UA-DCM: Uncertainty-aware Causal Decision Making via Effect Bound Decomposition

Hilaf Hasson; Md Musfiqur Rahman; Murat Kocaoglu; Ziwei Jiang

REVIEW 4 major objections 4 minor 3 references

UA-DCM: Uncertainty-aware Causal Decision Making via Effect Bound Decomposition

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Causal decision-makers can know with high probability when more observational data will never settle the best action, because the intersection of effect bounds across a confidence set isolates an irreducible non-identifiability region.

desk verdict The nonID/sample uncertainty decomposition is a real conceptual contribution, and the formal interval-intersection results are fine; the practical tri-state decision is on shakier ground than the paper's wording suggests. read the letter →

arxiv 2601.22736 v2 pith:7D5D55MN submitted 2026-01-30 cs.LG cs.AI

Md Musfiqur Rahman , Ziwei Jiang , Hilaf Hasson , Murat Kocaoglu This is my paper · ORCID

classification cs.LGcs.AI

keywords causaldecisionmakingpartialidentificationeffectboundsunobservedconfoundingepistemicuncertaintyneuralmodelsconfidencesetsnon-identifiability

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that uncertainty in causal effect estimates from finite observational data can be split into two parts with different remedies. The inner 'nonID' region—the intersection of effect bounds over all distributions consistent with the data—cannot be narrowed by collecting more samples; only measuring new variables (such as instruments) or running an experiment can shrink it. The outer 'sample uncertainty' band—the part of the bounds that varies across the confidence set—does shrink as the sample grows. If the decision threshold falls inside the inner region, the paper's algorithm says 'observe' rather than 'collect', saving resources. If true, this gives practitioners a principled stopping signal for observational data collection in confounded settings.

What carries the argument

The central machinery is the intersection-vs-union decomposition of effect intervals over a confidence set of observational distributions, computed via the four quantities Lx = max_P min_S P(y|do(x)), Lx = min_P min_S ..., Ux = min_P max_S, and Ux = max_P max_S. Theorems 2 and 3 establish that effect sets are contiguous intervals (surjectivity via connectedness/intermediate value), so min/max suffice; Theorem 4 gives intersection = [Lx, Ux] and Corollary 1 gives union = [Lx, Ux]. The practical optimizer is a Lagrangian-relaxed deep causal model: it maximizes or minimizes the effect while a dual parameter enforces that the entailed distribution stays within epsilon of the target candidate; ep

What would settle it

Take a small bow graph (binary X, Y, U) with a known SCM. The true nonID interval [L*,U*] for P(y|do(x)) can be computed exactly by enumerating response-variable bounds (Balke-Pearl linear program). Run UA-DCM with finite samples of increasing size. If the reported inner region [Lx,Ux] does not converge to [L*,U*] asymptotically (either failing to cover values in I* or covering values outside it), then the optimization missed extreme SCMs and the claim 'more samples cannot resolve this' is not trustworthy.

Watch

Extended reading notes

Core claim

On the paper's own terms: for any graph with unobserved confounders and discrete observed variables, the causal effect P(y|do(x)) is not a point even with infinite data; it lies in an interval [L,U] over all SCMs compatible with the true observational distribution. With finite samples the observable distribution itself is only known up to a confidence set C; the paper defines the nonID uncertainty as the intersection of these intervals over P in C, and the sample uncertainty as their union minus that intersection. Theorems 4 and 6 prove the intersection is exactly [max_P min_S effect, min_P max_S effect] — the two inner bounds — so estimating four extreme quantities (min-max, max-max, min-mi

Load-bearing premise

The load-bearing premise is that the neural-network training and epsilon-net sampling actually find the true maximum and minimum causal effect for every distribution in the confidence set; the paper acknowledges optimization may be imperfect and yield smaller bounds than the true ones, in which case the 'irreducible' inner region could be under-estimated and the Observe/Collect decision could be wrong.

Editorial extensions

If this is right

If the decision threshold lies in the inner region [Lx, Ux], practitioners can stop collecting observational samples with high probability and instead measure instruments or confounders, or run a randomized study.
As sample size grows, the outer band shrinks and the inner region widens, so the algorithm's 'collect' advice naturally terminates when the threshold exits the widening inner region.
Adding an instrument variable shrinks the nonID interval (demonstrated on synthetic and Parents' Labor Supply data), giving a concrete route to resolve ambiguity.
For identifiable queries, the decomposition still yields finite-sample bounds, and the inner region collapses to a point as n goes to infinity, subsuming standard finite-sample effect bounds.
The decision rules extend to multiple actions via ATE; ambiguity is defined as 0 lying inside the ATE inner region, giving the same Return / Observe / Collect tri-state protocol.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same intersection/union decomposition could apply to any causal estimand with a continuous effect mapping, such as counterfactual probabilities or mediated effects, because the contiguous-interval argument relies on topology, not on P(y|do(x)) specifics.
An adaptive data-collection policy could be built on this: keep collecting while the threshold is in the outer band, then switch to variable measurement once it enters the inner region; the paper's framework provides the stopping criterion for the first phase.
A conservative modification—deliberately enlarging the inner region to account for neural optimization error—would make the 'observe' recommendation safe but weaken 'collect'; the current heuristic trades certified safety for tightness.
The epsilon-net heuristic is the main scalability bottleneck; replacing it with an exact or certified min-max optimizer would convert the 'high probability' claim into a rigorous guarantee.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The genuinely new thing here is the decomposition: take the intersection of causal-effect bounds over a confidence set of observational distributions, and you get an inner interval that cannot shrink with more data and an outer band that can. Given exact min/max over the confidence set, the main theorems are interval-arithmetic identities and the Propositions are straightforward monotonicity claims. That is a clean, useful way to frame "when should I stop collecting observations and instead measure new variables or run a trial." The Angrist-Evans demonstration is a nice sanity check: the inner region widens and the outer band shrinks with more samples, just as the theory says. The paper is also honest about its own limitations.

The soft spot is exactly where the practical claim lives. The high-probability "Return/Observe/Collect" decision requires the four quantities in Definition 6/12 to be computed reliably. Algorithm 2 samples a finite epsilon-net instead of searching the confidence set, and Section 5.3 admits "it is possible that we miss the true distribution." Algorithm 3 trains neural SCMs with a Lagrangian relaxation, and Appendix A admits the optimization "may be imperfect in practice, giving smaller bounds than the true ones." Smaller bounds make the estimated inner region too small, so the algorithm can output Collect or even Return when the true nonID region covers the threshold, or output Observe when the decision is still sample-limited. Corollary 3's 3-epsilon guarantee assumes the algorithm already found the true smallest max over all epsilon balls—exactly the unverified property—and it is proved only for a binary bow graph.

That said, this is a gap between the conceptual result and its approximation, not a flaw in the conceptual result itself. The theorems hold under exact optimization, and the authors are explicit that the heuristics are approximate. What is missing is any conservative correction, coverage check, or repeated-seed error analysis, and no code is released. The empirical evaluation is therefore illustrative rather than confirmatory.

Who gets value: researchers in partial identification, causal decision-making, and anyone thinking about data collection strategies under unobserved confounding. The paper introduces a framing worth engaging with even if the practical tool needs hardening. It deserves a serious referee, but the referee should push for either conservative bounds or a validation protocol that demonstrates the decisions are reliable in practice. I would send it to review, with the expectation of major revision.

Referee Report

4 major / 4 minor

Summary. The paper introduces a framework, UA-DCM, for decomposing the epistemic uncertainty in causal effect estimation from finite observational data into two components: a 'nonID' inner region that, with high probability, cannot be eliminated by collecting more samples, and an outer 'sample uncertainty' band that may be eliminated. For a confidence set C of observational distributions and a fixed graph G, the inner region is defined as the intersection over P in C of the intervals I_P of causal effects achievable by SCMs compatible with P; Theorems 4–6 identify this intersection with [max_P min_S f, min_P max_S f] for both P(y|do(x)) and ATE. The paper proposes to approximate the four extrema by an epsilon-net over C (Algorithm 2) and by training deep causal models with a Lagrangian relaxation (Algorithm 3), then uses the resulting bounds to output one of three moves: Return the best action, Observe additional variables, or Collect more samples. Experiments on synthetic and real labor-supply data illustrate the behavior of the method.

Significance. If the four extrema in Definition 6/12 could be computed exactly, or conservatively approximated with guarantees, the proposed decomposition would be practically valuable: it gives a principled tri-state decision and can tell practitioners when further observational data collection is futile. The interval-arithmetic identities in Theorems 4–6 are correct and do not depend on the neural approximation. The paper is also commendably transparent about its main limitations, explicitly admitting that the epsilon-net may miss the true distribution and that DCM training may give smaller bounds than the true ones. However, because these two admissions concern the exact quantities that drive the tri-state decision, the central 'with high probability' claim is not established for the practical algorithm as stated.

major comments (4)

[Section 5.3, Algorithms 2–3; Appendix A] The central practical claim is unsupported by any approximation guarantee. The decision rules in Theorem 5 and Theorem 7 are correct only when the four quantities in Eq. (1)/(2) are exact. But Section 5.3 states 'since we are sampling the joint distribution only a finite number of times, it is possible that we miss the true distribution', and Appendix A states 'neural network optimization may be imperfect in practice, giving smaller bounds than the true ones.' Both errors make the estimated inner region a strict subset of the true intersection, so a threshold in the true nonID region can be misclassified as outer or unambiguous, leading to a wrong Collect or Return instead of Observe. No finite-sample or optimization error bound is provided.
[Appendix D, Corollary 3] Corollary 3 is the only nearby error guarantee, but it does not cover the algorithm. It is proved only for the binary bow graph, and its proof assumes that the algorithm has already found the smallest max over all epsilon balls — exactly the property that is unverified because Algorithm 2 samples finitely many joint distributions and Algorithm 3 solves a non-convex Lagrangian relaxation. Thus Corollary 3 cannot support the paper's 'with high probability' conclusion for the general algorithm.
[Definition 6, Theorems 4–6] The paper does not discuss the possibility that the inner region is empty. The intersection over P in C of the intervals I_P equals [max_P min_S f, min_P max_S f], but if max_P min_S f > min_P max_S f, this 'interval' is empty. This can occur when the confidence set is wide. Theorem 5 and Algorithm 1 assume Lx <= Ux; the empty case is undefined. The paper should either state conditions under which nonemptiness is guaranteed or explicitly handle the empty inner region in the decision rules.
[Propositions 1 and 3; Section 5.1] The statement that nonID uncertainty 'cannot be reduced by increasing the number of samples' is potentially misleading. The estimated inner region [Lx,Ux] is the intersection over C, and as C shrinks with more data this intersection expands toward the true nonidentifiability interval I_P. The paper's own real-data experiment reports the inner region width increasing from 0.67 to 0.73 as the sample size grows. What is truly non-reducible is the limiting interval for the true distribution; the finite-sample inner region is only a lower bound on that limit. This distinction should be stated explicitly.

minor comments (4)

[Algorithm 1] The notation in Algorithm 1 is very hard to read: Lx, \underline{L}_x, and \bar{L}_x are visually nearly identical in the rendered text, and the branch conditions in lines 11–16 are therefore ambiguous. Please rewrite with clear subscripts/superscripts and explicitly define which of the four quantities is used in each condition.
[Section 6] The experimental section reports single runs without repeated-seed variance or sensitivity analysis with respect to epsilon_s, alpha, and lambda. Given that the algorithm is a heuristic, reporting distribution over seeds and a sensitivity study would materially strengthen the empirical claims.
[Appendix D.1] The proof of Theorem 2 invokes the open mapping theorem after constructing h_G. Since identifiable causal estimands are rational functions of the observed distribution, continuity and the intermediate value theorem directly give surjectivity on a connected domain; the proof is more complicated than necessary.
[Throughout] Typos and small errors: 'neural net-basd' in the Conclusion, 'Corrolary' in Section 5.2, 'contains contains' in Section 6.3, and inconsistent capitalization of 'nonID' vs 'non-id'.

Circularity Check

1 steps flagged · score 2.0 of 10

The uncertainty decomposition is largely a definitional construction: nonID uncertainty is defined as an intersection over the confidence set, so its monotonicity under more data follows by definition. The min-max optimization framework and experiments add independent content, and no fitted-parameter circularity was found.

self definitional [Definition 5 and Proposition 1 (Section 5.1); proof in Appendix D.3]
"Define the set S_P = {S an SCM | P_obs(S) = P, S |= G} and I_P = {P_S(y|do(x)) | S ∈ S_P}. Then define the intersection ∩_{P∈C} I_P to be the nonID uncertainty... Proposition 1. The uncertainty in the causal decision-making problem due to nonID uncertainty in P(y|do(x)) (i.e., inner region) cannot be reduced by increasing the number of samples in the data."

The 'cannot be reduced with more data' property is built into the definition: nonID uncertainty is defined as the intersection over the confidence set C. Since more samples give a smaller confidence set, the intersection over a smaller set is a superset, so the inner region can only expand. Proposition 1 (and its ATE analog, Proposition 3) is therefore an immediate monotonicity consequence of Definition 5/8, not an independent first-principles result. The same definitional status carries into the 'cannot be improved with more data' part of Theorem 5 and Theorem 7.

full rationale

The mathematical decomposition is self-contained as a construction: given a confidence set C, the intersection ∩_{P∈C} I_P is well-defined, and the min-max identities in Theorem 4/6 are correct as standard interval-intersection equalities. The paper's central 'nonID cannot be reduced' claim is a direct consequence of the definition of nonID uncertainty as an intersection over C, combined with the fact that more samples shrink C; it is a definitional monotonicity, not a fitted-parameter prediction. No parameter is fitted to a subset of the data and then 'predicted' as a closely related quantity. The practical algorithm relies on a DCM representability theorem (Theorem 1) cited from prior work by the authors, but that theorem is not the target result and the core decomposition does not depend on it; the citation is therefore not load-bearing in a circular way. The paper's admitted optimization gaps (Section 5.3: 'it is possible that we miss the true distribution'; Appendix A: 'neural network optimization may be imperfect in practice, giving smaller bounds than the true ones') are correctness or robustness risks, not circularity. The real-data demonstrations are consistent with the framework but do not independently validate the definitional monotonicity. Overall, the paper is not circular in a harmful sense; score 2 reflects one minor self-definitional step while the algorithmic contribution remains substantive.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The framework's mathematical decomposition depends on exact causal-effect optimization for every distribution in a valid confidence set; in practice this is replaced by heuristic epsilon-net sampling and non-convex DCM training. The listed axioms make explicit the discrete semi-Markovian setting, the representational assumption on neural causal models, and the statistical assumptions behind the confidence set.

free parameters (4)

epsilon_s (epsilon-net covering width) = not reported
Algorithm 1/2 partition each Hoeffding confidence interval into intervals of width epsilon_s; the paper says smaller values give tighter estimates but require more candidates. It is user-chosen and directly controls the estimated bounds.
alpha (error probability) = not clearly reported
Controls the Hoeffding/Bonferroni confidence set. The text uses alpha/m per dimension in one passage, but no exact experimental value is stated.
lambda (ATE objective weight) = not reported
Algorithm 3 balances the distribution-matching loss with log ATE during training; chosen by hand without a sensitivity analysis.
number of sampled candidate distributions = 150-200
Experiments sample more than 150 (or 130) distributions from the epsilon-net and train a separate DCM for each; finite sampling introduces an unquantified approximation error.

assumptions (7)

domain assumption Variables are discrete
Assumption 1; the confidence-set construction is described for binary/discrete conditional probabilities, simplifying the epsilon-net but limiting generality.
domain assumption SCM is semi-Markovian
Assumption 2; unobserved confounders are modeled through shared exogenous noise in the DCM framework, excluding general latent structures.
domain assumption The ADMG is known
Assumption 3; the neural architectures and bounds are arranged according to the given graph, and the practitioner must already know the confounding structure.
domain assumption DCMs are representative enough to realize every compatible SCM
Definition 3 and Theorem 1 (cited from the authors' prior work) assert that sufficiently large neural causal models can match observational distributions and identifiable interventional distributions; if capacity or training is insufficient, estimated bounds are not true bounds.
standard math Causal estimand is a continuous rational function with positive denominators
Theorems 2/3 require continuity and connectedness of the image to conclude surjectivity onto [a,b]; the proofs assume P(pa(y))>0 and treat probability distributions as continuous variables.
domain assumption Hoeffding+Bonyferroni confidence intervals cover the true distribution with high probability
Section 4 and Appendix C.3 rely on IID sampling and a product-of-intervals confidence set; coverage holds only under those sampling assumptions.
domain assumption Observing an instrument reduces the nonID region
Section 5.3's Observe move assumes instrumental variables shrink the set of compatible SCMs, citing Balke/Pearl/Neuberg; the real-data experiment uses same-sex sibling as an instrument.

how reviews work

0 comments

Cite this review

Pith. "Pith review of UA-DCM: Uncertainty-aware Causal Decision Making via Effect Bound Decomposition." pith.science (2026). https://pith.science/paper/7D5D55MN

@misc{pith2026260122736,
  author       = {Pith},
  title        = {Pith review of: UA-DCM: Uncertainty-aware Causal Decision Making via Effect Bound Decomposition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7D5D55MN}},
  note         = {Machine review of arXiv:2601.22736}
}

read the original abstract

Causal inference from observational data can provide strong evidence for finding the best action in a decision-making scenario without having to perform expensive randomized trials. The causal effect of an action is often not pointwise identifiable even with infinite data due to unobserved confounding factors. Furthermore, having only finitely many samples adds another layer of uncertainty to causal effect estimation. Several existing methods can be used to obtain upper and lower bounds to the causal effect, ranging from symbolic methods to the more recent neural network-based approaches, which implicitly incorporate both sources of uncertainty. However, these methods do not inform whether collecting more samples may or may not help identify the best action from observational data, leaving experts in the dark about their data collection strategies. We address this problem with a novel framework that can distinguish the range of causal effect values that might be eliminated by collecting more samples from the range of values that, with high probability, cannot be eliminated with more observational samples. We show that this partitioning can be obtained by solving max-min and min-max optimization problems. We leverage neural causal models to approximately recover this decomposition in practice. We demonstrate via experiments on synthetic and real-world datasets that our algorithm can determine when collecting more samples will not help determine the best action. Our framework can help practitioners decide when to resort to non-observational studies or seek to measure some of the unmeasured confounders for optimal decision-making.

Figures

Figures reproduced from arXiv: 2601.22736 by the authors.

**Figure 1.** Left: The case of infinite observational data. An identifiable causal effect can be estimated pointwise as in (a). Otherwise, we obtain bounds as in (b). Right: Given finite samples, we cannot pointwise estimate since the true distribution lies in a continuum of a confidence set. We can only obtain an interval as in (c). When the effect is not identifiable, we obtain a larger bound as in (d). Our goal is to analyze … view at source ↗

**Figure 2.** The proposed inner and outer causal effect bounds. In many scenarios, we need to compare the bounds of the causal effect P(y|do(x)) with a scalar quantity such as the marginal P(y). Suppose we need to make the critical decision of choosing an action X = x for the population level to improve the outcome. We can determine that by comparing the bound of P(Y |do(x)) with the probability P(Y ): the model is biased if P(Y… view at source ↗

**Figure 4.** Experiment Results on Parents Labor Supply Dataset [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Causal graphs used in experiments E.2 Experiment detail for section 6.1 Setup: In the first experiments, we consider the causal graph shown in Figure 5a where we have 2 instrument variables (I1, I2), one unobserved confounder (U), a treatment (X) and an outcome variabl…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 1 linked inside Pith

[1]

@esa (Ref

\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...
[2]

\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...
[3]

o baum and Peter G \

@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

arXiv 2011

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

@esa (Ref

\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

[2] [2]

\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

[3] [3]

o baum and Peter G \

@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

arXiv 2011