본문 바로가기 주메뉴 바로가기
국회도서관 홈으로 정보검색 소장정보 검색

결과 내 검색

동의어 포함

목차보기

Title page 1

Contents 1

Abstract 2

1. Introduction 3

2. Description of the research problem 4

3. Inference under noise addition 5

3.1. Truncated modification 7

3.2. Discrete Noise Addition 10

3.3. Likelihood approach based on Z = X + ϵ 13

3.4. Predicting X from a Released Z 15

3.5. Simulation Results 15

3.5.1. MLE: Gaussian Noise 15

3.5.2. MLE: Laplace Noise 18

3.5.3. Bayes: Gaussian Noise 21

3.5.4. Bayes: Laplace Noise 23

3.5.5. Conditional Inference on X 25

4. Synthetic Data 37

4.1. Plug-In Sampling (PLS) 38

4.2. Posterior Predictive Sampling (PPS) 43

5. Multiple imputations 52

5.1. Noise Addition 52

5.2. Plug-In Sampling (PLS) 52

5.3. Posterior Predictive Sampling (PPS) 52

6. Comparison of three methods and concluding remarks 55

References 56

Tables 9

Table 3.1. Monte Carlo estimates of mean(U/n), variance, and MSE of θ = U/n under Gaussian noise for θtrue ∈ {0.1, 0.3, 0.5, 0.8} and... 9

Table 3.2. Monte Carlo estimates of mean(U/n), variance, and MSE of θ = U/n under Laplace noise with standard deviation σ (scale b = σ/√2)... 10

Table 3.3. Simulated mean and variance of the maximum-likelihood estimator θ under Gaussian noise with standard deviations σ = 0.5, 1.0,... 17

Table 3.4. Simulated mean and variance of the maximum-likelihood estimator θ under Laplace noise with standard deviations σ = 0.5, 1.0,... 20

Table 3.5. Simulated mean and variance of the Bayes estimate θBayes under Gaussian noise with standard deviations σ = 0.5, 1.0, 2.0,... 23

Table 3.6. Simulated mean and variance of the Bayes estimate θBayes under Laplace noise for each true value θtrue ∈ {0.1, 0.3, 0.5, 0.8},... 25

Table 3.7. Simulated posterior probabilities P(true_x / Z) under Gaussian noise for standard deviations σ = 0.5, 1.0, 2.0, 5.0 and sample sizes... 27

Table 3.8. Posterior probabilities P(X = x / Z) for n = 20 under Gaussian noise with σ = 5 28

Table 3.9. Simulated posterior probabilities P(true_x / Z) under Laplace noise for standard deviations σ = 0.5, 1.0, 2.0, 5.0 and sample sizes... 30

Table 3.10. Posterior probabilities P(X = x / Z) for n = 20 under Laplace noise with σ = 5 for various true parameter values θ 31

Table 3.11. Bayesian posterior predictive performance under Gaussian noise: for each true value θtrue ∈ {0.1, 0.3, 0.5, 0.8}, sample size... 33

Table 3.12. Bayesian posterior predictive probabilities under Gaussian noise (σ = 5, n = 20): we report the probability of the MAP estimate,... 34

Table 3.13. Bayesian posterior predictive performance under Laplace noise: for each true parameter θtrue ∈ {0.1, 0.3, 0.5, 0.8}, sample size... 36

Table 3.14. Bayesian posterior predictive probabilities under Laplace noise (σ = 5, n = 20): for each true parameter θ ∈ {0.1, 0.3, 0.5, 0.8}... 37

Table 4.1. 100 × True Probability P(ZPLS = x / X = x) by n and X/n 39

Table 4.2. Percentage of PLS Simulations where X is fixed and Z = X by θ and n 40

Table 4.3. Percentage of PLS Simulations where Z = X by θ and n 40

Table 4.4. True Probability P(ZPPS = x / X = x) by n and X/n 49

Table 4.5. Percentage of PPS Simulations where X is fixed and Z = X by n and X/n 49

Table 4.6. Percentage of PPS Simulations where Z = X by θ and n 49

Table 5.1. θ∗ Estimates based on PPS Imputations with Fixed θ∗ 53

Table 5.2. X Estimates based on PPS Imputations with Fixed θ∗ 54

Table 5.3. Percentage of Simulations with X = round(Z) based on m PPS Imputations 54

Figures 18

Figure 1. Empirical distributions of the standardized statistic U for the MLE under Gaussian noise 18

Figure 2. Empirical distribution of the standardized error U for the MLE under Laplace noise 21

Figure 3. UPLS Simulation Histogram by θ and n 41

Figure 4. Variance Stabilized UPLS Simulation Histogram by θ and n 42

Figure 5. UPPS Simulation Histogram by θ and n 50

Figure 6. Variance Stabilized UPPS Simulation Histogram by θ and n 51

초록보기

In this paper we consider the inferential problem for a Binomial proportion in situations when the exact number of units possessing an attribute under consideration is unavailable due to privacy reasons; however a synthesized version of this number is available.

The inference problem is addressed under three types of available information: noise added version and plug-in sampling based and posterior sampling based data.

A comparison of the three modes of data source is made based on inferential accuracy and a measure of privacy.