Understanding the Challenges and Uncertainties of Seroprevalence Studies for SARS-CoV-2

Abstract

SARS-CoV-2 continues to widely circulate in populations globally. Underdetection is acknowledged and is problematic when attempting to capture the true prevalence. Seroprevalence studies, where blood samples from a population sample are tested for SARS-CoV-2 antibodies that react to the SARS-CoV-2 virus, are a common method for estimating the proportion of people previously infected with the virus in a given population. However, obtaining reliable estimates from seroprevalence studies is challenging for a number of reasons, and the uncertainty in the results is often overlooked by scientists, policy makers, and the media. This paper reviews the methodological issues that arise in designing these studies, and the main sources of uncertainty that affect the results. We discuss the choice of study population, recruitment of subjects, uncertainty surrounding the accuracy of antibody tests, and the relationship between antibodies and infection over time. Understanding these issues can help the reader to interpret and critically evaluate the results of seroprevalence studies.

Keywords: COVID-19; SARS-CoV-2; antibody testing; coronavirus; seroprevalence.

Figure 1

Total confirmed COVID-19 cases in Republic of Ireland as of 6 April 2021. Official COVID-19 case numbers (HPSC [11] 6 April 2021) are based on RT-PCR positive tests only (a small number of which may be false positives). The true number of infections also includes an unknown number of people who were either never tested, or else falsely tested negative.

Figure 2

Importance of representative sampling. In this population, the rate of prior infection is higher among women. As a result, Sample 1, which is predominantly women, overestimates the prevalence of prior infection among the wider population. By contrast the prevalence in Sample 2, which contains approximately equal numbers of men and women, is much closer to that of the wider population.

Figure 3

Point and interval estimates of seroprevalence from two studies, SCOPI [24] in Ireland and REACT2 [51] in England. Based on a sample of 913 people in Dublin, the SCOPI study estimated seroprevalence at 3.1%. The corresponding interval estimate indicates that the true value of seroprevalence in the wider population in Dublin was likely to lie between 2.1% and 4.5%. By contrast, the REACT2 study in England enrolled a large number of participants (ca. 100k), and thus the corresponding confidence interval for seroprevalence is narrow—greater sample sizes give more precise estimates (provided they are indeed representative of the population).

Figure 4

Sensitivity and specificity of a SARS-CoV-2 antibody test. In a sample of previously infected people, some will correctly test positive (true positives) while others will incorrectly test negative (false negatives)—sensitivity refers to the proportion of previously infected people who correctly test positive. Similarly, never-infected people may correctly test negative (true negatives), or incorrectly test positive (false positives)—specificity is the proportion of never-infected people who correctly test negative.

Figure 5

The outcome of testing 1000 people, 50 of whom have previously been infected, using an antibody test with 80% sensitivity and 94% specificity. The true prevalence is thus 5% (50 out of 1000), while the apparent prevalence, i.e., the proportion of tests that give a positive result, is 9.7%.

Figure 6

Estimated sensitivity and specificity of four different antibody tests. Tests A, B, and C are based on data presented on the US Food and Drug Administration website (for three different commercially available tests), Test D is for illustrative purposes. Estimated sensitivity and specificity for Test A are both high, with a high degree of certainty (i.e., narrow interval estimates). For Tests B and C, estimated sensitivity and specificity are more uncertain, while for Test D these values are definitely lower.

Figure 7

Estimates of seroprevalence from 10,000 simulated studies, using different antibody tests and adjusting results for imperfect sensitivity and specificity. Each simulation selects a random sample of 2500 participants from a population with an overall prevalence of 10%. The extent to which estimated prevalence differs from true prevalence on average depends on the test used—lower and/or more uncertain values of sensitivity and specificity result in more uncertain estimates of prevalence.