Finite-History Estimand

• Finite-history estimand is defined as a target parameter that depends solely on a bounded segment of observed data, ensuring that causal effects are estimated without relying on future events.
• It is applied in various settings such as dynamic panel experiments, multistate clinical trials, and recurrent-event analyses by comparing recent treatment paths or finite histories.
• The concept distinguishes between target estimands and history-dependent nuisance models, enhancing real-world interpretability and preventing bias from conditioning on post-intervention outcomes.

Searching arXiv for the cited papers to ground the article in published work. Finite-history estimand denotes a target parameter whose definition depends on a bounded history rather than on an unbounded trajectory or on conditioning that reaches into the future. Across the relevant literatures, that bounded history may be the last $p+1$ treatment assignments in a panel experiment, the observable multistate process up to time $t$ or a finite horizon $\tau$, or the recurrent-event and survival history realized before death (Bojinov et al., 2020, Viviano et al., 2021, Bühler et al., 2022, Ragni et al., 2024). Other papers use closely related finite-history structure without introducing a separate finite-history estimand as such: coalition-based interventional queries in causal explainability, history-dependent nuisance models in off-policy evaluation, and finite belief histories in Theory-of-Mind benchmarking (Parafita et al., 24 Sep 2025, Zhou et al., 28 May 2025, Tang et al., 2024).

1. Conceptual scope

The most stable feature of a finite-history estimand is that the target is indexed by a bounded segment of a process. In panel experiments, the target compares outcomes under alternative recent treatment paths of length $p+1$, while holding earlier history fixed (Bojinov et al., 2020). In dynamic observational panels, the target is a mean potential outcome under a specified finite treatment history $d_{1:T}$, or under only the last $h$ treatment assignments when a finite exposure window is of interest (Viviano et al., 2021). In multistate clinical-trial methodology, the relevant history is the observed process up to $t^{-}$ or a fixed horizon $\tau$, and the recommended estimands are marginal features such as state occupancy, cumulative incidence, survival, restricted mean time, or utility-weighted summaries rather than post-randomization conditionals (Bühler et al., 2022). In recurrent-event settings with death, the patient’s history enters through the event count before $t$ and the alive time $T_D \wedge t$ (Ragni et al., 2024).

A key negative characterization is equally important. Several papers explicitly separate finite-history structure from finite-history estimands. The off-policy evaluation work keeps the estimand fixed as the ordinary policy value and varies only the history length used in the estimated behavior policy inside the weights (Zhou et al., 28 May 2025). The do-SHAP work uses finite coalitions $t$0 in a static DAG, but does not define a temporal finite-history estimand (Parafita et al., 24 Sep 2025). The Theory-of-Mind benchmark introduces finite belief history as a reasoning requirement rather than a formal statistical estimand (Tang et al., 2024).

This suggests that “finite-history estimand” functions less as a single universal definition than as a structural criterion: the target depends on a bounded, explicitly delimited history and avoids reliance on future paths or unbounded memory.

2. Dynamic causal effects in panels and treatment paths

In panel experiments, the finite-history estimand is formalized through path-indexed potential outcomes. With non-anticipation, $t$1 depends only on assignments up to time $t$2, so it can be written as $t$3. The basic dynamic causal effect compares two histories,

$t$4

and the lag-$t$5 version restricts attention to a recent path of length $t$6: $t$7 Its population versions average over units, time, or both, including

$t$8

The associated Horvitz–Thompson-type estimator is unbiased over the randomization distribution, and the paper develops finite-population asymptotics and conservative as well as exact randomization-based inference (Bojinov et al., 2020).

Dynamic covariate balancing adopts the same path-based logic in an observational panel. The central mean potential outcome is

$t$9

with causal contrast

$\tau$0

The paper also isolates a finite exposure window $\tau$1 by comparing

$\tau$2

Identification is built from no anticipation, sequential ignorability, recursive projections of potential-outcome conditional means on past histories, and dynamic covariate balancing rather than a fully specified propensity score model (Viviano et al., 2021).

These formulations make finite history a property of the causal target itself. They also clarify why standard regression summaries can fail: the panel-experiment paper shows that linear fixed-effects estimators do not recover a causally interpretable estimand if there are dynamic causal effects and serial correlation in assignments (Bojinov et al., 2020). A plausible implication is that bounded-history path comparisons are not merely descriptive conveniences; they are often required to preserve causal meaning under carryover and sequential dependence.

3. Multistate and recurrent-event formulations

In multistate clinical-trial methodology, the relevant process history is

$\tau$3

and transition intensities are defined by

$\tau$4

The paper’s central argument is that these dynamic conditional features should generally not serve as primary trial estimands. Instead, primary estimands should be marginal process features defined from what has happened by time $\tau$5 or $\tau$6: state occupancy $\tau$7, cumulative incidence $\tau$8, survival $\tau$9, restricted mean time

$p+1$0

and utility-weighted means such as

$p+1$1

The recommended estimand is the comparison of these marginal features across treatment arms, for example at a clinically chosen $p+1$2 (Bühler et al., 2022).

The rejection of history-conditioned intensities as primary estimands is substantive rather than semantic. The paper shows that even if true conditional intensities are proportional, the marginal intensity ratio can become time-dependent because conditioning on being event-free at time $p+1$3 introduces post-randomization selection. The same paper therefore emphasizes “real-world” interpretability, explicit handling of intercurrent events, and avoidance of targets that condition on future information or on counterfactual principal strata (Bühler et al., 2022).

The patient weighted while-alive estimand is a direct finite-history example in recurrent events with death. Its causal target is

$p+1$4

where

$p+1$5

and the causal contrast is

$p+1$6

When $p+1$7 is the identity, the estimand is the patient-level mean of the number of recurrent events up to $p+1$8 divided by time alive up to $p+1$9. Identification uses the g-formula

$d_{1:T}$0

under consistency, exchangeability given $d_{1:T}$1, and positivity (Ragni et al., 2024).

Here, finite history is not merely a time cutoff. The numerator counts only events before death and before $d_{1:T}$2, while the denominator uses the alive time $d_{1:T}$3. This makes the estimand explicitly dependent on a finite realized event history while avoiding extrapolation to unobserved post-death paths.

4. Localized interventional queries in structural causal models

The do-SHAP framework provides a static-DAG analogue of finite-history dependence. Its core value function is the interventional query

$d_{1:T}$4

where $d_{1:T}$5 is a coalition of features for the factual instance $d_{1:T}$6, and the do-Shapley attribution for a feature $d_{1:T}$7 is

$d_{1:T}$8

The paper emphasizes that $d_{1:T}$9 is an interventional query on the causal model, not a purely associational quantity (Parafita et al., 24 Sep 2025).

The work does not define a temporal finite-history estimand. Its finite-history relevance comes from two structural facts. First, the estimand depends on a finite coalition $h$0, which the paper describes as a finite intervention history over selected variables. Second, many coalitions can be reduced without changing the estimand. If $h$1 is a frontier between $h$2 and $h$3, then

$h$4

The paper defines an irreducible subset $h$5 of a coalition $h$6 and shows that $h$7. Computationally, this is implemented through the Frontier-Reducibility Algorithm, topological scanning, caching, and a bitmask encoding $h$8 (Parafita et al., 24 Sep 2025).

The main methodological contribution is estimand-agnostic causal inference. Instead of deriving a bespoke estimand for each query, a single proxy SCM $h$9 with graph $t^{-}$0 is trained so that

$t^{-}$1

and then identifiable interventional queries are estimated by sampling from the intervened learned model. The paper states that this makes do-SHAP feasible on complex graphs (Parafita et al., 24 Sep 2025).

A plausible implication is that finite-history ideas extend beyond temporal data: in a DAG, only a finite, graph-determined subset of interventions may be causally effective for the target variable.

5. Finite history in nuisance models rather than in the estimand

Off-policy evaluation in reinforcement learning provides a sharp counterexample to the idea that any bounded-history object defines a finite-history estimand. The target throughout is the standard policy value

$t^{-}$2

The paper assumes the true behavior policy is Markov, $t^{-}$3, but studies estimators that replace it by a history-dependent estimate

$t^{-}$4

where $t^{-}$5 is a $t^{-}$6-step history and $t^{-}$7 is nested (Zhou et al., 28 May 2025).

The paper’s theoretical result is a bias-variance decomposition for ordinary importance sampling and related estimators. For ordinary and sequential IS, and often for doubly robust estimation, conditioning on longer history decreases asymptotic variance while increasing finite-sample bias; for marginalized IS, longer history worsens asymptotic mean squared error (Zhou et al., 28 May 2025). The finite-history variable is therefore $t^{-}$8, but $t^{-}$9 indexes the nuisance model used inside the weights, not the target policy value.

This distinction is central. A finite-history estimand changes the target functional. A finite-history nuisance specification changes how the same target is estimated. The paper explicitly treats the latter, not the former (Zhou et al., 28 May 2025).

6. Related non-statistical usage and interpretive boundaries

The Theory-of-Mind literature supplies another adjacent usage. The taxonomy of zero, finite, and infinite belief history defines finite belief history as a setting in which the model must reason over a finite sequence of prior belief states to infer another agent’s latest belief. In the “Pick the Right Stuff” benchmark, the finite-history condition is created by adding a third room that shows a random previous snapshot, described indirectly as an “nth-to-last” earlier state; the model must reconstruct which past state the user observed and infer the resulting belief (Tang et al., 2024).

The paper does not use “estimand” in the formal statistical sense. The closest target quantity is the user’s latest belief about the item’s position given a finite history of observations and updates, measured by average score rather than by a parametric functional. Empirically, performance on Zero Belief History is consistently better than on Finite Belief History, with reported average scores of $\tau$0 and $\tau$1, respectively (Tang et al., 2024).

Taken together, the cited works support three interpretive boundaries. First, finite history may index the causal target itself, as in lag-$\tau$2 treatment-path effects, multistate marginal features up to $\tau$3 or $\tau$4, and while-alive recurrent-event summaries (Bojinov et al., 2020, Bühler et al., 2022, Ragni et al., 2024). Second, finite history may instead index the estimation strategy or nuisance model, as in history-dependent importance-weight denominators (Zhou et al., 28 May 2025). Third, finite history may describe the informational burden of a reasoning task rather than an estimand at all, as in finite belief-history benchmarks (Tang et al., 2024).

The most defensible general definition is therefore restrictive: a finite-history estimand is a target parameter that is explicitly determined by a bounded observed, intervened, or treatment history. Under that definition, the term excludes many history-dependent procedures, even when bounded memory plays a central practical role.