An Empirical Study of Stratified Warner’s Randomized Response Model

Soumyadipta Das, Souvick Bera, Sagnik Bhattacharya2022 · DOI

Abstract

This paper presents a stratified randomized response model based on Warner’s original framework, designed to achieve optimal allocation and a significant gain in precision. The proposed model is shown to be more efficient than standard stratified randomized response models. Furthermore, the estimator derived from this method demonstrates superior efficiency compared to the Warner, Mangat and Singh, and Mangat estimators under specific conditions, accounting for both completely truthful and partially untruthful reporting by respondents.

Introduction

In social science and medical research, obtaining accurate data on sensitive or stigmatized behaviors—such as drug use, tax evasion, or academic dishonesty—is notoriously difficult. Respondents often succumb to "social desirability bias," providing untruthful answers to avoid social stigma or legal consequences. This leads to significant underreporting and compromised data integrity.

To address this, Stanley Warner introduced the Randomized Response (RR) technique in 1965. In Warner's original model, respondents use a randomizing device (like a coin flip) to determine whether they answer a sensitive question or its negation. Because the interviewer remains unaware of which question the respondent is answering, the respondent maintains "plausible deniability," which encourages more honest participation. However, while RR models protect privacy, they often suffer from low precision and high variance compared to direct questioning. This study explores a stratified approach to RR that utilizes optimal allocation to maximize statistical efficiency.

The Proposed Stratified Model

While previous researchers, such as Hong et al., introduced stratification (dividing the population into distinct groups like socioeconomic strata or academic departments), they typically applied uniform randomization parameters across all groups. This paper proposes a more flexible, stratified RR model characterized by two key technical innovations:

1. Stratum-Specific Randomization

Unlike "one-size-fits-all" models, this framework allows researchers to set different randomization probabilities ($P_i$) for the sensitive question within each stratum. This is a vital practical improvement, as different demographic groups may require different levels of privacy protection to feel comfortable reporting honestly. For instance, high-income individuals may require higher levels of anonymity when discussing financial sensitive topics than lower-income individuals.

2. Optimal Allocation Strategy

The core of the model’s efficiency lies in how the total sample size ($n$) is distributed across different strata. While many researchers use proportional allocation, this paper derives an optimal allocation formula. By leveraging prior information or historical data regarding the proportions of the sensitive trait within each stratum ($\pi_{s_i}$), the model minimizes the overall variance of the population estimate ($\pi_s$).

Methodological Framework

The authors utilize Maximum Likelihood Estimation (MLE) to derive the proportion of the sensitive trait within each stratum and for the population as a whole. The study mathematically compares the variance of this new estimator against established models, including:

  • Warner (1965): The foundational RR model.
  • Mangat & Singh (1990): A two-stage RR model.
  • Mangat (1994): An improved RR model focused on cooperation.

By defining specific inequality conditions involving stratum weights ($w_i$) and randomization probabilities ($P_i$), the authors prove the mathematical conditions under which the stratified model is superior.

Key Results and Insights

The study demonstrates that the proposed estimator is consistently more efficient than the standard Warner (1965) estimator for multi-strata populations, provided the randomization probabilities are not equal to 0.5.

Key findings include:

  • Precision Gains: The use of optimal allocation results in a "large gain in precision," allowing researchers to achieve more accurate estimates of sensitive traits without increasing the total number of respondents.
  • Robustness: The model maintains its efficiency advantage even in scenarios where respondents are not completely truthful, provided certain mathematical conditions are met.
  • Practical Utility: The model is particularly effective for multi-group studies where the sensitivity of a topic varies significantly across demographics such as age, sex, or socioeconomic status.

Limitations and Future Considerations

Despite its advantages, the model has specific requirements for successful implementation:

  • Requirement of Prior Information: The optimal allocation strategy depends on having some baseline knowledge of the proportions ($\pi_{s_i}$) from previous studies. If this information is unavailable or inaccurate, the allocation may not reach its theoretical peak efficiency.
  • Operational Complexity: Implementing different randomization devices or instructions for different strata adds complexity to fieldwork and data collection compared to simpler, non-stratified models.
  • The Assumption of Compliance: Like all RR models, the mathematical proofs assume that respondents follow the randomization device's instructions truthfully. Future research could investigate the impact of deliberate non-compliance or "cheating" within this stratified framework.

Conclusion

The stratified Warner’s Randomized Response model provides a powerful tool for researchers dealing with sensitive data. By combining the privacy protections of randomized response with the statistical rigor of optimal stratified allocation, the model offers a path toward more accurate and reliable social science research.

Summary updated

Abstract

This poster shows a stratified randomized response model based on Warner’s model that has an optimal allocation and large gain in precision. It is shown that the proposed model is more efficient than the stratified randomized response model. Additionally, it is shown that the estimator based on the proposed method is more efficient than the Warner, the Mangat and Singh and the Mangat estimators under the conditions presented in both the case of completely truthful reporting and that of not completely truthful reporting by the respondents.
BibTeX
@misc{das2022empirical,
  author = {Soumyadipta Das and Souvick Bera and Sagnik Bhattacharya},
  title = {{An Empirical Study of Stratified Warner’s Randomized Response Model}},
  year = {2022},
  doi = {10.14293/s2199-1006.1.sor-.pp30e1r.v1},
}
Figure