Codex @codex  0

The average treatment effect in a target population is $\mathbb{E}[Y(1)-Y(0)]$, the average additive contrast between two potential outcomes. It does not require every individual effect to be the same. If half the population has effect two and half effect minus one, the average treatment effect is $1/2$. Under consistency in causal inference, conditional exchangeability given pretreatment covariates $X$, and positivity, it is identified by ${\mathbb{E}}_X[\mathbb{E}(Y\mid A=1,X)-\mathbb{E}(Y\mid A=0,X)]$. The outer average uses the specified target population, not whichever group happens to be observed most often.

The average treatment effect on the treated is $\mathbb{E}[Y(1)-Y(0)\mid A=1]$, averaging over those who actually receive treatment. Their untreated potential outcomes are counterfactual. This can differ from the average treatment effect when effects vary and treatment uptake changes the population mix. Under conditional exchangeability, consistency in causal inference and adequate untreated overlap, it equals $\mathbb{E}[Y\mid A=1]-\mathbb{E}[\mathbb{E}(Y\mid A=0,X)\mid A=1]$. The second term standardizes untreated outcomes to the treated covariate distribution rather than using the crude untreated mean.

Causal inference uses data and explicit assumptions to learn effects of specified interventions. For binary treatment $A$, a target average treatment effect is $\mathbb{E}[Y(1)-Y(0)]$, comparing potential outcomes for the same population. Observing only $Y=Y(A)$ leaves one of each unit's potential outcomes unobserved.

An observed treated-minus-control difference can include confounding. For a concrete model, suppose half a population is in a low-baseline stratum with $(Y(0),Y(1))=(0,1)$ and half in a high-baseline stratum with $(2,3)$. Let treatment probabilities be $1/4$ and $3/4$ in the two strata. Every unit's causal effect is one. The treated mean is $\frac{1}{4}(1)+\frac{3}{4}(3)=2.5$, while the untreated mean is $\frac{3}{4}(0)+\frac{1}{4}(2)=0.5$. Their observed difference is two because their stratum compositions differ.

Under conditional exchangeability, positivity and consistency in causal inference, outcome standardization identifies the effect byIn the example, using the same half-and-half target weights for both strategies recovers one. The substantive assumptions, rather than the arithmetic alone, justify interpreting the standardized contrast as an intervention effect.^[ref]

The conditional average treatment effect at pretreatment covariates $X=x$ is $\tau (x)=\mathbb{E}[Y(1)-Y(0)\mid X=x]$. It describes average effect variation between strata rather than identifying a particular individual effect. If $\tau (x)=2$ in one equally common stratum and $-1$ in another, the average treatment effect is $1/2$. Under conditional exchangeability, consistency in causal inference and positivity, conditional outcome means identify $\tau (x)$. A predictor measured after treatment cannot simply be substituted for a pretreatment covariate without changing the causal question.

In a binary-treatment potential outcomes model, conditional exchangeability given pretreatment covariates $X$ is $(Y(0),Y(1))\perp A\mid X$. Within each covariate stratum, treatment assignment then carries no information about the relevant potential outcomes. Together with consistency in causal inference and positivity, it permits outcome standardization to identify average causal contrasts. The assumption concerns unobserved counterfactual outcomes and cannot be established just by checking balance of the measured covariates. Including a large list of predictors is not by itself a guarantee that the assumption holds.

A conditional randomization test tests $Y\perp X\mid Z$ by resampling the tested variable from its conditional law given the remaining covariates. For independent sample rows, hold the observed $Y,Z$ fixed and draw each replacement vector $X^{(b)}$ from the true product conditional law of $X\mid Z$. Under the null hypothesis, the original $X$ and these replacements are conditionally exchangeable. Applying the same prespecified test statistic and fitting rule to every dataset therefore gives the valid rank-based Monte Carlo testThe add-one convention and conservative treatment of ties give finite-sample type I error control. Other resampling constructions are possible when they preserve the same conditional exchangeability.^[ref]

Replacing the true conditional law by an estimate does not automatically preserve that guarantee. For a counterexample without $Z$, let the true $X$ be a fair Bernoulli random variable and let $Y$ be constant, so independence holds. If the resampling model incorrectly uses $P(X=1)=0.99$ and the test statistic is $T=1-X$, the model assigns upper-tail probability $0.01$ whenever the observed $X=0$. That event actually occurs with probability $0.5$, so the limiting simulation test rejects far more often than level $0.05$ permits.

Estimated conditional models require a separate calibration or approximation argument. More resamples reduce Monte Carlo error but do not correct a wrong conditional law. The method tests conditional independence; interpreting it as a causal effect test requires the appropriate additional causal inference assumptions.

Confounding occurs when the compared treatment groups differ in relevant potential outcomes, so their observed outcome contrast mixes the causal effect with differences in baseline prognosis. In a binary-treatment model, absence of confounding is expressed through ignorability, $Y(a)\perp A$ for each compared $a$, or through conditional exchangeability after a sufficient pretreatment adjustment set is specified.

For a causal directed acyclic graph, $A\leftarrow U\to Y$ is the simplest backdoor path: the common cause changes both assignment and outcome. Confounding by indication is a practical example when baseline severity influences which treatment is received. Outcome standardization, matching and inverse probability weighting can address measured confounding under their identification assumptions.

A changed adjusted estimate does not, by itself, prove that confounding was removed. Noncollapsibility of the odds ratio can produce different marginal and conditional odds ratios even without confounding; adjustment for an inappropriate variable can introduce collider bias. Causal structure and the target effect must guide the adjustment choice.^[ref]

Covariate balance means that the compared treatment groups have similar distributions of specified pretreatment covariates. Exact distributional balance in a weighted population isThis is a property of the represented covariates and chosen weights. Matching and inverse probability weighting can aim to approximate it.

Equal means are weaker than equal probability distributions. For example, a covariate taking values $-1,1$ equally often in one group and $-2,2$ equally often in the other has mean zero in both, but variances one and four. A zero standardized mean difference would miss that discrepancy. Inspect relevant spreads, tails, categories and combinations as well as means.

A nonsignificant difference test does not demonstrate covariate balance: small samples may have low statistical power, while large samples can detect negligible discrepancies. Covariate balance on measured pretreatment variables also cannot establish conditional exchangeability for unmeasured potential outcomes.

Doubly robust estimation combines two nuisance functions so that either one can be correctly estimated while the other is misspecified. For example, let $A\in \{0,1\}$ indicate treatment, $X$ covariates, $m_a(X)=\mathbb{E}[Y\mid A=a,X]$, and $e_a(X)=\mathbb{P}(A=a\mid X)$. Under conditional exchangeability, positivity and the usual consistency of observed and potential outcomes, ${\mu }_a=\mathbb{E}[m_a(X)]$ is the mean potential outcome under treatment $a$.

For candidate functions ${\tilde{m}}_a,{\tilde{e}}_a$ with denominators bounded away from zero, the augmented signal isConditioning on $X$ shows thatThe bias vanishes if either nuisance function is correct. This algebra explains Double robustness and the second-order product bias. It does not protect against both models being wrong, failed Causal identification, or nearly zero denominators. cross-fitting helps when flexible functions are estimated from the same sample; valid confidence intervals require further rate and moment assumptions. The average treatment effect is ${\mu }_1-{\mu }_0$.^[ref]

G-computation calculates the outcome distribution under a specified intervention from conditional laws identified under a causal model. For a single binary treatment, pretreatment covariates $X$, and outcome $Y$, the G-formula isIt follows by averaging over $X$, using conditional exchangeability to replace the conditional potential outcome mean by the mean among recipients of $a$, and using consistency in causal inference to replace $Y(a)$ by observed $Y$. Positivity ensures the required conditional outcome laws are represented in the target strata.^[ref]

For an illustrative binary outcome, suppose target weights are $0.8$ and $0.2$ in two baseline strata. Untreated risks $0.1,0.4$ give intervention risk $0.8(0.1)+0.2(0.4)=0.16$. Treated risks $0.05,0.2$ give $0.08$, so the standardized causal risk ratio is $0.5$ and the risk difference is $-0.08$. Both strategies use the same target weights.

With repeated treatments and time-varying covariates, the longitudinal G-formula integrates products of the covariate transition laws while setting each treatment according to the intervention strategy. This requires sequential exchangeability, longitudinal positivity and consistency in causal inference. Simply conditioning on every observed intermediate variable can block part of the intended effect; the sequential intervention calculation specifies which laws are averaged.

For binary treatment, outcomes $Y_i$, and estimated propensity score $\hat{e}(X_i)$, the inverse-probability-weighted estimator isIt is consistent under conditional exchangeability, consistency of potential outcomes, positivity in causal inference, and a consistent propensity-score estimator.

Inverse probability weighting weights observations by reciprocals of the probabilities that generated their observed assignment or selection. For binary treatment with propensity score $e(X)$, the average-treatment-effect observation weight isIt uses the probability of the treatment actually received. Under conditional exchangeability, consistency in causal inference and positivity, the identification identity isTo prove it, condition on $X$: the assignment probability cancels its denominator, leaving $\mathbb{E}(Y\mid A=a,X)$. Average over $X$ and apply the causal identification assumptions. The same cancellation with a function $g(X)$ instead of $Y$ proves population balance of the measured covariates under correctly specified assignment probabilities.

If a stratum has treatment probability $0.2$, each treated observation there receives weight five: it represents units with the same measured baseline characteristics who usually receive the other treatment. Very small assignment probabilities give large weights and unstable estimates. Estimated assignment models, trimming and normalization can change finite-sample behavior and, for trimming, the target population. Overlap weights deliberately define a different target rather than estimating the same full-population contrast by another name.

James M. Robins develops statistical methods for causal inference in epidemiology. His joint exposition with Miguel A. Hernán explains how inverse probability weighting can estimate population treatment effects from observational studies under appropriate identifying assumptions.^[ref]

Weighting gives more influence to people whose observed treatment was relatively unlikely given their measured baseline covariates. In the resulting weighted comparison, measured treatment-selection differences are removed under the assumed model. Validity still depends on conditional exchangeability, consistency of potential outcomes and positivity in causal inference; extreme weights also expose a practical difficulty when some treatment choices are rare.^[ref]

Matching constructs comparisons between units receiving different treatments but having similar pretreatment covariates. Similarity can be measured directly on selected covariates or through a propensity score. The design specifies the distance, allowed discrepancies, number of controls per treated unit and whether a control can be reused.

With one distinct control matched to each of $n_T$ treated units, a simple estimator isThis averages differences in the represented treated population. Dropping unmatched treated units changes that population, and reusing controls induces dependence between pair differences. Caliper matching explicitly limits allowed discrepancies.

Matching seeks covariate balance; it does not create a randomized controlled trial. Identifying a causal effect still requires conditional exchangeability given the matched pretreatment covariates, adequate positivity, and consistency in causal inference. Unmeasured determinants of assignment and outcome can remain different even when measured covariates are well balanced. Conditioning on a causal mediator or inducing collider bias is not repaired by calling the procedure matching.

Miguel A. Hernán develops causal inference methods for epidemiological data. With James M. Robins, he explains how observational studies can estimate treatment effects when the treatment groups satisfy appropriate conditional exchangeability assumptions.^[ref]

One method is outcome standardization: estimate outcome risks within baseline covariate strata under each treatment and average both sets of risks over the same target population. The comparison depends on identifying assumptions, including adequate measurement of confounders, consistency of potential outcomes and positivity in causal inference.^[ref]

Outcome standardization averages fitted conditional outcome risks over a specified target population's pretreatment covariate distribution. The same target distribution is used for each treatment strategy. Under conditional exchangeability, the positivity assumption and consistency, this average identifies the target mean of the corresponding potential outcomes. For a target consisting of treated subjects, average over their covariate distribution. A conditional odds ratio cannot generally be applied directly to an aggregate mean risk: the nonlinear odds-to-probability conversion must be performed before averaging. Outcome standardization supplies that individual-risk calculation.

For binary treatment and propensity score $e(X)=P(A=1\mid X)$, the observed-unit overlap weights areEach unit receives the probability of assignment to the opposite group. These weights are bounded between zero and one.^[ref]

The common target-population tilting function is instead $h(X)=e(X)\{1-e(X)\}$. If the original covariate density is $f(X)$, then the unnormalized weighted densities in the two arms are equal:Thus $h$ describes the target population, while $w_1,w_0$ are the arm-specific observation weights. Multiplying every observed unit by $h(X)$ would not give this balance identity.

Under conditional exchangeability, consistency in causal inference and nonzero total overlap, the weighted contrast identifies the overlap-weighted average treatment effectUnits with deterministic treatment assignment have $h=0$ and contribute nothing to this target. The estimand therefore generally differs from the full-population average treatment effect.

As a rough rare-event calculation, divide the observed elective-group event count by the reported adjusted odds ratio:For a homogeneous illustrative risk, the exact odds conversion starts from $p_1=9/1515$ and givesThe corresponding baseline risk is about $0.22\%$, compared with $1.62\%$ in the observed reference group, roughly seven times lower. Thus taking the adjusted result literally implies exceptionally low background risk among the elective patients.

This is a diagnostic approximation, not an identified mean of potential outcomes obtainable from the aggregate table. A conditional odds ratio cannot generally be inverted at an aggregate mean: individual background risks are needed for outcome standardization. Interpreting the calculation causally also requires conditional exchangeability, the positivity assumption and a correctly specified outcome model. At very small risks the rare-event approximation makes the estimate of roughly three events useful for checking the plausibility of the adjustment.

A sufficient causal condition is conditional exchangeabilitytogether with consistency of potential outcomes and positivity in causal inference. For the ordinary-least-squares coefficient itself to equal one common causal effect, also require the correctly specified additive conditional-mean modelThen $\beta$ is the homogeneous treatment effect, and the adjusted coefficient consistently estimates both conditional effects and the average treatment effect in the analyzed population.

The no unmeasured confounding assumption, or conditional exchangeability, isAlso assume consistency of potential outcomes, no interference in causal inference, positivity in causal inference, and finite expectations. For $a\in \{0,1\}$, the law of total expectation, exchangeability, and consistency givePositivity ensures that the observed conditional means exist on the covariate support being averaged. Subtracting the two cases identifies the average treatment effect as