Differences in oral HPV prevalence estimation via Bayesian and frequentist methodologies from people living with HIV in Puerto Rico

Introduction: Puerto Rico faces significant public health challenges due to a high rate of HPV-related cancers and a high prevalence of HIV, consistently ranking among the U.S. states and territories with the most AIDS cases. While population-based data on oral HPV infection for Puerto Rico remains scarce, studies focusing on high-risk groups have found a 12.5% prevalence rate among drug users in the San Juan metropolitan area. Logistic regression, a common method for frequentist estimation, often fails to account for the accuracy of a screening method. Studies confirm that Bayesian statistical models overcome these limitations by incorporating test performance. Despite these advantages, the use of Bayesian prevalence estimation is rare. Specifically, they have proven useful in scenarios where screening methods are suboptimal; limited sample size, or the core assumptions of frequentist models are unable to be met. These challenges are commonly encountered in studies addressing site-specific HPV prevalence, making Bayesian prevalence estimation a particularly well-suited and recommended alternative. Objective : To compare point estimates of oral HPV prevalence derived from frequentist and Bayesian models. Methods: HPV genotyping data from an on-going study titled Multi-omics Predictors of Oral HPV Outcomes among People Living with HIV in Puerto Rico (P20GM148324) was used to estimate population prevalence. HPV status of this data was determined using the SPF10-LiPA25 method. R-statistical software and was used for prevalence estimation. The rjags library in R will provide the connection to run Markov Chain Monte Carlo (MCMC) simulations in JAGS software, Rogan-Gladen (RG) model will be utilized as the frequentist model since it compensates for diagnostic misclassifications. The comparison metrics to determine the best estimate will include error distribution for the point estimate and length of the confidence interval, or credible interval for the Bayesian case. Results: The Bayesian estimation was 0.121 ± 0.0261 (CrI95%: 0.073 - 0.176) whilst the Rogan-Gladen model was 0.117 ± 0.0465 (CI95%: 0.032 - 0.212). While estimates are similar for both models, there is a slight improvement for both standard errors (SE) and corresponding intervals for the Bayesian estimation compared to the RG estimation. Conclusion: The Bayesian method outperformed in both SE and Interval estimations as expected. A disadvantage of applying RG estimation is lack of interval estimation in the default function; intervals were manually calculated. Likewise, the number of iterations in MCMC impacts the SE for the Bayesian approach, further research must be done to determine if a cut-off point should be considered.