acceptodds
Under review as a conference paper at ICLR 2027

Neural Proximal Causal Inference for Longitudinal Data with Unmeasured Confounding

Abstract

Unmeasured confounders pose a fundamental challenge in estimating causal effects of time-varying treatments from observational data. Proximal Causal Inference (PCI) addresses this through proxy variables, but estimation under PCI requires solving recursively coupled ill-posed integral equations. Existing longitudinal methods make this tractable only through parametric models that cannot capture nonlinear trajectory dynamics. We introduce Neural-PCI, a deep learning estimation framework grounded in established proxy-based identification and, to our knowledge, the first implementable neural method for the full -stage recursion. A joint Maximum Moment Restriction (MMR) formulation enforces all moment conditions through a single closed-form kernel loss, and yields three semi-parametric estimators with complementary robustness properties. We show that stage-wise sequential estimation incurs error growth exponential in the time horizon , while our joint formulation gives additive error aggregation that scales linearly in on average. Experiments on synthetic data and the semi-synthetic MIMIC-III benchmark show that Neural-PCI consistently outperforms parametric PCI, latent variable, and sequential-randomization baselines across horizons.

Primary area:
causal reasoning
Keywords:
Causal inference, Proximal causal inference, Longitudinal data, Maximum moment restriction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MVg9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Neural-PCI, a neural framework for estimating longitudinal causal effects using outcome and treatment proxies to address unmeasured confounding. It jointly estimates recursively coupled bridge functions through maximum moment restriction losses and constructs outcome-regression, weighting, and targeted estimators. The paper presents error analyses and synthetic and semi-synthetic experiments supporting its proposed approach.

Critical strengths

The paper tackles an important practical challenge: making longitudinal proximal causal estimation flexible and computationally implementable beyond restrictive parametric bridge models. Its shared recurrent architecture and joint estimation framework provide a concrete approach to fitting coupled bridge equations, with a dedicated stabilization strategy for treatment bridges.

Critical weaknesses
  1. It seems that the identification assumptions are not demonstrated to hold in the synthetic experimental setting. For Assumption A.6, the paper mentioned that “The condition holds automatically when Ut{U_t} is Markov and structural equations depend on U only through its current value”. I’m not sure if this is true in general. For example, at t=1t=1, A.6 requires \begin{equation} U_0\perp \bigl(Y^{\bar a},W_0,Z_0,W_1,Z_1\bigr)\mid U_1,A_0,X_0,X_1. \end{equation} Consequently, it requires U0⊥W0∣U1,A0,X0,X1U_0 \perp W_0\mid U_1,A_0,X_0,X_1. Consider this simple model: \begin{equation} U_0 \sim N(0,1), \quad U_1 = \frac{1}{2}U_0 + \epsilon, \quad W_0 = U_0 + \eta \end{equation} where ε,η\varepsilon,\eta are independent standard normal noises, independent of U0U_0. Let treatment A0A_0 be independently randomized and X0,X1X_0,X_1 be constants. However, in this simple case, A.6 fails:
\operatorname{Cov}(U_0,W_0\mid U_1,A_0,X_0,X_1) =\operatorname{Cov}(U_0,W_0\mid U_1) = \operatorname{Var}(U_0\mid U_1)= 1-\frac{(1/2)^2}{1+(1/2)^2}= \frac45>0. \end{equation}$$ Even if $A_0$, $X_0$, and $X_1$ contain more information about $U_0$, as the synthetic DGP in Appendix H demonstrated, I believe they do not automatically eliminate the dependence. 2. Theorem F.1 controls bridge-function $L^2$ error, whereas Theorem G.1 controls moment discrepancies. The conversion in Remark G.3 is therefore essential to comparing these bounds, but its proposed control by two neighboring discrepancies appears unjustified. For example, take $T=3$ and add the same nonzero constant $c$ to all nonterminal bridges, keeping $H_3=Y$ fixed. Provided these shifted functions belong to the hypothesis class, both $d_{k_0}$ and $d_{k_1}$ remain zero, while the $L^2$ error of $h_1$ equals $|c|>0$, contradicting the proposed inequality at $t=1$. 3. Section 2.2 motivates the method through two difficulties: ill-posed bridge estimation and recursive error propagation. However, the experiments do not seem to sufficiently distinguish these two advancements. The nonlinear semi-synthetic benchmark tests flexibility against misspecified parametric baselines, but does not establish stability under increasingly ill-posed bridge estimation. Figure 2(a) supports joint estimation empirically, but different regularization-selection criteria and the absence of stagewise bridge-error diagnostics limit attribution of its advantage specifically to reduced recursive error propagation.
Questions for authors
  1. For Assumption A.5, the synthetic DGP would require, among other things, \begin{equation} Z_t\perp W_{t-1}\mid U_t,\bar A_{t-1},\bar X_t, \end{equation} since its outcome proxies are unaffected by treatment interventions. However, Appendix H contains the path \begin{equation} Z_t\leftarrow Z_{t-1}\leftarrow U_{t-1}\rightarrow W_{t-1}, \end{equation} which is not blocked by the stated conditioning set. Thus, the graphical structure does not guarantee this conditional independence. Could the authors explain why the required conditional independence holds under their specific structural equations?

  2. Why does the equivalence in Appendix E.1 transfer double robustness to the targeted estimator?

  3. Why does Neural P-IPW improve as the horizon increases? Its MAE decreases from 0.054 at T=3T=3 to 0.022 at T=20T=20, despite the increasing difficulty of cumulative weighting. What would be an intuitive interpretation of this result?

  4. What explains the abrupt deterioration of MSMM P-OR at T=20T=20? Its MAE rises from 0.076 at T=10T=10 to 2.276 at $T=20, while MSMM P-IPW changes from 0.303 to 0.390.

3 comments

Thank you for a careful review. W1, W2 and Q1 are correct, and each is corrected below; for W3 we ran additional experiments it calls for. The identification formulas (Eqs. 2 and 4) and the Neural P-OR and P-IPW rows of Table 1 are unchanged.

W1 and Q1. Assumption A.6 and the open path

Both points are correct and share one cause: a past proxy such as W0W_0 depends on a past confounder U0U_0, so conditioning on the current UtU_t cannot separate them; the Markov property of UU does not help. The claim that A.6 "holds automatically" was wrong, and A.6 will be removed.

We return to the history form of Ying et al. (2023). A.2, A.5, A.7 and A.8 will be stated with the full latent history Uˉt\bar U_t, and A.6 replaced by the forward condition Zˉt−1⊥(Ut,Xt)∣Aˉt−1,Uˉt−1,Xˉt−1\bar Z_{t-1}\perp(U_t,X_t)\mid\bar A_{t-1},\bar U_{t-1},\bar X_{t-1}, used only in the treatment-bridge induction, so identification follows Ying et al. (2023). In the synthetic DGP, the path Zt←Zt−1←Ut−1→Wt−1Z_t\leftarrow Z_{t-1}\leftarrow U_{t-1}\rightarrow W_{t-1} is blocked once Uˉt\bar U_t is conditioned, and the forward condition holds by d-separation.

W2. Remark G.3

You are right: its control by two neighbouring discrepancies was unjustified, as your example shows. The revised conversion still gives bridge-function L2L^2 error, as Theorem F.1 does. It splits each stage-(t+1)(t+1) error into the part that stage tt can reproduce exactly and the rest. Stage tt copies the reproducible part without raising dktd_{k_t}, so it is passed back. The rest cannot be copied; it shows up in dktd_{k_t}, so its effect is charged to stage tt, once. Your constant shift is reproducible: stage 2, whose target is YY, sees it through dk2∝∣c∣d_{k_2}\propto|c|, and it passes back unchanged, so the bound for h1h_1 uses dk2d_{k_2} and holds.

For the comparison: sequential estimation fits each stage after the next is frozen, so it passes the whole error back, enlarged by up to κ\kappa per stage (Theorem F.1). A fit with small joint loss keeps every dktd_{k_t} small, so the rest stays small at every stage and only the reproducible part is passed back. When reproducible errors are not enlarged in total as they are passed back and the constants of each stage are bounded, the joint bridge error grows at most linearly in TT. Details: our response to ojfN (W1).

W3. Separating ill-posedness from error propagation

We claim one advance, not two: a single ill-posed moment condition is solved by MMR (Mastouri et al., 2021; Kompa et al., 2022), which we take as given. Our contribution is estimating the chain jointly. To address your concerns on the fairness of the comparison in figure 2(a), we ran additional analysis as follows:

Same selection criterion. Both arms now select the regularization weight from the same grid by the validation MMR loss. Updated MAE of sequential NMMR at T=3,5,7,10,20T=3,5,7,10,20: 0.013, 0.042, 0.261, 0.276, 0.598. Against the unchanged joint arm (Neural P-OR in Table 1), the arms agree at T=3,5T=3,5, but from T=7T=7 the sequential error is 5–8 times larger and still rising, so tuning does not explain the gap.

Where the error arises. Our simulator gives the true ATE but not the true bridges, so we check each stage's moment condition instead: on target-regime trajectories, the mean stage residual H^t+1−H^t\hat H_{t+1}-\hat H_t should be zero. In the sequential arm these means have the same sign at almost every stage and add up. In the joint arm they sit in one to three stages and partly cancel.

So the gap comes from error accumulating along the chain, not from a single inversion.

Q2. Does E.1 transfer double robustness?

The targeted output of Eq. (18) contained the observed-treatment indicator and Q^t\hat Q_t, so it was not a function of the outcome-side history, as Appendix E requires. The revised estimator is Eq. (7) with stage-wise self-normalized weights, obtained by shifting each stage's outcome bridge by a constant, and it is doubly robust; the formula, derivation and results are in our response to ojfN (W2).

Q3. Neural P-IPW improving with TT

We would not read it as an improvement. The Monte Carlo SE of the ground truth is 0.018–0.026 (to be added to Table 1) and the P-IPW errors span only 0.022–0.055, so the fair reading is "does not degrade". A likely reason: the RMS normalization (Eq. 14) and the Hájek form (Eq. 17) use only the relative structure of the weights, so P-IPW is insensitive to the scale drift that grows with TT.

Q4. MSMM P-OR at T=20T=20

This was an error in our implementation of the baseline: its future-step enumeration was capped at 12 steps, a limit we kept from the reference code, which affects only the two MSMM cells at T=20T=20. Without the cap, MSMM P-OR is 0.396 and MSMM P-DR is 0.399; results at T≤10T\le10 are unchanged. Table 1 and Section 5.2 will be corrected. Neural-PCI still has the best number in every column.

We hope these corrections address your concerns, and we would be glad to answer any further question.

Thank you for your thorough response. Your clarifications and the revisions to the manuscript have sufficiently addressed my previous concerns. I appreciate the considerable effort you have devoted to addressing these issues. While I recognize and appreciate these improvements, I believe the paper would still benefit from further refinement and polishing. With this in mind, I will respectfully maintain my original score.

Thank you for your reply. We are glad that our clarifications and revisions have addressed your concerns. In the revised version, we will integrate the corrected results into the main text, give complete proofs of the corrected arguments in the appendix, and revise the presentation throughout for clarity. If there are particular parts you would like us to refine, we would be grateful to know, and we will address them. Thank you again.

Reviewer_ojfNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the non-parametric estimation of longitudinal causal effect with unobserved time-varying confounders, in the proximal setting. To address the challenge of error propagations in the nested moment restrictions, it applies the MMR technique, combined with gradient blocking, to solve the moment restrictions simultaneously while preserving the temporal order. Simulated experiments show that the method is effective in long horizons.

Critical strengths

The paper studies an important problem in causal inference and successfully combines existing techniques to address the "curse of horizon" problem common in the setting.

The experiment shows that the Neural-PCI outperforms parametric, latent-variable and sequential-randomization baselines. Further ablation studies on Kernel design strengthen the empirical results.

Critical weaknesses

I have two major concerns:

  1. Remark G.3 is false, so the horizon-separation claim (contribution iii) is unproved.** Remark G.3 states ∥h^t−ht∗∥L2≤τdkt2+dkt−12\|\widehat h_t-h_t^*\|_{L^2}\le\tau\sqrt{d^2_{k_t}+d^2_{k_{t-1}}}.

Corollary G.4 uses it to turn Theorem G.1's moment-risk bound into an average L2L^2 error of order $\tau^2 \sqrt{T/n} $. However, we show that the claim might not be true by a counterexample:

Set T=3T=3 and choose nonzero constant cc. Add cc to each nonterminal bridge, retaining the terminal boundary:

At stage 0, δ1−δ0=0\delta_1-\delta_0=0. At stage 1, δ2−δ1=0\delta_2-\delta_1=0. Therefore both conditional moments vanish, and dk02=dk12=0d_{k_0}^2=d_{k_1}^2=0. At the interior stage t=1t=1, however, ∥δ1∥2=∣c∣\|\delta_1\|_2=|c|. The claimed inequality becomes

which fails for every finite τ\tau.

  1. The targeted P-DR estimator is robust only to a correct outcome bridge, not doubly robust. A counterexample is given below:

Setup (T=1T=1, target a=1a=1). Let A∼Bernoulli(1/2)A\sim\mathrm{Bernoulli}(1/2), U=W=Z=X=0U=W=Z=X=0 (constants), and Y0=Y1=Y=1Y^0=Y^1=Y=1. Then θ1=E[Y1]=1\theta^1=E[Y^1]=1. Nuisances. The treatment bridge is correct, Q0=2Q_0=2. The outcome bridge is misspecified, with fitted class {0}\{0\}, so H^01=0\widehat H_0^1=0. Double robustness requires consistency in this case. Step 1: Eq. (18). With ξ1=0\xi_1=0,

Step 2: Eq. (19).

which is uniquely minimized at ϵ^=−1/2\hat\epsilon=-1/2. Hence H~01=A\widetilde H_0^1=A. However,

Questions for authors

See weaknesses. Can you either show that the counterexamples are wrong or provide ways to fix the proofs?

3 comments

Thank you for two precise counterexamples. Both are correct, and both are fixed: we repair the bound behind contribution (iii) and correct the targeted P-DR estimator. Formal proofs will be in the revised paper.

W1. Remark G.3 and contribution (iii)

  • Your example is correct. Add cc to each non-terminal bridge, retaining the terminal boundary. Both conditional moments vanish at stages 0 and 1, yet at the interior stage t=1t=1 the bridge is off by ∣c∣|c|, so the inequality fails for every finite τ\tau.
  • The revision. Our proof in Appendix G has three links. Theorem G.1 makes the stage-wise moment discrepancies dktd_{k_t} small. Remark G.3 was meant to turn small discrepancies into small bridge error. Corollary G.4 turns the stage-0 bridge error into the ATE bound. Remark G.3, the middle link, fails: it looked only at stages tt and t−1t-1, but a wrong bridge at stage t+1t+1 can be reproduced exactly by a wrong bridge at stage tt, and the stage-tt moment then vanishes.

In the new middle link, each stage-(t+1)(t+1) error is split into the part that stage tt can reproduce exactly and the rest. Stage tt copies the reproducible part without raising dktd_{k_t}, so it is passed back. The rest cannot be copied; it shows up in dktd_{k_t}, so its effect is charged to stage tt, once. This bounds the bridge error at every stage for any fitted bridges. Theorem G.1 is unchanged.

  • Your example under the revised proof. A constant is the same function at every stage, so your shift is a reproducible error, handled in two steps. First, stage 2 sees it: its target is the terminal boundary, so its moment is δ3−δ2=−c\delta_3-\delta_2=-c and dk2d_{k_2} is proportional to ∣c∣|c|. The ill-posedness constant τ2\tau_2 is the largest ratio of a bridge error to the moment it produces, so ∥δ2∥2=∣c∣≤τ2dk2\Vert\delta_2\Vert_2=|c|\le\tau_2 d_{k_2}. Second, stage 1 has zero moment, so δ1\delta_1 is δ2\delta_2 passed back one stage, and a constant passes back unchanged. Hence ∥δ1∥2=∣c∣≤τ2dk2\Vert\delta_1\Vert_2=|c|\le\tau_2 d_{k_2}: the right-hand side is no longer 00, and the repaired bound holds for every cc. Your example also shows why: the error at stage tt is controlled by the discrepancies of stage tt and the later stages, up to the terminal boundary, not by those of stage t−1t-1.
  • Contribution (iii), restated. Sequential estimation fits each stage after the next is frozen, so it passes the whole error of a stage back, enlarged by up to κ\kappa per stage (Theorem F.1). A fit with small joint loss keeps every dktd_{k_t} small, so the part that cannot be reproduced stays small at every stage, and only the reproducible part is passed back. When reproducible errors are not enlarged in total as they are passed back (yours has factor one) and the constants of each stage (the ill-posedness τt\tau_t and how strongly a non-reproducible error shows in dktd_{k_t}) are bounded, the joint bridge error grows at most linearly in TT.

W2. Targeted P-DR

Your counterexample is correct: the targeted estimator was robust only to a correct outcome bridge, not doubly robust, and in your setup it returns 1/2. The perturbation of Eq. (18) contains the observed-treatment indicator and Q^t\hat Q_t, so the targeted bridge is no longer a function of the outcome-side history, and Case (ii) of Appendix E does not apply.

  • Revised estimator. Each fitted outcome bridge is shifted by a scalar so that the weighted stage residuals of Eq. (7) average to zero, with weights wt=1(Aˉt=aˉt)Q^tw_t=\mathbf 1(\bar A_t=\bar a_t)\hat Q_t: \begin{equation} \hat\theta^a_{\rm PDR}=\frac1n\sum_i\hat H^a_{0,i}+\sum_{t=0}^{T-1}\frac{N_t}{\bar w_t}, \end{equation} with wˉt\bar w_t and NtN_t the sample means of wtw_t and wt(H^t+1a−H^ta)w_t(\hat H^a_{t+1}-\hat H^a_t). This is Eq. (7) with stage-wise self-normalized weights.

  • Why it is doubly robust. If the outcome bridge is correct, the population NtN_t is zero. If the treatment bridge is correct, the shifted bridges are still functions of the outcome-side history, so Case (ii) of Appendix E applies, wˉt→1\bar w_t\to1, and the correction removes the bias exactly. The limiting wˉt\bar w_t must be non-zero; we will state this. In your example wˉ0=N0=1\bar w_0=N_0=1 and θ^1=1\hat\theta^1=1.

  • Results (same seeds and hyperparameters). The revised P-DR gives 0.031/0.019/0.022/0.023/0.035 for T=3,5,7,10,20T=3,5,7,10,20 and 0.006 on MIMIC, against 0.017/0.013/0.039/0.057/0.042 and 0.043 before: better at T≥7T\ge7 and on MIMIC, within Monte Carlo uncertainty (≈0.02\approx0.02) at T=3,5T=3,5.

We hope these corrections resolve both soundness concerns, and we would be glad to answer any further question.

The paper has too many theoretical issues (also see reviewer MVg9), which substantially questions the quality of the paper. My suspicion is that these proofs are AI-generated and not examined by human. I maintain my score.

Thank you for reading our response. The theoretical issues raised in the reviews are three: 1) Remark G.3 (your W1 and Reviewer MVg9's W2), 2) the double robustness of our targeted P-DR, which adapted the targeting step of DeepACE to the proximal setting (your W2 and Reviewer MVg9's Q2), and 3) Assumption A.6 (Reviewer MVg9's W1 and Q1). Each was a real error: an established argument was carried one step further without re-checking a condition it needs. The idea behind the paper is simple: sequential plug-in freezes each stage and passes its error back, while joint fitting can still correct it, and Figure 2(a) shows the resulting gap. The three errors lie in the formal arguments, not in this idea or in that gap. We correct each in our responses, and the identification formulas and the Neural P-OR and P-IPW results are unchanged. The corrected bound is also more informative than the original: it shows which part of the error joint fitting removes. We thank you for the counterexamples that led to it, take full responsibility for all content, and would be glad to discuss any step of the corrected arguments that you still find doubtful.

Reviewer_Vz1tOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces a deep learning-based method for estimating the effects of time-varying treatments under unmeasured confounding. The basic idea is to build on proximal causal identification. The main contribution is a joint Maximum Moment Restriction objective that combines moment conditions across all time steps, avoiding stage-wise plug-in estimation. The paper provides theoretical error bounds and evaluates the framework on synthetic trajectories and a semi-synthetic benchmark, reporting improved treatment-effect estimation across different horizons.

Critical strengths

The paper addresses an important problem of longitudinal causal inference: learning flexible bridge functions under an unmeasured confounder. The architecture incorporates the direction of the bridge recursions, and the treatment-kernel design addresses a concrete training difficulty. Experiments compare several relevant method families.

I want to clarify that I am not in this area, so I cannot accurately identify all the strengths.

Critical weaknesses

I want to clarify that I am not in this area, so I cannot identify the weaknesses.

Questions for authors

How to verify the assumptions of Z and W in practice?

General causal inference has some assumptions (consistency, exchangeability, and overlap). How do the defined neural networks satisfy these assumptions?

1 comment

Thank you for the careful summary and for two questions that many readers will share. In short: (Q1) the proxy assumptions cannot be fully verified from data; they are chosen from subject-matter knowledge and then checked for warning signs, much as an instrument is. (Q2) Consistency, exchangeability and overlap are assumptions about the data, not properties of the networks; the networks solve equations that are valid under them.

Q1. How can the assumptions on Z and W be verified in practice?

Assigning covariates to the roles ZZ (treatment-inducing proxy), WW (outcome-inducing proxy) and XX (adjustment covariates) is a design decision, as choosing an instrument is. Every proximal method, ours included, assumes a correct assignment; our contribution begins afterwards, at estimation. You are right that the paper left the practical question to the reader. We will add the following guidance to Section 2 and Appendix L.

  • Designation from subject-matter knowledge. A ZZ may influence treatment but has no direct effect on the outcome (A.4). A WW is related to the outcome but is not affected by ZZ or by the concurrent treatment (A.3). Both must carry information about the unmeasured confounder. Clinically, a vital sign that triggers treatment (e.g., blood pressure) is a natural ZZ, and a laboratory marker of an organ the treatment does not act on (e.g., creatinine) is a natural WW. Our MIMIC-III benchmark follows this logic (Appendix I).
  • Checks that can raise a warning.
  1. Relevance. A near-zero conditional dependence between ZtZ_t and WtW_t given (Aˉt−1,Xˉt)(\bar A_{t-1},\bar X_t) signals weak proxies, as a weak first stage signals a weak instrument. A kernel measure such as conditional HSIC is preferable to a linear one, which can miss nonlinear dependence; the kernel conditional independence (KCI) test (Zhang et al., 2011) gives a formal version of this check.

  2. Agreement of the two routes. The outcome-bridge estimator (P-OR) and the treatment-bridge estimator (P-IPW) reach the same target by different routes (Eqs. 2 and 4). A disagreement beyond sampling variability points to a violation. We already use this agreement for tuning (Section 5.1). Because tuning picks the setting where the two estimates agree best, a gap that remains after tuning is a strong warning, and we will report it as a diagnostic.

  3. Direct effect of ZZ. Some residual association between ZtZ_t and YY is expected, so this is not a test. An association as strong as that of the treatment itself, however, suggests that ZtZ_t acts on the outcome directly.

  • Sensitivity analysis. Move a covariate between WW, ZZ and XX and report how the estimate changes. This shows how much the conclusion depends on the assignment. We will report it for the MIMIC-III designation in Appendix I.

Q2. Consistency, exchangeability and overlap

The networks do not "satisfy" these assumptions, and no estimator could. They are conditions on the data-generating process, stated in Appendix A: consistency is A.1, overlap is positivity (A.2), and the role of exchangeability is taken by proximal latent randomization (A.5). The networks solve the moment equations (1) and (3), and Theorems A.11–A.12 show that, under the assumptions, any solution identifies the target. We will state the three assumptions in Section 2.1, before the bridge equations.

  • Exchangeability is where we differ from sequential-randomization methods such as DeepACE, which require treatment to be independent of the potential outcomes given the observed* history. We require this only given the observed history and the unmeasured* confounder history. This permits confounders that are never measured, and it is why proxies are needed.
  • Overlap matters most for the treatment-bridge route, as it does for inverse-probability weighting. Near-violations appear as large cumulative weights QT−1Q_{T-1}. Our normalizations (Section 3.3, Eq. 17) stabilize the estimate; they do not replace the assumption.
  • Consistency enters through the terminal anchor HT=YH_T=Y of the recursion.

Changes prompted by the other reviews

For transparency: following Reviewer MVg9, Appendix A will state the assumptions in the history form of Ying et al. (2023), which the answers above use; the identification formulas are unchanged. Following Reviewer ojfN, we corrected the doubly robust estimator and the error bound of Section 4; both corrections will be in the final version (details in that thread). The Neural P-OR and P-IPW results are unchanged.

The final version will also add a notation table and a plain-language paragraph on the two recursions for better presentation. We would be glad to clarify any point further.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.