From Conditional Predictions to a Population Contrast
Regression adjustment offers a way to compare treatment and control while accounting for measured pretreatment covariates. Its central operation is straightforward: use a fitted outcome model to predict each person's outcome under treatment and under control, then average each set of predictions over the same covariate distribution. The resulting contrast is not generally the treatment coefficient from a regression table. It is a standardized comparison whose interpretation depends on both causal assumptions and the adequacy of the outcome model.
Let \(A\) be a binary treatment, with \(A=1\) denoting treatment and \(A=0\) control. Let \(X\) denote the measured pretreatment covariates, and let \(Y(1)\) and \(Y(0)\) be the potential outcomes under treatment and control. The observed outcome is \(Y\). For \(a\in\{0,1\}\), define the conditional outcome mean \(m_a(x)=E[Y\mid A=a,X=x]\). Regression standardization estimates these conditional means, predicts them for each covariate value in a target population, and averages the predictions.
The empirical formula targets the covariate distribution of the \(n\) individuals being averaged over. In an observational sample, this may be the full study sample; a different target population would require averaging over that population's covariate distribution. The definition specifies the calculation, but by itself it does not make the contrast causal.
Assumptions for a Causal Interpretation
For the standardized conditional means to identify the average causal contrast, the standard adjustment assumptions must hold for the covariates \(X\). First, consistency requires that a person's observed outcome equals the potential outcome corresponding to the treatment actually received. Second, conditional exchangeability requires \(\{Y(0),Y(1)\}\perp A\mid X\): within covariate values, treatment assignment is independent of the potential outcomes. Third, positivity requires both treatment options to occur with positive probability at the covariate values represented in the target population: \(0<P(A=1\mid X=x)<1\).
Under these assumptions, the observed conditional mean in each treatment group identifies the corresponding potential-outcome mean: \(E[Y(a)\mid X=x]=E[Y\mid A=a,X=x]\). Averaging over \(X\) then identifies the population contrast. In addition, the fitted outcome model must estimate these conditional means adequately over the relevant covariate support. Causal assumptions cannot rescue an incorrectly specified outcome model; conversely, an excellent predictive fit cannot establish exchangeability or consistency.
The first equality averages conditional contrasts over the covariate distribution. The third equality uses conditional exchangeability and consistency to replace each potential-outcome mean with an observed conditional mean. Positivity is needed so those observed means are defined by data at the covariate values being averaged. In practice, the means are estimated, so correct model form and sufficient data also matter.
Calculating the Standardized Contrast
Worked Example: Adjustment Across Two Covariate Strata
Suppose a hypothetical observational study includes 100 people. \(X=0\) denotes a lower-risk baseline group and \(X=1\) a higher-risk group. The table gives the number of people and the mean outcome in each treatment-by-\(X\) cell. The outcome is a symptom score; higher values mean worse symptoms.
| Baseline group | Treated: count | Treated: mean score | Control: count | Control: mean score |
|---|---|---|---|---|
| \(X=0\) | 30 | 8 | 20 | 6 |
| \(X=1\) | 10 | 12 | 40 | 10 |
A saturated model for the two covariate groups gives fitted conditional means equal to the cell means: \(\hat{m}_1(0)=8\), \(\hat{m}_0(0)=6\), \(\hat{m}_1(1)=12\), and \(\hat{m}_0(1)=10\). Each treatment group occurs in each stratum, so the table contains direct information for all four conditional means. The proportions with \(X=0\) and \(X=1\) in the full sample are both \(50/100=0.5\).
Standardize the treatment predictions over that common distribution: \(\hat{\mu}_1=0.5(8)+0.5(12)=10\) symptom-score points, and \(\hat{\mu}_0=0.5(6)+0.5(10)=8\) points. Therefore, \(\hat{\tau}=10-8=2\) points. If consistency and conditional exchangeability hold, and the cell means represent the conditional means, the estimate is a two-point higher average score under treatment for this sample's covariate distribution.
For comparison, the unadjusted treated mean is \([30(8)+10(12)]/40=9\), and the unadjusted control mean is \([20(6)+40(10)]/60=8.667\). Their difference is \(0.333\) points, not 2. The crude comparison mixes treatment groups with different proportions of the baseline-risk groups; standardization instead compares predictions using the same 50:50 distribution.
Model Form and Treatment-Effect Differences
With continuous covariates, regression provides a compact way to estimate conditional means across many covariate values. The model must represent the relationship between \(X\), treatment, and the conditional outcome mean well enough for the intended predictions. A linear model that omits curvature or important interactions can give biased predictions even if the causal assumptions hold. In particular, a model that includes treatment but no treatment-by-covariate interaction imposes a constant conditional treatment contrast on its modeled scale.
Worked Example: Standardizing a Model With an Interaction
Consider a hypothetical study of a treatment's effect on a symptom score, with baseline severity \(X\) measured in points. Suppose a fitted conditional-mean model is \(\hat{m}(A,X)=50+4A+2X+3AX\). Here \(A\) is 0 or 1, and all predictions are symptom-score points. Assume the form is adequate over the observed covariate range and that the causal assumptions are credible.
For control, set \(A=0\), giving \(\hat{m}_0(X)=50+2X\). For treatment, set \(A=1\), giving \(\hat{m}_1(X)=54+5X\). The fitted conditional contrast is therefore \(\hat{m}_1(X)-\hat{m}_0(X)=4+3X\); it varies with baseline severity.
Suppose the four study participants have \(X=0,1,2,3\). Their predicted treatment outcomes are 54, 59, 64, and 69, so \(\hat{\mu}_1=(54+59+64+69)/4=61.5\). Their predicted control outcomes are 50, 52, 54, and 56, so \(\hat{\mu}_0=(50+52+54+56)/4=53\). The standardized contrast is \(61.5-53=8.5\) symptom-score points. Equivalently, the four predicted individual contrasts are 4, 7, 10, and 13, whose mean is \((4+7+10+13)/4=8.5\).
If the interaction term were omitted, the fitted model would constrain the treatment contrast to be 4 points for every severity value. That may be a useful approximation in some settings, but it is a substantive restriction, not a consequence of regression adjustment. The predictions and final standardized contrast depend on which functional form is used.
A linear regression can be fit to a continuous outcome, but the standardization idea is not limited to linear models. For other outcomes, one can fit a suitable conditional-mean model and predict the outcome on its original scale before averaging. The key is to average the conditional outcome predictions on the scale of the target contrast, rather than automatically interpreting a model coefficient on a link or transformed scale as the desired average effect.
Overlap and Extrapolation
Positivity has a practical counterpart: overlap. At a given covariate value, there must be observations in both treatment groups to support comparison there. When one group is absent or rare, the outcome model must predict what that group would have experienced in a region where it was not observed. Such extrapolation may be driven largely by the chosen functional form rather than by data. A model can still return a number, but the number may be highly assumption-dependent.
Worked Example: A Prediction Outside Observed Support
In a hypothetical study, 100 people have baseline severity \(X=0\) or \(X=10\). There are 40 treated and 40 control participants with \(X=0\), and 20 treated but no control participants with \(X=10\). Thus, treatment and control overlap at \(X=0\), but not at \(X=10\). Suppose a fitted model produces \(\hat{m}_0(X)=10+2X\) and \(\hat{m}_1(X)=14+2X\).
The sample has 80% of people at \(X=0\) and 20% at \(X=10\). The model predicts average treatment outcome \(0.8(14)+0.2(34)=18\), and average control outcome \(0.8(10)+0.2(30)=14\). The resulting standardized contrast is \(18-14=4\) outcome units.
The prediction \(\hat{m}_0(10)=30\) is unsupported by any control observations at \(X=10\). The observed data cannot determine the control mean for that group without an additional modeling assumption. For example, a conditional contrast of 4 units at \(X=0\) and 14 units at \(X=10\) would yield \(0.8(4)+0.2(14)=6\) units over this sample's covariate distribution, while remaining compatible with the observed treatment outcomes and the available control outcomes at \(X=0\). Thus, the model's 4-unit result is not secured by the data at \(X=10\); it depends on extrapolating the assumed model there. The lack of overlap also means positivity fails for this full target distribution.
Useful diagnostics include comparing covariate distributions by treatment group, identifying regions with few observations in either arm, and checking whether the standardized predictions rely on values far from observed data. Outcome-model diagnostics can reveal patterns such as curvature or unequal residual spread, but good residual behavior alone does not prove that the model is correct for unobserved treatment-covariate combinations. Nor does overlap establish conditional exchangeability: it addresses whether both treatment states are represented, not whether all relevant common causes have been measured.
What Regression Adjustment Does—and Does Not—Do
Stratified adjustment and regression standardization share the same basic logic: estimate treatment-specific conditional means and average their contrast over a common covariate distribution. Stratification makes comparisons within explicit groups and can be transparent when there are few discrete covariates. Regression can represent continuous covariates and many covariates more compactly, but its conclusions can be sensitive to functional-form choices and extrapolation. A raw regression treatment coefficient is not automatically the standardized contrast, especially when the model includes treatment interactions or uses a nonlinear link.
Propensity-score approaches model treatment assignment rather than the outcome and use that information to construct adjusted comparisons. They provide a different route to adjustment, not a way to bypass the same need for credible measured-covariate adjustment and adequate support. Regardless of approach, causal interpretation requires the relevant assumptions; statistical adjustment cannot establish them from the fitted model alone.
A common mistake is to say, “The regression includes the covariates, so its treatment coefficient is the adjusted causal effect.” That reasoning is incomplete. The coefficient may represent a conditional contrast at a reference value, or a model-scale contrast, rather than an average over the study population. The correction is to state the target contrast, generate treatment-specific predictions for the same covariate values, average each set, and subtract. Then assess whether assumptions and overlap support a causal interpretation.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- In one sentence, describe the averaging step that distinguishes standardization from simply reporting a treatment coefficient.
- A sample has two equally common covariate values. Fitted treatment contrasts are 3 and 9 outcome units. What is the standardized contrast?
- Why can a model produce a standardized estimate even when there are no control observations for some covariate values, and what is the problem with that estimate?
- Does good covariate overlap establish conditional exchangeability? Explain.
Solutions. (1) Predict the outcome under treatment and control for the same observed covariate values, average each set of predictions, and take their difference. (2) The standardized contrast is \(0.5(3)+0.5(9)=6\) outcome units. (3) The fitted model can extrapolate the missing conditional mean, but the prediction depends on model assumptions not checked by observations in that region; positivity is not satisfied there. (4) No. Overlap indicates that both treatment states are represented across covariate values; it does not show that measured covariates suffice to make treatment conditionally exchangeable.