arxiv: 2605.12847 · v1 · submitted 2026-05-13 · 📊 stat.ME

Recognition: unknown

Never Too LATE: A Fully Stochastic Update to the Potential Outcome Framework

Hanti Lin

Authors on Pith no claims yet

Pith reviewed 2026-05-14 19:07 UTC · model grok-4.3

classification 📊 stat.ME

keywords stochastic potential outcomesinstrumental variableslocal average treatment effectcausal Bayes netsRubin causal modelcompliersdegree of compliance

0 comments

The pith

The instrumental variable estimand identifies the degree-of-compliance-weighted average treatment effect when potential outcomes are modeled as stochastic Bernoulli parameters.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper replaces deterministic potential outcomes in the Rubin model with stochastic versions expressed as Bernoulli parameters in separate probability spaces. These parameters link to observed data through the factorization rule of a causal Bayes net, avoiding any assumption that every counterfactual settles a unique determinate outcome. It defines the Degree-of-compliance-weighted Average Treatment Effect (DATE) and proves that this quantity equals the standard instrumental variable estimand once the usual LATE assumptions are restated in terms of the Bernoulli parameters. The classic deterministic LATE result appears as a special case, which means existing IV practice has been estimating the DATE in a general stochastic setting all along.

Core claim

Stochastic potential outcomes are introduced as Bernoulli parameters in their own small probability spaces and connected to observables via the factorization rule of a causal Bayes net. The paper defines the Degree-of-compliance-weighted Average Treatment Effect (DATE) and shows that, under assumptions analogous to monotonicity, exclusion restriction, and instrument relevance but rewritten for the stochastic parameters, the DATE equals the usual IV estimand. The classic LATE identification result emerges as the deterministic special case, so existing IV practice can be reinterpreted as having estimated the DATE without assuming the unique-parallel-universe view.

What carries the argument

The Degree-of-compliance-weighted Average Treatment Effect (DATE), constructed from stochastic potential outcomes as Bernoulli parameters and identified through causal Bayes net factorization.

If this is right

Standard IV methods identify the DATE even when potential outcomes remain irreducibly stochastic.
The classic LATE result is recovered exactly when the Bernoulli parameters become deterministic 0-1 values.
Existing instrumental-variable applications can be reinterpreted as targeting the DATE without invoking unique parallel universes.
The framework preserves identification while dropping the requirement that every counterfactual condition settles a single outcome.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach may extend naturally to domains with intrinsic randomness such as clinical trials with biological variability.
It offers a way to maintain causal identification while aligning statistical models more closely with philosophical doubts about deterministic counterfactuals.
Similar weighting constructions could be applied to other estimands such as the stochastic average treatment effect on the treated.

Load-bearing premise

The causal Bayes net factorization rule connects the stochastic Bernoulli parameters to observables, and the monotonicity, exclusion, and relevance conditions continue to hold when restated directly in terms of those parameters.

What would settle it

In a population where individual compliance degrees and potential-outcome probabilities can be measured directly, an IV estimate that systematically deviates from the computed DATE would falsify the claimed equality.

Figures

Figures reproduced from arXiv: 2605.12847 by Hanti Lin.

**Figure 1.** Figure 1: A coarse-grained possible state of the world, where the red branch represents what actually happens in that state The second sentence says that something stochastic holds in this state of the world: under the counterfactual antecedent take = 0, i’s being cured is governed by a chance of 0.5 as if tossing a fair coin, and nothing further is said about which way it would actually have gone—after all, “that … view at source ↗

**Figure 2.** Figure 2: A fine-grained possibility, where the red branch represents the actual universe, and the blue branch represents the unique parallel universe under the counterfactual condition of not taking the treatment • for each possible intervention of interest, a unique counterfactual universe in which the outcome under that intervention is fully settled—expressed by a simple, non-stochastic counterfactual, and rep… view at source ↗

**Figure 3.** Figure 3: A causal Bayes net for instrumental variable estimation Indiv, represents the individual randomly chosen from the population under study. Accordingly, Cure denotes whether the randomly chosen person is cured or not. The design of this confounding variable is intended to make it very fine-grained, capturing all confounding factors (at least at a slice of time, as the present setup still does not model chang… view at source ↗

read the original abstract

In the classic potential outcome framework, the local average treatment effect (LATE) and its identification via an instrumental variable are stated in a deterministic setting at the individual level: each individual has settled potential outcomes such as ``cured if treated''. Several authors have proposed working instead with \emph{stochastic} potential outcomes -- counterfactual probabilities of the form ``the chance of being cured if treated'' -- but the integration of stochastic potential outcomes with the LATE machinery raises an issue. It is a metaphysical issue: in a stochastic setting, the standard joint-probability definitions of compliers and the LATE assume what I will call the \emph{unique-parallel-universe view}, which asserts that, in any genuinely possible state of the world, every counterfactual condition settles a unique determinate outcome even when the underlying causal disposition is irreducibly chancy. The statistician Dawid (2000) doubts the plausibility of this view; the philosopher Lewis (1973) develops a reductio argument against it. I propose a fully stochastic update to the Rubin causal model that drops the assumption of the unique-parallel-universe view: stochastic potential outcomes are introduced as Bernoulli parameters in their own (small) probability spaces, and are connected to observables via the factorization rule of a causal Bayes net. Within this framework, I define a Degree-of-compliance-weighted Average Treatment Effect (DATE) and prove that, under assumptions analogous to those used for the LATE but rewritten for the fully stochastic setting, the DATE equals the usual IV estimand. The classic LATE identification result emerges as a deterministic special case. Existing IV practice can therefore be reinterpreted: it has been estimating the DATE all along, in a general stochastic setting, without assuming the unique-parallel-universe view.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a stochastic rewrite of LATE with a new DATE quantity that recovers the IV estimand, but the proof sketch leaves open whether rewritten monotonicity actually blocks negative compliance degrees.

read the letter

This paper replaces fixed potential outcomes with stochastic Bernoulli parameters in separate probability spaces and links them to data through a causal Bayes net factorization. It defines DATE as a compliance-degree-weighted average of individual treatment-effect parameters and claims this equals the standard IV ratio under monotonicity, exclusion, and relevance restated in terms of the Bernoulli parameters. The classic deterministic LATE appears as the special case where the parameters are 0 or 1 with probability 1. That move directly answers the unique-parallel-universe objection raised by Dawid and Lewis, and it lets existing IV practice be re-read as estimating DATE without extra metaphysical assumptions. The framework is internally consistent on its own terms and the citation pattern to the LATE literature is appropriate. The soft spot is the identification step. The abstract asserts the equality holds once the assumptions are translated, yet the factorization rule alone does not force every unit’s compliance degree (difference in treatment probabilities under Z=1 versus Z=0) to be non-negative. If the population distribution over those parameters admits negative values, the weighted sum picks up an extra term proportional to the negative weights, exactly as defiers do in the deterministic case. No part of the provided abstract shows how the rewritten monotonicity rules this out or why the equality survives anyway. Readers working on causal foundations will find the setup useful for thinking about chancy counterfactuals. The paper is coherent enough to merit referee time; a serious editor should send it out so the proof can be checked in detail, especially the monotonicity translation.

Referee Report

1 major / 1 minor

Summary. The paper proposes a fully stochastic update to the Rubin causal model by treating potential outcomes as Bernoulli parameters in small probability spaces, connected to observables via the factorization rule of a causal Bayes net. It defines a Degree-of-compliance-weighted Average Treatment Effect (DATE) as a compliance-degree-weighted average of individual treatment-effect parameters and claims to prove that, under assumptions analogous to monotonicity, exclusion restriction, and instrument relevance but rewritten in terms of the stochastic Bernoulli parameters, the DATE equals the standard IV estimand. The classic LATE identification result is recovered as a deterministic special case, allowing reinterpretation of existing IV practice as estimating the DATE without assuming the unique-parallel-universe view.

Significance. If the central identification result holds rigorously, the work supplies a coherent metaphysical foundation for stochastic potential outcomes within the LATE/IV framework, potentially broadening applicability to settings with irreducible randomness while preserving the validity of standard IV estimators. It explicitly credits the deterministic special case and avoids parameter-fitting or ad-hoc adjustments, which strengthens its claim to unify deterministic and stochastic approaches.

major comments (1)

[identification theorem / DATE=IV equality] The central identification result (DATE equals IV estimand) is stated to hold under a rewritten monotonicity assumption on the Bernoulli parameters, but the factorization rule of the causal Bayes net does not by itself constrain the population distribution to exclude negative compliance degrees. If stochastic defiers are admitted, the weighted sum can differ from the IV ratio by a term proportional to the negative weights, exactly as in the deterministic case; no derivation is provided showing that an expectation-based monotonicity condition suffices to restore equality.

minor comments (1)

The abstract and introduction could clarify the precise measure-theoretic construction of the small probability spaces for the Bernoulli parameters to avoid ambiguity about how they factor into the joint distribution.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful and constructive report. The central identification result is indeed the heart of the paper, and we welcome the opportunity to clarify the role of the monotonicity assumption in the fully stochastic setting.

read point-by-point responses

Referee: The central identification result (DATE equals IV estimand) is stated to hold under a rewritten monotonicity assumption on the Bernoulli parameters, but the factorization rule of the causal Bayes net does not by itself constrain the population distribution to exclude negative compliance degrees. If stochastic defiers are admitted, the weighted sum can differ from the IV ratio by a term proportional to the negative weights, exactly as in the deterministic case; no derivation is provided showing that an expectation-based monotonicity condition suffices to restore equality.

Authors: We agree that the factorization rule of the causal Bayes net alone does not rule out negative compliance degrees; an explicit monotonicity assumption is required, exactly as in the deterministic LATE theorem. In the manuscript this assumption is imposed directly on the individual-level Bernoulli parameters: for each unit i, the treatment probability under Z=1 is at least as large as under Z=0. This condition is stated in terms of the parameters themselves (not their expectations), which by construction excludes negative compliance degrees before any averaging occurs. The DATE is then defined as the compliance-degree-weighted average of the individual treatment-effect parameters, and the identification argument shows that the IV ratio equals this weighted average because the negative-weight terms are absent. We acknowledge that the main-text presentation is concise and that a fully expanded, line-by-line derivation of the expectation step (showing how the parameter-level monotonicity carries through to the population moments) is not spelled out in the current version. We will add this derivation, together with an explicit statement that the monotonicity is imposed on the parameters rather than on their expectations, in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity: DATE defined from stochastic parameters and derived equal to IV estimand

full rationale

The paper defines stochastic potential outcomes as Bernoulli parameters in separate probability spaces, connects them to observables via the factorization rule of a causal Bayes net, and defines DATE as the compliance-degree-weighted average of individual treatment-effect parameters. It then derives that DATE equals the IV estimand under assumptions rewritten for the stochastic setting (analogous to monotonicity, exclusion restriction, and relevance). This is a mathematical derivation from the model primitives rather than a self-definitional equivalence, a fitted input renamed as prediction, or a load-bearing self-citation chain. The classic LATE is recovered as a deterministic special case, confirming the result extends the framework without presupposing the target equality by construction. No steps reduce the central claim to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

The central claim rests on translating standard IV assumptions to stochastic terms and using causal Bayes net factorization, with new entities for the stochastic model.

axioms (2)

domain assumption The factorization rule of a causal Bayes net connects stochastic potential outcomes to observables.
Invoked to link the new stochastic elements to data.
domain assumption Assumptions analogous to monotonicity, exclusion restriction, and relevance hold when rewritten for the stochastic Bernoulli parameters.
Required for the DATE to equal the IV estimand.

invented entities (2)

Stochastic potential outcomes as Bernoulli parameters no independent evidence
purpose: To model counterfactual probabilities in their own probability spaces.
New representation of potential outcomes to drop deterministic assumptions.
Degree-of-compliance-weighted Average Treatment Effect (DATE) no independent evidence
purpose: To serve as the target causal parameter identified by IV in the stochastic setting.
Invented quantity that generalizes the LATE.

pith-pipeline@v0.9.0 · 5615 in / 1480 out tokens · 63591 ms · 2026-05-14T19:07:27.895827+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 1 canonical work pages

[1]

Identification of Causal Effects Using Instrumental Variables

Angrist, J. D., Imbens, G. W., and Rubin, D. B. (1996) “Identification of Causal Effects Using Instrumental Variables”,Journal of the American Statistical Association, 91(434): 444-455

1996
[2]

Causal Inference without Counterfactuals

Dawid, A. P. (2000) “Causal Inference without Counterfactuals”,Journal of the American Statistical Association95(450): 407-424

2000
[3]

Interpretation and Choice of Effect Measures in Epidemiologic Analysis

Greenland, S. (1987) “Interpretation and Choice of Effect Measures in Epidemiologic Analysis”, American Journal of Epidemiology, 125(5): 761-768. H´ ajek, A. (2014) “Probabilities of Counterfactuals and Counterfactual Probabilities”,Journal of Applied Logic, 12(3): 235-251. 23 H´ ajek, A. (manuscript) “Most Counterfactuals Are False”, Research School of S...

1987
[4]

Potential Outcome and Directed Acyclic Graph Approaches to Causality: Relevance for Empirical Practice in Economics

Imbens, G. W. (2020) “Potential Outcome and Directed Acyclic Graph Approaches to Causality: Relevance for Empirical Practice in Economics”,Journal of Economic Literature58(4): 1129-1179

2020
[5]

Identification and Estimation of Local Average Treatment Effects

Imbens, G. W. and Angrist, J. (1994) “Identification and Estimation of Local Average Treatment Effects”,Econometrica62, 467-476

1994
[6]

Lewis, D. K. (1973)Counterfactuals, Blackwell

1973
[7]

The Logic of Counterfactuals and the Epistemology of Causal In- ference

Lin, H. (2025 manuscript) “The Logic of Counterfactuals and the Epistemology of Causal In- ference”, arXiv:2405.11284 [cs.AI].https://arxiv.org/abs/2405.11284

work page arXiv 2025
[8]

(2009),Causality: Models, Reasoning, and Inference, 2nd Edition, Cambridge Univer- sity Press

Pearl, J. (2009),Causality: Models, Reasoning, and Inference, 2nd Edition, Cambridge Univer- sity Press

2009
[9]

The Probability of Causation under a Stochastic Model for Individual Risk

Robins, J. and Greenland, S. (1989) “The Probability of Causation under a Stochastic Model for Individual Risk”,Biometrics, 45(4): 1125-1138

1989
[10]

Comment on ‘Causal Inference without Counterfactu- als’ by A. P. Dawid

Robins, J. M. and Greenland, S. (2000) “Comment on ‘Causal Inference without Counterfactu- als’ by A. P. Dawid”,Journal of the American Statistical Association, 95(450): 431-435

2000
[11]

Estimating Causal Effects of Treatments in Randomized and Nonrandom- ized Studies

Rubin, D. B. (1974) “Estimating Causal Effects of Treatments in Randomized and Nonrandom- ized Studies”,Journal of Educational Psychology66: 688-701

1974
[12]

Probabilities of Counterfactuals Are Counterfactual Probabilities

Santorio, P. (2025) “Probabilities of Counterfactuals Are Counterfactual Probabilities”,Journal of Philosophy

2025
[13]

Instrumental Variable Estimation with a Stochastic Monotonicity Assumption

Small, D. S., Tan, Z., Ramsahai, R. R., Lorch, S. A., and Brookhart, M. A. (2017) “Instrumental Variable Estimation with a Stochastic Monotonicity Assumption”,Statistical Science, 32 (4), 561-579

2017
[14]

N., & Scheines, R

Spirtes, P., Glymour, C. N., & Scheines, R. (2000)Causation, Prediction, and Search, MIT Press

2000
[15]

A Theory of Conditionals

Stalnaker, R. C. (1968) “A Theory of Conditionals”, in Harper, W. L., Pearce, G. A., & Stal- naker, R. C. (eds.)Ifs: Conditionals, Belief, Decision, Chance and Time, Springer Nether- lands: 41-55

1968
[16]

A Defense of Conditional Excluded Middle

Stalnaker, R. C. (1981) “A Defense of Conditional Excluded Middle”, in Harper, W. L., Pearce, G. A., & Stalnaker, R. C. (eds.)Ifs: Conditionals, Belief, Decision, Chance and Time, Springer Netherlands: 87-104

1981
[17]

Stochastic Counterfactuals and Stochastic Suffi- cient Causes

VanderWeele, T. J. and Robins, J. M. (2012) “Stochastic Counterfactuals and Stochastic Suffi- cient Causes”,Statistica Sinica, 22(1): 379-392. 24

2012
[18]

Defending Conditional Excluded Middle

Williams, J. R. G. (2010) “Defending Conditional Excluded Middle”,Noˆ us, 44(4): 650-668. 25

2010