THE PRIMACY OF PRIORS: THE EFFECT OF BAYESIAN PRIOR SELECTION ON THE ESTIMATED RISK OF INDUCED VZV DISSEMINATION FROM THE SHINGLES VACCINE IN IMMUNOCOMPROMISED PATIENTS

Purpose:  The number of immunocompromised individuals (from underlying disease or immunosuppressive therapy) is increasing, resulting in a growing subpopulation that is vulnerable to infections for which vaccines exist, such as shingles. Shingles is caused by the reactivation of the varicella zoster virus (VZV) in patients who have had chickenpox. Immunocompromised individuals have an increased risk of shingles and its severe complications including potentially fatal dissemination. A live-attenuated vaccine that prevents shingles is available for individuals age 50 and over but is contraindicated in immunocompromised individuals. The safety concern is that the live, attenuated vaccine could acutely induce a severe vaccine virus strain of shingles in individuals with compromised immune systems. This analysis estimates the risk of vaccine-induced dissemination in immunocompromised populations. 

Method: We developed a series of Bayesian models in JAGS to estimate the distribution of the posterior probability of dissemination from live attenuated vaccine based on a series of 62 cases in hematopoietic cell transplant recipients and patients with hematologic malignancies. Our models were each run with three chains, each with 21,000 iterations (1,000 for burn-in). Three different priors were used in our analysis: skeptical (μ=0.9, σ=0.09), flat (μ=0.5, σ=0.29), and enthusiastic (μ=0.1, σ=0.09) Beta distributions. We compared our models on mean posterior estimate and model fit. 

Result: Our results show that the prior selection is extremely important in determining the posterior distribution of risk of vaccine-induced dissemination, as expected for our limited data set. When using a skeptical prior, similar to what the Food and Drug Administration might consider, our mean posterior estimate of risk was 12.5% versus just 1.4% when using an enthusiastic prior, and 1.6% with a flat prior. The DIC and pD for each model also show that different priors have different effects on model fitting. 

Conclusion: To the best of our knowledge, this is the first analysis to quantify the risk of vaccine-induced disseminated shingles using a Bayesian approach. Our results support those of large-scale retrospective analyses and suggest the shingles vaccine is safe and effective in preventing shingles for immunocompromised individuals; however, the volatility of the posterior estimate given differing prior distributions indicates that more research is needed to understand the risk.