Statistical Challenges when Analyzing SARS-CoV-2 RNA Measurements Below the Assay Limit of Quantification in COVID-19 Clinical Trials

Abstract

Most clinical trials evaluating COVID-19 therapeutics include assessments of antiviral activity. In recently completed outpatient trials, changes in nasal SARS-CoV-2 RNA levels from baseline were commonly assessed using analysis of covariance (ANCOVA) or mixed models for repeated measures (MMRM) with single-imputation for results below assay lower limits of quantification (LLoQ). Analyzing changes in viral RNA levels with singly-imputed values can lead to biased estimates of treatment effects. In this paper, using an illustrative example from the ACTIV-2 trial, we highlight potential pitfalls of imputation when using ANCOVA or MMRM methods, and illustrate how these methods can be used when considering values <LLoQ as censored measurements. Best practices when analyzing quantitative viral RNA data should include details about the assay and its LLoQ, completeness summaries of viral RNA data, and outcomes among participants with baseline viral RNA ≥LLoQ, as well as those with viral RNA <LLoQ.

Figure 1:

Distribution of SARS-CoV-2 RNA from nasopharyngeal swabs in Active and Placebo arms by study visit in overall cohort (A and B) and among those with vRNA ≥LLoQ at Baseline/Day 0 (C and D). Levels of SARS-CoV-2 RNA (log₁₀ copies/ml) with horizontal line = median, box=interquartile range, whiskers = minimum/maximum (A and C); results below the LLoQ are plotted using an imputed value of 1 log₁₀ copies/ml. Proportion with quantifiable SARS-CoV-2 RNA (green) and unquantifiable (purple) (B and D). LLOQ = Lower Limit of Quantification.

Figure 2:

Quantile-quantile (Q-Q) plot for linear regression model for censored data for change in vRNA from baseline to Day 3 (A) and to Day 7 (B), both models included an indicator variable for treatment versus placebo and adjusted for baseline vRNA Standardized residuals (for the non-censored observations) calculated by dividing the residuals by their standard deviation (estimated from the fitted model). Quantiles for a standard normal distribution plotted on the x-axis take account of censored residuals.