Pointwise Inference for Conditional Treatment Effects

Why Pointwise Inference Needs a Fixed Target

A conditional average treatment effect (CATE) describes an average effect among units with a specified covariate profile. An estimate alone does not indicate how precisely that effect has been estimated. Pointwise inference adds uncertainty quantification for one target profile, such as a patient with a prespecified set of baseline characteristics. The word pointwise matters: the resulting interval is designed to cover the effect at that profile, not all profiles simultaneously.

Let $W\in\{0,1\}$ denote treatment, $X$ the covariates measured before treatment, and $Y(1)$ and $Y(0)$ the potential outcomes under treatment and control. For a fixed profile $x$, the target is

\tau(x)=E[Y(1)-Y(0)\mid X=x].

This is a conditional mean effect, not an individual treatment effect. For continuous covariates, it is understood as a value of a suitably regular conditional mean function at $x$; meaningful estimation requires data support near that profile.

Definition: A pointwise confidence interval for $\tau(x)$ is an interval constructed for one fixed covariate profile $x$, with a stated repeated-sampling coverage level under the estimator's assumptions. A 95% pointwise interval has approximately 95% coverage for that target across repeated samples; it does not assign a 95% probability to the fixed parameter being in the realized interval.

Identification and Conditions for Inference

For a causal interpretation, the observed data must identify the potential-outcome contrast. A standard set of conditions is consistency, conditional exchangeability, and overlap. Consistency means the observed outcome is $Y=WY(1)+(1-W)Y(0)$. Conditional exchangeability requires $(Y(1),Y(0))$ to be independent of $W$ given $X$. Overlap requires the treatment probability $e(x)=P(W=1\mid X=x)$ to be bounded away from zero and one in the region of interest. Under these conditions,

\tau(x)=E[Y\mid W=1,X=x]-E[Y\mid W=0,X=x].

Identification alone does not guarantee a valid confidence interval. Inference also depends on how the CATE estimator is constructed, whether its bias is negligible or corrected, and whether its standard error accounts for its sampling variability. Forest-based inference typically relies on properties such as honesty or sample splitting, sufficient observations near the target, regularity of the conditional effect, and an appropriate variance estimator. Orthogonal-score methods additionally require suitable nuisance estimates, overlap, and conditions on local smoothing or other ways of estimating the conditional mean. These are asymptotic procedures: a reported interval need not have its nominal coverage in a small sample or at a profile with little support.

Forest-Based Standard Errors

A causal forest can represent a CATE estimate at $x$ as a weighted average of observed outcomes or effect-related quantities, where observations similar to $x$ tend to receive more weight. An honest forest separates the data used to choose tree splits from data used to estimate effects within leaves. This separation helps control adaptive bias and supports the forest's variance calculation. The forest's standard error should be the one appropriate to the fitted estimator and its sampling design, rather than an ordinary standard error for a simple sample mean.

When the estimator is approximately normal after standardization, a two-sided 95% pointwise interval takes the form

\widehat{\tau}(x)\ \pm\ 1.96\,SE\{\widehat{\tau}(x)\}.

More generally, the multiplier is the relevant quantile of the estimator's limiting distribution. A normal approximation is useful only when its regularity conditions are reasonably credible, including adequate local information and sufficiently small bias relative to the standard error. The interval's width reflects uncertainty at this one profile; it does not account for searching over many profiles and reporting the most striking result.

Worked Example: A Forest Interval at a Prespecified Profile

Suppose an honest causal forest estimates the effect of a hypothetical program on annual medical spending for a prespecified profile $x$. The estimate is $\widehat{\tau}(x)=-\$240$ per person per year, and the forest's estimated standard error is $\$100$ per person per year. Assume the normal approximation and negligible-bias conditions are adequate for this illustration.

The 95% pointwise interval is

-\$240 \pm 1.96(\$100) = -\$240 \pm \$196 = [-\$436,-\$44]\text{ per person per year}.

The estimated effect is a reduction of $\$240$ per person per year. The interval is compatible with reductions ranging from $\$44$ to $\$436$ per person per year under the model and inference assumptions. It excludes zero, but this conclusion concerns the prespecified profile and does not establish the effect for every profile in the population.

Orthogonal Scores for CATE Inference

An alternative route uses an orthogonal pseudo-outcome. Define the outcome regressions $m_w(x)=E[Y\mid W=w,X=x]$ for $w=0,1$, and the propensity score $e(x)=P(W=1\mid X=x)$. With these nuisance functions, the score

\Gamma = m_1(X)-m_0(X) +\frac{W}{e(X)}\{Y-m_1(X)\} -\frac{1-W}{1-e(X)}\{Y-m_0(X)\}

has conditional expectation $E[\Gamma\mid X=x]=\tau(x)$ when the nuisance functions are correct and the identification conditions hold. The residual terms adjust the outcome-regression contrast using the observed treatment assignment. In practice, nuisance functions are estimated, often with cross-fitting: nuisance models are fitted on other folds of the data before scores are calculated for a held-out fold. This reduces the direct impact of overfitting on the score evaluation.

Key result: The orthogonal score is locally insensitive, to first order, to small errors in the nuisance functions at their true values. This property can reduce the impact of nuisance estimation error. It does not eliminate the need for overlap, adequate data near $x$, control of smoothing bias, or a standard error appropriate to the full estimator.

To see the source of this property, consider the expected score conditional on $X=x$. If the outcome regressions are perturbed by small amounts $\delta_1$ and $\delta_0$, the first term changes by $\delta_1-\delta_0$. To first order, the treated residual term changes by $-\delta_1$, and the negative control residual term changes by $+\delta_0$. The sum of these first-order changes is zero. A small perturbation in $e(x)$ likewise has no first-order effect when the outcome regressions are correct, because the conditional residual means are then zero. The remaining nuisance-error effects are second order under suitable regularity conditions; orthogonality is not a claim that they vanish exactly in a finite sample.

To estimate the conditional mean of the scores at a fixed profile, one can use local weights $a_i(x)$ that sum to one and give greater weight to observations near $x$. A generic local estimate is

\widehat{\tau}(x)=\sum_{i=1}^{n} a_i(x)\widehat{\Gamma}_i, \qquad \sum_{i=1}^{n}a_i(x)=1.

Here $\widehat{\Gamma}_i$ uses fitted nuisance functions, preferably obtained by cross-fitting. For continuous covariates, this weighted average generally smooths over a neighborhood, so its finite-sample target may differ from the exact value at $x$. Valid pointwise inference requires the smoothing bias to be negligible relative to the standard error, or requires a method that estimates and corrects the bias. Variance estimation must also reflect the local weights and score variability; treating estimated pseudo-outcomes as fixed without justification can understate uncertainty.

Worked Example: Computing Orthogonal Scores and a Local Estimate

Consider four hypothetical observations, with treatment $W_i$, outcome $Y_i$, fitted outcome regressions $m_1(X_i),m_0(X_i)$, and fitted propensity $e(X_i)$ shown below. Suppose these fits are cross-fitted and the observations receive local weights $0.40,0.30,0.20,0.10$ for a target profile $x$.

Observation	$W_i$	$Y_i$	$m_1(X_i)$	$m_0(X_i)$	$e(X_i)$	$a_i(x)$
1	1	8	7	5	0.50	0.40
2	0	4	6	4	0.50	0.30
3	1	5	6	3	0.25	0.20
4	0	2	5	2	0.50	0.10

For observation 1, the score is $7-5+(1/0.50)(8-7)=4$. For observation 2, it is $6-4-(1/0.50)(4-4)=2$. For observation 3, it is $6-3+(1/0.25)(5-6)=-1$. For observation 4, it is $5-2-(1/0.50)(2-2)=3$.

The local estimate is therefore

\widehat{\tau}(x) =0.40(4)+0.30(2)+0.20(-1)+0.10(3) =1.60+0.60-0.20+0.30 =2.30.

The estimate is 2.30 outcome units for the target profile under this local procedure. Four observations are far too few to support reliable asymptotic inference here; the calculation illustrates score construction and weighting, not a defensible standard error or confidence interval. In an application, uncertainty estimation and local-bias control would need to be justified using the full sample and the chosen estimator.

Profile Selection Changes the Inferential Question

A pointwise interval presumes that the target profile $x$ was fixed independently of the effect estimates, or that the inferential procedure otherwise accounts for how it was chosen. If an analyst examines many profiles and reports one with an unusually large estimated effect, the reported estimate is selected partly because its sampling error happened to be favorable. An ordinary pointwise interval at that selected profile does not generally retain its nominal coverage.

Worked Example: Searching Across Profiles

Imagine testing ten prespecified profiles when the true effect is zero at each. For illustration, suppose each estimate is normally distributed, each pointwise 95% interval uses the usual two-sided 1.96 cutoff, and the ten standardized estimates are independent. For any one profile, the probability that its interval excludes zero is 0.05. The probability that none of the ten intervals excludes zero is $0.95^{10}$. Thus, the probability that at least one excludes zero is

1-0.95^{10}\approx 0.4013.

So in this simplified example, searching across ten profiles gives about a 40.1% chance of finding at least one apparently nonzero effect, even though all ten true effects are zero. The independence assumption is only for this illustration; real profile estimates may be correlated. The example shows why a pointwise 95% interval cannot be interpreted as 95% coverage after selecting the most striking profile.

Practical Checks and Common Pitfalls

Before interpreting a pointwise interval, identify its target, estimator, uncertainty calculation, and assumptions. In particular, check whether the profile lies in a region with enough treated and control observations, whether extreme propensity scores make the score unstable, and whether the method has addressed bias from adaptive fitting or local smoothing. A small standard error does not by itself establish that these conditions hold.

Confusing pointwise with simultaneous coverage. A 95% pointwise interval concerns one profile. It is not a 95% guarantee that intervals cover effects across a whole range of profiles.
Ignoring profile selection. Choosing a profile because its estimated effect is large changes the inferential problem. A prespecified-profile interval does not automatically account for that search.
Treating orthogonality as immunity. Orthogonal scores reduce first-order sensitivity to nuisance errors, but weak overlap, poor local support, smoothing bias, and finite-sample instability remain relevant.
Overstating causal interpretation. The interval is for a causal CATE only when the identification assumptions are credible; otherwise, it describes an associational conditional contrast.

Key takeaway: Pointwise inference quantifies uncertainty for a CATE at a fixed profile. A valid interpretation requires both an estimator-specific standard error and attention to identification, local support, bias, and whether the profile was selected after examining the results.

Check Your Understanding

Answer these questions using the definitions and reasoning in this tutorial.

What does a 95% pointwise confidence interval claim about repeated sampling, and what does it not claim about the realized interval?
Name the three standard identification conditions used here and explain why identification is not sufficient by itself for valid inference.
In the orthogonal score, what role do the propensity score and the two outcome regressions play?
Why does orthogonality not remove the need to address local smoothing bias when estimating a CATE at a continuous covariate profile?
An analyst selects the profile with the largest estimated effect among many profiles. Why may an ordinary pointwise interval at that profile undercover?