: $$\alpha_0 = E_{\pi}\{\mu(0, R)\} = \sum_r \pi_r \mu(0, r).$$ Because ABCs identify properly global (intercept and slope) parameters, the group-specific coefficients are interpretable as the difference between the group-specific x effect (or intercept) and the global/group-averaged x effect (or intercept). There is no need to elevate any single (reference) group.

ABCs may also use the population group proportions, if those are known and passed to the function.

Interpeting the lmabc output

Revisiting the output from lmabc(y ~ x + race + x:race), we summarize the main conclusions:

• The global x effect is significant and positive (${\hat\alpha}_1 = 1.09$, $SE({\hat\alpha}_1) = 0.04$).
• The x:race interaction effects show that the group-specific x effect is significantly larger than average for race = A (${\hat\gamma}_A = 0.42$, $SE({\hat\gamma}_A) = 0.05$), significantly smaller than average for race = B (${\hat\gamma}_B = -0.44$, $SE({\hat\gamma}_B) = 0.09$) and race = D (${\hat\gamma}_D = -0.14$, $SE({\hat\gamma}_D) = 0.05$), and somewhat smaller than average for race = C (${\hat\gamma}_C = -0.18$, $SE({\hat\gamma}_C) = 0.09$).
• The group-specific slopes are computed by summing the relevant coefficients, for example $\mu(x+1, A) - \mu(x,A) = \alpha_1 + \gamma_A = 1.51$ is the group-specific x effect for race = A.
• Similar interpretations apply for the intercept coefficients.

Categorical-categorical interactions

The case of categorical-categorical interactions is arguably even more cumbersome for default (RGE) methods—and yet ABCs offer an even cleaner solution. Consider the lm output for two models that feature the categorical covariates race and sex: the main-only model y ~ race + sex and the group-modified model, y ~ race + sex + race:sex.

Default output: lm(y ~ race + sex) Estimate Std. Error t value Pr(>|t|) (Intercept) 9.58 0.12 81.70 0.00 raceB -7.69 0.21 -37.32 0.00 raceC -14.38 0.22 -66.12 0.00 raceD -7.50 0.19 -39.23 0.00 sexvv -0.02 0.16 -0.09 0.93

Default output: lm(y ~ race + sex + race:sex) Estimate Std. Error t value Pr(>|t|) (Intercept) 9.59 0.12 76.84 0.00 raceB -7.71 0.23 -33.89 0.00 raceC -14.63 0.31 -47.72 0.00 raceD -7.40 0.27 -27.22 0.00 sexvv -0.11 0.31 -0.35 0.72 raceB:sexvv 0.08 0.54 0.15 0.88 raceC:sexvv 0.46 0.47 0.98 0.33 raceD:sexvv -0.05 0.42 -0.13 0.90

In both cases, the reference groups for race (here, A) and sex (here, uu) are absent from all main and interaction effects. The output again is misleading: even for the simpler model y ~ race + sex, the main effects require consideration of both reference groups. For example, the intercept estimates $\alpha_0 = \mu(race = A, sex = uu)$, i.e., the expected response for race = A and sex = uu. Similarly, the main effects such as sexvv estimate $\mu(race = A, sex = vv) - \mu(race = A, sex = uu)$, i.e., the difference in the expected responses between sex = vv and sex = uu but only for race = A. When the reference groups are set at the usual values (White for race and Male for sex), these parametrizations are clearly inequitable. Finally, we see that the standard errors for the main effects increase when the race:sex interaction is added to to the model.

ABCs completely circumvent these issues. Consider the same two models, but now subject to ABCs:

ABC output: lmabc(y ~ race + sex) Estimate Std. Error t value Pr(>|t|) (Intercept) 3.54 0.07 52.49 0.00 raceA 6.03 0.10 58.32 0.00 raceB -1.66 0.16 -10.13 0.00 raceC -8.35 0.16 -53.33 0.00 raceD -1.46 0.11 -13.49 0.00 sexuu 0.01 0.07 0.09 0.93 sexvv -0.01 0.09 -0.09 0.93

ABC output: lmabc(y ~ race + sex + race:sex) Estimate Std. Error t value Pr(>|t|) (Intercept) 3.54 0.07 52.41 0.00 raceA 6.03 0.10 58.23 0.00 raceB -1.66 0.16 -10.11 0.00 raceC -8.35 0.16 -53.25 0.00 raceD -1.46 0.11 -13.47 0.00 sexuu 0.01 0.07 0.09 0.93 sexvv -0.01 0.09 -0.09 0.93 raceA:sexuu 0.02 0.04 0.36 0.72 raceB:sexuu 0.00 0.08 0.04 0.97 raceC:sexuu -0.23 0.20 -1.19 0.24 raceD:sexuu 0.11 0.17 0.67 0.50 raceA:sexvv -0.08 0.22 -0.36 0.72 raceB:sexvv -0.01 0.34 -0.04 0.97 raceC:sexvv 0.13 0.11 1.19 0.24 raceD:sexvv -0.03 0.05 -0.67 0.50

First, each group is present in both the main effects and interactions. There is no need to consider reference groups, and thus no single (race or sex) group is elevated above the others.

Second, all main effect estimates—including the (Intercept), all race effects, and both sex effects—are identical between the models that do and do not include the race:sex interaction. Unlike the previous setting with continuous-categorical interactions (x:race), this estimation invariance is exact. Importantly, this result makes no requirements on the true data-generating process or the categorical covariates, which here are highly dependent.

Similarly, the main effect standard errors—again for the (Intercept), all race effects, and both sex effects—are (almost) identical between the two models.

We summarize these (provable) invariance properties of ABCs for categorical-categorical interactions:

1. The estimated (Intercept), race, and sex effects under y ~ race + sex and y ~ race + sex + race:sex are identical; and
   
2. The standard errors of (Intercept), race, and sex under y ~ race + sex and y ~ race + sex + race:sex are
   1. Nearly identical when the race:sex effect is small or
   2. Smaller for the group-modified model when the race:sex effect is large.

Again, the analyst may include (race:sex) interaction effects “for free”: the main effect estimates are unchanged, and the statistical power for the main effects can only increase. Thus, we require that all main effects are included for any categorical-categorical or categorical-continuous interactions.

Interpreting ABCs with care

By design, ABCs leverage the (sample or population) categorical proportions to provide 1) more equitable output, 2) greater statistical efficiency, and 3) more interpretable parameters and estimates with properly global (i.e., group-averaged) main effects. However, the group-specific coefficients must be interpreted carefully in the context of the abundances $\{\pi_r\}$.

For instance, consider the simple model y ~ sex,

ABC output: lmabc(y ~ sex)

	Estimate	Std. Error	t value	Pr(>|t|)
(Intercept)	3.54	0.22	16.22	0
sexuu	1.73	0.20	8.84	0
sexvv	-2.15	0.24	-8.84	0

With two groups, ABCs imply that one effect must be positive and the other must be negative, as the two must average to zero. These effects are partly determined by the group abundances, mean(sex=="uu") = 0.55 and mean(sex=="vv") = 0.45. Because vv has a lower proportion, its estimated coefficient must be higher (in absolute value) to satisfy ABCs. Thus, we cannot merely interpret the vv effect as “larger than” the uu effect.

Fortunately, the standard errors inflate proportionally: hence, the t value statistics (Estimate / Std. Error) are equal and opposite. Similarly, the p-values will be identical for these (sexuu and sexvv) main effects.

Even with these caveats, ABCs offer an appealing parametrization of this ANOVA model. First, the estimated intercept exactly equals the sample mean, mean(y) = 3.54. Second, the sex-specific coefficients exactly equal the difference between the group-specific means and the overall mean, mean(y[sex=="uu"]) - mean(y) = 1.73. This is certainly a natural way to parametrize the group-specific and global effects for this model.

Additional details about lmabc

The lmabc package includes implementations for many common methods: summary, coef, print, plot, predict, logLik, vcov, and more.

lmabc also includes methods for generalized linear models (GLMs) with categorical covariates (glmabc) and penalized (lasso and ridge) regression with cross-validation (cv.penlmabc). These methods, like lmabc, can handle multiple continuous and categorical covariates and their interactions. We note a few points:

• The invariance properties of ABCs remain valid for multiple continuous and categorical covariates and their interactions. The conditions change slightly (see the reference below) but approximate invariance applies quite generally.
• For GLMs (glmabc), ABCs offer equitability and interpretability, but estimation invariance applies only for OLS. This work is currently under development.
• For penalized (ridge or lasso) regression, ABCs are immensely valuable. Because these penalized estimators “shrink” the coefficients toward zero, default RGE estimates of group-specific effects are statistically biased toward the reference group. This is especially concerning for protected groups (race, sex, religion, etc.) but also implies that 1) estimates and predictions depend on the choice of reference group and 2) differences between group-specific x effects and reference group x effects are attenuated and obscured. ABCs resolve these critical limitations, again by providing efficient estimation and shrinkage toward properly global coefficients. The statistical properties of these estimators are currently under development.

References

Kowal, D. (2024). Facilitating heterogeneous effect estimation via statistically efficient categorical modifiers. https://arxiv.org/abs/2408.00618

Kowal, D. (2024). Regression with race-modifiers: towards equity and interpretability. https://doi.org/10.1101/2024.01.04.23300033