REVIEW 4 major objections 5 minor 2 references
Analysis of regression discontinuity designs using censored data
T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read Regression discontinuity designs can estimate causal effects from censored survival data.
desk verdict Useful first bridge between censored-outcome transformations and RD designs, but the data-driven bandwidth used in practice is not covered by the stated theory—a gap the paper itself admits. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the censoring unbiased transformation: a function of the observed data whose conditional expectation given the forcing variable equals the conditional expectation of the true log survival time. The paper uses $Y_{\mathrm{IPCW}}=\Delta Y/G(T)$ and $Y_{\mathrm{DR}}=\Delta Y/G(T)+\int_0^{\widetilde T}\{Q_Y(u,W)/G(u)\}\,dM_G(u)$, where $Q_Y(u,W)$ is the conditional mean of log survival given survival past $u$ and $M_G$ is the censoring martingale. The transformation is what lets censored observations be handed to the ordinary local linear RD algorithm; the martingale term is what recovers information from censored cases and produces the claimed efficiency gain. Around this core sit the usual RD pieces: local linear fits on each side of the cutoff, a modified cross-validation bandwidth selector, and sandwich or nearest-neighbor variance estimators.
What would settle it
Simulate the sharp RD design with right censoring, compute the proposed estimator under the paper's data-driven bandwidth, and compare the empirical coverage of nominal 95% intervals with coverage when the bandwidth is fixed at $h = c n^{-1/5}$. If the data-driven version departs materially from 0.95 while the fixed-bandwidth version does not, the point of failure is the assumption that the estimated bandwidth can be ignored in the asymptotics.
Extended reading notes
Core claim
The central claim is that the causal RD estimand is identifiable and estimable from censored survival data by local linear regression on a transformed outcome. For log survival time $Y$, the inverse-probability transform is $Y_{\mathrm{IPCW}}=\Delta Y/G(T)$, and the doubly robust transform adds a martingale integral that uses information from censored observations. Theorem 1 and its sharp-design corollary state that, under regularity conditions, $n^{2/5}(\hat{\tau}-\tau-\phi)$ converges in distribution to a centered normal law, where $\phi$ is the asymptotic bias and the variance matrix is built from boundary variances and covariances of the transformed outcome and treatment assignment. Theorem 2 states that the doubly robust estimator using the true censoring and failure-time models has asymptotic variance no larger than that of the IPCW estimator or of a doubly robust estimator built from an incorrect failure-time model. The simulations and the PLCO prostate-cancer analysis are presented as illustrations of the method in use.
Load-bearing premise
The asymptotic theory treats the data-driven bandwidth from the paper's cross-validation selector as if it were the fixed bandwidth of order $n^{-1/5}$; the paper does not prove that the estimated bandwidth is asymptotically equivalent to that fixed sequence.
Editorial extensions
If this is right
- Both sharp and fuzzy RD estimators can be computed by applying existing local linear RD software to the transformed outcomes, so the method does not require a new estimation algorithm.
- In the fuzzy design, the ratio-of-jumps estimator identifies the complier average treatment effect on log survival time, giving an instrumental-variable interpretation under censoring.
- With correctly specified nuisance models, the doubly robust transform dominates inverse probability weighting in asymptotic variance and shows smaller bias in the paper's simulations.
- In the PLCO example, the method finds no statistically significant effect of the PSA $\geq 4$ ng/ml threshold on mortality or first cancer incidence.
Reading between the lines
- A practical rule the paper leaves implicit: use the doubly robust transform when a trustworthy failure-time model is available, but prefer the simpler IPCW transform when neither nuisance model can be relied on, because IPCW consistency requires only the censoring model.
- Because the main theorem fixes $h\sim n^{-1/5}$ while the implementation uses a data-driven bandwidth, a natural next test is whether the bandwidth selector preserves the claimed coverage; the paper does not settle this.
- The same censoring-unbiased transformation trick could be applied to other boundary estimands, such as distributional, quantile, or multi-cutoff RD, wherever a transformation with the right conditional expectation exists.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes estimators of sharp and fuzzy regression discontinuity treatment effects when the outcome is a censored survival time. It applies inverse probability censoring weighted and doubly robust censoring-unbiased transformations to the log survival time, then uses local linear RD estimators on the transformed outcomes. The main theoretical results (Theorem 1, Corollary 1) claim n^{2/5}-asymptotic normality with explicit bias and variance after replacing unknown censoring and survival distributions by uniformly consistent estimators. Theorem 2 claims an efficiency ordering in favor of the DR estimator with correctly specified models. Bandwidth selection adapts the Ludwig–Miller cross-validation criterion; inference uses plug-in or nearest-neighbor variance estimators and the bootstrap. The methods are evaluated in simulations and applied to the PLCO prostate cancer screening data.
Significance. If the distributional results were valid for the implemented estimator, the paper would make a useful contribution: it extends RD designs to censored outcomes, brings doubly robust efficiency improvements into the RD setting, and offers a practical implementation route through existing software such as rdrobust. The paper has clear strengths: explicit regularity conditions, an extension of the Hahn, Todd and Van der Klaauw proof strategy, reliance on established doubly robust transformation results, and a real-data illustration. However, the load-bearing inferential claims are not supported by the stated theorems because the data-driven bandwidth is used without asymptotic justification and the bias term in Theorem 1 is not addressed in the confidence-interval construction. In addition, the fuzzy RD efficiency theorem appears to contain a sign error in its covariance condition. These issues currently prevent the manuscript from supporting its inferential claims at the level stated.
major comments (4)
- [§4.3–4.4, Theorem 1 (condition R8)] The asymptotic theory assumes a fixed bandwidth with h∼n^{−1/5} (condition R8) and treats h as nonrandom in Lemmas 1–6. The implemented procedure in Section 4.3 chooses ĥ by the adapted Ludwig–Miller criterion, and Section 4.4 plugs ĥ into both the estimator and the variance formulas. No theorem shows ĥ/h→1 in probability or that the limiting distribution in Theorem 1 survives data-dependent bandwidth selection; the manuscript’s Section 7 states that this asymptotic theory is 'currently under investigation.' Because every simulation in Section 5 and the PLCO analysis in Section 6 use ĥ, the coverage probabilities in Tables 1–2 and the standard errors in Table 3 are not justified by the stated theory. The fuzzy case is worse: the bandwidth is min{ĥ_DR, ĥ_Z}, and the minimum of two data-dependent bandwidths is not covered by an argument that applies to each separately.
- [§4.4 and Theorem 1] Theorem 1 gives n^{2/5}(τhat−τ−φ) → N(0,Σ) with a nonzero bias φ of order n^{−2/5}. The variance estimator in Section 4.4 estimates only the asymptotic variance; the reported confidence intervals are centered at τhat with no estimation or correction for φ. Since the bandwidth is of the optimal order n^{−1/5}, the bias contributes a first-order term to coverage error, and no undersmoothing or explicit bias-correction argument is supplied. Thus the normal-approximation intervals in Tables 1–2 are not a consequence of Theorem 1 even if the bandwidth were fixed.
- [Supplementary Materials, definitions of Σ_IPCW_FRD and Σ_DR_FRD; §4.4] The variance formulas attached to Theorem 1 in the Supplement have first coefficient 1/τZ in the fuzzy RD variance, whereas the delta-method variance of the ratio τhat_Y/τhat_Z is (1/τZ^2)Var(τhat_Y) − 2τY/τZ^3 Cov + (τY^2/τZ^4)Var(τhat_Z). Section 4.4 itself uses 1/τhat_Z^2, matching the delta method. Consequently the Σ definitions in the Supplement do not correspond either to the delta method or to the implemented variance estimator. If the Supplement definitions were used for inference, the confidence intervals would be wrong.
- [Theorem 2 (fuzzy RD) and its proof in the Supplement] The fuzzy RD efficiency claim is not established by the given proof. With the delta-method variance (or even the Supplement’s version), the difference AVar(DR)−AVar(IPCW) contains a term −2τY/τZ^3 times the difference in the boundary covariances η_DR−η_IPCW. The theorem assumes η_DR≤η_IPCW, so that difference is nonpositive, which makes the covariance term positive (for τY,τZ>0). Hence the displayed reasoning 'we have ΣDR≤ΣIPCW' does not follow; smaller covariance with the treatment indicator increases, not decreases, the variance of the ratio estimator. The covariance condition appears to have the wrong sign, and the theorem as stated is unsupported.
minor comments (5)
- [§4.2] The text says 'Let G0 and S0 be true distributions of failure and censoring times, respectively,' but the Supplement and subsequent use define G0 as censoring and S0 as failure; please correct the wording.
- [§4.4] There are several typos, including 'resepctively', 'addtion', and 'in addtion'; the manuscript should be carefully proofread.
- [§4.4 and §5] The nearest-neighbor variance estimator depends on a number of neighbors K that is never specified in the main text or simulations; please clarify the choice of K.
- [§5] The bootstrap in the simulations uses only B=50 resamples for bootstrap standard errors and coverage; this is small for coverage estimation and should be stated with a justification.
- [§6] The data-driven RD plots are described as useful even though the theory of Calonico, Cattaneo and Titiunik (2015a) does not directly apply to the transformed responses; the paper should either justify this or present the plots as purely descriptive.
Circularity Check
No significant circularity: the censored-outcome transformations are imported from independent prior work, the local-linear asymptotics extend Hahn et al. without presupposing the target, and no fitted parameter is renamed as a prediction.
full rationale
The paper's principal derivation is self-contained in the relevant sense. The IPCW and DR transformations are not defined in terms of the RD estimand; their key property E(Y_DR|W)=E(logT|W) is quoted from Steingrimsson, Diao and Strawderman (2019) under conditions (C1)-(C5), which do not include the RD treatment effect or bandwidth. The local-linear asymptotic results in Theorem 1 and Corollary 1 are proved in the supplement by adapting Lemmas 1-6 of Hahn, Todd and Van der Klaauw (1999) to the transformed outcomes, with the bias and variance expressions derived from the stated regularity conditions rather than presupposed. Theorem 2's efficiency comparison follows from variance formulas previously established by Steingrimsson et al. (2019) and Suzukawa (2004), which are external, independent results and are not self-citations by the present authors. The paper does have a genuine gap: Section 4.3 replaces the fixed bandwidth h of Theorem 1 (condition R8) with a data-driven CV minimizer, and Section 4.4 plugs that minimizer into the variance estimator, while Section 7 concedes that asymptotic theory for the adapted LM bandwidth "is currently under investigation." This is a correctness/coverage gap, not circularity: the bandwidth is a nuisance tuning parameter, the estimand is not defined as the minimizer of the CV criterion, and the missing result is an asymptotic equivalence or rate for h-hat, not a reduction of the conclusion to an input. No fitted constant is relabeled as a prediction, and no load-bearing argument reduces to a self-citation chain.
Assumptions & free parameters
free parameters (2)
- Local linear bandwidth h =
data-driven via modified Ludwig-Miller cross-validation (Section 4.3)
- LM cross-validation trimming quantile xi =
0.5 in simulations (Section 5)
assumptions (7)
- domain assumption Random censoring: potential failure times, treatment, and forcing variable are independent of censoring time C (Condition 2).
- domain assumption Continuity of potential outcome means E(Y(1)|W=w) and E(Y(0)|W=w) at the cutoff (Condition 1).
- domain assumption No manipulation of the forcing variable (local randomization, Section 2).
- domain assumption Fuzzy RD monotonicity: Z(w*) is nondecreasing in w* at w0 (Condition 6).
- ad hoc to paper Uniform consistency of nuisance estimators: Ĝ -> G0 and Ŝ -> S* (Conditions C6-C7).
- standard math Double robustness property of the DR transformation from Steingrimsson et al. (2019).
- ad hoc to paper Bandwidth h ~ n^{-1/5}, compact symmetric kernel, and technical condition (R9) controlling remainders from estimated nuisance functions.
Cite this review
Pith. "Pith review of Analysis of regression discontinuity designs using censored data." pith.science (2026). https://pith.science/paper/RMVSWW6G
@misc{pith2026190803646,
author = {Pith},
title = {Pith review of: Analysis of regression discontinuity designs using censored data},
year = {2026},
howpublished = {\url{https://pith.science/paper/RMVSWW6G}},
note = {Machine review of arXiv:1908.03646}
}
read the original abstract
In medical settings, treatment assignment may be determined by a clinically important covariate that predicts patients' risk of event. There is a class of methods from the social science literature known as regression discontinuity (RD) designs that can be used to estimate the treatment effect in this situation. Under certain assumptions, such an estimand enjoys a causal interpretation. However, few authors have discussed the use of RD for censored data. In this paper, we show how to estimate causal effects under the regression discontinuity design for censored data. The proposed estimation procedure employs a class of censoring unbiased transformations that includes inverse probability censored weighting and doubly robust transformation schemes. Simulation studies demonstrate the utility of the proposed methodology.
Figures
Reference graph
Works this paper leans on
-
[1]
Abadie, A. and Imbens, G. W. (2006). Large sample properties of matching estimators for average treatment effects. Econometrica 74, 235-267. Andriole, G. L., Crawford, E. D., Grubb III, R. L., Buys, S. S., Chia, D., Church, T. R., Fouad, M. N., Gelmann, E. P., Kvale, P. A., Reding, D. J. and Weissfeld, J. L. (2009). Mortality results from a randomized pros...
work page 2006
-
[455]
Bai, X., Tsiatis, A. A. and O’Brien, S. M. (2013). Doubly-robust estimators of treatment- specific survival distributions in observational studies with stratified sampling. Biomet- rics 69, 830-839. Bor, J., Moscoe, E., Mutevedzi, P., Newell, M. L. and B¨ arnighausen, T. (2014). Regres- sion discontinuity designs in epidemiology: causal inference without ra...
work page 2013
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.