Treatment Effect Analysis

• Treatment effects refer to causal differences in outcomes between alternative exposure scenarios, encompassing individual and summarized population effects, and focusing on key metrics like the Average Treatment Effect (ATE) and Conditional Average Treatment Effect (CATE).
• Treatment effect methodologies address dynamic causal contrasts through transition hazards, mediating effects, distribution changes, and handling missing data due to intermediate or exposure-dependent events.
• Analysis techniques target complex treatment pathways through dynamic regression, individual-level forecasting, and Recovery control units through causal-forecast methods, emphasizing advanced estimators for point and distributional treatment parameters.

Treatment effects are causal contrasts between outcomes under alternative treatment regimes. In the potential-outcomes framework, the individual treatment effect is $Y(1)-Y(0)$, while common population-level estimands include the average treatment effect (ATE), the average treatment effect on the treated (ATT), the conditional average treatment effect (CATE), and the sample average treatment effect (SATE). Treatment effects may instead concern transition hazards, entire outcome distributions, tail behavior, dynamic trajectories, equilibrium spillovers, or effects among counterfactually defined subgroups. Their interpretation depends on the target population, treatment timing, interference structure, identifying assumptions, and the distinction between point and partial identification.

1. Potential outcomes, estimands, and identification

Let $Y(1)$ and $Y(0)$ denote potential outcomes under treatment and control, and let $A\in\{0,1\}$ be the observed treatment. Consistency gives $Y=Y(A)$; under randomization or conditional unconfoundedness, the marginal potential-outcome means and distributions can be related to observed treatment groups. The ATE is

$\operatorname{ATE}=E[Y(1)-Y(0)].$

The ATT is

$\operatorname{ATT}=E[Y(1)-Y(0)\mid A=1],$

and the CATE is

$\operatorname{CATE}(x)=E[Y(1)-Y(0)\mid X=x].$

The SATE averages individual effects over the realized study units rather than over a broader population. Causal identification generally requires consistency, treatment-exchangeability or randomization, positivity, and an appropriate no-interference condition such as SUTVA. Under observational assignment, a standard condition is

$(Y(1),Y(0))\perp A\mid X,$

together with

$0<P(A=1\mid X=x)<1.$

Randomization identifies marginal potential-outcome distributions and mean effects, but it does not identify the joint distribution of $Y(1)$0. Consequently, the individual treatment effect, its distribution, the exact proportion of winners, and the identities of the worst-affected individuals generally remain unidentified. This missing pairing problem is central to the distinction between ATEs, distributional effects, and individual-level effects (Kaji et al., 2023).

The estimand must be specified before selecting an estimator. A mean effect concerns average outcome change; a quantile treatment effect compares corresponding marginal quantiles; a CATE describes heterogeneity explained by observed covariates; and an individual treatment effect concerns a cross-world contrast that is ordinarily unobserved. These objects are not interchangeable. In particular, the quantile treatment effect

$Y(1)$1

does not generally describe the effect on the same individuals unless rank invariance is imposed.

A related distinction concerns model-based and model-free definitions. A model-based treatment effect is a stable parameter describing how treatment transforms an outcome distribution, such as an additive location shift or a multiplicative rate effect. A model-free effect is an average of individual or counterfactual outcome differences. These coincide in additive location models, but can differ when the treatment acts on a nonlinear distributional parameter. For example, a common additive effect on the log-odds scale does not imply a common difference in probabilities; the latter depends on baseline risk. Thus a sample average can vary with the composition of the sampled units even when a distributional treatment parameter is stable (Battey et al., 21 Jan 2026).

2. Dynamic, mediated, and post-treatment-defined effects

Treatment effects are often defined in settings where outcomes are realized after treatment and intermediate events or actions intervene. Conditioning on such variables can change the target population and invalidate a causal comparison.

For transition outcomes, let $Y(1)$2 indicate a transition in period $Y(1)$3 under treatment $Y(1)$4. The conditional transition probability is

$Y(1)$5

with survival probability

$Y(1)$6

The Average Treatment Effect on Treated Survivors in period $Y(1)$7 is

$Y(1)$8

The two terms refer to the same population: individuals who would survive under treatment. By contrast, the observed control survivors satisfy $Y(1)$9, generally defining a different population. Random assignment guarantees comparability at treatment assignment, not after treatment has altered survival. The resulting dynamic-selection problem means that randomization point identifies the instantaneous effect but generally not later conditional hazards. Nonparametric bounds can be derived from observed transition and survival probabilities, with intervals widening as survival falls (Vikström et al., 2017).

Cumulative effects and conditional transition effects answer different questions. A cumulative effect may concern

$Y(0)$0

or an expected-duration contrast $Y(0)$1. These quantities aggregate period-specific transition probabilities through survival products, but do not reveal the evolution of the treatment effect on the hazard. Two interventions may have the same duration effect while having different sequences of instantaneous effects.

Delayed outcomes also create intermediate-action problems. Let $Y(0)$2 be the focal treatment and $Y(0)$3 a vector of post-treatment actions. The total effect is

$Y(0)$4

whereas a controlled or partial effect fixes the actions at $Y(0)$5:

$Y(0)$6

A regression that controls linearly for post-treatment actions can combine direct effects, mediated effects, and selection terms. Its treatment coefficient may be negative even when $Y(0)$7 for every action configuration. Strong sign preservation holds automatically for convex averages of controlled effects, but not generally for conventional short, long, or interacted regressions. A regression with fixed effects for every joint action configuration produces a nonnegative weighted average of controlled effects under joint exogeneity (Bugni et al., 2023).

When the outcome is undefined after a post-treatment event, such as death, the contrast

$Y(0)$8

is not generally an average causal effect because the survivor populations differ. The survivor average causal effect targets individuals who would survive under both treatments:

$Y(0)$9

This principal-stratum effect is well defined but concerns an unobserved group and commonly requires cross-world assumptions. Conditional separable effects instead conceptualize treatment as components $A\in\{0,1\}$0 and $A\in\{0,1\}$1, where $A\in\{0,1\}$2 affects the post-treatment event and $A\in\{0,1\}$3 affects the outcome without affecting that event. The conditional separable effect at $A\in\{0,1\}$4 is

$A\in\{0,1\}$5

Its conditioning population is defined by a single event-related intervention and can, in principle, be observed in a component-randomized trial. Identification uses randomization, consistency, positivity, partial isolation, modified-treatment assumptions, and dismissible-component conditions (Stensrud et al., 2020).

A related post-treatment estimand is the Treatment Reactive Average Causal Effect (TRACE), defined among units who would exhibit a specified post-treatment value under treatment:

$A\in\{0,1\}$6

Observed conditioning on $A\in\{0,1\}$7 is generally biased because treated and control observations with $A\in\{0,1\}$8 correspond to different latent response types. TRACE can be expressed in terms of the overall treatment effect, the probability of treatment-reactive units, and an assumed or bounded effect among non-reactive units. It is therefore typically partially identified rather than directly estimable by conditioning on observed $A\in\{0,1\}$9 (Hazlett et al., 10 May 2025).

3. Distributional, heterogeneous, and risk-sensitive effects

The mean is only one feature of an outcome distribution. Distributional regression models treatment effects on location, scale, skewness, kurtosis, zero inflation, and other parameters. Generalized additive models for location, scale and shape (GAMLSS), also called distributional regression, specify

$Y=Y(A)$0

with separate predictors for each distributional parameter. Treatment can therefore change the conditional mean, dispersion, tail risk, poverty probability, or inequality measures independently. Marginal treatment effects are obtained by predicting the treatment and control distributions at common covariate values and comparing a functional $Y=Y(A)$1, such as a variance, quantile, Gini coefficient, Atkinson index, Theil index, or poverty probability (Hohberg et al., 2018).

Conditional and unconditional distributional effects differ. A conditional inequality effect measures dispersion among units with specified characteristics; an unconditional effect integrates conditional distributions over the covariate distribution. A program can increase mean consumption and reduce vulnerability while leaving conditional inequality unchanged or increasing it. In the Progresa reanalysis, treatment increased average food consumption, reduced vulnerability, and increased variance, while the Gini, Atkinson, and Theil effects were statistically indistinguishable from zero (Hohberg et al., 2018).

Treatment-effect heterogeneity is itself only partially identified when based on individual gains. Sharp bounds on the distribution of $Y=Y(A)$2 can be obtained from the marginal distributions of $Y=Y(A)$3 and $Y=Y(A)$4 using Fréchet–Hoeffding or Makarov bounds. For example,

$Y=Y(A)$5

and

$Y=Y(A)$6

These bounds are nonparametrically sharp: narrower conclusions require additional restrictions on the dependence between potential outcomes, such as rank invariance, rank similarity, or monotonicity.

The CATE provides a different risk measure. Define

$Y=Y(A)$7

The lower-tail conditional value at risk of individual effects satisfies the sharp upper bound

$Y=Y(A)$8

The right-hand side, CATE-CVaR, is the average CATE among the covariate-defined lower tail. If it is negative, the lower-tail average individual effect must also be negative. Lower bounds require restrictions on residual heterogeneity, such as

$Y=Y(A)$9

or

$\operatorname{ATE}=E[Y(1)-Y(0)].$0

Cross-fitted influence-function estimators permit inference even when CATE, propensity, and outcome regressions are estimated using flexible machine learning (Kallus, 2022).

Extreme-regime effects target tail behavior rather than average outcomes or finite quantiles. If conditional potential outcomes belong to generalized extreme-value domains of attraction with tail indices $\operatorname{ATE}=E[Y(1)-Y(0)].$1 and $\operatorname{ATE}=E[Y(1)-Y(0)].$2, the extreme treatment effect is

$\operatorname{ATE}=E[Y(1)-Y(0)].$3

A positive value represents a larger tail index and, under the paper’s convention, heavier or slower-decaying extremes. Estimation combines tail unconfoundedness, local pooling through an $\operatorname{ATE}=E[Y(1)-Y(0)].$4-max-sampler, conditional GEV likelihoods, and neural-network representations of the location, scale, and shape functions. This estimand is distinct from a high-quantile treatment effect: it concerns asymptotic tail shape (Aloui et al., 2023).

Functional outcomes introduce a further distinction between pointwise and shape-preserving effects. The Functional Average Treatment Effect is

$\operatorname{ATE}=E[Y(1)-Y(0)].$5

When curves are temporally misaligned, pointwise averaging can flatten shared features such as peaks, rebounds, and oscillations. Elastic shape analysis separates phase and amplitude using time warps and square-root slope functions. A phase-centered elastic mean defines population, individual, and local shape-based estimands. The populational estimand is identifiable by inverse-probability weighting, the local estimand through conditional outcome laws, and the individual estimand is not identifiable (Chabassier et al., 26 Oct 2026).

4. Time-varying effects, panels, and forecasting

Treatment effects may evolve over time, vary with exposure, or depend on treatment paths. State-space models represent baseline and treatment-effect coefficients as latent states that evolve through time. A representative dynamic regression is

$\operatorname{ATE}=E[Y(1)-Y(0)].$6

The treatment effect is

$\operatorname{ATE}=E[Y(1)-Y(0)].$7

State transitions for $\operatorname{ATE}=E[Y(1)-Y(0)].$8 can encode decay, persistence, growth, drift, seasonality, or local trends. Kalman filtering and smoothing estimate historical effects, while state propagation forecasts ATEs and CATEs at future time points. The method, called Causal Transfer, allows nonstationarity, measured confounding, heterogeneous effects, transfer across units, and temporal forecasting, but future effects remain model-dependent extrapolations rather than nonparametrically identified quantities (Li et al., 2018).

Panel identification under interactive fixed effects allows untreated outcomes to follow

$\operatorname{ATE}=E[Y(1)-Y(0)].$9

The interactive term $\operatorname{ATT}=E[Y(1)-Y(0)\mid A=1],$0 permits untreated paths to differ across individuals according to changing exposure to an unobserved factor. A time-invariant covariate whose coefficient is constant over time can instrument the pre-treatment outcome change that proxies for $\operatorname{ATT}=E[Y(1)-Y(0)\mid A=1],$1. With at least three periods, factor variation, rank, and selection restrictions, the ATT can be identified without requiring the number of periods to grow. Difference-in-differences is recovered when $\operatorname{ATT}=E[Y(1)-Y(0)\mid A=1],$2 is constant, while individual-specific linear trends correspond to $\operatorname{ATT}=E[Y(1)-Y(0)\mid A=1],$3. The method therefore relaxes parallel trends at the cost of requiring a substantively credible time-invariant-effect covariate (Callaway et al., 2020).

In complex treatment-path designs, no-anticipation and parallel trends identify actual-versus-status-quo effects. These compare the observed treatment path with the path in which each unit retains its period-one treatment. The normalized event-study effect can be expressed as a nonnegative weighted average of marginal current and lagged treatment effects. It does not generally evaluate treatment paths that were never implemented. A random-coefficients distributed-lag model can provide broader policy interpretation, but it imposes linearity in current and lagged treatments, a known maximum lag, no lag interactions, and time-invariant effects. Standard distributed-lag two-way fixed-effects coefficients can mix lags and use negative or treatment-effect-correlated weights (Chaisemartin et al., 11 Aug 2025).

When no control group exists, counterfactual outcomes can be forecast from individual pre-treatment histories. The Forecasted Average Treatment effect estimator is

$\operatorname{ATT}=E[Y(1)-Y(0)\mid A=1],$4

Its key requirement is average forecast unbiasedness:

$\operatorname{ATT}=E[Y(1)-Y(0)\mid A=1],$5

Basis-function regressions containing