Is the COVID-19 Antibody Seroprevalence in Santa Clara County really 50-85 fold higher than the number of confirmed cases?

Conclusions In this rather long post, I undertook a Bayesian analysis of the seroprevalence survey in Santa Clara county. Contrary to the author’s assertions, this formal Bayesian analysis suggests a much lower seroprevalence of COVID19 infections. The estimated fold increase is only 20 times higher (point estimate), rather than 50-80 fold higher and with a 95% credible interval 0.75 -55 fold for the most probable model. The major driver of the prevalence, is the specificity of the assay (not the sensitivity), so that particular choices for this parameter will have a huge impact on the estimated prevalence of COVID19 infection based on seroprevalence data. Whereas the common parameter and the shifted-along-ROC models yield similar estimates for the same prior (the uniform in probability scale is equivalent to standard logistic in the logit scale), a minor change to the priors used by the shifted-along-ROC model, leads to results that are qualitatively and quantitatively different. The sensitivitity of the estimated prevalence to the prior, suggests that the assay performance data are not sufficient to determine either the sensitivity or the specificity with precision. Even the minor shrinkage implied by changing the prior from the standard logistic to the standard normal , provides a small, yet crucial protection against overfitting and leads to a model with extremely sharp marginal posterior likelihood. Are the Bayesian analysis results reasonable? ...
Source: The Health Care Blog - Category: Consumer Health News Authors: Tags: COVID-19 Health Policy Christos Argyropoulos Pandemic public health santa clara Source Type: blogs