2.3 Conditional average treatment effect (CATE)

Conditional average treatment effect (CATE)

The conditional average treatment effect (CATE) estimates how a treatment effect differs across subgroups defined by observed characteristics. While the ATE gives you one number summarizing the treatment effect for an entire population, CATE answers a more targeted question: what is the treatment effect for individuals with these specific characteristics?

This matters because treatments rarely work the same way for everyone. A drug might help older patients more than younger ones. A marketing campaign might convert urban customers but not rural ones. CATE gives you the tools to detect and quantify that kind of variation.

Definition of CATE

CATE is the expected difference in potential outcomes for individuals who share a particular set of covariate values. Formally:

$\tau(x) = \mathbb{E}[Y(1) - Y(0) \mid X = x]$

where $Y(1)$ and $Y(0)$ are the potential outcomes under treatment and control, and $X$ represents observed covariates (age, gender, medical history, etc.).

The function $\tau(x)$ maps each covariate profile $x$ to a treatment effect. When $\tau(x)$ is the same for all $x$, the treatment effect is homogeneous and CATE collapses to the ATE. When $\tau(x)$ varies, you have heterogeneous treatment effects.

CATE vs. ATE

The ATE averages over the entire population:

$\text{ATE} = \mathbb{E}[Y(1) - Y(0)]$

You can think of the ATE as the average of CATE across the covariate distribution: $\text{ATE} = \mathbb{E}[\tau(X)]$. So the ATE is a single summary, while CATE is a function that tells you how the effect changes with $X$.

This distinction has practical consequences. If a treatment has an ATE of zero, that could mean it helps no one, or it could mean it helps some subgroups and hurts others in ways that cancel out. Only CATE can distinguish between these two very different scenarios.

Heterogeneous treatment effects

Heterogeneous treatment effects means the treatment effect varies across individuals or subgroups. Recognizing this heterogeneity is what motivates CATE estimation in the first place.

• A blood pressure drug might reduce systolic pressure by 15 mmHg in patients over 60 but only 3 mmHg in patients under 30.
• A job training program might raise earnings substantially for workers without a college degree but have little effect on those who already have one.

Identifying which subgroups benefit most (or are harmed) allows for better resource allocation and more personalized decision-making.

Estimating CATE

Estimating CATE is harder than estimating the ATE because you need enough data within each subgroup to get reliable effect estimates. Two broad families of methods are commonly used.

Regression methods for CATE

The simplest approach uses regression with interaction terms between the treatment indicator and covariates.

For example, with a single covariate $X$:

$Y = \beta_0 + \beta_1 T + \beta_2 X + \beta_3 (T \times X) + \epsilon$

Here $\beta_1 + \beta_3 x$ gives the estimated treatment effect at covariate value $x$. The interaction term $\beta_3$ captures how the effect changes with $X$.

This extends to generalized linear models and multilevel models. The main limitation is that you need to specify which interactions to include, and linear models may miss complex nonlinear patterns in $\tau(x)$.

Machine learning approaches to CATE

ML methods can discover heterogeneity patterns without requiring you to pre-specify interaction terms. Common approaches include:

• Causal forests (a variant of random forests designed specifically for treatment effect estimation)
• Bayesian additive regression trees (BART)
• Meta-learners (S-learner, T-learner, X-learner), which use standard ML models as building blocks for CATE estimation

These methods handle high-dimensional covariates and complex interactions well, but they require careful tuning and validation to avoid overfitting, especially when sample sizes are modest.

Challenges in estimating CATE

• Limited subgroup sample sizes. Splitting data by covariates reduces the effective sample size for each estimate, increasing variance.
• High-dimensional covariates. With many potential effect modifiers, the risk of finding spurious heterogeneity grows.
• Unmeasured confounders. If confounding differs across subgroups, CATE estimates can be biased even when the overall ATE estimate is not.
• Model selection. Choosing between methods (and validating results) is harder for CATE than for ATE, since you never observe both potential outcomes for the same individual. Cross-validation strategies need to be adapted for causal quantities.

Applications of CATE

Personalized medicine

CATE estimation is central to precision medicine. By estimating treatment effects conditional on a patient's genetic profile, comorbidities, and demographics, clinicians can recommend the treatment most likely to benefit that specific patient rather than relying on population-average evidence.

Targeted marketing campaigns

Marketers use CATE to estimate the uplift of a campaign for different customer segments. Rather than sending a promotion to everyone, you target the subgroup where the estimated CATE (incremental conversion probability) is highest, improving return on investment.

Policy evaluation and optimization

Policymakers can use CATE to determine which populations benefit most from an intervention. For example, estimating CATE for a job training program across education levels and regions can guide where to expand the program and where to reallocate resources.

Assumptions and limitations

CATE estimation inherits the core identification assumptions of the potential outcomes framework, and violations are especially consequential because they can distort the pattern of heterogeneity, not just the overall level.

Unconfoundedness assumption

Also called conditional ignorability, this assumption states:

$Y(1), Y(0) \perp\!\!\!\perp T \mid X$

In words: once you condition on the observed covariates $X$, treatment assignment $T$ is independent of the potential outcomes. This means there are no unmeasured confounders lurking behind the relationship between treatment and outcome.

If this assumption fails, your CATE estimates will be biased. Worse, the bias can vary across subgroups, potentially making you think a treatment helps one group when it actually doesn't.

Overlap assumption

Also called positivity, this requires:

$0 < P(T = 1 \mid X = x) < 1 \quad \text{for all } x$

Every covariate profile must have some probability of receiving treatment and some probability of receiving control. If certain subgroups almost always (or never) receive treatment, you're essentially extrapolating rather than estimating, and CATE estimates in those regions become unreliable.

Limitations of CATE estimates

• Finite sample bias. With limited data per subgroup, estimates can be noisy or systematically off.
• Model misspecification. Parametric models may impose the wrong functional form on $\tau(x)$; flexible models may overfit.
• Validation difficulty. You can't directly validate individual-level treatment effects since the fundamental problem of causal inference (you only observe one potential outcome per unit) still applies.
• Always interpret CATE estimates alongside sensitivity analyses and cross-validation results.

Advanced topics in CATE

Efficient estimation of CATE

Efficient estimators aim to minimize mean squared error by balancing bias and variance. Key approaches include:

• Doubly robust (DR) estimation combines an outcome model and a propensity score model. If either model is correctly specified, the estimator remains consistent.
• Targeted maximum likelihood estimation (TMLE) uses a targeted bias-reduction step on top of an initial outcome model.
• Efficient influence function-based estimators achieve the semiparametric efficiency bound under certain regularity conditions.

These methods are more robust to partial model misspecification than approaches that rely on a single model.

Nonparametric estimation of CATE

Nonparametric methods avoid strong assumptions about the functional form of $\tau(x)$:

• Kernel regression and local linear regression estimate the treatment effect locally around each point in covariate space.
• Spline-based methods fit flexible smooth functions.

These approaches can capture complex nonlinear heterogeneity but typically require larger sample sizes and can be computationally expensive in high dimensions.

CATE with instrumental variables

When unconfoundedness fails, instrumental variables (IVs) offer an alternative identification strategy. An instrument $Z$ must affect treatment assignment but have no direct effect on the outcome except through treatment.

• Two-stage least squares (2SLS) can be extended with interactions to estimate heterogeneous effects.
• The local average treatment effect (LATE) identifies the effect for compliers, and extensions estimate how LATE varies with covariates.
• Marginal treatment effect (MTE) methods trace out how the treatment effect varies with individuals' unobserved propensity to select into treatment, providing a richer picture of heterogeneity.

CATE in longitudinal settings

When treatments, covariates, and outcomes evolve over time, standard cross-sectional CATE methods can break down due to time-dependent confounding (where past treatment affects future covariates, which in turn affect future treatment).

Methods designed for this setting include:

• Marginal structural models (MSMs) with inverse probability of treatment weighting
• Structural nested mean models (SNMMs) that directly model the causal effect of treatment at each time point
• G-computation, which models the full sequence of conditional distributions to estimate counterfactual outcomes

Each of these accounts for the dynamic interplay between treatment and confounders over time, though they come with their own assumptions and data requirements.