acceptodds
Under review as a conference paper at ICLR 2027

Kernelized IV Confounding Screen for Testing Conditional Exogeneity

Abstract

Instrumental variables (IVs) provide a powerful route to causal identification when latent confounding is present, and conditional exogeneity is one of their most basic credibility requirements. Existing methods for testing IV conditional exogeneity typically rely on structural restrictions, such as linearity or discrete-support conditions, or require auxiliary side information, such as assignment mechanisms and negative controls. We therefore propose Kernelized IV Confounding Screen (KICS) for testing IV conditional exogeneity under more flexible structural conditions and without requiring auxiliary side information. The key idea of KICS is to compare the coefficient systems of an ordinary kernel regression and a candidate-IV-induced weighted kernel regression after removing the part of the candidate IV explained by observed covariates. Theoretically, at the population level, the two coefficient systems coincide under the null hypothesis and differ under detectable alternatives; the estimated coefficient discrepancy is asymptotically normal, enabling a Wald-type test for practical inference. Experiments in synthetic and real-data-based settings show that KICS detects conditional exogeneity violations in flexible nonlinear settings, remains competitive in settings favorable to specialized baselines, and yields real-data results consistent with prior IV assessments. Overall, KICS provides a practical approach to IV conditional exogeneity testing in settings where existing methods are difficult to apply.

Primary area:
causal reasoning
Keywords:
Instrumental Variables, Conditional Exogeneity, Kernel Methods, Latent Confounding

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_78zyOpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3(+2)
Read review
Paper summary

This paper proposes KICS, a test for the conditional exogeneity of a candidate instrumental variable under more flexible structural conditions than previous work and without requiring auxiliary side information. The key idea is that, under conditional exogeneity, reweighting the outcome regression by the squared residual of the candidate IV leaves the regression coefficients unchanged, whereas a violation induces a systematic coefficient discrepancy. Theoretical and empirical results demonstrate that KICS correctly tests conditional exogeneity under specified assumptions, with broader applicability than existing approaches.

Critical strengths
  1. Testing IV conditional exogeneity is a long-standing challenge that limits the practical use of IV methods, and the proposed procedure addresses this gap with clear practical value.

  2. The core idea, reweighting the outcome regression by the squared residual of the candidate IV leaves coefficients stable under conditional exogeneity but shifts them under a violation, is intuitive and well motivated. While the regression-comparison principle itself builds on KRCD, its IV-specific formulation here is novel.

  3. The paper is clearly written and easy to follow.

  4. The theoretical analysis is thorough and, under the stated sufficient conditions, correct.

  5. The empirical evaluation includes both synthetic and real-data experiments, and the results are convincing.

Critical weaknesses
  1. Although adapting the KRCD principle to IV testing is well motivated, the IV setting introduces a subtlety that undermines the direct analogy. Specifically, because the regression input is V=(X,T)V = ( X, T ) and TT is a collider on the path Z→T←UYZ\rightarrow T\leftarrow U_Y, conditioning on TT makes ZZ and UYU_Y​ dependent even when the null Z⊥UY∣XZ \perp U_Y \mid X holds. Thus, in principle, the IV-weighted regression coefficients can differ from the ordinary ones even under conditional exogeneity.

  2. The authors avoid this issue by imposing Condition A3 in the appendix, which requires that RZ2R_Z^2 ​ and h(UY)h(U_Y) do not interact given VV. This condition is strictly stronger than the null hypothesis itself, is not directly testable since UYU_Y ​ is latent, and appears unlikely to hold in practical IV settings where the collider structure is inherent. In effect, the paper addresses the untestability of conditional exogeneity by assuming another untestable condition.

  3. To mitigate this concern, the authors should (i) move Condition A3 to the main text, explain when it is plausible to hold, and discuss whether any diagnostic exists for it; and (ii) add experiments that systematically compare performance when A3 holds versus when it is violated to varying degrees, to clarify the practical robustness of the test.

Questions for authors

Please see Weaknesses.

If the authors can fully address my concern regarding Condition A3, I would consider this a highly interesting and practically valuable work that makes a significant contribution to IV research, and I would be very willing to raise my score.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 78zy (1/4)

We thank the reviewer for the constructive feedback. Our responses are provided below.

[W1 & W2] Does the treatment collider invalidate KICS under a valid IV, and can A3 hold in such settings?

  • Reviewer's concern: The reviewer notes that conditioning on the collider Z→T←UYZ\to T\leftarrow U_Y may induce dependence between ZZ and UYU_Y even under IV exogeneity, potentially causing coefficient discrepancies. Although A3 addresses this issue, the reviewer further questions whether A3 can hold in practical IV settings involving a collider.

  • Key conclusion: The presence of a collider does not necessarily violate A3.

  • Theoretical analysis: Collider-induced dependence between ZZ and UYU_Y does not necessarily extend to the squared residual-IV weight RZ2R_Z^2. Specifically, if E[RZ2∣UY,T,X]=E[RZ2∣T,X]E[R_Z^2\mid U_Y,T,X]=E[R_Z^2\mid T,X], then Cov⁡(RZ2,h(UY)∣T,X)=0\operatorname{Cov}(R_Z^2,h(U_Y)\mid T,X)=0, so A3 holds even when ZZ and UYU_Y are conditionally dependent.

  • Construction and validation in a collider setting: To show that A3 can hold despite the collider and to evaluate KICS in this setting, we construct a collider-IV scenario.

    • Data-generating process (DGP): We construct the IV using a random sign SS and a positive random magnitude AA. Here S∈{−1,1}S\in\lbrace -1,1\rbrace is balanced, AA is nonconstant, and S,A,U,XS,A,U,X are mutually independent, with U∼N(0,1)U\sim N(0,1) and X∼N(0,I3)X\sim N(0,I_3). The DGP is

      Zγ=SA(1+γU)Z_\gamma = SA(1+\gamma U)

      T∼Bernoulli⁡(σ(1.4SAκ+U))T \sim \operatorname{Bernoulli}\left(\sigma(1.4SA^\kappa+U)\right)

      Y=0.8T+0.5X1−0.3X2+0.2X3+U+εY.Y = 0.8T+0.5X_1-0.3X_2+0.2X_3+U+\varepsilon_Y.

    • Theoretical validation: At γ=κ=0\gamma=\kappa=0, the IV is valid, and both of the following hold:

      • Collider dependence is present. Because both SS and UU affect TT, conditioning on TT induces dependence between ZZ and UU: Cov⁡(Z,U∣T=1)≈−0.163≠0\operatorname{Cov}(Z,U\mid T=1)\approx-0.163\neq0.

      • A3 still holds. Here TT depends on the IV sign SS, but not its magnitude AA. Conditioning on TT can therefore associate SS with UU, but not A2A^2 with UU. Since RZ2=A2R_Z^2=A^2 and AA is independent of (T,X,U)(T,X,U), we have Cov⁡(RZ2,U∣T,X)=0\operatorname{Cov}(R_Z^2,U\mid T,X)=0. Thus, A3 holds exactly.

      This shows that A3 can hold in an IV setting with a collider.

    • Experimental validation: To further evaluate KICS's detection ability in this setting, we fix κ=0\kappa=0 and increase γ\gamma to progressively introduce dependence between the IV and the latent confounder UU. Using the same KICS implementation (250 repetitions, linear kernel, P=5P=5, α=0.05\alpha=0.05), we obtain:

      Table 1. KICS detection rates in the collider-IV setting.

      γ\gammaIV statusDetection rate
      0.00Valid0.028
      0.05Invalid0.116
      0.10Invalid0.520
      0.15Invalid0.900
      0.20Invalid0.996
      0.25Invalid1.000
    • Experimental conclusion: In this collider setting, KICS maintains a low false-rejection rate for a valid IV, while its detection rate rises sharply as exogeneity is violated. These results show that KICS can operate effectively in collider settings satisfying A3.

  • Paper Revision. We have revised Section 4.1 (p. 5, lines 260–269) and added the collider construction and detection experiment in Appendix D.1 (pp. 22–23, lines 1177–1213; Table 5 on p. 23), with the formal A3 statement in Appendix A.1.2 (p. 14, lines 726–736).

AuthorsOpenReview ↗

Response to Reviewer 78zy (2/4)

[W2 & W3] Is there an observable diagnostic for A3?

  • Reviewer's concern: A3 involves the unobserved latent confounder UYU_Y, making it difficult to assess directly. The reviewer also asks whether a practical diagnostic is available.

  • Observable diagnostic: To address this concern, we provide an observable conditional-moment representation of A3 and propose a cross-fitting-based diagnostic.

    • Observable conditional moment: Under the outcome model Y=g(T,X)+h(UY)+εYY=g(T,X)+h(U_Y)+\varepsilon_Y and the noise-orthogonality condition A4,

      Cov⁡(RZ2,Y∣T,X)=Cov⁡(RZ2,h(UY)∣T,X).\operatorname{Cov}(R_Z^2,Y\mid T,X)=\operatorname{Cov}(R_Z^2,h(U_Y)\mid T,X).

      This equality holds because g(T,X)g(T,X) is fixed given (T,X)(T,X) and contributes no conditional covariance, while A4 eliminates the corresponding noise contribution. Thus, the observable outcome YY can be used to assess the conditional-moment restriction underlying A3.

    • Diagnostic implementation: In practice, cross-fitting can be used to estimate the conditional means of RZ2R_Z^2 and YY given (T,X)(T,X). We can then examine whether the conditional expectation of the product of their residuals is close to zero. This allows us to diagnose the moment restriction underlying A3 without directly observing UYU_Y.

    • Diagnostic experiment: We further evaluate this diagnostic using the same collider DGP, fixing γ=0\gamma=0 and varying κ\kappa to progressively violate A3 while maintaining IV exogeneity. We use cross-fitting to estimate the conditional means and construct a treatment-stratified residual-product diagnostic. With 250 repetitions and α=0.05\alpha=0.05, we obtain:

      Table 2. Rejection rates of the observable A3 diagnostic.

      κ\kappaDiagnostic rejection rate
      0.000.012
      0.200.096
      0.400.392
      0.600.676
      0.800.856
      1.000.948
    • Diagnostic conclusion: The diagnostic exhibits a low rejection rate when A3 holds, while its rejection rate increases substantially as A3 is violated, reaching 0.948 at κ=1\kappa=1. These results demonstrate that the observable diagnostic can effectively detect violations of A3.

  • Paper Revision. We have revised Appendix D.2 (p. 23, lines 1216–1241; Table 6) to add the observable conditional-moment identity, cross-fitting diagnostic, and numerical evaluation.

AuthorsOpenReview ↗

Response to Reviewer 78zy (3/4)

[W3] How does KICS behave as A3 is progressively violated?

  • Theoretical insight: Violations of A3 need not induce large coefficient discrepancies. When the weighted Gram matrix is well-conditioned, the population discrepancy is bounded by a constant times E[c(V)2]\sqrt{E[c(V)^2]}, where c(V)=Cov⁡(RZ2,h(UY)∣V)c(V)=\operatorname{Cov}(R_Z^2,h(U_Y)\mid V). Thus, moderate A3 violations can remain within a bounded perturbation regime.

  • Sensitivity experiment: To evaluate the effect of A3 violations on KICS, we use the collider DGP introduced in our [W1 & W2] response.

    • Parameter control: We fix γ=0\gamma=0 so that the IV remains exogenous and increase κ\kappa to progressively violate A3. A3 holds at κ=0\kappa=0; as κ\kappa increases, treatment depends more strongly on the IV magnitude AA, inducing conditional dependence between the squared weight and the latent confounder given TT.

    • Experimental results: Using the same KICS configuration (250 repetitions, linear kernel, P=5P=5, α=0.05\alpha=0.05), we obtain:

      Table 3. KICS false-rejection rates under progressive violations of A3.

      κ\kappaDetection rate
      0.000.028
      0.100.028
      0.200.024
      0.300.028
      0.400.044
      0.500.052
      0.600.068
      0.700.088
      0.800.112
      0.900.116
      1.000.148
    • Experimental conclusion: All IVs remain valid, so the detection rates here measure false rejections. Relative to the nominal significance level α=0.05\alpha=0.05, these rates remain within the 95% Monte Carlo fluctuation range through κ=0.60\kappa=0.60, but exceed it from κ=0.70\kappa=0.70 onward. These results suggest that KICS exhibits some robustness to moderate violations of A3.

  • Paper Revision. We have revised Appendix D.3 (p. 24, lines 1242–1271; Table 7) to add the coefficient-discrepancy bound and A3-violation sensitivity experiment.

[W3] Main-text revisions to A3

  • We have moved Condition A3 to the main text and added discussions of its applicability in collider settings, observable diagnostics, the new diagnostic experiment, and the power and A3 sensitivity experiments.

  • Paper Revision. We have revised Section 4.1 (p. 5, lines 260–269; p. 6, lines 282–284) to present and explain the null-side A3 condition in the main text, with supporting details in Appendix A.1.2 (p. 14, lines 726–736) and technical analyses in Appendices D.1–D.3 (pp. 22–24, lines 1177–1271).

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

The paper proposes a test for the conditional exogeneity assumption in instrumental variable (IV) regression, allowing for potentially nonlinear relationships. The test compares the coefficients of a standard kernel regression with those of an IV-weighted version. Under conditional exogeneity, the variation in the instrument that remains after conditioning on the covariates should not affect the outcome regression, so the two sets of coefficients should be approximately equal. The authors construct a Wald-type test based on the difference and derive its asymptotic distribution.

Critical strengths

Testing the conditional exogeneity of an IV is an important problem. The procedure is simple and has a closed-form solution. The theoretical results are clearly derived and the experiments indicate good finite-sample performance.

Critical weaknesses
  1. I think there should be more discussion of what detectable alternatives are, perhaps by moving some of the conditions in Appendix A to the main text. Currently, it is not clear to me what the method uses to falsify conditional exogeneity. My intuition is that the variation in the instrument needs to be richer than that in the treatment in a certain sense, e.g., overidentification in the linear case.

  2. In most of the experiments, it seems that the confounding is joint between Y, T, and Z. A more realistic scenario would be one with confounding between Y and T in all scenarios, where only the effect of the confounder on Z is varied. If there is no confounding between Y and T, there is no need for an IV to begin with.

Questions for authors
  1. In practice, how do you select the kernel or the basis functions? Which kernel do you use in your experiments?
  2. I worry that testing for exogeneity violations in this manner invites post-selection inference issues when subsequent IV inference is conditioned on passing the exogeneity test. How do you recommend your method is used in practice?
  3. How sensitive are the results to estimation error in the nuisance function mm? Would it be possible to use a Neyman-orthogonal estimating equation for the IV-weighted estimator, as in the double machine learning literature?
  4. I think the acronym KRCD is never properly introduced.
10 comments
AuthorsOpenReview ↗

Response to Reviewer EMLw (1/3)

We thank the reviewer for the constructive feedback. Our responses are provided below.

[W1] What makes an exogeneity violation detectable by KICS?

  • Key conclusion: KICS detects exogeneity violations that induce systematic coefficient differences between ordinary and IV-weighted outcome regressions. This requires only a single candidate IV, without requiring richer IV variation than treatment variation or relying on classical overidentification.

    • Detection mechanism: Building on the regression-reweighting principle [1][2][3], we use the variation in the candidate IV unexplained by observed covariates XX to construct the squared residual weight w=RZ2w=R_Z^2. If this reweighting induces systematic changes in the outcome regression coefficients, KICS can detect the resulting discrepancy.

    • Detectability condition: An IV exogeneity violation is detectable by KICS if it produces a nonzero weighted residual signal in at least one kernel-basis direction. Let V=(T,X)V=(T,X) and e0=Y−E[Y∣V]e_0=Y-E[Y\mid V]. This condition can be expressed as:

      E[wb(V)e0]≠0.E[w b(V)e_0]\neq 0.

  • Manuscript revision: Following the reviewer's suggestion, we have moved the relevant conditions and explanations from Appendix A to the main text to clarify this detection mechanism.

  • Paper Revision. We have revised Section 4.1 (p. 5, lines 260–269; p. 6, lines 282–284) and clarified the detectability condition in Appendix A.1.2 (p. 14, lines 744–755).

[W2] Does treatment–outcome confounding persist under the valid-IV null?

  • Some original synthetic experiments indeed have settings where treatment–outcome confounding disappears when the candidate IV is valid. This makes the corresponding valid-IV settings less realistic, but these experiments still demonstrate violation detection and baseline comparisons. Our real-world experiment does not have this issue.

  • Additional experiment with persistent confounding: To address this concern, we added an experiment that keeps the effects of latent confounder UU on treatment and outcome fixed, while varying only its association with the candidate IV.

    • Data-generating process (DGP): Let S,U∈{−1,1}S,U\in\lbrace -1,1\rbrace be independent balanced binary variables, and let A>0A>0 be an independent nondegenerate random variable. The DGP is:

      Zγ=SA(1+γU),T=S+U+f(X),Z_\gamma=SA(1+\gamma U),\quad T=S+U+f(X),

      Y=0.7T+0.35T2+g(X)+0.8U+0.4εY,Y=0.7T+0.35T^2+g(X)+0.8U+0.4\varepsilon_Y,

      f(X)=0.30X1−0.25X2+0.15X32,f(X)=0.30X_1-0.25X_2+0.15X_3^2,

      g(X)=0.40X1−0.30X22.g(X)=0.40X_1-0.30X_2^2.

    • Confounding and IV validity: In all settings, UU continues to affect both TT and YY. The IV is valid at γ=0\gamma=0; when γ≠0\gamma\neq0, it depends on UU, producing a detectable weighted residual moment.

    • Experimental results: We evaluated KICS and the baseline methods with n=600n=600, 150 repetitions per setting, a quadratic polynomial kernel for KICS, P=15P=15, and significance level α=0.05\alpha=0.05. The detection rates are:

      Table 1. Detection rates under persistent treatment–outcome confounding.

      Methodγ=0.0\gamma=0.0γ=0.1\gamma=0.1γ=0.2\gamma=0.2γ=0.3\gamma=0.3γ=0.4\gamma=0.4
      KICS (ours)0.0270.9531.0001.0001.000
      LinNG-IV0.3600.2930.4530.6730.713
      IV-PIM1.0001.0001.0001.0000.880
      IV-GIN1.0001.0001.0001.0001.000
      Branson0.0470.0530.0530.0400.047
      NCF0.0930.0930.0870.0730.067
      AIT0.0330.0400.0400.0400.040
      Kitagawa0.0400.0400.0330.0330.033
      Sun0.0400.0400.0330.0330.033
    • Experimental findings: Under the valid-IV null, KICS achieves a detection rate of 0.027, below the nominal significance level of 0.05. As γ\gamma increases, its detection rate rises to 1. In contrast, the tested baseline implementations exhibit either elevated false-rejection rates or limited sensitivity to exogeneity violations. These results are consistent with the findings in the main paper regarding KICS's strong detection performance in general nonlinear settings.

  • Paper Revision. We have added the persistent-confounding experiment to Section 5.5 (p. 10, lines 500–520; Table 3), with the DGP, experimental settings, and baseline-applicability caveats in Appendix D.4 (p. 24, lines 1274–1295).

AuthorsOpenReview ↗

Response to Reviewer EMLw (2/3)

[Q1] How should kernels and basis functions be selected?

  • In our experiments, we use a quadratic polynomial kernel for continuous inputs and a linear kernel for discrete support. Basis functions are centered at PP selected sample points; Figure 4(c) reports sensitivity to PP.

  • In practice, we recommend choosing kernels based on the available structural knowledge:

    • With structural knowledge, choose a kernel that matches the approximate outcome regression structure. For example, use a linear kernel for approximately linear relationships and a polynomial kernel for low-order nonlinearities.
    • Without prior knowledge, a Gaussian RBF kernel is a flexible choice. Its universal approximation property allows continuous functions on compact domains to be approximated arbitrarily well. Kernel parameters can be selected on training data, followed by running KICS with the fixed kernel on an independent test set.
  • Paper Revision. We have revised Section 4.1 (p. 6, lines 299–303) and Appendix D.5 (p. 25, lines 1298–1305) with kernel and basis-selection guidance.

[Q2] How should KICS be used before downstream IV inference?

  • We recommend splitting the data into two parts: use the first part to screen candidate IVs with KICS, and the second, independent part to estimate causal effects and construct confidence intervals.

  • This avoids using the same data for both IV selection and causal effect inference.

  • Paper Revision. We have revised Section 4.1 (p. 6, lines 303–306) and Appendix D.5 (p. 25, lines 1306–1309) with the recommended independent data split.

[Q3] How sensitive is KICS to nuisance estimation error? Could Neyman orthogonality help?

  • Existing theory and additional experiment: The existing KICS theory controls estimation error in m(X)=E[Z∣X]m(X)=E[Z\mid X] through Condition C5. To assess finite-sample sensitivity to nuisance estimation error, we added a perturbation experiment.

    • Neyman-orthogonal extension: Following the reviewer's suggestion, we also implemented a Neyman-orthogonal extension and evaluated its performance in the same experiment.

    • Experimental setup and results: We perturb m^(X)\hat m(X) by δ\delta times its independent-validation RMSE, where δ=0\delta=0 indicates no additional bias. Each setting is repeated 250 times. Accuracy is defined as the average of the correct acceptance rate for valid IVs and the detection rate for invalid IVs.

      Table 2. Accuracy under nuisance-estimation perturbations.

      Perturbation δ\deltaOriginal KICS AccuracyOrthogonal Accuracy
      0.00.9760.974
      0.10.9700.974
      0.20.9700.974
      0.30.9680.974
      0.40.9680.974
      0.50.9680.974
      0.60.9680.972
      0.70.9640.972
      0.80.9600.972
    • Experimental findings: The results show that original KICS retains good finite-sample stability within the tested perturbation range, while orthogonalization further improves it.

  • Paper Revision. We have revised Appendix A.3 (p. 16, lines 837–857) and Appendix D.5 (p. 25, lines 1311–1333; Table 8), with a pointer in Section 4.1 (p. 6, lines 305–306).

[Q4] KRCD acronym

  • We have introduced the full name, Kernel Regression Confounder Detection (KRCD), at its first mention in the revised manuscript.

  • Paper Revision. We have revised Section 2.3 (p. 3, lines 134–141, specifically line 137) to expand KRCD at first mention.

References

[1] Buja et al. (2019). Models as Approximations II: A Model-Free Theory of Parametric Regression.

[2] Schultheiss, Bühlmann, and Yuan (2024). Higher-Order Least Squares: Assessing Partial Goodness of Fit of Linear Causal Models.

[3] Chen et al. (2026). Detecting Unobserved Confounders: A Kernelized Regression Approach.

AuthorsOpenReview ↗

Response to Reviewer EMLw (3/3)

For your convenience, we have included a summary table detailing the changes in the revised version of our paper.

Table 3. Revision summary of our paper.

CommentsLocation of Revisions in Our Revised Paper
[W1] DetectabilityClarified when residual-IV reweighting induces coefficient discrepancies and how the null-side assumptions differ from detectable alternatives. Section 4.1, pp. 5–6, lines 260–269 and 282–284; Appendix A.1.2, p. 14, lines 744–755.
[W2] Persistent treatment–outcome confoundingAdded an experiment retaining latent treatment–outcome confounding under the valid-IV null and comparing nine methods across exogeneity violations. Section 5.5, p. 10, lines 500–520 (Table 3); Appendix D.4, p. 24, lines 1274–1295.
[Q1] Kernel and basis selectionExplained kernel choices, sample-point basis construction, and independent-data tuning. Section 4.1, p. 6, lines 299–303; Appendix D.5, p. 25, lines 1298–1305.
[Q2] Downstream IV inferenceRecommended independent data splits for candidate screening and causal effect estimation. Section 4.1, p. 6, lines 303–306; Appendix D.5, p. 25, lines 1306–1309.
[Q3] Nuisance estimation and Neyman orthogonalitySpecified first-stage error control and added a perturbation study with an orthogonal extension. Appendix A.3, p. 16, lines 837–857; Appendix D.5, p. 25, lines 1311–1333 (Table 8).
[Q4] KRCD acronymExpanded the acronym on first use. Section 2.3, p. 3, line 137.

If you have any further concerns or questions, please feel free to raise them. We would be glad to continue the discussion and clarify them in more detail.

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

The authors propose a kernel based test of conditional exogeneity + the ad-hoc assumptions A3. They compare the difference between the regression vector of outcome Y onto treatment T and covariates and X, and of regression vector of Y onto T and X weighted by (Z - E[Z|X])^2.

Under conditional exogeneity + A3, this difference is zero, allowing for construction of a test.

Critical strengths

Clearly written. Significant problem.

Critical weaknesses

I can't seem to find a natural IV setting in which A3 holds while U | Z | X doesn't.

  • In linear structural equation + linear first stage, it does not in general.
  • One is when treatment is known to be exogenous, but then there's no point of testing for condirtional exogeneity.
  • Another one is when Z|X ~ Bernoulli(1/2) in which case the residual is automatically 1/4 everywhere and the difference of regression coefficient is 0, but then the test has no power.

In the synthetic experiments, whenever conditional exogeneity and A3 jointly hold or jointly fail, and as such they should be better presented as a test of A3 + cond. exo. as opposed to a test of cond exo over of a model where A3 holds.

Questions for authors

Did I miss a natural setting where the null holds under conditional exogeneity while not requiring A3 in a sense that defeats the purpose of testing for exogeneity.

9 comments
AuthorsOpenReview ↗

Response to Reviewer HyAN (1/3)

We thank the reviewer for the constructive feedback. Our responses are provided below.

[W & Q] Clarification on the Relationship Between A3 and Conditional Exogeneity

  • The reviewer's concerns mainly involve the following two misunderstandings:

    • A3 must also hold under the alternative: It is difficult to identify a natural setting where A3 holds but conditional exogeneity fails, raising concerns that A3 may undermine the practical purpose of testing exogeneity.
    • Simultaneous changes imply a joint test: In our synthetic experiments, A3 and conditional exogeneity always hold or fail together, suggesting that KICS is effectively a joint test of both conditions.
  • We would like to clarify these two points separately.

    • Response to Point 1: A3 is imposed only under the null hypothesis.

      • Reviewer's concern: It is difficult to identify a natural setting where A3 holds while conditional exogeneity is violated, raising questions about the practical role of A3.

      • Our clarification: A3 is only required to hold under conditional exogeneity (i.e., H0:Z⊥UY∣XH_0: Z\perp U_Y\mid X). As explicitly stated in Condition A3 (Appendix A.1.2, page 14):

        Under H0:Z⊥UY∣XH_0: Z\perp U_Y\mid X, assume that the IV-induced weight does not interact with the outcome-side latent component once VV is fixed, in the sense that

        E[R_Z^2h(U_Y)\mid V]=E[R_Z^2\mid V]E[h(U_Y)\mid V]. \tag{1}

      • Theoretical motivation for A3: A3 is not an arbitrarily introduced additional assumption, but rather a condition aligned with the core requirement of H0H_0. While H0H_0 requires the IV to be conditionally independent of the outcome-side latent confounder, A3 further concerns the corresponding conditional uncorrelatedness relation under squared-residual reweighting.

      • Sensitivity analysis of A3 violations

        • Experimental setup:

          • To evaluate the robustness of KICS to violations of A3, we further examine how progressively violating A3 affects KICS while maintaining IV exogeneity.

          • Data-generating process (DGP): Let S∈{−1,1}S\in\lbrace -1,1\rbrace be a balanced random sign, A>0A>0 a non-degenerate random magnitude, U∼N(0,1)U\sim N(0,1) a latent confounder, and X∼N(0,I3)X\sim N(0,I_3). Here, S,A,U,XS,A,U,X are mutually independent, and εY\varepsilon_Y is independent noise. The DGP is given by:

            Zγ=SA(1+γU),Z_\gamma = SA(1+\gamma U),

            T∼Bernoulli(σ(1.4SAκ+U)),T \sim \mathrm{Bernoulli}\left(\sigma(1.4SA^\kappa+U)\right),

            Y=0.8T+0.5X1−0.3X2+0.2X3+U+εY.Y = 0.8T+0.5X_1-0.3X_2+0.2X_3+U+\varepsilon_Y.

            where σ(⋅)\sigma(\cdot) denotes the sigmoid function.

          • Parameter control: We fix γ=0\gamma=0 to maintain conditional exogeneity, yielding Z=SAZ=SA and RZ2=A2R_Z^2=A^2. At κ=0\kappa=0, treatment depends only on the IV sign, not its magnitude, so Eq. (1) holds.

            As κ\kappa increases, treatment becomes increasingly dependent on the IV magnitude, inducing conditional correlation between RZ2R_Z^2 and the latent confounding component given VV, thereby progressively violating A3.

          • With n=600n=600, 250 replications, and a significance level of α=0.05\alpha=0.05, the results are as follows:

            Table 1. False-rejection rates of KICS under progressively stronger A3 violations.

            κ\kappaFalse-rejection rate
            0.000.028
            0.100.028
            0.200.024
            0.300.028
            0.400.044
            0.500.052
            0.600.068
            0.700.088
            0.800.112
            0.900.116
            1.000.148
        • Experimental findings: For κ≤0.60\kappa\leq0.60, the false-rejection rates remain within the 95% Monte Carlo fluctuation range around the nominal significance level, indicating that KICS exhibits robustness to moderate violations of A3.

  • (Continued in Part 2)

AuthorsOpenReview ↗

Response to Reviewer HyAN (2/3)

  • (Continued from Part 1)

    • Response to Point 2: The theoretical correspondence between A3 and conditional exogeneity explains their simultaneous changes.

      • Reviewer's concern: A3 and conditional exogeneity always hold or fail together in the synthetic experiments, suggesting that KICS is effectively a joint test of both conditions.

      • Our clarification:

        • Conceptual clarification: A3 is imposed only under H0H_0. Therefore, we believe that what the reviewer refers to as holding or failing together with conditional exogeneity is actually Eq. (1) within A3.

        • Theoretical correspondence: The central idea of A3 is that, when conditional exogeneity holds, it can manifest as a conditional uncorrelatedness relation under squared-residual reweighting (i.e., Eq. (1)). This establishes a clear theoretical correspondence between conditional exogeneity and Eq. (1), which also explains their simultaneous changes in our original synthetic experiments.

          • Under the null hypothesis: At zero confounding strength, conditional exogeneity holds, and Eq. (1) is trivially satisfied because the latent confounding component vanishes.
          • When conditional exogeneity is violated: As the latent confounder enters IV generation, conditional exogeneity is violated. In our original DGP, this dependence also violates Eq. (1).
        • Why KICS is not a joint test: KICS targets IV conditional exogeneity, while A3 is an auxiliary condition ensuring validity under the null. Under A3 and the other maintained conditions, KICS tests conditional exogeneity through the discrepancy between ordinary and IV-weighted regression coefficients, without separately testing Eq. (1). Thus, their simultaneous changes in our synthetic experiments do not imply a joint testing target.

  • Paper Revision. We have revised Section 4.1 (p. 5, lines 260–269; p. 6, lines 285–298) to clarify that A3 is required only under the valid-IV null and to distinguish it from the conditional-exogeneity testing target; see also Appendix A.1.2 (p. 14, lines 726–736). We have also added an A3 sensitivity analysis in Appendix D.3 (p. 24, lines 1242–1271; Table 7).

AuthorsOpenReview ↗

Response to Reviewer HyAN (3/3)

For your convenience, we have included a summary table detailing the changes in the revised version of our paper.

Table 2. Revision summary of our paper.

CommentsLocation of Revisions in Our Revised Paper
[W & Q] A3 under the alternativeClarified that A3 is required only under the valid-IV null, not under detectable alternatives, while retaining its original formulation and adding an explicit clarification after the equation. Section 4.1, p. 5, lines 260–269; Appendix A.1.2, p. 14, lines 726–736.
[W & Q] Joint-test interpretationClarified the distinction between the conditional-exogeneity testing target and the sufficient null-side moment condition. Section 4.1, p. 5, lines 260–269; p. 6, lines 285–298.
[W & Q] Sensitivity to A3 violationsAdded a discrepancy bound and a false-rejection experiment under progressively violated A3 while maintaining IV exogeneity. Appendix D.3, p. 24, lines 1242–1271; Table 7.

If you have any further concerns or questions, please feel free to raise them. We would be glad to continue the discussion and clarify them in more detail.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.