WorksMathematics
Differentially Private Confidence Intervals for Relative Risk and Odds Ratio
Abstract
Ratio statistics, such as relative risk and odds ratio, are central tools for comparing outcome probabilities across populations, particularly in clinical and epidemiological studies. However, when the underlying counts are released under differential privacy, the resulting ratio estimators can exhibit nontrivial bias and variance, and principled methods for statistical inference in this setting remain limited. In this work, we study estimation and inference for ratio statistics computed from counts perturbed by additive Laplace or Gaussian noise. We show that the natural plug-in estimators based on privatized counts are consistent, and we derive confidence intervals that explicitly account for both sampling variability and privacy-induced noise, making our approach appealing for moderate sample sizes. Our analysis provides transparent variance decompositions that clarify when privacy meaningfully affects inference. Empirical results on simulated and real datasets demonstrate that, for moderate privacy levels and sample sizes, the proposed confidence intervals closely match their non-private counterparts, while properly reflecting additional uncertainty in small-sample regimes. Our results provide both theoretical guarantees and practical guidance for conducting statistical inference with ratio statistics under differential privacy.
Introduction
In many practical settings, analysts do not design the privacy mechanism themselves. Instead, they receive counts that have already been perturbed to satisfy differential privacy (DP), for example, by a statistical agency, hospital, or data curator, and must carry out standard downstream analyses using these noisy releases. A common example is the analysis of contingency tables, where one wishes to compare event probabilities across two groups using ratio statistics such as the relative risk (RR) and the odds ratio (OR). These quantities are central in clinical trials, epidemiology, and public health, and they also arise broadly whenever event rates are compared across populations.
This raises a basic practical question: if the underlying counts have already been privatized, can one still form statistically meaningful estimates and confidence intervals (CIs) for RR and OR by plugging in the noisy counts? Since ratio statistics can be unstable when denominators are small, and since additive DP noise further perturbs the relevant counts, it is not a priori clear when such post-processed estimators remain reliable.
This paper is meant as a practical guide to that question. We study estimation and inference for RR and OR when the underlying counts are released under additive DP noise. We prove consistency of the private plug-in estimators (Lemma 1), derive asymptotically valid CIs under Gaussian perturbation (Theorems 2–4), and show how the same formulas extend naturally to Laplace noise by variance substitution. Our variance expressions separate sampling- and privacy-induced variability (Table 1), yielding a simple, practical rule: privacy has limited impact when effective cell counts are large, and becomes consequential mainly in sparse regimes. Our main message is simple: privacy has the largest effect precisely where inference was already difficult to begin with. Simulations and real-data examples (§ 4) support this picture, showing that, for moderate privacy levels and sample sizes, the private CIs are typically close to their non-private counterparts.
| Statistic | Sampling error | Perturbation error |
| Log RR | ||
| RR | ||
| Log OR |
Table 1. Decomposition of the asymptotic variances in Theorems 2–4 into sampling and perturbation components.
Related work. Our perspective is closely connected to a growing literature on inference from privatized low-dimensional summaries. A substantial body of work studies DP methods for frequency tables and related contingency-table problems [4, 12, 17, 7, 2]. More recently, the deployment of DP in official statistics, especially by the U.S. Census Bureau for the 2020 Census, has sharpened attention on the needs of analysts who must consume already-noised releases and understand how privacy propagates into downstream estimands [10, 11, 14]. Our work fits squarely into this consumer-of-DP-data viewpoint.
Most closely related to our work is a paper by Lin et al. [8], wherein they study the problem of constructing CIs around proportion estimates when using stratified random sampling techniques under DP constraints. In situations where the denominator of their proportion estimate (the size of the relevant sample) is private information, they consider adding Gaussian noise for DP to both the numerator and the denominator. As a result, they end up with a mathematical object—a ratio of two Gaussian random variables—that is also central to our analysis. They provide a detailed analysis of this ratio and derive an asymptotically valid CI. In § 3, we use proof techniques similar to those of [8] to obtain another DP-CI for the RR (Theorem 4).
A longer version of this work appears on arXiv [15]. The full version includes additional mechanisms and extensions, including direct perturbation of the ratio, transformed estimators, local-sensitivity-based approaches, and a more detailed treatment of bias and finite-sample behavior.
Preliminaries
Unless stated otherwise, denotes probability with respect to the joint distribution over all sources of randomness. We use for convergence in probability, and for convergence in distribution. Throughout the paper, non-private estimators are denoted by and their differentially private counterparts by .
The Relative Risk and Odds Ratio Statistics
The relative risk (RR) and odds ratio (OR) are fundamental measures for comparing outcome probabilities between two groups. The RR quantifies the multiplicative change in the probability of an event, while the OR compares the corresponding odds. When event probabilities are small, the two measures are numerically similar, but they may differ substantially when probabilities are moderate or large. The RR is often preferred for interpretability in clinical settings, whereas the OR arises naturally in logistic modeling and case–control studies; see Agresti [1] for background.
We suppose that and are independent binomial random variables, and , where and with . The sample sizes and are assumed known. Our goal is to estimate either or .
Let and denote the maximum likelihood estimators of and . The corresponding plug-in estimators are
These estimators are biased on finite samples but consistent for their respective population parameters; see, for example, van der Vaart [16].
To quantify uncertainty, one typically constructs CIs.
Definition 1 (Confidence Interval). Let be a distribution on and let be an i.i.d. sample from . A confidence interval for is a data-dependent interval satisfying
where probability is taken with respect to both the sampling randomness and any additional randomness in and .
Classical large-sample CIs for RR and OR are typically derived via normal approximations and the delta method; see Agresti [1, Ch. 2].
Differential privacy
We briefly review the notions from differential privacy that are used throughout the paper; see Dwork and Roth [6] for a comprehensive treatment.
Let be a data domain. An unordered dataset of size is represented as a tuple , where is assumed public. Two datasets are called neighbors, written , if they differ in exactly one entry. DP requires that the output distribution of a randomized algorithm not change too much when applied to neighboring datasets.
Definition 2 (Differential privacy). A randomized algorithm is -differentially private if for all neighboring datasets and all measurable sets , .
Smaller values of and correspond to stronger privacy guarantees. Two basic properties of DP are especially relevant in practice: privacy guarantees are preserved under arbitrary post-processing of the released output, and they degrade in a controlled way under composition of multiple private mechanisms; see Proposition 2.1 and Theorem 3.14 in [6]. The post-processing property is particularly important for our purposes, since throughout the paper we analyze ratio statistics constructed from already-privatized counts.
A standard way to ensure differential privacy for numerical queries is to add random noise calibrated to the query’s sensitivity. For , define the - and -sensitivities by , . The following two classical additive-noise mechanisms will be used throughout.
Theorem 1 (Laplace and Gaussian mechanisms). Let . Consider mechanisms of the form , where is a random noise vector independent of the data. Then the following guarantees hold.
Theoretical results
In this section, we characterize the joint impact of sampling variability and privacy-induced noise on the RR and OR. While classical CIs remain asymptotically valid when privacy perturbations are of a lower order than sampling error, they become insufficient in finite-sample regimes where privacy noise significantly alters the error distribution. We derive CIs for both RR and OR under Gaussian noise and establish their asymptotic validity. Finally, we demonstrate how this construction extends to the Laplace mechanism, providing a general framework for inference with privatized count data.
Setting
We consider independent binomial random variables and with and known sample sizes . Privatized counts are released as and , where are independent of and of each other.
Our analysis uses replacement adjacency: the sample sizes and are treated as public, and neighboring databases differ in the event of one individual. If are not public, one could instead estimate them privately, using part of the privacy budget. The RR and OR estimators could then be computed by post-processing these noisy counts. The resulting CIs would need to account for the additional uncertainty, but this extension does not conceptually change our results and is therefore omitted due to space limitations.
In this section, whenever we write , , , or , the asymptotic regime is understood to be , , .
We consider either (i) Laplace noise , or (ii) Gaussian noise , , with parameters calibrated according to Theorem 1. We consider and ; equivalently and . This allows privacy perturbations that may exceed the usual sampling fluctuations of order and , while still guaranteeing consistency. For example, the assumption implies that if we were to use the Laplace mechanism (Theorem 1), we assume that .
Denote and . To ensure ratios are well defined, all statistics are computed using projected counts and .1 Under for some fixed , together with and , we have . Thus, clipping affects the statistics with vanishing probability, and we omit the subscript for brevity.
With these definitions, we consider the private plug-in estimators
Lemma 1 (Consistency of the perturbed RR and OR). Under the standing assumptions of Subsection 3.1, and .
Proof. We write . By the weak law of large numbers, . Since , Chebyshev’s inequality implies . Hence , and similarly .
The mappings and are continuous at when . The result follows from the continuous mapping theorem.
Confidence Intervals
We derive asymptotically valid CIs under additive Gaussian noise. All results are stated under the standing assumptions of Subsection 3.1.
For a real-valued statistic and standard error estimate , define the Wald-type interval , where is the quantile of a standard normal.
With this in mind, we now state our main theorems for valid CIs under Gaussian noise addition.
Theorem 2. Under the standing assumptions of Subsection 3.1, define
Then the interval satisfies . Consequently, is an asymptotic CI for .
An analogous result can be derived for the OR statistic.
Theorem 3. Under the standing assumptions of Subsection 3.1, define
Then the interval satisfies . Consequently, is an asymptotic CI for .
The proofs of Theorems 2 and 3 are given in the Appendix. At a high level, the argument has three steps: first, clipping the noisy counts is asymptotically negligible under and , ; second, the privatized counts are asymptotically normal (binomial + independent Gaussian noise); and third, applying the delta method to the log-RR and log-OR yields asymptotic normality with the stated variances. The proposed standard-error formulas are then shown to be consistent plug-in estimators of these asymptotic variances, and Slutsky’s theorem gives the validity of the resulting CIs.
A Direct Normal Confidence Interval for the Relative Risk In addition to the log-scale construction above, we also establish a CI based on a direct normal approximation of the ratio itself. While related in spirit to the approach of [8], our result is tailored to the two-sample setting with known group sizes. For completeness, we state the result below; a full proof appears as Theorem 4.2 in our extended manuscript [15].
Theorem 4 (Direct normal CI for ). Under the standing assumptions of Subsection 3.1, let , and define
Then the interval satisfies .
Laplace perturbations. The preceding theorems are stated for Gaussian perturbations, for which the privatized counts remain asymptotically normal after variance adjustment. For Laplace perturbations, the same plug-in variance formulas can be used as a finite-sample approximation by replacing with . Formally, if and , so that the privacy perturbation is lower order than the sampling error, then the Laplace noise is asymptotically negligible and the resulting intervals remain asymptotically valid. When the Laplace noise is not lower order than the sampling error, asymptotic normality no longer follows from negligibility of the privacy perturbation. In this case, our interval is best understood as a moment-matched Gaussian approximation, obtained by replacing the Laplace noise with Gaussian noise of the same variance.
When does privacy materially affect inference?
A central concern for analysts using DP statistics is whether privacy-induced perturbation can severely reduce statistical power. Our variance formulas (Table 1) show that, in the settings we study, this concern is often less severe than it may first appear.
Table 1 decomposes the asymptotic variance into a sampling component and a perturbation component, for simplicity setting . The key structural feature is that the sampling terms scale like and , whereas the perturbation terms scale like and . Thus, for fixed , the sampling standard error decreases at rate , while the additional error due to privacy decreases faster, at rate .
This gap in rates has an important consequence. In high-power regimes, where the counts are large, privacy adds only a lower-order term to the overall uncertainty. Conversely, when the privacy term is non-negligible, the sampling variance is already large, meaning that the corresponding non-private statistic is itself unstable. Thus, a large privacy-induced contribution to the interval width should not be interpreted as privacy alone creating the difficulty; rather, it indicates a regime in which the underlying non-private inference is already fragile. Indeed, taking the ratio between the DP CI width (sampling plus perturbation error) and the non-private width (sampling error only), we obtain
Similarly, for the odds ratio,
Estimating privacy-induced inflation. The same variance decomposition can also be used as a simple diagnostic after the privatized table has been released. Since both the sampling component and the perturbation component in Table 1 are functions of the released noisy counts and the known noise variance, an analyst can estimate the relative contribution of privacy noise without needing to incur any additional privacy cost. In particular, one can compare the perturbation component to the sampling component. For Laplace perturbations, the same diagnostic is obtained by replacing the Gaussian variance by the Laplace variance. This provides a practical way to assess, from the released table itself, whether the privacy mechanism materially changes the inferential conclusion.
Empirical evaluation
We empirically study the bias of the private estimators, as well as the width and coverage of the proposed CIs, on both simulated and real data. Throughout the simulation study we set for simplicity. We also ran additional experiments with unequal group sizes and observed qualitatively similar behavior; these results are omitted for brevity.
Bias. In general, it is known that the RR and OR non-private estimators are biased. We examine how additive noise further affects this bias. We fix and consider three parameter configurations, representing decreasing, null, and increasing signal regimes, . For each configuration and each privacy level , we run independent repetitions. In each repetition, we generate fresh data and , and then independently perturb the counts using either Gaussian or Laplace noise. Gaussian noise is calibrated using the analytic Gaussian mechanism to satisfy -DP with , while Laplace noise uses scale (Theorem 1). We then compute the privatized plug-in estimator for RR or OR from the perturbed counts, with projection to , and estimate its expectation by averaging across repetitions. For reference, we also report the mean of the corresponding non-private plug-in estimator and the true parameter value.
Empirical Evaluation
Figure 1 depicts the bias of the private and non-private estimators. The private estimators exhibit a small upward bias relative to their non-private counterparts, particularly in the strong-privacy regime (small ). As expected, the bias decreases as increases. Across all three parameter settings, the private estimators remain close to the non-private ones, and the discrepancy becomes negligible for moderate privacy levels. Notably, under Laplace noise addition, the bias induced by perturbation is of a lower order than the inherent sampling bias. For example, when , the bias due to privacy (the gap between the dotted line and the blue line) is negligible, while the bias due to sampling is clearly manifested (the gap between the dotted line and the dashed line).

Figure 1. Empirical mean of the privatized plug-in estimator for RR and OR (1) under Laplace and Gaussian noise addition. The dashed horizontal line indicates the true parameter value. Top row: RR. Bottom row: OR.
Confidence interval width and coverage. We evaluate the empirical coverage and average width of CIs for both the RR and the OR as functions of the privacy parameter under the same setting as in the previous experiment. For RR we compare two interval constructions: one based on Theorem 2, and the second on Theorem 4. For OR, we use Theorem 3. A noiseless baseline (setting the noise variances to zero) is included for reference.
Figures 2(a) and 2(b) summarize the empirical coverage and mean width of the proposed CIs. Across all parameter configurations, the empirical coverage of the private intervals remains close to the nominal level . The behavior is stable even in asymmetric signal regimes, suggesting that the delta-method-based approximation remains accurate after noise addition.
As expected, the average width of the intervals decreases monotonically with . In the strong-privacy regime, perturbation variance dominates, leading to visibly wider intervals. As increases, the widths approach those of the noiseless baseline, reflecting the fact that the additional variance from privacy becomes negligible relative to sampling variability. The Gaussian mechanism yields systematically wider intervals than the Laplace mechanism. This is expected, since for small the Gaussian noise scale is proportional to , whereas the Laplace scale is proportional to . With , this constant factor is non-negligible at moderate sample sizes.

Figure 2a. Odds Ratio: Mean width and coverage for CI construction. The dashed horizontal line marks the nominal coverage .

Figure 2b. Relative Risk: Mean width and coverage for CI construction. The dashed horizontal line marks the nominal coverage .
Overall, the simulations confirm the theoretical predictions: even in these relatively small datasets, privacy-induced variability inflates interval width primarily when sampling variability is already substantial, while coverage remains well-controlled across privacy levels.
Real Data Application. In this subsection, we illustrate the practical behavior of the proposed CIs using real contingency tables. We study both the RR and the OR, and compare their non-private and private releases to understand how DP noise affects statistical inference in practice. Since the RR intervals from
Empirical Evaluation
Theorems 4 and 2 lead to very similar conclusions, we report only the interval of Theorem 2 for simplicity.
We consider three representative tables, displayed in Panels (a)–(c) of Table 2. These examples span three different regimes. Panel (a) is a large-sample rare-event table: the total sample size is in the tens of thousands, while the event counts themselves are only in the hundreds. Panel (b) is an intermediate-scale table, with all four cells in the hundreds and a moderate association between exposure and outcome. Panel (c) is a much smaller table, with one cell in the single digits and a strong treatment effect. This progression from large counts, to moderate counts, to very small counts allows us to highlight how the magnitude of the cell counts governs the robustness of private inference. For the case-control table in Panel (b), the odds ratio is the epidemiologically natural estimand; we also report the RR-style plug-in ratio only as a numerical benchmark for the behavior of ratio statistics under privatization.
| Panel | Group / Smoking / Treatment | Yes / Case | No / Control |
| (a) | Placebo | 189 | 10,845 |
| (a) | Aspirin | 104 | 10,933 |
| (b) | Ever smoked | 163 | 150 |
| (b) | Never smoked | 124 | 161 |
| (c) | Streptomycin | 51 | 4 |
| (c) | Control | 38 | 14 |
Table 2. Three contingency tables used in the experiments. (a) Table 2.1 in [1] (aggregated): Aspirin use vs. myocardial infarction. (b) Table 2 in [13]: tobacco smoking vs. breast cancer. (c) Table IX in [9]: MRC streptomycin RCT (tuberculosis).
For each table, we release the two event counts as a two-dimensional count query and add independent Laplace noise with scale to each coordinate. This vector release is -DP by the Laplace mechanism. We then construct CIs using Theorems 2 and 3, setting the perturbation variance to .
For each dataset and each privacy level, we record the average value of the private estimator and the average width of the corresponding CI. These averages are computed over independent noise draws, so as to isolate the effect of the privacy mechanism itself. The results are summarized in Table 3.
| Panel | Width | Width Priv | Width | Width Priv | |||||
| (a) | 0.5 | 0.551 | 0.298 | 0.546 | 0.299 | ||||
| (a) | 1 | 0.550 | 0.550 | 0.264 | 0.273 | 0.546 | 0.546 | 0.265 | 0.274 |
| (a) | 2 | 0.550 | 0.266 | 0.546 | 0.268 | ||||
| (b) | 0.5 | 1.179 | 0.435 | 1.420 | 1.126 | ||||
| (b) | 1 | 1.178 | 1.178 | 0.362 | 0.381 | 1.411 | 1.413 | 0.926 | 0.979 |
| (b) | 2 | 1.178 | 0.367 | 1.412 | 0.939 | ||||
| (c) | 0.5 | 1.282 | 140.928 | 9.086 | 61735.705 | ||||
| (c) | 1 | 1.269 | 1.271 | 0.461 | 0.668 | 4.697 | 6.966 | 13.977 | 1101.317 |
| (c) | 2 | 1.270 | 0.519 | 5.528 | 44.206 |
Table 3. Non-private and private point estimates and CI widths under Laplace-perturbed counts. For RR we use the log-scale CI of Theorem 2; for OR we use Theorem 3. Width denotes the noiseless CI width, and Width Priv denotes the average private CI width.
Table 3 shows that the impact of DP depends strongly on the scale of the underlying counts. In Panel (a), the total sample size is very large, and even the event counts are in the hundreds, so the perturbation has only a minor effect. Both and are essentially identical to their non-private counterparts across all privacy levels, and the increase in interval width is modest. In Panel (b), the effect of privacy is visible, but remains well-controlled. The private point estimates are very close to the non-private ones for both RR and OR, and the CIs widen only moderately. Thus, when the cell counts are in the hundreds, Laplace perturbation introduces some additional uncertainty, but does not qualitatively distort the inferential conclusions.
The most dramatic behavior appears in Panel (c). In this case, the additive noise is no longer negligible relative to the observed counts, and the ratio estimators become highly unstable under strong privacy. For RR, the average private estimate increases only slightly, from 1.269 to 1.282 at , but the average private CI width explodes from 0.461 to 140.928. The effect is even more pronounced for OR: the average private CI width becomes enormous. This instability is exactly what one expects when noise can move a small count close to the boundary, causing ratios and log-ratios to vary dramatically.
Conclusion
We studied estimation and inference for relative risk and odds ratio when the underlying event counts are released under additive DP mechanisms. We showed how to construct plug-in estimators and variance-adjusted CIs that explicitly account for both sampling variability and privacy-induced noise. Empirically, the proposed intervals preserve coverage across a range of privacy levels, with widths close to their non-private counterparts in well-powered settings. Our variance decompositions further clarify when privacy affects inference: privacy has modest effects in high-count regimes, while sparse regimes remain fragile, both with and without privacy. These findings support simple plug-in inference from privatized counts as a practical tool, while also clarifying its limitations under sparse data.
Acknowledgments. This work was supported in part by a gift to the McCourt School of Public Policy and Georgetown University, Simons Foundation Collaboration 733792, Israel Science Foundation (ISF) grant 2861/20, Apple, ERC grant 101125913, and a grant from the Israeli Council of Higher Education. Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them.
Disclosure of Interests. The authors have no competing interests to declare that are relevant to the content of this article.
Appendix
Proof of Theorem 2
Proof. We begin by showing that trimming the noisy counts has no asymptotic effect. Denote and , and define , . Since , we have . By the triangle inequality,
The first term tends to 0 exponentially fast by Hoeffding's inequality for binomials. The second tends to 0 because and , implying . Thus , and similarly . Consequently, , , and in particular
Next, we follow the construction of the delta method. Define . The function is twice continuously differentiable in a neighborhood of , with . Write
where is independent of . Since , , , and analogously for . To identify the asymptotic law of , write
where
Let
Then by the classical CLT, and is independent of . Hence, for each fixed , the characteristic function of is
Since , its characteristic functions converge to , uniformly for in compact sets; see, e.g., Lévy’s continuity theorem [5 Sec. 26]. Because , this gives . Therefore,
where we used . By Lévy’s continuity theorem,
The same argument applies to .
Let and . Define
Since and , a Taylor expansion around gives
By independence of the two counts and their perturbations, the leading linear term is asymptotically normal with variance . Moreover,
because and . Hence the Taylor remainder is .
Finally, by (2) and continuity of and on , . Slutsky’s theorem therefore implies
which yields
Exponentiating the endpoints preserves coverage for .
Proof of Theorem 3
Proof. The trimming argument is identical to the beginning of the proof of Theorem 2. In particular, clipping has no asymptotic effect, and
Moreover, by the same argument as in that proof,
where .
We now apply the delta method with . The function is twice continuously differentiable in a neighborhood of , with
Hence
Denote
The leading linear term has variance
and the remainder satisfies .
Indeed, on an event whose probability tends to one, and remain in fixed interior fractions of their ranges, and the second derivatives of are of order and . Moreover, since and , the displayed remainder is . By independence of the two counts and their perturbations, the leading linear term is asymptotically normal with variance
Therefore , and by (3) and continuity of and on , . Slutsky’s theorem then yields , and therefore
Exponentiating the endpoints preserves coverage for OR.
References
- [1]Agresti, A.: Categorical data analysis, vol. 792. John Wiley & Sons (2012)
- [2]Awan, J., Slavković, A.: Differentially private uniformly most powerful tests for binomial data. Advances in Neural Information Processing Systems 31 (2018)arxiv.org/abs/1805.09236
- [3]Balle, B., Wang, Y.X.: Improving the gaussian mechanism for differential privacy: Analytical calibration and optimal denoising. In: International Conference on Machine Learning. pp. 394–403. PMLR (2018)arxiv.org/abs/1805.06530
- [4]Barak, B., Chaudhuri, K., Dwork, C., Kale, S., McSherry, F., Talwar, K.: Privacy, accuracy, and consistency too: a holistic solution to contingency table release. In: Proceedings of the twenty-sixth ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems. pp. 273–282 (2007)DOI
- [5]Billingsley, P.: Probability and Measure. Wiley, 3 edn. (1995)
- [6]Dwork, C., Roth, A.: The algorithmic foundations of differential privacy. Foundations and Trends® in Theoretical Computer Science 9(3–4), 211–407 (2014)
- [7]Gaboardi, M., Lim, H., Rogers, R., Vadhan, S.: Differentially private chi-squared hypothesis testing: Goodness of fit and independence testing. In: International conference on machine learning. pp. 2111–2120. PMLR (2016)arxiv.org/abs/1602.03090
- [8]Lin, S., Bun, M., Gaboardi, M., Kolaczyk, E.D., Smith, A.: Differentially private confidence intervals for proportions under stratified random sampling. Electronic Journal of Statistics 18(1), 1455–1494 (2024)DOI
- [9]Medical Research Council Streptomycin in Tuberculosis Trials Committee: Streptomycin treatment of pulmonary tuberculosis. British Medical Journal 2(4582), 769–782 (1948). https://doi.org/10.1136/bmj.2.4582.769
- [10]Mervis, J.: Can a set of equations keep us census data private. Science 10 (2019)DOI
- [11]Mueller, J.T., Santos-Lozada, A.R.: The 2020 us census differential privacy method introduces disproportionate discrepancies for rural and non-white populations. Population Research and Policy Review 41(4), 1417–1430 (2022)
- [12]Rinott, Y., O’Keefe, C.M., Shlomo, N., Skinner, C.: Confidentiality and differential privacy in the dissemination of frequency tables. Statistical Science 33(3), 358–385 (2018)DOI
- [13]Rollison, D.E., Brownson, R.C., Hathcock, H.L., Newschaffer, C.J.: Case-control study of tobacco smoke exposure and breast cancer risk in delaware. BMC cancer 8(1), 157 (2008)
- [14]Santos-Lozada, A.R., Howard, J.T., Verdery, A.M.: How differential privacy will affect our understanding of health disparities in the united states. Proceedings of the National Academy of Sciences 117(24), 13405–13412 (2020)
- [15]Shoham, T., Liggett, K.: Differentially private ratio statistics. arXiv preprint arXiv:2505.20351 (2025)arxiv.org/abs/2505.20351
- [16]Van der Vaart, A.W.: Asymptotic statistics, vol. 3. Cambridge university press (2000)
- [17]Vu, D., Slavkovic, A.: Differential privacy for clinical trial data: Preliminary evaluations. In: 2009 IEEE International Conference on Data Mining Workshops. pp. 138–143. IEEE (2009)DOI