Cumulative Cross-World Weighted Effects

• Cumulative cross-world weighted effects are defined as weighted averages of counterfactual outcome differences, preserving the contrast between two static treatment regimes.
• They leverage cumulative, time-specific weighting to adjust for longitudinal nonpositivity and separate mechanistic contrasts from policy-relevant interventions.
• The methodology employs NPSEM, strong sequential randomization, and inverse probability weighting to accurately identify causal contrasts in dynamic treatment scenarios.

Searching arXiv for the cited papers and closely related longitudinal causal inference work. Cumulative cross-world weighted effects are longitudinal causal estimands for comparing two static treatment regimes when standard longitudinal positivity fails. They are defined as weighted averages of the potential outcome contrast $Y(\overline a_T)-Y(\overline a_T^\prime)$, with weights that accumulate over time and depend simultaneously on the treatment process under both counterfactual treatment histories (McClean et al., 14 Jul 2025). The construction is intended to preserve the contrast between the two regimes themselves rather than substituting an implementable stochastic policy, but this comes at the cost of targeting a non-implementable cross-world quantity (McClean et al., 14 Jul 2025). The topic also intersects with recent work on clone-censor-weighting (CCW), which shows that weighting procedures for treatment initiation windows can implicitly target hybrid or impossible interventions assembled from incompatible treatment histories, thereby clarifying why cross-world weighting is both interpretable in mechanistic terms and problematic in policy terms (Webster-Clark et al., 2024).

1. Formal longitudinal setup and estimand

The proposed framework is longitudinal with $T$ time points and observed data

$Z=(X_1,A_1,X_2,A_2,\dots,X_T,A_T,Y),$

where $X_t$ are time-varying covariates, $A_t\in\{0,1\}$ are binary treatments, and $Y$ is the final outcome (McClean et al., 14 Jul 2025). Treatment histories are denoted $\overline A_t=(A_1,\dots,A_t)$, covariate histories are $\overline X_t=(X_1,\dots,X_t)$, and the history before treatment at time $t$ is

$H_t=(\overline X_t,\overline A_{t-1})$

(McClean et al., 14 Jul 2025).

Potential outcomes are defined in an NPSEM: $T$0 (McClean et al., 14 Jul 2025). For a regime $T$1, the counterfactual outcome is $T$2, and the counterfactual covariate history is $T$3 (McClean et al., 14 Jul 2025).

The cumulative cross-world weighted effect is

$T$4

(McClean et al., 14 Jul 2025). Here the natural propensity scores are

$T$5

$T$6

(McClean et al., 14 Jul 2025). Thus the estimand is not merely a causal contrast between two regime-specific means; it is a weighted contrast whose weighting functional itself is cross-world.

A plausible implication is that the estimand should be understood as selecting, across time, the portions of the two counterfactual trajectories that remain jointly informative when positivity fails. The paper states this point more directly by emphasizing that the weights adapt “cumulatively across timepoints and simultaneously across both counterfactual treatment histories” (McClean et al., 14 Jul 2025).

2. Cumulative weighting across counterfactual worlds

The “cumulative” feature is the product structure

$T$7

(McClean et al., 14 Jul 2025). Weighting therefore responds at every time point to longitudinal nonpositivity rather than only at baseline.

The paper gives several explicit weighting choices (McClean et al., 14 Jul 2025):

Choice	$T$8	$T$9
No weighting	$Z=(X_1,A_1,X_2,A_2,\dots,X_T,A_T,Y),$0	$Z=(X_1,A_1,X_2,A_2,\dots,X_T,A_T,Y),$1
Weight toward only the target regime	$Z=(X_1,A_1,X_2,A_2,\dots,X_T,A_T,Y),$2	$Z=(X_1,A_1,X_2,A_2,\dots,X_T,A_T,Y),$3
Weight toward only the comparator regime	$Z=(X_1,A_1,X_2,A_2,\dots,X_T,A_T,Y),$4	$Z=(X_1,A_1,X_2,A_2,\dots,X_T,A_T,Y),$5
Overlap weighting	$Z=(X_1,A_1,X_2,A_2,\dots,X_T,A_T,Y),$6	$Z=(X_1,A_1,X_2,A_2,\dots,X_T,A_T,Y),$7
Trimming	$Z=(X_1,A_1,X_2,A_2,\dots,X_T,A_T,Y),$8	$Z=(X_1,A_1,X_2,A_2,\dots,X_T,A_T,Y),$9

A concrete trimming example is

$X_t$0

(McClean et al., 14 Jul 2025). A smooth trimming choice used in the empirical application is

$X_t$1

(McClean et al., 14 Jul 2025).

The key restriction on the weights is

$X_t$2

(McClean et al., 14 Jul 2025). This requirement ensures that if either relevant propensity score is zero, the combined weight is also zero. In substantive terms, paths unsupported in either counterfactual world are removed from the weighted functional.

This suggests that weighting is not simply a variance-control device. It determines which regions of longitudinal state space contribute to the estimand and therefore partly defines the target itself.

3. Identification without standard positivity

Identification is developed under NPSEM and strong sequential randomization: $X_t$3 (McClean et al., 14 Jul 2025). The paper notes that this is stronger than ordinary sequential exchangeability because the estimand depends on cross-world quantities (McClean et al., 14 Jul 2025).

Under NPSEM and strong sequential randomization, if positivity has held up to earlier times, then the natural propensity scores are identified by observed-data quantities: $X_t$4 with an analogous expression for $X_t$5 under $X_t$6 (McClean et al., 14 Jul 2025).

The main identification formula is

$X_t$7

$X_t$8

(McClean et al., 14 Jul 2025). The paper explicitly states that no standard positivity assumption is required for identification because the design of the weights replaces it (McClean et al., 14 Jul 2025).

The framework distinguishes positivity from overlap of covariate distributions through the density ratio

$X_t$9

and the condition

$A_t\in\{0,1\}$0

which the paper terms partial common support (McClean et al., 14 Jul 2025). Partial common support is weaker than full overlap, but the paper states that it ensures the weighted functional remains informative about the causal contrast (McClean et al., 14 Jul 2025).

4. Interpretation, non-implementability, and the cross-world problem

The paper presents cumulative cross-world weighted effects as mechanism-relevant but not policy-relevant in the sense of a feasible intervention (McClean et al., 14 Jul 2025). The reason is structural: the estimand depends on a product of weights evaluated under two counterfactual worlds simultaneously, and one cannot implement an intervention that generates both $A_t\in\{0,1\}$1 and $A_t\in\{0,1\}$2 for the same person (McClean et al., 14 Jul 2025).

The resulting tradeoff is explicit. If positivity holds, the ordinary causal contrast

$A_t\in\{0,1\}$3

is both interpretable and implementable. If positivity fails, implementable stochastic or modified policies can be identified, but those may conflate regime effects with downstream effects on intermediate treatments and covariates (McClean et al., 14 Jul 2025). The cumulative cross-world weighted effect instead preserves the pure contrast of the two static regimes, but only as a non-implementable cross-world quantity (McClean et al., 14 Jul 2025).

The paper also gives a null-preservation property: $A_t\in\{0,1\}$4 (McClean et al., 14 Jul 2025). This establishes that the estimand respects a strong form of causal nullity despite its non-implementable character.

Closely related concerns appear in the analysis of CCW for treatment initiation windows. There, a regimen such as “start treatment prior to day 30” is shown to estimate the potential outcome under a two-stage intervention where “A) prior to day 30, everyone follows the treatment start distribution of the study population and B) everyone who has not initiated by day 30 is forced to initiate on day 30” (Webster-Clark et al., 2024). When earlier initiators are used to represent those forced to initiate at the end of the window, the target may require assigning a day-30 starter the prior exposure history of someone who started earlier, producing a cross-world or impossible intervention (Webster-Clark et al., 2024). This parallel is conceptually important: in both settings, weighting may preserve a mechanistic contrast while moving away from interventions that could be literally enacted.

5. Partial common support and collapse to zero

A central substantive limitation is that identification alone does not guarantee informativeness. The paper states that if for some $A_t\in\{0,1\}$5,

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

then

$A_t\in\{0,1\}$7

regardless of the true potential outcome contrast (McClean et al., 14 Jul 2025). The argument is that when the relevant conditioning event has probability zero, the corresponding propensity score is set to zero, so the combined weight is zero on those paths; if no common support remains, the entire weighted functional vanishes (McClean et al., 14 Jul 2025).

Conversely, under positive-probability overlap at each timepoint, the paper states that the functional can preserve sign: if

$A_t\in\{0,1\}$8

then

$A_t\in\{0,1\}$9

and if

$Y$0

then

$Y$1

(McClean et al., 14 Jul 2025). Partial common support is therefore not required for algebraic identification, but it is required for substantive interpretability (McClean et al., 14 Jul 2025).

This suggests that cumulative cross-world weighted effects answer a restricted version of the original causal question: they remain informative only on the longitudinal region where both counterfactual trajectories retain some common support. Outside that region, the estimand is intentionally silent.

6. Estimation theory and machine-learning implementation

For the identified target component $Y$2, the paper defines the backward recursive regression

$Y$3

$Y$4

(McClean et al., 14 Jul 2025). It also defines inverse weights

$Y$5

$Y$6

(McClean et al., 14 Jul 2025).

Under smooth weights and boundedness and moment assumptions, the uncentered efficient influence function is

$Y$7

(McClean et al., 14 Jul 2025). The paper notes that $Y$8 is the usual sequential-regression residual term and $Y$9 captures uncertainty from estimating the weights and propensity scores; the EIF includes terms involving both $\overline A_t=(A_1,\dots,A_t)$0 and $\overline A_t=(A_1,\dots,A_t)$1, plus derivative terms

$\overline A_t=(A_1,\dots,A_t)$2

$\overline A_t=(A_1,\dots,A_t)$3

(McClean et al., 14 Jul 2025). A notable feature is that the EIF contains a term propagating information from the non-target regime through the density ratio $\overline A_t=(A_1,\dots,A_t)$4 (McClean et al., 14 Jul 2025).

The doubly robust-style algorithm uses: propensity estimation for both regimes, density-ratio estimation for $\overline A_t=(A_1,\dots,A_t)$5, backward regression for $\overline A_t=(A_1,\dots,A_t)$6, plug-in EIF evaluation on held-out data, and cross-fitting or sample splitting (McClean et al., 14 Jul 2025). The estimator is

$\overline A_t=(A_1,\dots,A_t)$7

(McClean et al., 14 Jul 2025). Under the stated conditions,

$\overline A_t=(A_1,\dots,A_t)$8

(McClean et al., 14 Jul 2025). The nuisance estimators need only converge at roughly $\overline A_t=(A_1,\dots,A_t)$9 rate because the bias is second order and the remainder is $\overline X_t=(X_1,\dots,X_t)$0 (McClean et al., 14 Jul 2025).

To avoid direct density-ratio estimation, the paper uses the identity

$\overline X_t=(X_1,\dots,X_t)$1

so that $\overline X_t=(X_1,\dots,X_t)$2 can be estimated through standard binary regressions for conditional probabilities (McClean et al., 14 Jul 2025). The paper explicitly characterizes this reformulation as machine-learning friendly (McClean et al., 14 Jul 2025).

7. Relation to treatment-window CCW and empirical illustration

The connection to CCW clarifies a broader family resemblance among longitudinal weighting estimands. In CCW, each eligible individual is cloned into the regimens under study, clones are censored when they become incompatible with the regimen, and inverse probability of censoring weights reweight the remaining uncensored person-time (Webster-Clark et al., 2024). For a “start by day 30” regimen, those who have not initiated by day 30 are censored at day 30, and the weights are based on the inverse of the conditional probability of remaining uncensored: $\overline X_t=(X_1,\dots,X_t)$3 (Webster-Clark et al., 2024).

The paper on CCW emphasizes that when exposure effects are time-varying, ignoring exposure history when estimating IPCW can estimate the risk under an impossible intervention and can create selection bias (Webster-Clark et al., 2024). Its “limited CCW” versus “all initiator CCW” distinction formalizes this point: the all-initiator approach can estimate the risk under an impossible intervention when exposure effects vary over time because it retroactively gives day-30 starters the prior exposure history of earlier starters (Webster-Clark et al., 2024). The simplifying assumptions under which this problem disappears are “No exposure effect,” “Exposure effect begins after the end of the period,” and “Instantaneous effect,” with different implications for whether earlier initiators may be included in IPCW and whether the distribution of treatment timing can be ignored (Webster-Clark et al., 2024).

A plausible implication is that cumulative cross-world weighted effects and CCW illuminate the same conceptual boundary from different directions. CCW shows how cross-world structure can arise inadvertently from weighting within treatment initiation windows; cumulative cross-world weighted effects make that structure explicit and treat it as the target of inference.

The empirical illustration in the cumulative cross-world weighting paper uses the wagepan dataset, following Vella and Verbeek, with worker data from 1980–1987 and analysis focused on 1980–1983 (McClean et al., 14 Jul 2025). The treatment is union membership each year, the target regimes are always unionized $\overline X_t=(X_1,\dots,X_t)$4 and never unionized $\overline X_t=(X_1,\dots,X_t)$5, and the outcome is log wage in 1983 (McClean et al., 14 Jul 2025). Estimation uses smooth trimming,

$\overline X_t=(X_1,\dots,X_t)$6

together with a doubly robust-style estimator, five-fold cross-fitting, and SuperLearner nuisance estimation with linear model, GLM, lasso, regression tree, and random forest learners (McClean et al., 14 Jul 2025).

The reported effect estimate is

$\overline X_t=(X_1,\dots,X_t)$7

with 95% CI

$\overline X_t=(X_1,\dots,X_t)$8

(McClean et al., 14 Jul 2025). On the log-wage scale, the paper states that this corresponds to roughly a 22% increase in wages in 1983 attributable to always being in a union versus never being in a union, under the paper’s assumptions (McClean et al., 14 Jul 2025). The authors note, however, that the no-unmeasured-confounding assumption may be questionable and that sensitivity analysis would be valuable (McClean et al., 14 Jul 2025).

Cumulative cross-world weighted effects therefore occupy a specific niche in longitudinal causal inference: they are cumulative products of time-specific weights, cross-world because they use propensity information from two counterfactual treatment histories, identifiable without standard positivity, mechanistically faithful but not implementable, and substantively informative only under partial common support (McClean et al., 14 Jul 2025).