acceptodds
Under review as a conference paper at ICLR 2027

BGM-IV: An AI-Powered Bayesian Generative Modeling Approach for Instrumental Variable

Abstract

Instrumental-variable (IV) regression enables causal estimation under endogeneity, but modern IV problems often involve nonlinear structural effects and high-dimensional covariates. Existing methods typically operate in observed or generic learned feature spaces, and they often yield point estimates without uncertainty quantification. We introduce BGM-IV, a Bayesian generative modeling approach that performs nonlinear IV regression through posterior inference in a causally structured latent space. BGM-IV separates covariate variation by the role in the treatment and outcome mechanism, and accounts for endogeneity through an IV-integrated pseudo-likelihood that averages over instrument-induced treatment variation. The resulting model provides both structural-function estimates and predictive intervals for outcomes under intervention. Across various benchmark datasets, BGM-IV outperforms existing nonlinear IV methods overall, with significant gains in high-dimensional settings, while achieving near-nominal predictive coverage. These results highlight structured latent generative modeling as a flexible approach to uncertainty-aware IV inference with rich covariates. The code of BGM-IV is available at https://github.com/anonymous.

Primary area:
causal reasoning
Keywords:
Causal inference, Instrumental Variables, Bayesian Generative Models, Latent Variable Models, Uncertainty Quantification

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_B67POpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3(+2)
Read review
Paper summary

This paper proposes BGM-IV for nonlinear instrumental-variable estimation with high-dimensional covariates. It partitions the latent representation into four blocks and incorporates an IV-integrated pseudo-likelihood into a tempered quasi-posterior. Training alternates between updating the covariate, treatment, and outcome generators and the observation-specific latent states. Beyond point estimation, the method constructs predictive intervals for outcomes under intervention. Experiments on three versions of a synthetic demand-design benchmark, including vector and image covariates, report improvements in structural MSE and evaluate predictive coverage.

Critical strengths

Predictive intervals extend the method beyond point estimation. Providing model-based uncertainty information for outcomes under intervention is a useful addition to structural-mean estimation in nonlinear IV problems.

Critical weaknesses
  • The causal justification for the tempered quasi-posterior needs further explanation. The citation in Section 3.4 suggests a generalized-Bayes interpretation. While this can justify a loss-based update, it does not by itself establish that the chosen loss targets the intended causal quantity. The sensitivity analysis for the tempering parameter provides empirical evidence, but does not resolve this theoretical question.

  • The IV-integrated density loss need not recover the correct structural mean. Section 3.2 motivates this objective as a way to address endogeneity, but the usual IV conditions alone do not guarantee this property.

    For a simple example without ZZ, let W,U,ηW,U,\eta be independent standard normal variables and define

    X=W+U,Y=−2X+U+η.X=W+U,\qquad Y=-2X+U+\eta.

    The error ϵ=U+η\epsilon=U+\eta satisfies E[ϵ∣W]=0E[\epsilon\mid W]=0, while E[ϵ∣X]=X/2E[\epsilon\mid X]=X/2. The instrument is relevant and excluded from the outcome equation. The moment equation E[Y−aX∣W]=(−2−a)W=0E[Y-aX\mid W]=(-2-a)W=0 uniquely identifies the true slope a=−2a=-2.

    The true conditional distributions are X∣W=w∼N(w,1)X\mid W=w\sim N(w,1) and Y∣W=w∼N(−2w,2)Y\mid W=w\sim N(-2w,2). With the correct first stage and the working outcome model pω(y∣x)=N(ax,s2)p_\omega(y\mid x)=N(ax,s^2), s2>0s^2>0, the paper's IV-integrated pseudo-likelihood in Eq. (3) becomes pIV(y∣w)=∫pω(y∣x)pϕ(x∣w) dx=N(y;aw,a2+s2)p_{\mathrm{IV}}(y\mid w)=\int p_\omega(y\mid x)p_\phi(x\mid w)\,dx=N(y;aw,a^2+s^2). Its population negative log-likelihood, up to a constant, is

    +\frac{2+(a+2)^2}{2(a^2+s^2)}.$$ Holding $s^2$ fixed gives $$\left.\frac{\partial R}{\partial a}\right|_{a=-2} =-\frac{2}{4+s^2}+\frac{4}{(4+s^2)^2}<0.$$ Thus, slightly increasing $a$ above $-2$ lowers the loss, although this working family contains the true intervention distribution $N(-2x,2)$. This establishes a failure of mean recovery within this linear-Gaussian family, not inconsistency of the full neural estimator. Appendix A analyzes the integrated EGM outcome loss and provides conditional structural-mean guarantees for final prediction. These results do not establish that optimizing Eq. (8) recovers the structural mean. My concern is this missing link, rather than an error in the conditional results: the predictive intervals are useful in principle, but the density objective may introduce structural-mean bias.
Questions for authors

Could the authors explain mathematically under what conditions the IV-integrated pseudo-likelihood addresses endogeneity bias and recovers the structural mean, particularly in light of the example above?

8 comments
AuthorsOpenReview ↗

Response to Reviewer B67P (1/2): the objective is revised, and the example in the review is solved

Thank you for the precise analysis. We agree that the original objective need not recover the structural mean under the IV assumptions alone. We have revised outcome estimation within BGM-IV’s generative framework by fitting the structural mean through the IV moment restriction and estimating the noise variance separately from structural residuals. With this change, the example in the review is solved exactly, and the revised procedure recovers both the true structural slope and the noise variance at the population level. Appendix A contains the link between the estimation objectives and their structural targets that the review found missing.

1. The example and the objective of the original submission. We agree. For the objective of the original submission (Eq. 8 there), the IV conditions alone do not guarantee recovery of the structural mean. Maximizing the IV-integrated density is a Kullback-Leibler projection of the law of Y∣W,AY\mid W,A onto the laws of m(X~,aY)+sξm(\tilde X,a_Y)+s\xi, with X~\tilde X drawn from the first stage independently of the Gaussian ξ\xi. For a fixed sufficient representation, a correct first stage, and completeness, correct specification of the integrated reduced-form density at m0m_0 is sufficient for mean recovery. Matching the reduced-form variance at m0m_0 requires

Var(ϵ∣W,A)+2 Cov(m0(X,AY),ϵ∣W,A) ≥ 0.\mathrm{Var}(\epsilon\mid W,A)+2\thinspace\mathrm{Cov}\big(m_0(X,A_Y),\epsilon\mid W,A\big)\ \ge\ 0 .

The IV assumptions do not imply this. In the example the left-hand side is 2+2⋅(−2)=−22+2\cdot(-2)=-2, the constraint s2≥0s^2\ge0 binds, and the population optimum is a∗=1−7≈−1.65a^\ast=1-\sqrt7\approx-1.65 with s2=0s^2=0, in agreement with the sign of the derivative in the review.

2. Revision: separate estimation of the structural mean and the noise scale. We retain the Gaussian structural outcome model and revise how its mean and noise scale are estimated. The revised IV term is a Gaussian pseudo-likelihood centered on the structural mean integrated over the first-stage treatment distribution (Eqs. 3-4):

pIV(y∣w,z)=N(y; μˉω,ϕ(w,z), σIV2),μˉω,ϕ(w,z)=∫μω(x,z0,z1) pϕ(x∣w,z0,z2) dx.p_{\mathrm{IV}}(y\mid w,z)=\mathcal N\big(y;\ \bar\mu_{\omega,\phi}(w,z),\ \sigma_{\mathrm{IV}}^2\big),\qquad \bar\mu_{\omega,\phi}(w,z)=\int\mu_\omega(x,z_0,z_1)\thinspace p_\phi(x\mid w,z_0,z_2)\thinspace dx .

The outcome mean network μω\mu_\omega minimizes the precision-weighted integrated squared loss (y−μˉω,ϕ(w,z))2/2σω2(z0,z1)(y-\bar\mu_{\omega,\phi}(w,z))^2/2\sigma_\omega^2(z_0,z_1) with the noise scale held fixed (Eq. 8a). The noise scale σω2(z0,z1)\sigma_\omega^2(z_0,z_1) is fitted in a separate step to the residual at the observed treatment, with the mean held fixed (Eq. 8b), and it supplies the precision weights of the mean update. A Gaussian pseudo-likelihood parametrized by a conditional mean is consistent for that mean without requiring the distribution to be correctly specified (Gourieroux et al., 1984), which underlies the mean update.

3. The example under the revised objective. For the working mean m(x)=axm(x)=ax, the revised population objective is Rc(a)=2E[c]+(a+2)2E[cW2]R_c(a)=2\mathbb E[c]+(a+2)^2\mathbb E[cW^2]. It is minimized at the structural slope a=−2a=-2 for every weight cc, and the noise-scale step then returns Var(U+η)=2\mathrm{Var}(U+\eta)=2, so the fitted interventional law is N(−2x,2)\mathcal N(-2x,2), which is the true one. Running the complete BGM-IV procedure on this design (n=5,000n=5{,}000, ten repetitions, with one independent standard normal covariate added because the method takes covariates) gives:

slope (truth −2-2)noise s.d. (truth 2≈1.41\sqrt2\approx1.41)
BGM-IV, revised objective−2.02±0.05-2.02\pm0.051.421.42
OLS on the same samples−1.50-1.50-
2SLS on the same samples−2.00-2.00-

4. The missing link (Appendix A).

  • Proposition A.2 (revised). With a correctly specified treatment model, for every bounded positive weight that is a function of (W,V)(W,V), the population minimizers of the weighted integrated squared loss are exactly the solutions of the IV moment equation. Under completeness this is the structural mean. The statement covers the initialization (Eq. 20) and the training update (Eq. 8a).
  • Lemma A.2 (new). The population target of the noise-scale step is the conditional second moment of the residual at the observed treatment. At the structural mean it equals Var(ϵ∣AY)\mathrm{Var}(\epsilon\mid A_Y) and contains no endogeneity covariance.
  • Remark A.3 (new) explains why the outcome mean and noise scales are updated separately: the IV-integrated residual has conditional variance Var(Y∣W,V)\mathrm{Var}(Y\mid W,V), which contains the first-stage spread of the structural mean and the endogeneity covariance, so a noise scale fitted to it would not target the variance of the disturbance.

The initialization, the training update and the selection criterion (Eq. 22) now share one population target, and Corollary A.1 gives the conditional consistency of the selected covariate-only predictor as before. Section 3.3 states this in the main text.

AuthorsOpenReview ↗

Response to Reviewer B67P (2/2): the tempered quasi-posterior (Weakness 1)

5. Weakness 1: what the tempered quasi-posterior is over, and which loss it uses. The quasi-posterior of Section 3.4 (Eq. 5) is a distribution over the unit-level latent state zz. The causal quantity is carried by the mean μω\mu_\omega of the outcome generator, and in the revision both use the same IV-moment residual. The outcome generator minimizes the precision-weighted IV-integrated squared loss, whose population minimizer is the structural mean (Proposition A.2). The latent quasi-posterior is

q(z∣x,y,v,w) ∝ p(z) pθ(v∣z) pϕ(x∣w,z0,z2) exp⁡{−γ (y−μˉω,ϕ(w,z))22σIV2},q(z\mid x,y,v,w)\ \propto\ p(z)\thinspace p_\theta(v\mid z)\thinspace p_\phi(x\mid w,z_0,z_2)\thinspace\exp\Bigl\lbrace -\gamma\thinspace\frac{(y-\bar\mu_{\omega,\phi}(w,z))^2}{2\sigma_{\mathrm{IV}}^2}\Bigr\rbrace,

that is, the posterior kernel of the generative model for (V,X)(V,X) multiplied by a Gibbs factor built from that same IV-moment loss. The loss in the generalized-Bayes update is therefore one whose population target is the structural mean, which is the property required in Weakness 1. With the density objective of the original submission this was not the case, so the two weaknesses had a common cause.

6. The causal content does not rest on the tempering. At γ=0\gamma=0 the latent update does not involve YY at all, and the structural mean is still learned through Eq. 8a. We added γ=0\gamma=0 to the sensitivity study (Table 4, Figure 5): relative to the default γ=0.01\gamma=0.01, its mean structural MSE is 1.031.03, 1.011.01 and 0.960.96 times as large at n=1,000n=1{,}000, 5,0005{,}000 and 10,00010{,}000, and the four values γ∈0,0.01,0.05,0.1\gamma\in0,0.01,0.05,0.1 differ by less than the repetition-to-repetition variability in every cell. Predictions and model selection use latent states re-inferred from the covariates alone (Lemma A.1), never the outcome-informed training states.

7. Effect on the benchmarks. All experiments were rerun with the revised objective. Relative to the original submission, the structural MSE of BGM-IV is lower on the demand design (geometric mean of the paired ratio revised/original: 0.700.70, 0.830.83 and 0.860.86 at n=1,000n=1{,}000, 5,0005{,}000 and 10,00010{,}000) and on the vector-proxy benchmark (0.830.83-0.910.91), and essentially unchanged on the image benchmark (0.970.97-1.001.00). This is the direction the analysis in the review predicts, and the comparison with the baselines is unchanged.

Since this concern was central to the assessment, we would be grateful if you would reconsider the rating in light of the revision. We are glad to clarify any remaining point.

Reference. Gourieroux, C., Monfort, A., & Trognon, A. (1984). Pseudo Maximum Likelihood Methods: Theory. Econometrica, 52(3), 681–700. https://doi.org/10.2307/1913471

We hope that our additional experiments and clarifications have addressed your major concerns. I would greatly appreciate your consideration of updating the assessment and score accordingly. We would also be happy to discuss and address any further questions that you may have. Thank you!

Reviewer_3VJ7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BGM-IV, an extension of the CausalBGM framework (Liu & Wong, JASA 2026) to nonlinear instrumental-variable (IV) regression with high-dimensional covariates. Covariates VV are generated from a low-dimensional latent Z=(Z0,Z1,Z2,Z3)Z=(Z_0,Z_1,Z_2,Z_3) that is partitioned by causal role; the treatment model is pϕ(x∣w,z0,z2)p_\phi(x\mid w,z_0,z_2) and the outcome model is pω(y∣x,z0,z1)p_\omega(y\mid x,z_0,z_1). Endogeneity is handled through an "IV-integrated pseudo-likelihood" pIV(y∣w,z)=∫pω(y∣x,z0,z1) pϕ(x∣w,z0,z2) dxp_{IV}(y\mid w,z)=\int p_\omega(y\mid x,z_0,z_1)\,p_\phi(x\mid w,z_0,z_2)\,dx, evaluated by Monte Carlo, which is a likelihood analogue of the DeepIV integrated loss. Training consists of an IV-aware EGM initialization (encoder, adversarial prior matching and an integrated squared loss computed by Gauss–Hermite quadrature), followed by alternating SGD on the three generators and per-unit gradient ascent on the latent states under a tempered quasi-posterior (γ=0.01\gamma=0.01). K=10K=10 restarts are selected by the training-sample IV-moment residual. At prediction time the latents are re-inferred from VV alone (MAP), g^(x,v)=μω(x,z^0(v),z^1(v))\hat g(x,v)=\mu_\omega(x,\hat z_0(v),\hat z_1(v)), and predictive intervals come from HMC draws of p(z∣v)p(z\mid v) mixed with N(μω,σω2)\mathcal N(\mu_\omega,\sigma^2_\omega).

Theory: exogeneity of covariate-only readouts; a latent NPIV moment and identification under completeness; the population target of the integrated squared loss; an oracle inequality for the multi-start selection; and L2L^2 consistency of the selected predictor under a candidate-approximation assumption and an ill-posedness modulus.

Validation: three variants of the airline demand design (low-dimensional V=(T,S)V=(T,S); 784-dimensional noisy proxies reduced to kk principal components; an MNIST image covariate), with 20 paired repetitions, against KIV, DeepIV, DFIV, DML-IV and FIL. BGM-IV has the lowest mean structural MSE in almost all configurations (e.g., 2.6–3x lower than DFIV on MNIST for n≥5000n\ge 5000), supported by Holm-adjusted paired tests, ablations (EGM initialization, IV pseudo-likelihood, latent partition), a γ\gamma sensitivity study, a weak-IV study and an empirical latent-exogeneity check. The 90/95/99% predictive intervals have near-nominal marginal coverage of new outcomes.

Critical strengths
  1. Careful, well-controlled empirical evidence in a setting that matters. High-dimensional covariates, as opposed to high-dimensional treatments or instruments, get little attention in nonparametric IV, and the evaluation is unusually disciplined. Every method sees identical training arrays. Every stochastic baseline gets the same 10-restart budget with its own training-side selection criterion. There are 20 paired repetitions, and the comparisons use Holm-adjusted Wilcoxon tests. The gains on the image benchmark are large and uniform: BGM-IV has lower error than DFIV in every repetition for n≥5000n\ge 5000. The vector-proxy sweep over kk shows that BGM-IV degrades gracefully as nuisance dimensions are added, while FIL and DFIV deteriorate sharply. This is the main reason I take the method seriously.

  2. A direct treatment of the "outcome-informed latent" risk. Using YY to update per-unit latent states is the most obvious way a model like this could re-introduce endogeneity. The paper keeps the training latents separate from the covariate-only readout that is used for prediction and selection. It also explains formally why the readout preserves E[ϵ∣W,A]=0E[\epsilon\mid W,A]=0 and checks this empirically with negative controls, positive controls and power calibration. The ablations are also well designed. The advantage of the IV-integrated term grows with ρ\rho, which is what one would expect if it is removing confounding bias.

Critical weaknesses

The uncertainty-quantification claim, one of the three stated contributions, is not supported and is arguably not well defined. (a) The predictive intervals omit the uncertainty in the estimated structural function. The paper gives only a point estimate of the structural function, g^(x,v)=μω^(x,z^0(v),z^1(v))\hat g(x,v)=\mu_{\hat\omega}(x,\hat z_0(v),\hat z_1(v)) (Eq. 12). The intervals of Sec. 3.6 are predictive intervals for a new outcome under intervention, Y(x)∣V=v=g0(x,v)+ϵY(x)\mid V=v=g_0(x,v)+\epsilon. A valid predictive interval for this quantity must account for two things: the disturbance ϵ\epsilon, and the error in the estimate of g0g_0. In Bayesian form, p(y∣x,v,D)=∫N{y; g(x,v), σ2}  π(g∣D) dg.p(y\mid x,v,\mathcal D)=\int \mathcal N\{y;\,g(x,v),\,\sigma^2\}\;\pi(g\mid\mathcal D)\,dg . Eq. (13) contains only the first. It is evaluated at a single fitted ω^\hat\omega, since the generator parameters are point estimates. Its two sources of spread are:

  • σω^2\sigma^2_{\hat\omega}, the outcome model's term for the disturbance ϵ\epsilon around the structural mean (whether it estimates Var(ϵ)\mathrm{Var}(\epsilon) correctly is a separate issue; see (b));
  • p(z∣v)∝p(z) pθ^(v∣z)p(z\mid v)\propto p(z)\,p_{\hat\theta}(v\mid z), which reflects how precisely the covariates determine the latent state under the fitted covariate model. It involves neither WW, XX nor YY. Neither term represents the estimation error μω^−g0\mu_{\hat\omega}-g_0. σω\sigma_\omega describes how YY scatters around the fitted mean, not how far the fitted mean may be from g0g_0. The intervals are therefore plug-in intervals that treat g^\hat g as if it were g0g_0. In NPIV this omission is first order, not a finite-sample refinement. g0g_0 is identified only through the integral equation E[Y∣W,V]=∫g0(x,V) dP(x∣W,V)E[Y\mid W,V]=\int g_0(x,V)\,dP(x\mid W,V), so its estimation error is governed by the ill-posedness of the conditional-expectation operator. This error grows as the instrument weakens, and g0g_0 is not identified when the instrument is irrelevant. A more accurate model of ϵ\epsilon neither reduces this error nor reflects it, and no term of Eq. (13) responds to the strength of identification. The Bayesian component of the method, posterior inference over zz, thus does not address the source of uncertainty that is specific to IV. CausalBGM, by contrast, places variational posteriors on the generator parameters and evaluates coverage of the causal estimand itself (Liu & Wong 2026b, Sec. 2.3 and 3.6). I would ask the authors to do one of two things. Either propagate parameter uncertainty (e.g., a posterior or ensemble over ω\omega, or a bootstrap over training samples) and report intervals for g0(x,v)g_0(x,v) itself, or describe the current intervals explicitly as plug-in predictive intervals conditional on g^\hat g.

(b) σω2\sigma^2_\omega is not an estimate of the interventional noise. Under additive separability the target is Y(x)∣V=v = g0(x,v)+ϵ∣V=vY(x)\mid V=v \,=\, g_0(x,v)+\epsilon \mid V=v. The pseudo-likelihood, however, models Y∣W,VY\mid W,V as μω(X~)+e\mu_\omega(\tilde X)+e with X~∼pϕ(⋅∣w,z)\tilde X\sim p_\phi(\cdot\mid w,z) independent of ee. The variance of that model is E[σω2]+Var(μω(X~))E[\sigma^2_\omega]+\mathrm{Var}(\mu_\omega(\tilde X)), while the true variance is Var(g0(X)∣W,V)+Var(ϵ∣W,V)+2 Cov(g0(X),ϵ∣W,V)\mathrm{Var}(g_0(X)\mid W,V)+\mathrm{Var}(\epsilon\mid W,V)+2\,\mathrm{Cov}(g_0(X),\epsilon\mid W,V). The endogeneity covariance term therefore either goes into σω2\sigma^2_\omega, biasing the interval widths in a direction set by the sign of the confounding, or distorts μω\mu_\omega.

(c) The evaluation cannot tell a calibrated interval from an inflated one. In all benchmarks Var(ϵ)=ρ2+(1−ρ2)=1\mathrm{Var}(\epsilon)=\rho^2+(1-\rho^2)=1 on the original outcome scale, so an oracle 95% interval for a new outcome has width 2×1.96≈3.92\times 1.96\approx 3.9. The reported widths are about 72 outcome units (Fig. 3c) and 99.5 outcome units (Appendix H.3). That is roughly 20–25 times the oracle width and on the same scale as the point-estimate RMSE (about 18 for the vector proxy at n=5000n=5000, k=8k=8; about 41 for the image benchmark at n=10,000n=10,000). Near-nominal marginal coverage therefore mostly shows that the bands are wide enough to absorb the error in the structural estimate. It is not evidence of a calibrated predictive distribution, and averaging over the grid can hide strong conditional mis-coverage (CausalBGM's own Fig. 3A shows calibration that depends on xx). The paper reports no band for g0(x,v)g_0(x,v), no proper scoring rule (interval score or CRPS) and no UQ baseline (for example deep ensembles or bootstrap versions of DFIV/DeepIV, quasi-Bayesian dual IV (Wang et al., NeurIPS 2021), or conformal approaches). The statement that BGM-IV "uniquely quantifies the uncertainty of its predictions" should either be substantiated along these lines or be substantially toned down.

Much of the method is also inherited from CausalBGM: the four-block partition, the G/H/F networks, the EGM initialization and the iterative latent updates. The new ingredients are essentially the instrument in the treatment model, a likelihood version of the DeepIV integration, the covariate-only readout and the multi-start selection, so the methodological novelty is narrower than presented.

Questions for authors

Q1. For a few (x,v)(x,v), please split the predictive variance in Eq. (13) into the latent-posterior part Varj[μω(x,z(j))]\mathrm{Var}_j[\mu_\omega(x,z^{(j)})] and Ej[σω2]E_j[\sigma^2_\omega], and report the learned σω\sigma_\omega against the true sd(ϵ)=1\mathrm{sd}(\epsilon)=1 as a function of ρ\rho. If σω\sigma_\omega tracks ρ\rho, that would confirm W1(b).

Q2. Can you report credible or confidence bands for g0(x,v)g_0(x,v) itself and their coverage? These could come from a parameter posterior, a Laplace or SWAG approximation, ensembles, or a bootstrap over training samples. Please also report conditional coverage by price region and customer group, interval scores, and simple UQ baselines, such as DFIV/DeepIV deep ensembles built from the 10 restarts you already train.

Q3. With an irrelevant instrument (λ=0\lambda=0 in the design of Table 10), g0g_0 is not identified. Do BGM-IV's point estimates and intervals reflect this, for example through wider bands, or does it still return confident estimates? This directly tests whether the UQ is IV-aware.

Q4. Under what conditions does the population maximizer of the IV-integrated log pseudo-likelihood satisfy the IV moment restriction? As an empirical complement, what happens if the stochastic-optimization stage keeps the integrated squared loss of Eq. (21) instead of Eq. (8)?

Q5. Please add γ=0\gamma=0 to Table 4 and Figure 5.

Q6. The EGM initialization alone accounts for a 17.7–18.9x reduction in MSE, which suggests the learned representation of VV does much of the work. Could you give DFIV and DeepIV the same covariate-only representation z^(v)\hat z(v) (or the EGM encoder output) as covariates? Relatedly, on the low-dimensional demand design with V=(T,S)V=(T,S), why is BGM-IV about 2.5x better than DFIV at n=1000n=1000, when no representation learning should be needed there?

Q7. Appendix F says each baseline uses "the configuration its implementation provides for the demand design". On the MNIST benchmark, do DeepIV, DFIV and DML-IV use convolutional feature extractors, as in their own image experiments, or MLPs on raw pixels? BGM-IV uses a tuned CNN and benchmark-specific settings, so please state the tuning budget for each method.

Q8. The block sizes ([3,2,1,2], [2,1,1,2], [1,1,0,5]) appear to have been chosen with knowledge of the DGP. How sensitive are the results to wrong block sizes, for example a Z2Z_2 or Z0Z_0 that is too small and would violate the first-stage sufficiency? Is there a data-driven rule? In CausalBGM, they used an SIR-based rule as heuristic.

Suggestions: Add a real-data or semi-synthetic application with genuinely high-dimensional covariates, for example Mendelian randomization with genomic covariates, or returns to schooling with rich controls. All three benchmarks are variants of a single DGP.

To delimit the scope, include DeepGMM/AGMM (cited but not compared) and a benchmark with a high-dimensional treatment (e.g., dSprites as in DFIV). Also add experiments for the binary-treatment variant described in Sec. 3.1. Appendix B tests only linear correlation. A nonlinear dependence test (e.g., HSIC or distance correlation) between the latent states and ξ\xi/UU would be more convincing. State clearly in Sec. 3 which components are inherited from CausalBGM/EGM and which are new.

8 comments
AuthorsOpenReview ↗

Response to Reviewer 3VJ7 (1/4): intervals for the structural function (W1(a), W1(c), Q2)

Thank you for the very detailed and constructive review! W1-W2 and Q1-Q8 are answered below in the numbering of the review.

W1(a), Q2. Intervals for g0(x,v)g_0(x,v) itself. We agree with this reading of the original submission, and we took the first of the two options. Section 3.6 now reports credible intervals for g0(x,v)g_0(x,v) (Eq. 13):

g^(x,v) ± z1−α/2slat2(x,v)+scand2(x,v).\hat g(x,v)\ \pm\ z_{1-\alpha/2}\sqrt{s_{\mathrm{lat}}^2(x,v)+s_{\mathrm{cand}}^2(x,v)} .

slat2s_{\mathrm{lat}}^2 is the variance of the structural value over HMC draws from p(z∣v)p(z\mid v). scand2s_{\mathrm{cand}}^2 is the variance across the K=10K=10 independently initialized candidates that the multi-start protocol already trains, which is the ensemble route suggested in the review. The two are combined through a normal approximation, and the interval is evaluated against the known g0g_0.

Coverage of the true structural function, averaged over the grid and 20 repetitions:

benchmark90%95%99%mean width (95%)
vector proxy, k=8k=8, n=5,000n=5{,}00090.995.398.956.9
image, n=50,000n=50{,}000, ρ=0.5\rho=0.593.796.398.7118.8

W1(c), Q2. UQ baselines, scoring rule, conditional coverage.

  • UQ baselines. We built deep ensembles from the ten runs that DFIV, DeepIV and DML-IV already train (ensemble mean ±\pm normal quantile ×\times standard deviation across runs). At n=5,000n=5{,}000 their 95% intervals cover g0g_0 at 62-80%, 47-63% and 41-55% of the grid points, against 92.7-99.9% for BGM-IV (Figure 3b, Table 17).
  • Interval score. Table 17 reports the 95% interval score (Gneiting and Raftery, 2007). BGM-IV has the lowest score for seven of the nine values of kk.
  • Conditional coverage. Table 18 reports coverage within four price ranges, four time ranges and the seven customer groups. It stays within 93.2-96.4% on the vector-proxy benchmark and 93.3-97.7% on the image benchmark, and the width follows the error of the subgroup (on the vector-proxy benchmark it grows from 28.7 for customer group 1 to 83.5 for group 7, in line with an RMSE of 8.4 and 21.8).
  • Wording. "Uniquely quantifies" was removed, and the abstract, introduction and conclusion now refer to coverage of the structural function.

On the width. The oracle width of 3.9 computed in the review applies to a predictive interval for a new outcome when g0g_0 is known. The intervals we now report are for g0g_0, so their width reflects its estimation error: at k=8k=8 the width is 56.9 for an RMSE of 15.4, close to the 2×1.96×15.4≈602\times1.96\times15.4\approx60 of a calibrated normal interval.

AuthorsOpenReview ↗

Response to Reviewer 3VJ7 (2/4): W1(b), Q1, Q3

W1(b). σω2\sigma_\omega^2 and the endogeneity covariance.
We agree for the objective of the original submission: a density pseudo-likelihood has to match the reduced-form variance of Y∣W,VY\mid W,V, so the endogeneity covariance goes either into σω2\sigma_\omega^2 or into μω\mu_\omega (Reviewer B67P's example is the second case). The revised method estimates the mean and the noise scale separately, which removes both channels. The outcome mean solves a precision-weighted IV moment equation that contains no variance parameter of its own, so its population solution is the structural mean for every positive weight (Proposition A.2). The noise scale is fitted in a separate step to the residual using the observed treatment, Y−μω(X,z0,z1)Y-\mu_\omega(X,z_0,z_1). Under additive separability ϵ=Y−g0(X,V)\epsilon=Y-g_0(X,V), so at the structural mean the population target of this step is Var(ϵ∣AY)\mathrm{Var}(\epsilon\mid A_Y), with no covariance term (Lemma A.2). On Reviewer B67P's design the fitted noise s.d. is 1.42 against the true 1.41.

Q1. Variance decomposition, and the learned noise scale against ρ\rho. Decomposition. In the revised interval the two parts are the latent-posterior variance and the between-candidate variance. Share of the latent part at nine grid points (prices 11.6 / 17.9 / 23.4, customer groups 1, 4 and 7, time fixed at 5.26), mean over 20 repetitions:

group 1group 4group 7whole grid
vector proxy (k=8k=8, n=5,000n=5{,}000)0.16 / 0.22 / 0.370.42 / 0.41 / 0.480.45 / 0.73 / 0.660.37
image (n=50,000n=50{,}000)0.12 / 0.18 / 0.700.16 / 0.36 / 0.640.18 / 0.37 / 0.560.26

The remainder is the between-candidate variance, and Table 16 gives the share for every kk.

Learned noise scale against ρ\rho. The fitted scale varies little with ρ\rho. On the demand design at n=10,000n=10{,}000 the fitted noise s.d. is 6.4, 6.3, 6.4, 6.4 and 6.3 at ρ=0.1\rho=0.1, 0.250.25, 0.50.5, 0.750.75 and 0.90.9, and s.d. decreases as nn grows. Its level is above sd(ϵ)=1\mathrm{sd}(\epsilon)=1 because, in finite samples, the residual at the observed treatment also contains the estimation error of the structural function. In the method σω\sigma_\omega serves as the precision weight of Eq. 8a, and it does not enter the intervals of Eq. 13.

Q3. Irrelevant instrument (λ=0\lambda=0). We ran this on the vector-proxy benchmark (k=8k=8, n=5,000n=5{,}000, 20 repetitions), the setting in which the intervals are evaluated in Figure 3, by setting λ=0\lambda=0 in the price equation of Table 10. The results are:

λ=1\lambda=1λ=0\lambda=0
structural MSE (divided by 10210^2)2.40288.9
mean width of the 95% interval56.9175.4

The point estimate degrades, as it must without identification, and the interval widens by a factor of 3.1: the independently initialized candidates disagree far more when the instrument carries no information.

AuthorsOpenReview ↗

Response to Reviewer 3VJ7 (3/4): Q4-Q7

Q4. When does the density pseudo-likelihood satisfy the IV moment restriction, and what if training keeps the integrated squared loss? Conditions. For a fixed sufficient representation and a correct first stage, correct specification of the integrated reduced-form density is one sufficient condition. Its population optimum then satisfies the IV mean restriction, and completeness identifies m0m_0. Matching the reduced-form variance at m0m_0 requires Var(ϵ∣W,A)+2Cov(m0(X),ϵ∣W,A)≥0\mathrm{Var}(\epsilon\mid W,A)+2\mathrm{Cov}(m_0(X),\epsilon\mid W,A)\ge0. The IV assumptions alone do not ensure this distributional compatibility. The revised objective instead directly targets the IV mean restriction through the integrated squared loss (Proposition A.2).

Empirical complement. We ran this variant. Keeping the integrated squared loss of the initialization (Eq. 21 of the original submission) in the stochastic-optimization stage, with everything else unchanged, gives a paired geometric-mean ratio to the original submission of 0.77, 0.90 and 0.90 on the demand design at n=1,000n=1{,}000, 5,0005{,}000 and 10,00010{,}000. The revised method retains the Gaussian structural outcome model and additionally estimates the noise scale in a separate step, which supplies the precision weight of the mean update (Eqs. 8a-8b). It gives 0.70, 0.83 and 0.86 on the demand design, 0.83-0.91 on the vector-proxy benchmark and 0.97-1.00 on the image benchmark. All tables and figures now report the revised method.

Q5. γ=0\gamma=0. We added γ=0\gamma=0 to Table 4 and Figure 5. Relative to γ=0.01\gamma=0.01 its mean MSE is 1.03, 1.01 and 0.96 times as large at n=1,000n=1{,}000, 5,0005{,}000 and 10,00010{,}000. With γ=0\gamma=0, BGM-IV still has a lower mean MSE than every baseline in all 15 configurations of the demand design.

Q6. Baselines given the BGM-IV representation, and DFIV with twice the sample. Shared representation (Appendix N, Table 20). We gave DFIV and DeepIV the covariate-only states z^(v)\hat z(v) of the selected BGM-IV model in place of their covariates, on the vector-proxy benchmark (k=8k=8, n=5,000n=5{,}000). Structural MSE (divided by 10210^2):

original covariatesBGM-IV representation
DeepIV12.895.00
DFIV7.167.33
BGM-IV2.40-

Even with the BGM-IV representation, neither baseline reaches BGM-IV: the error of DeepIV is still 2.1 times that of BGM-IV (5.00 against 2.40), and DFIV does not improve (7.33 against 7.16). BGM-IV has a lower error than both hybrids in all 20 repetitions. The representation alone therefore does not explain the full advantage of BGM-IV, and the remaining difference is due to the estimator.

Low-dimensional design (Table 21). One factor is the use of the sample: DFIV splits the training sample between its two stages, while BGM-IV uses every observation in both. Giving DFIV twice the sample reduces its error, averaged over ρ\rho, from 10.08 at n=1,000n=1{,}000 to 3.98 at n=2,000n=2{,}000, against 2.80 for BGM-IV at n=1,000n=1{,}000. BGM-IV at n=1,000n=1{,}000 still has the lower error in 86 of 100 paired repetitions.

Q7. Baseline architectures and configurations on the image benchmark. All baselines use convolutional networks, and none is an MLP on raw pixels (now stated in Appendix F). DeepIV and DFIV use the image feature extractor of the DFIV code base: two 3×33\times3 convolution layers with 64 filters, max pooling, and dense layers of widths 128 and 64. The image encoder of BGM-IV uses the same convolutional feature extractor. FIL uses the convolutional network described in its paper, and DML-IV the convolutional network of its released code.

On configurations: every method uses one fixed configuration per benchmark, shared across all sample sizes and confounding levels. For the baselines it is the configuration of the authors' implementation. For BGM-IV it is the one in Appendix E, and Appendices C and M report its sensitivity to γ\gamma and to the latent partition. Every stochastic method receives the same budget of ten runs with its own training-side selection.

Reviewer_kp5zOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a Bayesian generative modeling framework for nonlinear instrumental variable regression with high-dimensional covariates. The proposed method learns a structured latent representation that partitions covariates into four parts. The framework alternates between optimizing the generative models and inferring latent variables, using posterior inference to produce estimates and prediction intervals. The authors establish identification and conditional consistency results under assumptions of representation sufficiency, completeness, and related conditions.

Critical strengths
  • I think the paper proposed an interesting framework for dealing with high-dimensional covariates, which is quite useful in practice.
  • The experiments demonstrate good performance of the proposed framework in various settings.
Critical weaknesses
  • Although the paper provides a thorough overview of the existing works on IV, I think the paper should also be positioned with respect to some existing works with related methodologies. For example, the NCM (Xia et al., 2021) and DCM (Rahman et al., 2024) model the exogenous distribution by learning the mapping from a prior distribution and fitting the observational data. Some recent works that model the interventional distribution, such as DIV (Holovchak et al., 2025), should also be discussed. More importantly, the paper has limited discussion of prior work on structural latent space modeling (Liu et al., 2024). This is important to address to clearly understand the paper's main contribution.

  • The identification assumption is not sufficiently clear in the main paper. The IV assumption in Section 2 is formulated conditional on the observed covariates V, whereas the generative model in Equations (1) and (2) operates on the partitioned latent representation ZZ. The distinction here needs some clarification. Although Lemma A.1 establishes that the covariate representations preserve conditional mean exogeneity, it is not clear that the learned representation is sufficient for recovering the structural outcome function.

Ref:

Xia, Kevin, et al. "The causal-neural connection: Expressiveness, learnability, and inference." Advances in Neural Information Processing Systems 34 (2021): 10823-10836.

Rahman, Md Musfiqur, and Murat Kocaoglu. "Modular learning of deep causal generative models for high-dimensional causal inference." arXiv preprint arXiv:2401.01426 (2024).

Holovchak, Anastasiia, et al. "Distributional instrumental variable method." arXiv preprint arXiv:2502.07641 (2025).

Questions for authors
  • Could the authors please provide some intuition for why integrating the outcome density first and then using it for quasi-posterior inference would be better than the objective in DeepIV? It would be interesting to understand the key difference between the proposed approach and existing work. Also, Equation (3) describes the outcome distribution conditioned on the latent variables; would it be possible to have an ablation study on the latent variable modeling as well as the proposed objective for quasi-posterior inference?
5 comments
AuthorsOpenReview ↗

Response to Reviewer kp5z (1/2): positioning and identification

Thank you for the constructive review. Below is our point-by-point response.

1. Positioning with respect to NCM, DCM, DIV and structured latent-space modeling. The revised introduction places BGM-IV within the broader literature on generative causal inference, and Section 3 explicitly distinguishes the inherited architecture from our methodological contributions.

  • NCM (Xia et al., 2021) and modular DCM (Rahman and Kocaoglu, 2024). These are general frameworks for causal inference with neural structural causal models, rather than methods developed for IV regression. They do not address IV regression with a continuous treatment and high-dimensional covariates.
  • DIV (Holovchak et al., 2025). DIV focuses on distributional IV inference and is not specifically designed for structured representation learning with high-dimensional covariates. BGM-IV explicitly learns a causally partitioned latent representation of the observed covariates and addresses endogeneity through IV conditional moment restrictions. Its contribution is to integrate this representation learning with structural-mean estimation, covariate-only prediction, and uncertainty quantification for the structural function.
  • CausalEGM (Liu et al., 2024) and CausalBGM. These methods provide structured latent representations for causal inference under unconfoundedness. BGM-IV adapts this architecture to settings where treatment remains endogenous after covariate adjustment. The central change is an IV-integrated outcome objective that learns the structural mean from instrument-induced treatment variation, with the noise scale estimated separately. This extension is supported by identification and conditional-consistency results under the stated assumptions. We further develop uncertainty quantification of BGM-IV by combining latent-posterior uncertainty with an ensemble approximation to model-estimation uncertainty, yielding credible intervals for the structural function.

2. IV assumptions on VV versus a model on ZZ. You are right that Lemma A.1 gives exogeneity only. The revised Section 3.3 makes the logic explicit:

  1. Identification is a statement about observables. Under the conditions of Section 2, the distribution of (X,Y,V,W)(X,Y,V,W) identifies g0(x,v)g_0(x,v), and no latent variable is involved.
  2. The latent representation is part of the estimator. It is a learned function of the covariates, A=φ(V)A=\varphi(V), and BGM-IV searches over structural functions of the form μ(x,aY(v))\mu(x,a_Y(v)) and treatment models of the form k(x∣w,aX(v))k(x\mid w,a_X(v)).
  3. Exogeneity is automatic. Since AA is a function of VV, iterated expectations give E[ϵ∣W,A]=0\mathbb E[\epsilon\mid W,A]=0 from E[ϵ∣W,V]=0\mathbb E[\epsilon\mid W,V]=0 (proved in Lemma A.1).
  4. Sufficiency of the representation is an assumption. That g0(x,V)=m0(x,AY)g_0(x,V)=m_0(x,A_Y), with a treatment model that depends on VV through AXA_X, is Assumption A.1. It is of the same kind as the realizability of learned features in DFIV.
  5. Approximation error and conditional consistency. Corollary A.1 accounts for representation error through the final-candidate approximation condition. Consistency requires the stated approximation, first-stage, generalization, quadrature, and inverse-stability conditions. The selection criterion compares candidates through IV-moment risk (Eq. 15), translating this risk into structural-function error additionally requires inverse stability.

Empirical evidence. On the demand design, the outcome-relevant blocks (z^0,z^1)(\hat z_0,\hat z_1) of the covariate-only readout, computed on fresh units, recover the time covariate with R2=0.9998R^2=0.9998 and the customer group with accuracy 0.99940.9994 (20 models at ρ=0.9\rho=0.9, n=5,000n=5{,}000). The partition study (Appendix M) points the same way: removing the shared block Z0Z_0 increases the error by a factor of 6.0-10.9.

AuthorsOpenReview ↗

Response to Reviewer kp5z (2/2): relation to DeepIV, and ablations

3. Why integrate the outcome density, and how does the objective differ from DeepIV? We have revised the outcome objective of BGM-IV. BGM-IV retains its Gaussian structural outcome model, with the mean and noise scale now estimated separately (Eqs. 8a-8b). The revised IV term is a Gaussian pseudo-likelihood centered on the first-stage-integrated structural mean (Eqs. 3-4). Under the stated assumptions, its precision-weighted mean objective shares the population target of DeepIV’s integrated squared loss (Proposition A.2).

The key difference is how IV estimation is integrated with representation learning and uncertainty quantification. DeepIV estimates the treatment distribution and then fits a response network. BGM-IV jointly learns a generative model of covariates, treatment and outcome in a causally partitioned latent space. The structural mean is learned through the IV moment restriction, while the noise variance is estimated separately from residuals at the observed treatment and supplies precision weights for mean estimation (Eqs. 8a-8b). The IV pseudo-likelihood also guides latent-state inference during training. By combining latent-posterior and ensemble uncertainty, BGM-IV provides structural-function intervals with empirical coverage, rather than just a point estimate. Empirical experiments further supported that BGM-IV's coverage is substantially closer to the nominal level than that of the DeepIV ensembles.

The representation-transfer experiment shows why this distinction matters. On the vector-proxy benchmark (k=8k=8, n=5,000n=5{,}000), giving DeepIV the BGM-IV representation reduces its structural MSE by about 61%, demonstrating the value of the learned representation by BGM-IV. BGM-IV still achieves a 52% lower MSE than DeepIV using that representation and performs better in all 20 repetitions (Appendix N). Thus, supplying the learned representation as fixed inputs improves DeepIV substantially but does not reproduce the performance of the complete BGM-IV estimator. The image benchmark provides complementary evidence of the complete method’s effectiveness, with 70-72% lower MSE than DeepIV at n=5,000n=5{,}000 (Table 1).

4. Ablations of the latent-variable modeling and of the objective. All ablations were rerun with the revised objective (Table 2, Appendices C, K and M).

Latent-variable modeling.

  • Structured versus unstructured latent space. With five nuisance covariates, replacing the partitioned latent space by an unstructured one of the same total dimension increases the structural MSE by a factor of 1.44-1.64. The nuisance information is captured by the nuisance block (R2=0.96R^2=0.96) and not by the coordinates read by the treatment and outcome generators (R2≈0R^2\approx0). This supports the role of the partition in separating nuisance variation from the information used for estimation.
  • Contribution of individual blocks and sensitivity to their sizes. Block sizes (Appendix M, Table 19): removing the shared block Z0Z_0 increases the error by a factor of 6.0-10.9, and removing the outcome-specific block Z1Z_1 by a factor of 1.39-1.41. Enlarging the blocks does not hurt (0.76-0.84). The default partition does not match these true dimensions. These results suggest that BGM-IV can achieve strong performance without prior knowledge of the true latent dimensions.

Objective.

  • IV integration in structural-mean estimation. Replacing the IV-integrated mean by the outcome mean using the observed treatment, under amplified confounding, increases the structural MSE by a factor of 1.02, 1.48 and 2.93 at ρ=0.1\rho=0.1, 0.50.5 and 0.90.9. The benefit of IV integration therefore grows as confounding increases.
  • Separate estimation of the mean and the noise scale. The revised objective lowers the structural MSE relative to the original submission on the demand design (ratio 0.70, 0.83 and 0.86 at n=1,000n=1{,}000, 5,0005{,}000 and 10,00010{,}000) and on the vector-proxy benchmark (0.83-0.91).
  • Outcome information in latent inference. Tempering exponent of the quasi-posterior: γ=0\gamma=0 is now included, and γ∈0,0.01,0.05,0.1\gamma\in0,0.01,0.05,0.1 perform similarly. Relative to the default γ=0.01\gamma=0.01, setting γ=0\gamma=0 gives mean structural MSE ratios of 1.03, 1.01 and 0.96 at n=1000,5000,10000. It shows performance is insensitive to removing the outcome contribution from latent inference in these settings (Table 4, Figure 5).
  • Precision weighting: with heteroscedastic outcome noise, removing them increases the structural MSE by a factor of 1.16, 1.25 and 1.37 at ρ=0.1\rho=0.1, 0.50.5 and 0.90.9 (Table 2). In this setting BGM-IV also has the lowest error among the methods compared: 4.92, 4.19 and 4.23, against 6.42, 6.11 and 5.15 for DFIV and 10.74, 9.20 and 8.65 for FIL (Table 15).

We hope this clarifies the contribution and the identification argument. If it addresses your concerns, we would be grateful if you would reconsider the rating, and we are glad to answer any further question.

We hope that our additional experiments and clarifications have addressed your major concerns. I would greatly appreciate your consideration of updating the assessment and score accordingly. We would also be happy to discuss and address any further questions that you may have. Thank you!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision (revised manuscript and code uploaded)

We thank the three reviewers for their careful and constructive feedback. In response, we have modified the outcome objective, rerun all experiments, and substantially expanded the empirical evaluation. A revised manuscript has been uploaded, with changes marked in blue, and the supplementary code has been updated to the revised method for reproducibility. All section, equation, figure, and table references refer to the revised manuscript unless stated otherwise.

1. One small change to the method: separate estimation of the outcome mean and noise scale (Reviewer B67P; Reviewer 3VJ7, W1(b) and Q4; Reviewer kp5z, question). Following Reviewer B67P's analysis, we retain the Gaussian structural outcome model and revise how its mean and noise scale are estimated. The revised IV term is a Gaussian pseudo-likelihood centered on the structural mean integrated over the first-stage treatment distribution (Eqs. 3-4). The mean is updated through a precision-weighted integrated squared loss, with the noise scale held fixed (Eq. 8a). The noise scale is fitted separately to residuals using the observed treatment, with the mean held fixed (Eq. 8b). Under the stated assumptions, Proposition A.2 and Lemma A.2 establish the population targets of these updates as the structural mean and the conditional variance of the structural disturbance, respectively. This separation avoids fitting the structural noise variance to reduced-form variation that also contains first-stage variation and endogeneity covariance (Remark A.3).

2. All experiments were rerun, and the conclusions are unchanged. After the small change, BGM-IV still attains the lowest mean structural MSE in 15/15 configurations of the demand design, 35/36 of the vector-proxy benchmark and 15/15 of the image benchmark. Relative to the original submission its error is lower on the demand design and on the vector-proxy benchmark (geometric mean of the paired ratio revised/original: 0.70-0.86 and 0.83-0.91) and unchanged on the image benchmark (0.97-1.00).

3. The intervals now target the structural function (Reviewer 3VJ7, W1 and Q1-Q3). Section 3.6 reports credible intervals for g0(x,v)g_0(x,v) itself (Eq. 13). They combine the variance over the latent posterior p(z∣v)p(z\mid v) with an ensemble approximation to model uncertainty, quantified by the variance across the K=10K=10 independently initialized candidates that the method already trains. This follows the ensemble approach suggested by Reviewer 3VJ7. The 95% intervals cover the true structural function at 92.7-99.9% of the grid points on the vector-proxy benchmark and at 96.3% on the image benchmark, whereas deep ensembles of DFIV, DeepIV and DML-IV built from their ten runs cover it at 41-80% (Figure 3b). Appendix L adds interval scores and coverage within subgroups.

4. Extensive new experiments we have conducted suggested by the reviewers.

  • γ=0\gamma=0 in the sensitivity study (Table 4, Figure 5).
  • Sensitivity to the latent partition, with a training-side rule for choosing it (Appendix M).
  • Baselines given the BGM-IV representation, and DFIV with twice the sample: BGM-IV keeps the lower error in both comparisons (Appendix N).
  • Two semi-synthetic benchmarks with real covariates, IHDP and PM-CMR, and high-dimensional versions of them: BGM-IV has the lowest error on all 8 datasets and in all 20 high-dimensional configurations (Section 4.4, Appendix O).
  • DeepGMM as an additional baseline: BGM-IV has the lower error in all 580 paired repetitions (Appendix P).
  • A nonlinear (HSIC) version of the latent-exogeneity check (Appendix B).
  • An ablation of the precision weights under heteroscedastic noise (Table 2, Appendix K).
  • The architectures of the baselines on the image benchmark (Appendix F): all of them use convolutional networks.

Besides the extensive experiments, we have also clarified the methodological contributions relative to prior work and the assumptions under which the learned latent representation supports structural-mean identification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.