Why Covariate Adjustment Can Identify Causal Effects
In an observational comparison, people who receive a treatment may differ from those who do not. A raw difference in their mean outcomes can therefore combine a treatment effect with pre-existing differences between the groups. If those differences are captured by observed pre-treatment covariates, a comparison within covariate values can have a causal interpretation under specific assumptions. This tutorial derives the assumptions and the resulting adjustment formulas.
Let \(D\) be a binary treatment indicator, with \(D=1\) for treatment and \(D=0\) for control. Let \(Y(1)\) and \(Y(0)\) denote the potential outcomes for an individual under treatment and control, respectively. The observed outcome is \(Y\). Let \(X\) be a set of measured covariates determined before treatment. The target population is the population over which the effects are defined.
These are different targets: the ATE averages over the covariate distribution of the whole target population, while the ATT averages over the covariate distribution of treated individuals. Both concern causal contrasts, but neither is automatically equal to the observed difference \(E[Y\mid D=1]-E[Y\mid D=0]\).
Assumptions for Identification
Identification asks whether a population quantity can be uniquely expressed in terms of the distribution of observed data under stated assumptions. Two conditions are central here. The first makes treatment and control individuals comparable after conditioning on \(X\); the second ensures that both treatment states occur at the relevant covariate values.
This is also called conditional exchangeability or selection on observables. It is an assumption about the treatment assignment mechanism and potential outcomes, not a consequence of fitting a model. It is more plausible when \(X\) includes relevant pre-treatment causes of both treatment choice and the outcome. Conditioning on a variable affected by treatment does not satisfy the intended role of a pre-treatment adjustment set.
Overlap is also called positivity. For the ATE, the condition is needed across the covariate support of the target population. Requiring it over that full support is sufficient for identifying the ATT as well, though the ATT only needs control outcomes to be available over covariate values represented among the treated. Overlap is a population condition; a finite sample may still contain few or no observations in one group in some regions.
We also use consistency: for each individual, the observed outcome equals the potential outcome under the treatment actually received, so \(Y=DY(1)+(1-D)Y(0)\). This presumes a well-defined treatment and no interference between individuals’ treatment assignments. Without these conditions, the observed outcome may not correspond to the potential outcome indexed by \(D\).
Deriving the ATE Adjustment Formula
Write the conditional mean difference at covariate value \(x\) as \(\tau(x)=E[Y\mid D=1,X=x]-E[Y\mid D=0,X=x]\). Under the assumptions, this difference equals the conditional average treatment effect. The steps below show why and then average that contrast over the target population’s distribution of \(X\).
By consistency, among individuals with \(D=1\), the observed \(Y\) is \(Y(1)\). Thus \(E[Y\mid D=1,X=x]=E[Y(1)\mid D=1,X=x]\).
Conditional independence implies \(E[Y(1)\mid D=1,X=x]=E[Y(1)\mid X=x]\). Applying the same reasoning to control individuals gives \(E[Y\mid D=0,X=x]=E[Y(0)\mid X=x]\).
The within-\(x\) observed difference is \(E[Y(1)-Y(0)\mid X=x]\). Averaging over the target population’s covariate distribution gives the ATE.
The last expectation is over the covariate distribution in the target population, not over the treated group alone. Overlap ensures that the conditional means for both treatment states are defined for the covariate values being averaged. In a discrete-strata setting, this expectation is a weighted sum; with continuous covariates, it is an integral over their target-population distribution.
Deriving the ATT Adjustment Formula
For the ATT, the target population is restricted to individuals with \(D=1\). Their treated potential outcomes are observed by consistency. Their control potential outcomes are not observed, but conditional independence allows the mean control potential outcome for treated individuals at a given \(x\) to be recovered from control individuals with the same \(x\). The covariate values must be averaged according to the treated group’s distribution.
In the final equality, consistency identifies the treated mean, while conditional independence identifies the treated group’s missing control mean from the conditional control mean. The formula averages the same conditional treatment contrast used for the ATE, but uses the covariate distribution among treated individuals rather than the population distribution.
Worked Example: ATE and ATT From Two Covariate Strata
Worked Example: ATE and ATT From Two Covariate Strata
Hypothetical data. Suppose a target population contains 100 individuals in two pre-treatment strata, defined by a risk category. Within each stratum there are individuals in both treatment groups. Outcomes are measured in points, and the conditional means are:
| Risk stratum | Population count | Treated count | Control count | Mean outcome if treated | Mean outcome if control |
|---|---|---|---|---|---|
| Low | 60 | 20 | 40 | 11 | 8 |
| High | 40 | 30 | 10 | 18 | 12 |
Assume consistency, conditional independence given risk stratum, and overlap. Overlap is present in both strata because each has at least one treated and one control individual; the population treatment probabilities are \(20/60\) in the low-risk stratum and \(30/40\) in the high-risk stratum, both strictly between zero and one.
The low-risk conditional contrast is \(11-8=3\) points, and the high-risk contrast is \(18-12=6\) points. The ATE weights these contrasts by the population proportions in each stratum:
Thus, under the assumptions, the identified ATE is 4.2 points in this hypothetical population. For the ATT, the treated group has 20 low-risk and 30 high-risk individuals, so its stratum proportions are \(20/50=0.4\) and \(30/50=0.6\). Therefore:
The ATT exceeds the ATE here because treated individuals are more concentrated in the high-risk stratum, where the conditional contrast is larger. This does not mean that risk category itself causes the treatment effect difference; the example only shows how differing target-population weights can produce different averages.
Worked Example: Why the Raw Difference Is Not the ATE
Worked Example: Why the Raw Difference Is Not the ATE
Use the same hypothetical stratum counts and conditional means. The overall treated mean is \(0.4(11)+0.6(18)=4.4+10.8=15.2\) points, because 40% of treated individuals are low-risk and 60% are high-risk. The overall control mean is \(0.8(8)+0.2(12)=6.4+2.4=8.8\) points, because 80% of controls are low-risk and 20% are high-risk.
The raw treated-minus-control difference is therefore \(15.2-8.8=6.4\) points. It is not the ATE of 4.2 points or the ATT of 4.8 points. The two groups have different stratum compositions, so this crude comparison combines within-stratum outcome contrasts with differences in which strata the groups represent.
For the ATE, reweight the conditional contrasts using the whole population’s proportions: \(0.6(3)+0.4(6)=4.2\) points. For the ATT, use the treated proportions: \(0.4(3)+0.6(6)=4.8\) points. Conditional independence supports the within-stratum contrasts; the choice of weights determines which population effect is being summarized.
Worked Example: What Happens When Overlap Fails
Worked Example: What Happens When Overlap Fails
Hypothetical setting. Suppose all low-risk individuals receive control and all high-risk individuals receive treatment. In the observed data, the low-risk control mean is 8 points and the high-risk treated mean is 18 points. The other two means—low-risk treated and high-risk control—are not observed.
For the low-risk stratum, \(P(D=1\mid X=\text{low})=0\); for the high-risk stratum, \(P(D=1\mid X=\text{high})=1\). Thus overlap fails in both strata. There is no within-stratum observed comparison from which to identify either conditional treatment contrast.
For example, if the low-risk treated mean were 9 points, its contrast would be \(9-8=1\) point; if it were 13 points, the contrast would be \(13-8=5\) points. Both possibilities are consistent with the observed low-risk control data. The observed data alone therefore do not determine the missing contrast. Without additional assumptions or information, the adjustment formulas cannot identify the ATE across both strata. For an ATT, the missing control outcome among high-risk treated individuals also prevents identification if no high-risk controls are observed.
What the Assumptions Do—and Do Not—Establish
Conditional independence and overlap are substantive conditions, not automatic features of a covariate-adjustment calculation. Conditional independence cannot generally be verified from observed outcomes alone: for each individual, only one potential outcome is observed. Researchers must use subject-matter knowledge, study design, and sensitivity analysis to assess whether relevant confounding may remain. Overlap can be examined in observed data, but apparent sample overlap does not prove population overlap.
If conditional independence fails, a within-\(X\) treated-control difference may still reflect unmeasured differences in potential outcomes. If overlap fails, a conditional mean required by the formula is unavailable for some target covariate values. Restricting the target population can sometimes address lack of support, but it changes the estimand and requires explicit justification.
Identification is also distinct from estimation and inference. The formulas describe population quantities in terms of observed-data distributions. A sample estimator must estimate those conditional means and weights, and its uncertainty must be assessed separately. The formulas alone do not guarantee precise estimates, particularly when treatment or control observations are scarce in some covariate regions.
Check Your Understanding
Use the definitions and assumptions above to answer the following.
- In one sentence, distinguish the population over which the ATE is averaged from the population over which the ATT is averaged.
- In the two-stratum example, why is the ATE 4.2 points while the ATT is 4.8 points?
- State what conditional independence implies about the treated group’s mean potential control outcome at a fixed covariate value.
- Suppose no treated individuals occur in a covariate stratum that belongs to the target population. Which identification condition fails, and why is the within-stratum contrast unavailable?
- Does observing a covariate-adjusted difference establish conditional independence? Explain briefly.