Tutorials › Causal Inference › Conditional Average Treatment Effects as Prediction Targets

Causal Learning and Policy Decisions · Tutorial 948 of 972

Conditional Average Treatment Effects as Prediction Targets

Learn to specify a conditional treatment-effect target precisely and distinguish it from predicting outcomes or estimating effects for individuals.

Advanced 10 min read

What You'll Learn

  • Define the conditional average treatment effect using potential outcomes and covariates
  • Distinguish a subgroup-average causal effect from an individual treatment effect
  • Explain why predicting outcomes is not the same as predicting treatment effects
  • Specify the target population and covariates that index a conditional effect
  • Compare conditional effects on difference, ratio, and odds-ratio scales
  • Interpret how a treatment contrast and outcome scale shape the estimand

Why a Conditional Treatment Effect Is a Prediction Target

A causal effect need not be summarized by one number for an entire population. The effect of a treatment may differ across values of baseline covariates: for example, the average effect might be larger in one age group than another. A conditional average treatment effect (CATE) describes that pattern as a function of covariates. It is a target for learning how average causal effects vary across covariate-defined groups—not a claim that the effect for each particular person can be observed or predicted exactly.

The word “prediction” requires care here. Predicting an outcome under treatment and predicting a treatment effect are different tasks. The first concerns an outcome level; the second concerns a contrast between outcomes that would occur under different treatment options. Before interpreting either, specify the population, the treatment contrast, and the outcome scale.

Definition: Let \(A\) denote a treatment, with treatment options \(a=1\) and \(a=0\), and let \(Y(1)\) and \(Y(0)\) denote the potential outcomes under those options. Let \(X\) be a vector of covariates measured before treatment. In a specified target population \(P^*\), the additive conditional average treatment effect is $$ \tau_{P^*}(x)=E_{P^*}[Y(1)-Y(0)\mid X=x]. $$ It is the average potential-outcome difference among members of the target population with covariates \(X=x\).

The target population is part of the definition. It may be all eligible patients in a health system, people in a particular region, or members of a study population. The covariate value \(x\) identifies a conditional group within that population. If a covariate pattern is absent from the target population, the corresponding conditional mean has no direct target-population interpretation there.

An additive CATE uses the outcome’s original units. If \(Y\) is symptom score in points, then \(\tau_{P^*}(x)=2\) means that, among target-population members with covariates \(x\), the average potential score under treatment is two points higher than the average potential score under control. The sign and substantive desirability depend on how the outcome is defined.

Conditional Averages Are Not Individual Effects

For a particular unit, the individual treatment effect is \(Y(1)-Y(0)\). This quantity compares two potential outcomes for that unit, but only one treatment option is received, so the two outcomes are not jointly observed for that person. The CATE instead averages this difference across units with the same covariate value. It describes a group-level average, even when \(X\) includes many measured characteristics.

The conditional average can be expressed as a difference between conditional potential-outcome means. This is a useful identity, not an identification claim: it does not say that either potential-outcome mean, or their difference, can be recovered from observed data without additional assumptions.

$$ \begin{aligned} \tau_{P^*}(x) &=E_{P^*}[Y(1)-Y(0)\mid X=x]\\ &=E_{P^*}[Y(1)\mid X=x]-E_{P^*}[Y(0)\mid X=x]. \end{aligned} $$

The second equality follows from linearity of conditional expectation. It shows that a CATE is a contrast of two conditional average potential outcomes. It is not the difference between two observed outcomes for a single person, and it is not itself an outcome prediction. Whether observed data identify this target is a separate question requiring conditions beyond its definition.

Worked Example: A CATE and a Target-Population Average

Suppose a hypothetical program is evaluated for a continuous outcome measured in points. In a specified target population, the conditional mean potential outcomes are as follows. The covariate \(X\) indicates whether a person is in the younger or older age group.

Age group\(E[Y(1)\mid X]\) (points)\(E[Y(0)\mid X]\) (points)CATE (points)
Younger13103
Older1486

For the younger group, the conditional effect is $$ \tau_{P^*}(\text{younger})=13-10=3\text{ points}. $$ For the older group, it is $$ \tau_{P^*}(\text{older})=14-8=6\text{ points}. $$ These are group-average causal contrasts in the stated target population, not effects known for each younger or older person.

Now suppose 75% of this target population is younger and 25% is older. Its marginal average treatment effect is the average of the conditional effects using those target-population proportions: $$ 0.75(3)+0.25(6)=2.25+1.50=3.75\text{ points}. $$ The conditional effects are 3 and 6 points, while the marginal effect is 3.75 points. Changing the target population’s age composition could change this marginal average even if the two conditional effects stayed the same.

The proportions and conditional effects must refer to the stated target population for this calculation to describe that population. This example defines and averages causal targets; it does not establish how to estimate them from observed data.

Outcome Prediction Is Not Effect Prediction

For each treatment option, an outcome prediction can be represented by a conditional mean potential outcome, such as \(E_{P^*}[Y(1)\mid X=x]\). A CATE is a contrast between such means. Predicting one outcome mean well does not, by itself, ensure that the contrast is estimated well: errors in the two means may fail to cancel, or may cancel in ways that obscure poor predictions for either mean.

The distinction also matters for interpretation. A group with a high predicted outcome under treatment need not have a large treatment effect. Its predicted outcome under control could also be high. Conversely, a group with a modest treated outcome could have a substantial effect if its untreated outcome would have been lower.

Worked Example: High Outcome Level Does Not Imply a Large Effect

Consider two hypothetical covariate groups in the same target population, with an outcome measured in points. Suppose the conditional mean potential outcomes are:

GroupMean under treatmentMean under controlAdditive CATE
Group A90 points88 points2 points
Group B70 points60 points10 points

For Group A, the predicted mean under treatment is 90 points, but the effect is $$ 90-88=2\text{ points}. $$ For Group B, the predicted mean under treatment is only 70 points, but the effect is $$ 70-60=10\text{ points}. $$ Thus Group A has the higher treatment-outcome prediction, while Group B has the larger additive treatment effect. Ranking groups by their predicted treated outcomes would not rank them by CATE.

These comparisons concern conditional means of potential outcomes. They do not say that every person in Group B gains 10 points, or that any one person’s effect can be read from the group means.

The Treatment Contrast and Outcome Scale Matter

A CATE requires a specified contrast between treatment options and a specified scale. For a continuous outcome, an additive contrast is often expressed as a mean difference in the outcome’s original units. For a binary outcome, the conditional mean is a risk, and one might compare risks by subtraction or division. These are different estimands, even when they use the same treatment groups and covariates.

For example, the risk difference is the treated risk minus the control risk. The risk ratio is the treated risk divided by the control risk. An odds ratio compares the odds, where the odds for risk \(r\) are \(r/(1-r)\). An effect on one scale cannot generally be translated into a constant effect on another scale across covariate values. The scale should therefore be selected and reported as part of the causal question, not treated as a cosmetic choice after estimation.

Worked Example: One Binary Outcome, Several Conditional Contrasts

Suppose a hypothetical target-population subgroup has a 20% risk of an adverse event under treatment and a 10% risk under control. These are conditional risks for the same covariate-defined group.

On the risk-difference scale, $$ 0.20-0.10=0.10, $$ so the treatment-associated causal contrast is an increase of 0.10, or 10 percentage points, in the subgroup’s average event risk.

On the risk-ratio scale, $$ \frac{0.20}{0.10}=2.00. $$ The treated risk is twice the control risk. On the odds-ratio scale, the conditional odds are \(0.20/0.80=0.25\) under treatment and \(0.10/0.90\approx0.1111\) under control, giving $$ \frac{0.20/0.80}{0.10/0.90}=\frac{0.25}{0.1111}\approx2.25. $$ These values do not conflict: they quantify the same pair of conditional risks on different scales. The risk difference is in percentage points, while the risk ratio and odds ratio are unitless.

Specifying a CATE Target Precisely

A statement such as “estimate heterogeneous treatment effects” is incomplete until it identifies what the effect is conditional on and which population and contrast are intended. In particular, covariates should be defined on a meaningful pre-treatment scale. Conditioning on variables measured after treatment can change the causal question and is not automatically appropriate for a baseline CATE.

A useful target specification states the following elements explicitly:

  • Target population: the people or units to whom the conditional effect refers.
  • Treatment contrast: the treatment options being compared, including relevant versions or levels when needed.
  • Outcome: the outcome definition, measurement time, and units.
  • Effect scale: for example, an additive difference, risk difference, or ratio.
  • Conditioning covariates: the baseline covariate vector \(X\) and the values or groups \(x\) of interest.

These choices define what is to be learned; they do not guarantee that it can be learned from the available data. Identification, estimation, and validation are distinct tasks. In particular, a prediction algorithm can produce a smooth function of \(X\), but the fact that its output varies with covariates does not by itself show that the variation represents causal effect heterogeneity.

1
Name the population.
State which target population supplies the conditional averages and where the covariate patterns of interest occur.
2
Define both treatment options and the outcome.
Specify the contrast and the outcome’s measurement, timing, and units.
3
Choose the scale and conditioning variables.
Write the conditional contrast as a function of baseline covariates and interpret its values on the chosen scale.

Common Mistakes and Why the Distinctions Matter

A frequent mistake is to call a fitted conditional outcome mean a CATE. If a model predicts \(E[Y\mid A=1,X=x]\), it is making an outcome prediction for the treated condition, not necessarily estimating the causal contrast \(E[Y(1)-Y(0)\mid X=x]\). The correction is to define both potential-outcome means and the contrast between them, then address separately whether the data and assumptions support that causal interpretation.

Another mistake is to treat the CATE as an individual effect. A conditional mean averages over the units in a covariate-defined group; it does not reveal the two potential outcomes for each person. Even if the CATE is known exactly, individual effects can vary within that group.

Finally, reporting “the CATE” without an outcome scale or target population leaves the estimand ambiguous. A risk difference and risk ratio can differ in magnitude and interpretation, while a marginal average can change with the target population’s covariate distribution. Clear specification is essential before comparing, transporting, or using conditional effects for decisions.

Key takeaway: A CATE is a target-population conditional average causal contrast, indexed by covariates. It is neither an individual treatment effect nor a prediction of an outcome level. Its meaning depends on the population, treatment options, outcome, and effect scale.

Check Your Understanding

Answer each question by naming the relevant estimand or distinction.

  1. In one sentence, define the additive CATE using potential outcomes, a target population, and covariates.
  2. Why does a conditional average treatment effect not reveal the treatment effect for each individual in that covariate group?
  3. A subgroup’s predicted mean outcome is higher under treatment than another subgroup’s. Does that establish that its CATE is larger? Explain.
  4. For a binary outcome, what is the risk difference when the conditional risks are 0.30 under treatment and 0.20 under control? What is the risk ratio?
  5. Name five elements that should be specified to define a conditional treatment-effect target clearly.
Need help applying this method to your own data?

Our team supports biostatistics, statistical programming and pharmacometric analyses. Send us a request and we will reply by email.

Request an analysis
Clinical Biostats