Researchers and clinicians often rely on molecular assays like PCR to identify and monitor viral infections, instead of the resource-prohibitive gold standard of viral culture. However, it remains unclear when (if ever) PCR measurements of viral load are reliable indicators of replicating or infectious virus. The recent popularity of PCR protocols targeting subgenomic RNA for SARS-CoV-2 has caused further confusion, as the relationships between subgenomic RNA and standard total RNA assays are incompletely characterized and opinions differ on which RNA type better predicts culture outcomes. Here, we explore these issues by comparing total RNA, subgenomic RNA, and viral culture results from 24 studies of SARS-CoV-2 in non-human primates (including 2167 samples from 174 individuals) using custom-developed Bayesian statistical models. On out-of-sample data, our best models predict subgenomic RNA positivity from total RNA data with 91% accuracy, and they predict culture positivity with 85% accuracy. Further analyses of individual time series indicate that many apparent prediction errors may arise from issues with assay sensitivity or sample processing, suggesting true accuracy may be higher than these estimates. Total RNA and subgenomic RNA showed equivalent performance as predictors of culture positivity. Multiple cofactors (including exposure conditions, host traits, and assay protocols) influence culture predictions, yielding insights into biological and methodological sources of variation in assay outcomes-and indicating that no single threshold value applies across study designs. We also show that our model can accurately predict when an individual is no longer infectious, illustrating the potential for future models trained on human data to guide clinical decisions on case isolation. Our work shows that meta-analysis of in vivo data can overcome longstanding challenges arising from limited sample sizes and can yield robust insights beyond those attainable from individual studies. Our analytical pipeline offers a framework to develop similar predictive tools in other virus-host systems, including models trained on human data, which could support laboratory analyses, medical decisions, and public health guidelines.