Why Use a Pooled Outcome Model?
A conditional treatment effect asks how the expected outcome contrast between treatment and control varies with measured covariates. One way to estimate it is to model outcomes separately in the two treatment groups. An alternative is to fit one pooled outcome model using all observations, with both treatment and covariates as inputs. This approach is called an S-learner: the same fitted model is used to predict outcomes under each treatment condition.
Pooling can make efficient use of the full sample and lets a flexible model represent complex relationships. But it also creates a distinctive risk: if the fitted model is regularized, the treatment indicator or its interactions with covariates may be shrunk, excluded, or otherwise underrepresented. The model can then predict outcomes reasonably well overall while estimating treatment-effect differences poorly.
The subtraction is essential. A fitted outcome level \(\hat f(1,x)\) is not itself a treatment effect; it is the model’s predicted outcome under treatment for covariate value \(x\). The S-learner predicts both treatment conditions at the same \(x\), then takes their difference.
When Does the Prediction Have a Causal Interpretation?
Define the conditional average treatment effect (CATE) as \(\tau(x)=E[Y(1)-Y(0)\mid X=x]\), where \(Y(1)\) and \(Y(0)\) are the potential outcomes under treatment and control. To identify this causal contrast from observed data, assume consistency, conditional exchangeability, and positivity. Consistency means the observed outcome equals the potential outcome corresponding to the treatment actually received. Conditional exchangeability means \((Y(1),Y(0))\) is independent of \(A\) given \(X\). Positivity means both treatment conditions have positive probability at covariate values in the target population.
Under these assumptions, the observed conditional mean \(m(a,x)=E[Y\mid A=a,X=x]\) equals \(E[Y(a)\mid X=x]\), so \(\tau(x)=m(1,x)-m(0,x)\). An S-learner estimates the two conditional means with one model. Its causal interpretation therefore depends both on the identification assumptions and on the model estimating those conditional means adequately. A pooled model does not, by itself, remove confounding or guarantee correct predictions.
Deriving the Contrast in a Linear S-Learner
A simple model makes clear how treatment effect heterogeneity enters. For one covariate \(X\), consider a linear outcome model with a treatment-by-covariate interaction:
Here \(A\) is coded 0 for control and 1 for treatment. Under this model, the predicted control outcome at \(X=x\) is \(\beta_0+\beta_Xx\). The predicted treated outcome is \(\beta_0+\beta_A+\beta_Xx+\beta_{AX}x\). Subtracting the control prediction from the treated prediction gives
The main treatment coefficient \(\beta_A\) is the predicted treatment effect at \(X=0\), not necessarily the average effect. The interaction coefficient \(\beta_{AX}\) determines how the predicted effect changes with \(X\). If the interaction is omitted, the model imposes a constant treatment effect, even if the true effect varies with \(X\).
Worked Example: Contrasting Predictions From an Interaction Model
Suppose a hypothetical study uses a linear S-learner for a continuous outcome measured in points. Its fitted model is \( \hat f(A,X)=10+2A+3X+1.5AX \), where \(X\) is a centered symptom score. Assume consistency, conditional exchangeability, and positivity are plausible for the population and covariate values considered, and that this linear model adequately describes the conditional means.
At \(X=2\), the control prediction is \( \hat f(0,2)=10+2(0)+3(2)+1.5(0)(2)=16 \) points. The treated prediction is \( \hat f(1,2)=10+2(1)+3(2)+1.5(1)(2)=21 \) points. Thus \( \hat\tau(2)=21-16=5 \) points.
The formula gives the same result: \( \hat\tau(2)=2+1.5(2)=5 \) points. At \(X=-1\), it gives \( \hat\tau(-1)=2+1.5(-1)=0.5 \) points. Under the stated causal and model assumptions, the predictions estimate conditional effects at those covariate values. They do not establish that treatment causes a five-point increase for every individual with \(X=2\).
How Regularization Changes the Contrast
Regularization adds a penalty to a model-fitting objective to discourage complexity or large coefficients. For example, ridge regression minimizes a squared-error loss plus a penalty proportional to the sum of squared coefficients. Lasso regression uses a penalty proportional to the sum of absolute coefficient values and can set some coefficients exactly to zero. Tree-based models also regularize, for example by limiting depth or requiring a minimum number of observations for a split.
In an S-learner, the treatment indicator is one input among the covariates. A regularized model may find that using the treatment indicator contributes little to its overall prediction objective, particularly when the outcome is noisy or the treatment groups are imbalanced. It may therefore shrink the treatment coefficient toward zero. Heterogeneity often requires additional structure, such as interaction terms or treatment-specific splits. Those terms may be penalized even more strongly because they add complexity.
This can produce shrinkage of the treatment contrast: predictions under treatment and control become more similar. Shrinkage can reduce variance when genuine effects are small or weakly estimated, but it can also pull real effects toward zero. Regularization is a prediction strategy, not a guarantee of unbiased conditional effects.
Worked Example: Shrinkage of an Interaction Coefficient
Consider a hypothetical linear S-learner whose treatment-effect prediction is \( \hat\tau(x)=0.8+\hat\gamma x \), with effects measured in points and \(x\) a standardized covariate. Suppose that, after accounting for the other model terms, the unpenalized estimate of the interaction coefficient is 1.2. To illustrate a ridge penalty, assume the interaction coefficient minimizes the following one-dimensional objective:
The first term represents fit to the data in this simplified coefficient calculation; the second penalizes a large interaction. Differentiating and setting the derivative to zero gives \(2(\gamma-1.2)+2\lambda\gamma=0\). Therefore, \( (1+\lambda)\gamma=1.2 \), so \( \hat\gamma=1.2/(1+\lambda) \).
If \(\lambda=3\), then \( \hat\gamma=1.2/4=0.3 \). The regularized treatment-effect prediction is \( \hat\tau(x)=0.8+0.3x \), rather than the unpenalized \(0.8+1.2x\). At \(x=2\), the unpenalized prediction is \(0.8+1.2(2)=3.2\) points, whereas the penalized prediction is \(0.8+0.3(2)=1.4\) points. This calculation isolates how a penalty can suppress estimated heterogeneity; the coefficient values are hypothetical, and the one-dimensional objective is an illustration rather than a claim about every fitting algorithm.
A treatment coefficient can also be penalized. If the model uses \( \hat f(A,X)=\hat\beta_0+\hat\beta_AA+\hat\beta_XX+\hat\beta_{AX}AX \), then the predicted contrast is \( \hat\beta_A+\hat\beta_{AX}x \). Shrinkage of \(\hat\beta_A\) attenuates the overall level of the contrast, while shrinkage of \(\hat\beta_{AX}\) flattens its change across \(x\). If both are heavily shrunk, the model may predict almost no treatment effect at any \(x\), regardless of whether the data-generating effect is truly zero.
Why a Good Outcome Fit Can Miss Heterogeneity
Outcome prediction and treatment-effect prediction are different targets. The outcome model is typically fitted to predict \(Y\), not directly to minimize error in \(\hat\tau(X)\). If treatment effects vary across subgroups but the variation affects only a modest fraction of observations, capturing those differences may produce little improvement in overall outcome prediction. A regularized pooled model may then prefer a simpler model that predicts outcomes adequately but gives nearly constant treatment effects.
Worked Example: An Additive Model Masks Subgroup Differences
Suppose a hypothetical randomized study has four equally sized covariate-treatment cells. The covariate \(X\) has two values, low and high. The observed mean outcomes, measured in points, are as follows. Assume treatment was randomized within the covariate groups, there are observations in every cell, and the displayed cell means represent the conditional means of interest.
| Covariate group | Control mean | Treatment mean | Within-group contrast |
|---|---|---|---|
| Low | 0 | 0 | 0 points |
| High | 0 | 4 | 4 points |
A model with a treatment main effect and a covariate main effect, but no interaction, imposes a common treatment contrast. Because the groups have equal sizes, the overall control mean is \((0+0)/2=0\) points and the overall treatment mean is \((0+4)/2=2\) points. The least-squares common treatment contrast is therefore \(2-0=2\) points. The additive S-learner predicts a two-point treatment effect in both groups.
The group-specific contrasts in the table are instead 0 points and 4 points. Thus the additive model misses the difference between groups even though its common contrast matches the average of the two group contrasts. Adding an interaction or allowing a suitable model split could represent the pattern, but regularization may discourage that extra complexity. In practice, noise and unequal cell sizes can make this tradeoff still more consequential.
This example also separates average effect accuracy from heterogeneity accuracy. A model can recover a useful overall contrast while misranking or flattening subgroup effects. That distinction matters when conditional estimates are used to describe which covariate profiles appear to benefit more, rather than simply to summarize an average.
Practical Interpretation and Common Pitfalls
An S-learner is attractive when a single flexible model can borrow information across treatment groups. But every predicted contrast is an extrapolation from the fitted model: for each \(x\), the model predicts both \(A=1\) and \(A=0\), including the condition not observed for a particular individual. Positivity and adequate covariate support are therefore important, and model flexibility cannot compensate for a lack of support.
- Confusing fitted outcome with treatment effect. The prediction \(\hat f(1,x)\) is an outcome prediction under treatment. The conditional effect prediction is \(\hat f(1,x)-\hat f(0,x)\).
- Assuming that a pooled model is automatically causal. Pooling does not ensure exchangeability, positivity, or correct conditional-mean estimation.
- Interpreting a small estimated effect as proof of no effect. A penalty may have shrunk the treatment indicator or interactions. A small contrast can reflect limited information or model restrictions as well as a genuinely small effect.
- Reading an interaction coefficient without its scale or reference value. In the linear model, \(\beta_A\) is the contrast at \(X=0\), and \(\beta_{AX}\) is the change in that contrast per unit increase in \(X\).
- Equating good outcome prediction with good effect estimation. Overall outcome accuracy does not ensure that subgroup contrasts are accurately represented.
When fitting an S-learner, inspect how the model encodes treatment and possible effect modifiers, and understand which terms its regularization penalizes. The substantive interpretation should remain conditional on the identification assumptions and the model’s ability to estimate both treatment-specific conditional means. Regularization can stabilize estimates, but its effect on the treatment contrast needs to be considered explicitly.
Check Your Understanding
Use the definitions and calculations in this tutorial to answer the following questions.
- For a fitted model \(\hat f(A,X)\), write the S-learner estimate of the conditional treatment effect at \(X=x\). Explain why a prediction under treatment alone is not that effect.
- In the linear model \(f(A,X)=\beta_0+\beta_AA+\beta_XX+\beta_{AX}AX\), derive the predicted treatment contrast and state what \(\beta_A\) represents.
- Using \(Q(\gamma)=(\gamma-1.2)^2+\lambda\gamma^2\), derive the minimizing value of \(\gamma\). What happens to this value as \(\lambda\) increases?
- In the subgroup example, what is the average treatment contrast across equally sized low and high groups? Why does an additive model miss the subgroup difference?
- Name the identification assumptions needed to interpret a contrast of conditional outcome means as a CATE, and explain why regularization does not establish those assumptions.