Pith. sign in

REVIEW 3 major objections 5 minor 18 references

Valid causal inference from private synthetic data needs causal moments preserved and privacy noise modeled, not generic fidelity alone.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 12:42 UTC pith:5YC5V2K6

load-bearing objection Solid methods paper: causal moments as the DP workload, honest RMSE/coverage tradeoff, and a real theory-to-experiment gap on re-fit DR. the 3 major comments →

arxiv 2607.08122 v1 pith:5YC5V2K6 submitted 2026-07-09 cs.LG

Workload-Preserving Differentially Private Synthetic Data for Causal Inference via Maximum-Entropy Calibration

Amir Asiaee , Kaveh Aryan This is my paper
classification cs.LG
keywords differential privacysynthetic datacausal inferenceaverage treatment effectdoubly robust estimationmaximum-entropy calibrationnoise-aware multiple imputationworkload design
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that differentially private synthetic tables built from generic low-dimensional queries can look distributionally faithful yet still break causal estimates such as the average treatment effect, because those estimands depend on treatment-arm balance and outcome moments that generic marginals need not keep. The authors define causal workloads: a fixed set of queries that measure the orthogonal moments used by doubly robust estimators, release them once under differential privacy, and either plug the noisy moments into stable estimators or reconstruct a maximum-entropy synthetic distribution for reuse. Theory decomposes ATE error into sampling, privacy, workload-approximation, Monte Carlo, and calibration terms, and the same release can support ATE, ATT, and subgroup analyses without further privacy spend. Empirically, pairing causal workloads with noise-aware multiple imputation is the only private pipeline that achieves near-nominal confidence-interval coverage at strict privacy budgets across the main benchmarks, while naive analysis of synthetic rows undercovers severely. The practical lesson is a tradeoff: generic fidelity can win on point accuracy when privacy is loose, but valid causal inference requires preserving the causal moments and propagating the privacy noise rather than treating synthetic rows as real.

Core claim

A single differentially private release of causal-workload moments—treatment-arm feature masses and outcome-weighted feature moments—supports both direct stable moment-map estimation and maximum-entropy synthetic data for ATE (and related weighted effects), with ATE error controlled by an explicit five-term decomposition; noise-aware multiple imputation on that release is what restores calibrated uncertainty, whereas treating synthetic rows as real data produces invalid intervals.

What carries the argument

Causal workloads: DP queries built from the orthogonal-score moments of doubly robust estimators, released once, then either used by Lest-stable moment maps or reconstructed by maximum-entropy calibration into synthetic records, with ATE error decomposed into sampling, privacy, workload-approximation, Monte Carlo, and calibration terms, and confidence intervals formed by noise-aware multiple imputation (NA+MI).

Load-bearing premise

The chosen finite feature map for covariates is rich enough that the retained first-order causal moments nearly identify the target, so leftover approximation bias does not dominate once privacy noise shrinks.

What would settle it

On a benchmark with known ATE, replace the causal-moment workload by a generic marginal workload of equal privacy cost and analyze both with the same noise-aware intervals: if the causal release no longer uniquely recovers near-nominal coverage at strict privacy budgets while matching or beating generic methods on coverage-calibrated error, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • One DP synthetic table built from causal moments can answer ATE, ATT, and subgroup questions with no extra privacy cost, which direct private ATE releases cannot do.
  • At strict privacy budgets, analysts who need valid intervals should prefer causal workloads plus noise-aware multiple imputation over generic synthetic data with naive inference.
  • As privacy budgets relax, generic workloads may still win on point RMSE, so release design should follow whether the goal is calibrated causal uncertainty or broad distributional fidelity.
  • SNR thresholding and an explicit calibration gap let practitioners trade controllable solver residual against approximation bias instead of treating reconstruction as exact.
  • Adaptive selection of which causal feature groups to measure (CAUSAL-AIM) is an operating-point tool: helpful on some datasets and budgets, harmful when each adaptive round spends more budget than the targeting gain.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Any select–measure–reconstruct private synthesizer, not only the graphical-model family used here, could be retargeted by swapping the measured query set for orthogonal-score moments of the intended causal class.
  • The same moment-first design likely extends to other pathwise-differentiable targets (e.g., policy values or mediation functionals) once their influence-function moments replace the ATE blocks.
  • Coverage falling as privacy relaxes is a general warning for private synthetic inference: when noise shrinks, unmodeled approximation bias becomes visible, so bias diagnostics belong in the release, not only in the paper’s appendix.
  • Hybrid budgets that split spend between causal moments and a few generic marginals may be the practical default when both point accuracy and calibrated coverage matter.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes causal workloads for differentially private synthetic data: query sets built from treatment-arm feature masses and outcome-feature moments that appear in orthogonal (doubly robust) scores. Noisy answers can be used either as direct moment-map estimators or reconstructed via maximum-entropy calibration (Private-PGM style) into reusable synthetic microdata. Theory gives Lipschitz stability for projected ridge and clipped cell IPW/AIPW maps, Gaussian moment accuracy, and a five-term ATE error decomposition (sampling, privacy, workload approximation, Monte Carlo, calibration). The authors introduce CAUSAL-AIM (adaptive selection) and NA+MI (noise-aware multiple imputation) for intervals, and argue that one release supports ATE, ATT, and subgroups without extra privacy spend. Empirically, on IHDP, Twins, ACIC, and LaLonde, Causal+NA+MI is the only private method with near-nominal 95% coverage at ε≤1 (99.8–100%), while naive synthetic analyses stay ≤35.2%; generic workloads (MST/AIM) often win point RMSE as privacy relaxes.

Significance. If the claims hold, the paper cleanly articulates a useful design principle: DP synthetic data for causal inference should preserve the moments of the orthogonal score, not only generic marginals, and must propagate DP noise into uncertainty. The multi-estimand reuse property is a genuine structural advantage over one-shot DP ATE estimators. Strengths include explicit finite-sample decompositions with proofs in Appendix E, honest reporting of the RMSE–coverage tradeoff, ablations (workload dimension, MI draws, n_syn, overlap), multi-estimand reuse evidence, and a public reproducibility repository. The work is a solid contribution to DP synthesis and private causal inference even if some theory-to-pipeline gaps remain.

major comments (3)
  1. [§5 Theorem 3; §7.1; App. L] Theorem 3 and the stability results (Theorem 1, Proposition 1, Corollary 1) bound estimators that are Lipschitz functions of the released workload answer q (ridge plug-in; clipped cell IPW/AIPW). The experimental pipeline instead runs standard DR/AIPW with nuisances re-fit on sampled synthetic rows (Section 7.1, Algorithm 4). Appendix L acknowledges that the theorem does not fully analyze every fitted-nuisance routine on synthetic microdata. This is load-bearing for the central coverage claim (Causal+NA+MI near-nominal at ε≤1): NA+MI correctly re-samples the DP moments, but each synthetic-table DR estimate still treats nuisance-fitting error as ordinary sampling variance. Please either (i) prove that re-fitted DR on max-ent samples inherits the moment-map guarantee under stated conditions, (ii) report the direct q-route estimators as the primary theory-aligned method with matching covera
  2. [§7–8; App. F; App. L] The coverage–privacy paradox (coverage falling as ε grows on ACIC; Appendix L) is attributed to workload approximation bias Approx(φ;S) once DP noise shrinks. Appendix F gives a bias-aware NA+MI interval with a coverage guarantee under a conservative diagnostic, but the main tables and figures report uncorrected NA+MI. Because approximation bias is the paper’s own weakest modeling assumption (finite φ / L=5 bins; Proposition 2), the bias-aware correction or at least the diagnostic ratio Approx̂(φ)/σ_NA+MI should appear in the main experimental results, not only as an appendix proposal, so readers can see when the method is privacy-dominated versus approximation-dominated.
  3. [Prop. 1–2; App. A; §7] Proposition 2 and the default 4p workload establish first-order sufficiency only for nuisances projected onto span(φ). Experiments use concatenated one-hot / quantile-bin features and then re-fit DR on continuous-style synthetic tables after reconstruction (Appendix A). For concatenated features the Gram is not diagonal, and cell-level IPW stability (Proposition 1) does not automatically transfer. Please clarify which experimental configurations are literally covered by Proposition 1 versus which rely on post-reconstruction DR as a heuristic, and add a controlled check where the estimator is exactly the moment-map IPW/AIPW of Proposition 1 on the same release.
minor comments (5)
  1. [§7.2 Table 1] Table 1 and Figures 2–4 make the RMSE–coverage tradeoff clear; consider adding a short main-text sentence that Causal+NA+MI is not recommended when point RMSE alone is the goal at ε≥2, to match the practical takeaway in §7.2.
  2. [Figure 5; App. K] CAUSAL-AIM helps substantially on ACIC (Figure 5) but fails to recover fixed-workload calibration on IHDP (Appendix K). The operating-point message is stated; a one-line decision rule in the main text (when to prefer fixed vs adaptive) would help practitioners.
  3. [App. K Table 3] The direct DP ATE comparison (Table 3) uses output-perturbation proxies weaker than the cited PrivATE / OA methods; the caveat is present but easy to miss. Flag more prominently that this is not a head-to-head with the published estimators.
  4. [§4.2] Notation switches between eq, q̃, and similar for the noisy release; a single symbol table early in §4 would reduce friction.
  5. [Figure 13] Figure 13 (fidelity vs causal utility) is a strong conceptual figure; consider promoting a compact version to the main text if space allows.

Circularity Check

0 steps flagged

No significant circularity: ATE bounds and coverage claims rest on standard DP sensitivity, Lipschitz stability, and external known-ATE benchmarks, not on quantities defined from the method’s own fit.

full rationale

The derivation chain is: (i) define a causal workload as the first-order arm and outcome-feature moments used by projected orthogonal scores (Def. 1, Prop. 2); (ii) release them with the Gaussian mechanism and bound coordinate error by sensitivity (Thm. 2); (iii) convert moment error into ATE error for estimators that are Lipschitz in those moments (Thm. 1, Prop. 1, Cor. 1); (iv) compose sampling, privacy, approximation, Monte Carlo, and calibration terms for the synthetic route (Thm. 3); (v) propagate DP noise via NA+MI posterior draws of the measured moments (Alg. 4). None of these steps defines the target ATE in terms of the estimator’s own fitted output, nor does the paper fit a free parameter on a subset of the evaluation and then call a closely related quantity a prediction. Empirical claims are checked against external semi-synthetic and experimental benchmarks (IHDP, Twins, ACIC, LaLonde, ACS) with known ground-truth ATE, not against self-generated labels. Self-citations (CausalWrap, reward-guided generation, multi-national HIV synthesis) appear only as related work contrasting prior-injection generators with measurement design; they are not invoked as uniqueness theorems or as the sole support for the main bounds. The theory–experiment gap noted by the skeptic (moment-map Lipschitz bounds vs. DR with nuisances re-fit on synthetic rows) is a coverage/correctness concern, not a circular reduction of a claimed prediction to its inputs. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The method rests on standard causal identification and DP measurement facts, plus modeling choices that define the workload and reconstruction. Free parameters are mostly discretization/MI/solver knobs that affect Approx and CalGap. Invented entities are algorithmic constructs (causal workload, Causal-AIM, NA+MI), operationally defined rather than postulated physical objects; their independent handle is empirical coverage/RMSE on public benchmarks.

free parameters (6)
  • feature map ϕ / quantile bins L
    Default L=5 quantile bins (and concatenated one-hot encoding) sets the workload dimension and approximation bias; not derived from data-generating process, chosen by practice and ablated.
  • MI draws M
    Default M=20 imputation draws for NA+MI; ablation shows small M often suffices, but the reported intervals depend on this choice.
  • synthetic sample size nsyn
    Default nsyn=10n (ablation says nsyn=n often enough); affects Monte Carlo term and compute.
  • SNR threshold τ_SNR and calibration ridge α_cal
    Optional filters/regularizers controlling which noisy moments are matched; defaults off in main pipeline but recommended as safeguards.
  • outcome clip bound B and propensity clip
    B=5 after public standardization and DR propensity clip [0.02,0.98] bound sensitivity and stability; implementation choices that enter constants.
  • CAUSAL-AIM rounds K and score/measurement budget split
    Adaptive selection consumes privacy per round; operating point is data-dependent and hand-chosen.
axioms (6)
  • domain assumption Unconfoundedness and positivity: (Y(1),Y(0)) ⊥ T | X and e(x)∈[η,1−η] a.s. (Assumption 1).
    Required for ATE identification as E[m1(X)−m0(X)]; standard causal assumption, not proved here.
  • domain assumption Bounded outcomes and features: Y clipped to [−B,B], ∥ϕ(X)∥2≤ϕmax, enabling finite ℓ2 sensitivity of the workload.
    Used for Gaussian mechanism scale and Lipschitz constants in Theorems 1–3.
  • standard math Estimator Lipschitz/stability of moment-map ATE estimators (ridge plug-in; clipped IPW/AIPW) w.r.t. workload coordinates.
    Proved under conditioning/clipping assumptions in Thm. 1 / Prop. 1; converts moment error into ATE error.
  • standard math Gaussian mechanism (ε,δ)-DP accuracy and post-processing immunity for reconstruction/sampling.
    Standard DP toolkit; Thm. 2 is a union-bound Gaussian tail application.
  • ad hoc to paper First-order sufficiency of the 4p causal workload for projected orthogonal-score targets in the span of ϕ (Prop. 2).
    Principled design claim for this paper’s workload; not a minimax lower bound, and Gram moments are optional extras.
  • domain assumption Max-entropy / Private-PGM-style reconstruction approximately matches measured moments up to a controllable CalGap.
    Reconstruction family is fixed to AIM/Private-PGM-like solvers; exact matching may be infeasible under noise.
invented entities (3)
  • causal workload Qϕ independent evidence
    purpose: DP query set of arm-specific feature masses and outcome-feature moments intended to preserve orthogonal-score structure.
    Defined operationally in Def. 1; independent evidence is empirical ATE/coverage behavior versus generic workloads, not an external physical prediction.
  • CAUSAL-AIM independent evidence
    purpose: Adaptive private selection of feature groups by estimated ATE-error reduction.
    Algorithmic variant of AIM with causal utility; performance is operating-point dependent across datasets.
  • NA+MI (noise-aware multiple imputation for causal DP synth) independent evidence
    purpose: Propagate known Gaussian DP noise through max-entropy reconstruction and Rubin’s rules for causal CIs.
    Specialization of existing noise-aware MI ideas to causal workloads; calibrated coverage is the external check.

pith-pipeline@v1.1.0-grok45 · 34821 in / 4081 out tokens · 45606 ms · 2026-07-10T12:42:38.642428+00:00 · methodology

0 comments
read the original abstract

Workload-based differentially private (DP) synthetic data methods privately measure aggregate queries and post-process the noisy answers into synthetic records. Generic workloads can achieve strong distributional fidelity, but causal estimands such as the average treatment effect (ATE) depend on treatment-arm balance and outcome moments that generic marginals need not preserve. We propose causal workloads: DP query sets designed around the orthogonal moments used by doubly robust causal estimators. The released workload can be used directly by stable moment-map estimators or reconstructed by maximum-entropy calibration into reusable synthetic data; our theory decomposes ATE error into sampling, privacy, workload-approximation, Monte Carlo, and calibration terms. We also introduce Causal-AIM, an adaptive workload selector, and a noise-aware multiple-imputation (NA+MI) procedure for confidence intervals from DP synthetic data. Because the workload is released once, the same DP synthetic table can support ATE, ATT, and subgroup analyses without additional privacy spending. Empirically, causal workloads are most useful at strict privacy budgets and for calibrated uncertainty, while generic workloads often retain an advantage for point RMSE as privacy relaxes. The broader lesson is a tradeoff: distributional fidelity can help point accuracy, but valid causal inference requires preserving causal moments and propagating DP noise rather than treating synthetic rows as real.

Figures

Figures reproduced from arXiv: 2607.08122 by Amir Asiaee, Kaveh Aryan.

Figure 1
Figure 1. Figure 1: End-to-end pipeline. The data holder chooses the fea [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: ATE RMSE versus privacy budget (ε) on four bench￾marks (nrep = 500, δ = 1/n2 ). MST + naive DR leads on RMSE at ε ≥ 2; at ε = 0.5, Causal + NA+MI matches or beats it on half the benchmarks, and on IHDP at ε = 1. The value of causal workload methods lies chiefly in enabling calibrated coverage (Fig￾ure 3). 1000, ε = 0.5), Causal + NA+MI has coverage 1.00 while MST + naive DR has coverage 0.07, and NA+MI als… view at source ↗
Figure 3
Figure 3. Figure 3: Empirical 95% CI coverage across privacy budgets and [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: ATE RMSE of CAUSAL-AIM (adaptive) versus a fixed causal workload on ACIC DGP 7 (nrep = 100). The left panel overlays all privacy budgets: horizontal lines mark the fixed work￾load and curves show CAUSAL-AIM as a function of adaptive rounds K; the right panel shows the most frequently selected fea￾tures. At ε ≥ 1, CAUSAL-AIM reduces RMSE by 78−96% over the fixed workload at K ≤ 3 (53−91% at K=5). At ε = 0.5… view at source ↗
Figure 6
Figure 6. Figure 6: Benchmark bias comparison: Mean absolute ATE error ( [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: ACS study: ATE RMSE heatmaps on the ACS semi-synthetic study ( [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: ACS study: Empirical 95% CI coverage on the same ACS grid as Figure 7 ( [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Workload dimension on IHDP at ε = 1 (nrep = 200), with and without SNR thresholding (τSNR = 3). Without thresholding, RMSE decreases with p in the tested range (p ≤ 101); with τSNR = 3, RMSE is roughly flat, with lower bias and more variance; the bias-dominated regime of Remark 2 is not reached on this dataset. K.6 CALIBRATION-RIDGE SENSITIVITY On IHDP at ε = 1 (nrep = 200), RMSE / coverage across max-entr… view at source ↗
Figure 10
Figure 10. Figure 10: MI draws ablation: empirical 95% CI coverage versus number of imputation draws [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Synthetic sample size ablation: ATE RMSE versus the ratio [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Overlap ablation on ACIC DGP 7 at ε = 1 (nrep = 200). The positivity constant η varies from 0.3 (strong overlap) to 0.01 (near violation). As overlap weakens, MST + naive DR maintains low RMSE (0.60–0.62) but poor coverage (0.21–0.28), while Causal workload + NA+MI has higher RMSE (7.5–8.3) with near-nominal coverage (0.995–1.00) at every η. This confirms that the coverage–RMSE tradeoff persists across ov… view at source ↗
Figure 13
Figure 13. Figure 13: Marginal fidelity versus causal utility across all method–dataset– [PITH_FULL_IMAGE:figures/full_fig_p028_13.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · 3 internal anchors

  1. [1]

    Claire McKay Bowen and Fang Liu

    doi: 10.1111/rssb.12158. Claire McKay Bowen and Fang Liu. Comparative study of differentially private data synthesis methods.Statistical Science, 35(2):280–307,

    work page doi:10.1111/rssb.12158

  2. [2]

    Moritz Hardt, Katrina Ligett, and Frank McSherry

    doi: 10.1214/17-BA1085. Moritz Hardt, Katrina Ligett, and Frank McSherry. A sim- ple and practical algorithm for differentially private data release. InAdvances in Neural Information Processing Systems (NeurIPS), volume 25,

    work page doi:10.1214/17-ba1085

  3. [3]

    Preprint; PMCID: PMC13015641

    doi: 10.64898/2026.03.11.26348077. Preprint; PMCID: PMC13015641. Haim Kaplan, Shachar Schnapp, and Uri Stemmer. Differ- entially private approximate quantiles. InProceedings of the 39th International Conference on Machine Learning (ICML), volume 162 ofProceedings of Machine Learning Research, pages 10751–10761. PMLR,

    work page doi:10.64898/2026.03.11.26348077 2026

  4. [4]

    Model agnostic differentially private causal inference.arXiv preprint arXiv:2505.19589,

    Christian Janos Lebeda, Mathieu Even, Aurélien Bellet, and Julie Josse. Model agnostic differentially private causal inference.arXiv preprint arXiv:2505.19589,

    work page arXiv

  5. [5]

    Christos Louizos, Uri Shalit, Joris M

    doi: 10.1038/s41467-026-74492-0. Christos Louizos, Uri Shalit, Joris M. Mooij, David Sontag, Richard Zemel, and Max Welling. Causal effect inference with deep latent-variable models. InAdvances in Neural Information Processing Systems (NeurIPS), volume 30,

    work page doi:10.1038/s41467-026-74492-0

  6. [6]

    Winning the NIST Contest: A scalable and general approach to differentially private synthetic data

    Ryan McKenna, Gerome Miklau, and Daniel Sheldon. Win- ning the NIST contest: A scalable and general approach to differentially private synthetic data.arXiv preprint arXiv:2108.04978,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    Jeffrey W

    14778/3551793.3551817. Jeffrey W. Miller and David B. Dunson. Robust Bayesian inference via coarsening.Journal of the American Sta- tistical Association, 114(527):1113–1125,

    work page arXiv

  8. [8]

    Fengshi Niu, Harsha Nori, Brian Quistorff, Rich Caruana, Donald Ngwe, and Aadharsh Kannan

    doi: 10.1080/01621459.2018.1469995. Fengshi Niu, Harsha Nori, Brian Quistorff, Rich Caruana, Donald Ngwe, and Aadharsh Kannan. Differentially pri- vate estimation of heterogeneous causal effects. InPro- ceedings of the First Conference on Causal Learning and Reasoning (CLeaR), volume 177 ofProceedings of Ma- chine Learning Research, pages 618–633. PMLR,

    work page doi:10.1080/01621459.2018.1469995 2018

  9. [9]

    Differentially Private Covariate Balancing Causal Inference

    Yuki Ohnishi and Jordan Awan. Differentially private covariate balancing causal inference.arXiv preprint arXiv:2410.14789,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Trivellore E

    doi: 10.1055/a-2385-1355. Trivellore E. Raghunathan, Jerome P. Reiter, and Donald B. Rubin. Multiple imputation for statistical disclosure limi- tation.Journal of Official Statistics, 19(1):1–16,

    work page doi:10.1055/a-2385-1355

  11. [11]

    PrivATE: Differentially private confidence in- tervals for average treatment effects.arXiv preprint arXiv:2505.21641, 2025a

    Maresa Schröder, Justin Hartenstein, and Stefan Feuer- riegel. PrivATE: Differentially private confidence in- tervals for average treatment effects.arXiv preprint arXiv:2505.21641, 2025a. Maresa Schröder, Valentyn Melnychuk, and Stefan Feuer- riegel. Differentially private learners for heterogeneous treatment effects. InInternational Conference on Learn- ...

    work page arXiv

  12. [12]

    Benchmarking Differentially Private Synthetic Data Generation Algorithms

    Yuchao Tao, Ryan McKenna, Michael Hay, Ashwin Machanavajjhala, and Gerome Miklau. Benchmarking differentially private synthetic data generation algorithms. arXiv preprint arXiv:2112.09238,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    doi: 10.1214/16-STS592. Workload-Preserving Differentially Private Synthetic Data for Causal Inference via Maximum-Entropy Calibration (Supplementary Material) Amir Asiaee1 Kaveh Aryan2 1Department of Biostatistics, Vanderbilt University Medical Center, Nashville, Tennessee, USA 2Department of Informatics, King’s College London, London, UK A FEATURE ENCOD...

    work page doi:10.1214/16-sts592

  14. [14]

    as the calibration filter, and report the corrected intervalbτNA+MI ±z 1−α/2(σ2 NA+MI + \Approx(ϕ)2)1/2. If \Approx(ϕ)≪σ NA+MI, the correction is negligible and DP noise dominates; if \Approx(ϕ)≳σ NA+MI, the analysis is approximation-dominated, which is the diagnostic signature of the coverage–privacy paradox at highε. Connection to robust Bayes.The corre...

    work page 2016

  15. [15]

    5:D syn,(ℓ) ←n syn ancestral-sampling draws fromP syn,(ℓ) 6:(bτ (ℓ),bv(ℓ))←doubly robust ATE estimate and its variance estimate onD syn,(ℓ) 7:end for 8:¯τ←M −1P ℓbτ(ℓ);W M ←M −1P ℓbv(ℓ);B M ←(M−1) −1P ℓ(bτ(ℓ) −¯τ)2 9:T M ←W M + (1 + 1/M)BM ;ν M ←(M−1) 1 + WM (1+1/M)BM 2 ▷Rubin’s rules [Rubin, 1987] Ensure:Point estimate¯τand interval¯τ±t νM ,1−α/2 √TM G N...

    work page 1987

  16. [16]

    The bias-aware variant of Appendix F replaces TM by TM + \Approx(ϕ)2

    The between-imputation component BM is what widens the interval as ε decreases: smaller ε means larger DP noise, more dispersed posterior draws q(ℓ) S , and hence more variablebτ(ℓ). The bias-aware variant of Appendix F replaces TM by TM + \Approx(ϕ)2. H EXPERIMENTAL PROTOCOL AND SUPPORTING RESULTS H.1 FEATURE CONSTRUCTION AND DEFAULTS For each dataset, c...

    work page 2021

  17. [17]

    •Synthetic sample size: varyn syn/n∈ {1,2,5,10,50}

    •MI draws: varyM∈ {5,10,20,50,100}. •Synthetic sample size: varyn syn/n∈ {1,2,5,10,50}. •Overlap: vary the positivity constantηin the ACIC DGP and measure the effect on RMSE and coverage. I ACS SEMI-SYNTHETIC DGP DETAILS Outcome standardization constants.For DP measurement, each dataset’s outcome is standardized as (Y−m pub)/spub using the fixed public co...

    work page arXiv 2018

  18. [18]

    Rows index sample size n∈ {1000,5000,20000} and columns index ε∈ {0.5,1,2,5}

    is mixed: the proxies win on single-estimand RMSE on ACIC, while Causal + NA+MI wins on IHDP and attains full coverage everywhere; and the synthetic-data release supports multiple estimands at shared privacy 1000500020000 n 0.13 0.13 0.13 0.13 0.08 0.08 0.08 0.08 0.08 0.08 0.08 0.08 Non-private DR n 1.17 0.49 0.29 0.25 0.25 0.21 0.20 0.20 0.20 0.19 0.19 0...

    work page 2024