Pith.

Methodology · arXiv:2611.16278 v1 · submitted 12 Nov 2026 · 41 references

On Ignorability Beyond Missing at Random: Completion-Invariant Missingness for Nonmonotone Missing Data

Kosuke Morikawa, Yutaka Kano, Shonosuke Sugasawa

The paper says

“The missing-at-random assumption underlies most analyses of incomplete data. Under nonmonotone missingness, it is especially stringent, imposing pattern-specific restrictions with different conditioning sets. A strictly weaker normalization condition is known to justify maximizing the observed-data likelihood.”

more of the abstract

We show that every law satisfying this condition has a unique missing-at-random counterpart sharing its observed-data and complete-data distributions. Observed-data procedures developed under missing at random for identified functionals, including doubly robust estimators, therefore remain valid, with the same semiparametric efficiency bound. We then prove that the condition holds if and only if filling in the missing values from their conditional distribution under the true full model preserves the complete-data distribution, hence the name completion-invariant missingness. We use a characteristic kernel to compare the completed distribution with the fitted model and assess the joint compatibility of the specified full model and completion-invariant missingness. We derive the asymptotic null distribution with estimated parameters and establish bootstrap and multiplier calibration. Since a parametric full model alone need not identify completion-invariant missingness, we give identification conditions for Gaussian, Student t and finite mixture families. Simulations assess sensitivity to departures from each assumption, and applications to clinical laboratory measurements and childhood growth illustrate the diagnostic.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

The passes disagreed. Reader: holds up. Skeptic: agrees with the reader. Referee: major revision.

Solid theory, honest about limitations: CIM is a real relaxation of MAR with preserved efficiency, but the kernel test is a joint test of CIM plus correct model specification and cannot separate the two. This line is wrong

3 major objections · 7 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 41 references say.

The paper says, at Section 6.2, Table 2

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.77):

“Rejection indicates failure of either component or both and cannot distinguish between them.”

Pith says: Objection 1 of 3 · major

The simulation results in Table 2 reveal a critical practical limitation: under MCAR with a t10 law (CIM holds), the test rejects at 62% (n=500) and 100% (n=2000), while genuine CIM failure at γ=0.5 (B1) is detected only 5.2–8.2%. This asymmetry means that in practice, rejection is ambiguous and non-rejection is weak evidence for CIM.

more of objection 1

The authors acknowledge this entanglement in Section 8, but the severity of the issue deserves more prominent discussion. The test is hypersensitive to heavy tails and only weakly sensitive to the departures it is designed to detect. This does not invalidate the theoretical results, but it substantially limits the practical utility of the diagnostic as presented.

The paper says, at Section 4.2, Theorem 3

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The asymptotic null distribution of the test depends on the eigenvalues λ_j of the fitted covariance operator, which are not consistently estimated by the bootstrap in a simple way. The bootstrap calibration is justified, but the paper does not discuss how to ensure the nondegeneracy and continuity of the quantile condition (Assumption S1(iv)).

more of objection 2

In practice, if the limiting distribution is degenerate or has point masses, the bootstrap may not provide correct size control. The authors should comment on when this condition might fail and how practitioners can check it.

The paper says, at Section 5.2, Theorem 4

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The Gaussian identification condition (18) requires a tail bound on the complete-case probability π_1(y). This is a condition on the unobserved selection probability and cannot be verified from the observed data. The authors acknowledge this, but the practical relevance of the theorem is limited because it provides a sufficient condition that is not checkable.

more of objection 3

The alternative identification results (Theorem 5, Proposition 2) rely on exclusive variables and monotonicity, which are also strong assumptions. The paper would benefit from a clearer discussion of the trade-offs between these identification strategies and their plausibility in applications.

7 minor objections

The paper says, at Section 2, Definition 1

Pith says: minor objection 1 of 7

The definition of CIM uses the notation π_r(y(r)) = P(R=r | Y(r)=y(r)). It would be helpful to explicitly state that this is defined PY,θ0-almost surely, as done later.

The paper says, at Section 3, Theorem 1

Pith says: minor objection 2 of 7

The statement of Theorem 1 is dense. A brief intuitive explanation of the three equivalent conditions and their implications would improve readability.

show 5 more

The paper says, at Section 4.2, Algorithm 1

Pith says: minor objection 3 of 7

The algorithm description is clear, but it would be helpful to note that the bootstrap calibration requires B to be large enough for stable p-value estimation. The paper uses B=499 or 1999; a brief comment on the choice would be useful.

The paper says, at Section 6.1, Table 1

Pith says: minor objection 4 of 7

The rejection rates under the null are generally close to 5%, but at n=2000 under A1 the rate is 6.7% with Monte Carlo SE 0.8%. This is slightly elevated. The authors might comment on whether this reflects a small-sample bias in the bootstrap calibration or is within expected Monte Carlo variation.

The paper says, at Section 7.1, SUPPORT2 analysis

Pith says: minor objection 5 of 7

The analysis uses a bivariate Gaussian working model for transformed laboratory measurements. Given the strong rejection, it would be informative to see a sensitivity analysis using a more flexible full model (e.g., a mixture) to assess whether the rejection is driven by model misspecification rather than CIM failure. The authors mention this possibility in Section 8 but do not explore it in the application.

The paper says, at Supplementary Material, Section S2.6

Pith says: minor objection 6 of 7

The closed-form Gaussian kernel calculations are very useful. It would be helpful to also provide the formulas for the cross inner products at different fitted parameters, which are needed for the bootstrap calibration.

The paper says, at General

Pith says: minor objection 7 of 7

The paper uses the term 'completion-invariant missingness' (CIM). While the name is motivated by Theorem 2, it might be confused with 'complete-case missingness' or other terms. The authors should ensure the acronym is clearly defined at first use.

What would settle it

Run the fitted-kernel diagnostic and direct-likelihood target estimates at n=500 and n=2000 under a misspecified-but-moment-matched full model with MCAR — a skew-normal or three-component Gaussian mixture matched in mean and covariance to the generating N4(0,Σ) law of Experiment A — using the Section S4.2 setting (B=499 centred refits, kernel bandwidths 1,2,4). Compare rejection rates against those for the CIM-failure mechanisms B1 and B2 at γ=0.5 and γ=1 in Table 2. If the moment-matched misspecified model is rejected at rates comparable to the γ=0.5 CIM-failure scenarios (5.2–8.2%), while the estimator bias under that CIM failure (≈0.08 for E(Y4), Table S2) exceeds its standard error, then the diagnostic cannot serve as a reliable gate for the inference the transfer theorem certifies, and non-rejection at practical sample sizes would not constitute evidence for CIM.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 13 Nov 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 2 min 30 s · 46,112 tokens read, 22,474 written
  3. Editor wrote the verdict above 24 s · 47,196 tokens read, 464 written
  4. Referee major revision: 3 major objections, 7 minor 18 s · 46,688 tokens read, 1,424 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This paper is worth reading. The core idea—completion-invariant missingness (CIM)—is a genuine relaxation of MAR that still preserves the sampling distribution of every observed-data statistic and the semiparametric efficiency bound for identified functionals. The proof of Theorem 1 and Corollary 1 is clean and, as far as I can tell, correct. The kernel-based diagnostic is a natural extension, and the authors are upfront about its limitations: it tests the joint null of CIM plus correct full-model specification, and rejection can come from either. The simulations bear this out cleanly—the test is hypersensitive to heavy tails (t10 under MCAR rejects 100% at n=2000) and only weakly sensitive to genuine CIM failure at moderate sample sizes (B1, γ=0.5, rejection 8.2% at n=2000). The authors acknowledge this entanglement in Section 4 and the concluding remarks, so it is not a hidden flaw. The identification results in Section 5 are a useful addition, though they require fairly strong conditions (exclusive variables, monotone selection probabilities, or tail bounds on the complete-case probability). The real-world applications are illustrative rather than definitive. The paper is methodologically sound, the code and data are archived, and the central theoretical claim is important. The main practical limitation is that the diagnostic cannot distinguish model misspecification from CIM failure, and the test's power against genuine CIM departures is modest unless the departure is strong or the sample is large. For a reader working on missing data, this paper provides a clear framework and a useful diagnostic, but the diagnostic should be interpreted with caution. I would send this to a serious referee; the theory is solid and the contribution is substantive, even if the practical tool has the acknowledged limitations.

The strongest objection

The paper's central claim is correct: I verified the key proof steps of Theorem 1 — the counterpart is a probability law preserving P_O; CIM ⟺ counterpart is MAR (the (ii)⇒(i) direction via F = Σ_r F_r and E[F·F_r] = 0); uniqueness of the MAR law fixing both marginals; and equality (7) of attainable pairs. Corollary 1's tangent-space argument is valid: since CIM and MAR induce the same set of observed-data laws, the canonical gradient of any pathwise differentiable ψ identified via (8)–(9) is the same, so the efficiency bound transfers. Theorem 2's completion characterization is also clean. | The genuinely load-bearing limitation, which the reader and authors identify, is exact full-model specification, and the paper's own numbers show it dominates the diagnostic's practical behaviour. Under MCAR with a t10 law, CIM holds, yet rejection is 62.0% (n=500) and 100% (n=2000) (Table 2); under a moment-matched Gaussian mixture, also misspecified, rejection stays 4.2–4.8%. Genuine CIM failure at γ=0.5 (B1) is detected only 5.2–8.2% at n=500–2000. The test is thus hypersensitive to heavy tails, blind to second-order-matched misspecification, and only weakly sensitive to the departures it exists to flag; non-rejection is weak evidence for CIM and rejection is ambiguous as to cause. On the inference side, Table S2 shows direct-likelihood estimators under CIM failure are inconsistent for complete-data targets (B1, γ=1: bias 0.17 for E(Y4), 2.4% coverage at n=2000), so the transfer — exactly true for every observed-data statistic's sampling distribution — does not make the fitted parametric complete-data estimates reliable, which additionally requires CIM and correct specification.

I read the paper in good faith as a theoretical contribution: Theorem 1 (CIM ⟺ MAR counterpart with both marginals preserved and uniqueness), Corollary 1 (efficiency bound transfer under identification), and Theorem 2 (completion characterization) are the load-bearing results. I verified the delicate part of the Theorem 1 proof — the (ii)⇒(i) direction, where E(F(Y)F_r(Y)) = 0 for each r requires the indicator-insertion and σ(Y(r))-measurability of F on {π_r > 0}, both of which hold — and the Corollary 1 tangent-space argument, which is valid because the induced observed-data models coincide. The sampling-distribution transfer to MAR-based procedures does not require correct specification or parameter distinctness, exactly as claimed. The weakest point, for the test and for complete-data inference, is correct full-model specification. The reader identified this; the authors acknowledge it in Sections 4 and 8; and the paper's own Tables 2 and S2 quantify it sharply. My contribution is to sharpen that acknowledgement into a precise claim: the entanglement is not merely philosophical. At practical sample sizes the test's power against moderate genuine CIM failure (5–8% for B1, γ=0.5) is comparable to its false-rejection rate under heavy-tailed MCAR (62–100% for t10), so rejection and non-rejection are both unreliable as CIM diagnostics unless the full model is effectively certain. Since this limitation is acknowledged, honestly reported, and does not touch the central transfer theorem or the efficiency result, the ACCEPT verdict stands. I recommend no change; the concern argues for care in interpretation rather than modification of the claims.

The referee's report

The paper makes a clear theoretical contribution by establishing the exact relationship between CIM and MAR, showing that the weakening from MAR to CIM costs nothing for inference on identified functionals of the complete-data marginal. The efficiency invariance result (Corollary 1) is important and cleanly proved. The kernel test is a principled diagnostic, and the identification results address a key practical concern. The paper is well-written and the proofs appear sound. The main limitation, which the authors acknowledge, is that the test assesses the joint null of correct full model and CIM, so rejection can be driven by model misspecification alone. The simulation results confirm this entanglement. The paper is a significant contribution to the missing data literature.

Pith's reading of the paper (a machine's summary, not the authors')

When data are incomplete, most analyses assume the missing values are missing at random (MAR), meaning the chance of missingness depends only on observed information. This assumption is often questionable, especially when different variables are missing for different people. This paper studies a weaker condition called completion-invariant missingness (CIM). Under CIM, the sum of the pattern-specific missingness probabilities equals one, rather than each pattern being individually MAR. The authors show that if CIM holds, any statistical procedure designed for MAR, such as doubly robust estimators, remains valid and achieves the same efficiency. They also characterize CIM through a completion property: filling in missing values from their conditional distribution under the true model preserves the complete-data distribution. To check whether CIM is compatible with a specified model, they develop a kernel-based test that compares the completed distribution with the model distribution. The test accounts for parameter estimation and uses bootstrap calibration. The paper also gives conditions under which CIM is identifiable from observed data for Gaussian, Student t, and mixture models. Simulations and applications to clinical data illustrate the methods.

Core claim. Every law satisfying completion-invariant missingness has a unique missing-at-random counterpart sharing its observed-data and complete-data distributions, so MAR-based procedures for identified functionals remain valid with the same semiparametric efficiency bound. This follows from Theorem 1 and Corollary 1.

The premise it rests on. The assumption that the full model is correctly specified is load-bearing for the test and for complete-data inference. If the parametric family used for fitting does not contain the true complete-data distribution, the direct-likelihood estimator can be inconsistent and rejection of the joint null can merely indicate model misspecification. The authors acknowledge this in Section 4 and the concluding remarks.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

The missing-at-random assumption underlies most analyses of incomplete data. Under nonmonotone missingness, it is especially stringent, imposing pattern-specific restrictions with different conditioning sets. A strictly weaker normalization condition is known to justify maximizing the observed-data likelihood. We show that every law satisfying this condition has a unique missing-at-random counterpart sharing its observed-data and complete-data distributions. Observed-data procedures developed under missing at random for identified functionals, including doubly robust estimators, therefore remain valid, with the same semiparametric efficiency bound. We then prove that the condition holds if and only if filling in the missing values from their conditional distribution under the true full model preserves the complete-data distribution, hence the name completion-invariant missingness. We use a characteristic kernel to compare the completed distribution with the fitted model and assess the joint compatibility of the specified full model and completion-invariant missingness. We derive the asymptotic null distribution with estimated parameters and establish bootstrap and multiplier calibration. Since a parametric full model alone need not identify completion-invariant missingness, we give identification conditions for Gaussian, Student t and finite mixture families. Simulations assess sensitivity to departures from each assumption, and applications to clinical laboratory measurements and childhood growth illustrate the diagnostic.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-13

The authors' main claim, as the classifier picked it out 0.69

“Observed-data procedures developed under missing at random for identified functionals, including doubly robust estimators, therefore remain valid, with the same semiparametric efficiency bound.”

How the main claim is stated: asserted 0.99

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.81 · human participants 0.77 · controlled comparison 0.03 · randomized assignment 0.02 · states quantitative result 0.20 · compares with prior results 0.71 · mentions code or data availability 0.02

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
The missing-at-random assumption underlies most analyses of incomplete data. background 1.00 0.04 0.02 asserted 1.00
Under nonmonotone missingness, it is especially stringent, imposing pattern-specific restrictions with different conditioning sets. background 0.98 0.07 0.03 asserted 1.00
A strictly weaker normalization condition is known to justify maximizing the observed-data likelihood. background 0.98 0.05 0.05 asserted 1.00
We show that every law satisfying this condition has a unique missing-at-random counterpart sharing its observed-data and complete-data distributions. result 0.90 0.94 0.44 asserted 1.00
Observed-data procedures developed under missing at random for identified functionals, including doubly robust estimators, therefore remain valid, with the same semiparametric efficiency bound. interpretation 0.77 0.51 0.35 asserted 0.99
We then prove that the condition holds if and only if filling in the missing values from their conditional distribution under the true full model preserves the complete-data distribution, hence the name completion-invariant missingness. result 0.85 0.92 0.28 asserted 1.00
We use a characteristic kernel to compare the completed distribution with the fitted model and assess the joint compatibility of the specified full model and completion-invariant missingness. method 0.92 0.93 0.05 asserted 1.00
We derive the asymptotic null distribution with estimated parameters and establish bootstrap and multiplier calibration. method 0.84 0.95 0.04 asserted 1.00
Since a parametric full model alone need not identify completion-invariant missingness, we give identification conditions for Gaussian, Student $t$ and finite mixture families. contribution 0.87 0.91 0.06 asserted 0.94
Simulations assess sensitivity to departures from each assumption, and applications to clinical laboratory measurements and childhood growth illustrate the diagnostic. contribution 0.71 0.95 0.05 asserted 0.99

What kind of paper the classifier reports this is: theoretical result 0.95 · evidence from proof or derivation 1.00

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-13: 1 request, 7,481 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (1)
Figure 1 of the paper
Figure 1. Counterpart map and observed- and complete-data marginals. The horizontal arrows labelled

Figures are the authors', reproduced from the paper.

References (41)
  1. [1] + ϕ(v + 1)}/2 and f1(v) = 3ϕ(v− 1)/4 + ϕ(v + 1)/4. Their ratio is 1/2 + expit(2v). With complete-or-missing observations, selection π1(v) = 1/4 + expit(2v)/2 under f0 gives the same observed law as MCAR with probability 1 /2 under f1. Hence bounded monotone selection alone doe... on Pith
  2. [11] f11(x1, x2;θ) = 1− ρ10f10(x1 |R= 10)f10(x1;θ) − ρ01f01(x2 |R= 01)f01(x2;θ) −ρ 00.(S11) Figure S1 compares the right side, reconstructed from incomplete patterns, with the left side estimated from complete cases. Their difference, multiplied by the fitted full density, locates ... on Pith
  3. [14] doi:10.1080/01621459.2020.1862669. Corrected author version: arXiv:1909.01848v4. W. Miao, P. Ding, and Z. Geng. Identifiability of normal and normal mixture models with nonig- norable missing data.Journal of the American Statistical Association, 111(516):1673–1683, on PithPith's reviewarXiv
  4. [15] doi:10.1080/01621459.2015.1105808. W. Miao, L. Liu, E. Tchetgen Tchetgen, and Z. Geng. Identification, doubly robust estimation, and semiparametric efficiency theory of nonignorable missing data with a shadow variable, on Pithdoi:10.1080/01621459.2015.1105808
  5. [16] URL https://arxiv.org/abs/1509.02556v3. arXiv:1509.02556, version on PithPith's reviewarXiv
  6. [23] doi:10.1080/01621459.2020.1796358. J. L. Schafer.Analysis of Incomplete Multivariate Data, volume 72 ofMonographs on Statistics and Applied Probability. Chapman & Hall, London, on Pithdoi:10.1080/01621459.2020.1796358
  7. [24] doi:10.1201/9781439821862. S. R. Seaman, J. C. Galati, D. Jackson, and J. B. Carlin. What is meant by “missing at random”? Statistical Science, 28(2):257–268, on Pithdoi:10.1201/9781439821862
  8. [25] doi:10.1214/13-STS415. B. K. Sriperumbudur, A. Gretton, K. Fukumizu, B. Sch¨ olkopf, and G. R. G. Lanckriet. Hilbert space embeddings and metrics on probability measures.Journal of Machine Learning Research, 11(50): 1517–1561, on Pithdoi:10.1214/13-sts415
  9. [28] doi:10.5705/ss.202016.0325. A. W. van der Vaart and J. A. Wellner.Weak Convergence and Empirical Processes: With Applications to Statistics. Springer, New York, on Pithdoi:10.5705/ss.202016.0325
  10. [31] = 1√n nXi=1 IFθ(Oi) +o p(1),EIF θ = 0,E∥IF θ ∥2 <∞. Let N (b)i be the number of times row i is selected in bootstrap replicate b of size n drawn uniformly with replacement from the observed rows. The refitted estimator obeys √n(bθ(b) − bθ) = 1√n nXi=1(N (b)i −1) IF θ(Oi) +o P ... on Pith
  11. [32] +O p(∥bθ−θ 0∥2). The derivative envelope and finite parameter dimension yield a law of large numbers in the corresponding finite product of Hilbert spaces. The empirical derivative therefore converges in operator norm to ˙gθ = E ˙ξ(O; θ0). Substitution of the estimator’s linea... on Pith
  12. [33] If A <0, the ratio tends to zero at both ends of R, which is incompatible with a positive nonconstant monotone function. This proves the Gaussian case. For any Student t density, its ratio at t and −t tends to one as t→ ∞, irrespective of its location, scale, and positive fini... on Pith
  13. [34] +E ∂ξ CCH (O;θ) ∂θ θ=θ0 IFθ(O). Each bootstrap refit recomputes both conditional means and uses its own centred statistic n∥bgCC,(b)H − bgCCH ∥2H and critical value. Rows observing only C contribute zero; normalization uses the original sample sizen. 28 For gk, conditional int... on Pith
  14. [36] = 0.7. Thus PO is exactly the observed-data law generated by PY,θ ′ and an MCAR mechanism with pattern probabilities (0 .1, 0.1, 0.1, 0.7). Under PY,R, π00 = 0 .7, while π11, π10, and π01 are the first three displayed probabilities. Hence F(0,0) = 0.2 + 0.2 √ 2 + 0.7−1 = 0.2 √... on Pith
  15. [37] S4.3 Target inference in Experiment B Table S2 summarizes estimation of the two targets in the 18 settings of Experiment B, using the same simulated datasets and bootstrap refits as the main-paper diagnostic. 40 Table S2: Experiment B: target inference across all 18 settings. ... on Pith
  16. [100] The regression and free-covariance parameter counts are 30, 51 and 78 for the four-laboratory linear, six-laboratory linear and four-laboratory fixed-knot specifications. The model forCremains unspecified throughout. Table S5 lists the 12 observed four-laboratory patterns; the... on Pith
  17. [216] in order 11 , 10, 01, 00; Table S3 gives original-scale summaries. Rows with both laboratories missing have observed likelihood one and zero centred embedding contribution but remain innand in the resampling. Table S3: SUPPORT2 summaries by laboratory observation pattern. Mean... on Pith
  18. [1974] E. J. Tchetgen Tchetgen, L. Wang, and B. Sun. Discrete choice models for nonmonotone non- ignorable missing data: Identification and inference.Statistica Sinica, 28(4):2069–2088, on Pith
  19. [1976] doi:10.1093/biomet/63.3.581. H. Sang, J. K. Kim, and D. Lee. Semiparametric fractional imputation using gaussian mixture models for handling multivariate missing data.Journal of the American Statistical Association, 117(538): 654–663, on Pithdoi:10.1093/biomet/63.3.581
  20. [1988] doi:10.1080/01621459.1988.10478722. R. J. A. Little. Pattern-mixture models for multivariate incomplete data.Journal of the American Statistical Association, 88(421):125–134, on Pithdoi:10.1080/01621459.1988.10478722
show 21 more
  1. [1990] doi:10.1214/aop/1176990862. A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Sch¨ olkopf, and A. Smola. A kernel two-sample test. Journal of Machine Learning Research, 13(25):723–773, on Pithdoi:10.1214/aop/1176990862
  2. [1993] doi:10.1080/01621459.1993.10594302. 14 G. Lu and J. B. Copas. Missing at random, likelihood ignorability and model completeness.The Annals of Statistics, 32(2):754–765, on Pithdoi:10.1080/01621459.1993.10594302
  3. [1994] doi:10.1080/01621459.1994.10476818. D. B. Rubin. Inference and missing data.Biometrika, 63(3):581–592, on Pithdoi:10.1080/01621459.1994.10476818
  4. [1995] doi:10.7326/0003-4819-122-3-199502010-00007. I. Kojadinovic and J. Yan. Goodness-of-fit testing based on a weighted bootstrap: A fast large-sample alternative to the parametric bootstrap.The Canadian Journal of Statistics, 40(3):480–501, on Pithdoi:10.7326/0003-4819-122-3-199502010-00007
  5. [1996] doi:10.1007/978-1-4757-2545-2. Vanderbilt Biostatistics. SUPPORT2 public data and variable documentation. https://hbiostat. org/data/repo/supportdesc, on Pithdoi:10.1007/978-1-4757-2545-2
  6. [1997] doi:10.1002/(SICI)1097-0258(19970115)16:1 <39::AID- SIM535>3.0.CO;2-D. J. M. Robins, A. Rotnitzky, and L. P. Zhao. Estimation of regression coefficients when some regressors are not always observed.Journal of the American Statistical Association, 89(427):846–866, on Pithdoi:10.1002/(sici)1097-0258(19970115)16:1
  7. [2004] doi:10.1214/009053604000000166. D. Malinsky, I. Shpitser, and E. J. Tchetgen Tchetgen. Semiparametric inference for nonmonotone missing-not-at-random data: The no self-censoring model.Journal of the American Statistical Association, 117(539):1415–1423, on Pithdoi:10.1214/009053604000000166
  8. [2005] doi:10.1111/j.1541-0420.2005.00377.x. P. J. Bickel, C. A. J. Klaassen, Y. Ritov, and J. A. Wellner.Efficient and Adaptive Estimation for Semiparametric Models. Johns Hopkins University Press, Baltimore, on Pithdoi:10.1111/j.1541-0420.2005.00377.x
  9. [2006] doi:10.1191/0962280206sm448oa. J. M. Robins and R. D. Gill. Non-response models for the analysis of non-monotone ignorable missing data.Statistics in Medicine, 16:39–56, on Pithdoi:10.1191/0962280206sm448oa
  10. [2008] doi:10.1111/j.1467-9868.2007.00640.x. K. Morikawa, Y. Kano, and S. Sugasawa. Reproducibility materials for On Ignorability Beyond Missing at Random: Completion-Invariant Missingness for Nonmonotone Missing Data, on Pithdoi:10.1111/j.1467-9868.2007.00640.x
  11. [2010] doi:10.1371/journal.pone.0009155. A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum likelihood from incomplete data via the EM algorithm (with discussion).Journal of the Royal Statistical Society: Series B (Methodological), 39 (1):1–38, on Pithdoi:10.1371/journal.pone.0009155
  12. [2011] doi:10.1093/biomet/asq073. W. A. Knaus, F. E. Harrell, J. Lynn, et al. The SUPPORT prognostic model: Objective estimates of survival for seriously ill hospitalized adults.Annals of Internal Medicine, 122(3):191–203, on Pithdoi:10.1093/biomet/asq073
  13. [2013] doi:10.4187/respcare.02786. M. L. A. De Kroon, C. M. Renders, J. P. Van Wouwe, S. Van Buuren, and R. A. Hirasing. The Terneuzen Birth Cohort: BMI changes between 2 and 6 years correlate strongest with adult overweight.PLoS ONE, 5(2):e9155, on Pithdoi:10.4187/respcare.02786
  14. [2016] doi:10.1080/00031305.2016.1164756. R. D. Gill, M. J. van der Laan, and J. M. Robins. Coarsening at random: Characterizations, conjectures, and counter-examples. In D. Y. Lin and T. R. Fleming, editors,Proceedings of the First Seattle Symposium in Biostatistics: Survival Analys... on Pithdoi:10.1080/00031305.2016.1164756
  15. [2018] doi:10.1080/01621459.2016.1256814. 15 R. Sundberg. Maximum likelihood theory for incomplete data from an exponential family.Scandinavian Journal of Statistics, 1(2):49–58, on Pithdoi:10.1080/01621459.2016.1256814
  16. [2019] doi:10.1080/07350015.2017.1302879. Y.-C. Chen. Pattern graphs: A graphical approach to nonmonotone missing data.The Annals of Statistics, 50(1):129–146, on Pithdoi:10.1080/07350015.2017.1302879
  17. [2020] doi:10.1093/biomet/asz061. C. Breunig. Testing missing at random using instrumental variables.Journal of Business & Economic Statistics, 37(2):223–234, on Pithdoi:10.1093/biomet/asz061
  18. [2022] doi:10.1214/21-AOS2094. M. D. Davis, B. K. Walsh, S. E. Sittig, and R. D. Restrepo. AARC clinical practice guide- line: Blood gas analysis and hemoximetry: 2013.Respiratory Care, 58(10):1694–1703, on Pithdoi:10.1214/21-aos2094
  19. [2023] doi:10.1111/biom.13666. 16 Supplementary Material This supplement accompanies the main paper and uses its notation; “full model” denotes a model for the complete-data distribution. Because Kano (2014) is written in Japanese, every result taken from that paper is restated here ... on Pithdoi:10.1111/biom.13666
  20. [2026] URL https://doi.org/10.5281/zenodo.24096545. Version 1.0.1. Software. R. F. Potthoff, G. E. Tudor, K. S. Pieper, and V. Hasselblad. Can one assess whether missing data are missing at random in medical studies?Statistical Methods in Medical Research, 15(3):213–234, on Pithdoi:10.5281/zenodo.24096545
  21. [2255] and 223 .7 ( n = 3006), respectively. These are model-based summaries after rejection of joint compatibility, not bias-corrected inference for the unknown complete-data means. S5.4 Conditional specifications and observation patterns The conditional cohort consists of the 3513 ... on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment