RRT Meets RDD: Statistical Strategies for Assuring Response Privacy in Telephone Surveys

Abstract

A WIDE variety of statistical techniques have been developed and incorporated into survey questions on sensitive traits. In the simplest case, they provide population estimates of these traits while assuring the privacy of individual responses. The original version, proposed by Stanley Warner (1965), is known as the related question randomized response technique (RRT). It consists of presenting the survey respondent with two complementary statements: "I have sensitive trait X" and "I do not have sensitive trait X." A randomizing device, such as a die, is used to chose a statement, and the respondent is asked to provide a true or false response to the one selected. The interviewer is not shown the die nor informed of which statement was answered. Assuming that respondents adhere to instructions, an unbiased estimate of the true proportion of individuals with trait X (TI) can be obtained as follows:where <f> is the probability of selecting the sensitive statement, N(yes) the number of respondents answering yes, and N the total number of Abstract Despite its growing role in survey research, the telephone survey has been largely neglected with respect to the design and testing of statistical procedures for assuring response privacy (e.g., randomized response techniques). This paper demonstrates that such procedures are no less feasible to administer in telephone surveys than in face-to-face surveys. Both randomized response techniques and alternatives to randomized response are considered.

Gender Differences in Marital Satisfaction: A Meta‐analysis

“…A trim and fill test for data censoring (Duval & Tweedie, ), which imputed 26 potentially missing samples yielding nonsubstantively different results, suggested the absence of significant publication bias. We also conducted several Orwin's fail‐safe N tests for data censoring (Orwin & Boruch, ) and found there would need to be 9,173 missing samples with a Hedges's g of 0.00 added to the included samples to nullify the statistical significance of the observed aggregated effect size ( g = 0.04, K = 226). Furthermore, there would need to be 155 missing samples with a Hedges's g of 0.21 added to the included samples to move the very small observed aggregated effect size ( g = 0.04, K = 226) to a small effect size ( g = 0.11), 723 missing samples with a Hedges's g of 0.46 to move the aggregated effect size to a moderate effect size ( g = 0.36), 1,404 missing samples with a Hedges's g of 0.76 to move the aggregated effect size to a large effect size ( g = 0.66), and 2,176 missing samples with a Hedges's g of 1.10 to move the aggregated effect size to a very large effect size ( g = 1.00).…”

Long-Term Anti-Hypertensive Therapy and Stroke Prevention: A Meta-Analysis

“…Funnel plots were used to visually assess for publication bias, while the Begg and Mazumdar [11] test was used to quantify the amount of publication bias. The Orwin failsafe N test [12] was used to determine the number of missing studies would be required to make the summary effect trivial.…”

Anesthetic Neuroprotection in Experimental Stroke in Rodents

“…Evidence for publication bias was sought by constructing funnel plots, performing Orwin’s fail-safe N test10, and determining Egger’s regression intercept11. Quantification of the impact of publication bias on summary effects was estimated with “trim and fill”12.…”