Comparison of different approaches in estimating the time-varying reproductive number for COVID-19

PURPOSE: The time-varying reproductive number (R(t)) is an indicator of transmissibility that has utility in evaluating public health interventions and assessing transmission factors. However, the R(t) may be biased by generation time misspecification, reporting delays, underestimation of cases, and day-to-day variations. We compared several methods of adjustments in developing an approach to estimating an unbiased R(t.) METHODS & MATERIALS: A meta-analysis of generations times was conducted to reduce misspecification. A probabilistic bias approach was compared to standardization by a test positivity of 5% in adjusting for underestimation. A Poisson deconvolution process using an incubation period of 5.2 days (95% CI: 4.9-5.5) and laboratory turnover times between 2-, 5- and 10-days was utilized to adjust for reporting delays. We compared smoothing (7- and 14-day moving averages), a generalized additive model (GAM), and a local regression (LOESS) model to adjust for day-to-day variation. The adjusted R(t) was compared to a crude R(t) by eyeballing, Mean Average Percentage Error (MAPE), and Mean Absolute Deviation (MAD). We estimated the R(t) using Malaysian COVID-19 daily case data from 7 March 2020-20 June 2021 utilizing Cori et al.’s method. RESULTS: We estimated a pooled serial interval of 4.95 days (95% CI: 4.62-5.29). The R(t) estimated using case counts adjusted for underestimation using standardization by test positivity (MAPE: 0.31; 95% CI: 0.30-0.49, MAD: 0.5; 95%CI: 0.5-0.54) were more volatile, exhibited larger peaks and wider confidence intervals, especially in periods of lower incidence, compared to the probabilistic bias approach (MAPE: 0.07; 95% CI: 0.06-0.07, MAD: 0.26; 95%CI: 0.26-0.28). GAM (MAPE: 1.85, 95% CI: 1.63-2.08) and LOESS (MAPE: 0.29, 95% CI: 0.29-0.29) models had smoothed out almost all variations in the R(t). Longer lab turnover periods created smoother R(t) with larger peaks and resulted in greater volatility in the estimates. CONCLUSION: Biases in the estimation of the R(t) may critically change its interpretation for public health interventions. It is important to adjust for these biases and understand the underlying limitations of these estimations; primarily when utilized within the context of pandemic control.