Debiasing Covid-19 prevalence estimates

Abstract

Timely, accurate epidemic figures are necessary for informed policy. In the Covid-19 pandemic, mismeasurement can lead to tremendous waste, in health or economic output. 'Random' testing is commonly used to estimate virus prevalence, reporting daily positivity rates. However, since testing is necessarily voluntary, all 'random' tests done in the field suffer from selection bias. This bias, unlike standard polling biases, goes beyond demographical representativeness and cannot be corrected by oversampling (i.e. selecting people without symptoms to test). Using controlled, incentivized experiments on a sample of all ages, we show that people who feel symptoms are up to 33 times more likely to seek testing. The bias in testing propensities leads to sizable prevalence bias: test positivity is inflated by up to five times, even if testing is costless. This effect varies greatly across time and age groups, making comparisons over time and across countries misleading. We validate our results using the REACT study in the UK and find that positivity figures have indeed a very large and time varying bias. We present calculations to debias positivity rates, but importantly, suggest a parsimonious way to sample the population bypassing the bias altogether. Our estimation is both real-time and consistently close to true values. These results are relevant for all epidemics, besides covid-19, when carriers have informative beliefs about their own status.

Wastewater-Based Epidemiology for COVID-19: Handling qPCR Nondetects and Comparing Spatially Granular Wastewater and Clinical Data Trends

“…Clinical-testing results can be further biased by various types of self-selection. 28 , 29 Though it is impossible to precisely determine the relative contributions of these factors and biases, context can suggest which are likely to have the greatest influence in a given instance. For example, an unexplained spike in wastewater—but not clinical—data for a zone housing disproportionate numbers of individuals with characteristics that could cause lower propensity to test (e.g., limited access to transportation; low English proficiency) could be a sign of the presence of infected individuals detected through WBE but not clinical testing.…”

“…WBE results can be affected by many factors, including variability in SARS-CoV-2 excretion rates, wastewater composition and temperature, average in-sewer travel time (coupled with viral decay in sewer lines), per-capita water use, autosampler settings, and movement of people in and out of sampling zones. Clinical-testing results can be further biased by various types of self-selection. , Though it is impossible to precisely determine the relative contributions of these factors and biases, context can suggest which are likely to have the greatest influence in a given instance. For example, an unexplained spike in wastewaterbut not clinicaldata for a zone housing disproportionate numbers of individuals with characteristics that could cause lower propensity to test (e.g., limited access to transportation; low English proficiency) could be a sign of the presence of infected individuals detected through WBE but not clinical testing.…”

Wastewater surveillance for COVID-19 response at multiple geographic scales: Aligning wastewater and clinical results at the census-block level and addressing pervasiveness of qPCR non-detects

“…WBE results can be affected by many factors, including variability in SARS-CoV-2 excretion rates [16], wastewater composition and temperature, average in-sewer travel time, per-capita water use [17], autosampler settings [18], and movement of people in and out of sampling zones. Clinical-testing results can be further biased by various types of selfselection [19,20]. Though it is impossible to precisely determine the relative contributions of these factors and biases, context can suggest which are likely to have the greatest influence in a given instance.…”

Pre-procedural testing improves estimated COVID-19 prevalence and trends

“…Ideally, to determine the latent prevalence of COVID-19 in a community, a large group of individuals would be sampled randomly and tested irrespective of factors positively or negatively associated with COVID-19. However, random sampling is logistically difficult under the constraints of a pandemic, and such studies are often affected by low participation rates and self-selection bias [3][4][5][6] . Prior studies have investigated alternative approaches to reduce bias in estimated prevalence, using mathematical models with randomized pooling [4][5]7 and online incentivized experiments to estimate prevalence based on key COVID-19 risk factors 3 .…”

“…However, random sampling is logistically difficult under the constraints of a pandemic, and such studies are often affected by low participation rates and self-selection bias 3-6 . Prior studies have investigated alternative approaches to reduce bias in estimated prevalence, using mathematical models with randomized pooling 4-5,7 and online incentivized experiments to estimate prevalence based on key COVID-19 risk factors 3 .…”