Misclassification bias in estimating clinical severity of SARS-CoV-2 variants – Authors' reply

We thank Christina Yek and colleagues for their Correspondence regarding our Article.¹ They note that people who test positive for SARS-CoV-2 generally have more severe disease than those who are infected but not tested. This finding could lead to the overestimation of absolute risks, but relative risks are not necessarily biased unless the proportion of detected severe cases differs systematically between variants. Citing modelling results that indicated a declining infection detection rate in the USA during the transition period between the dominance of the delta (B.1.617.2) and omicron (B.1.1.529) variants, possibly driven by increasing proportions of undetected infections in people with non-severe disease, Yek and colleagues hypothesise a mechanism for differential detection rates: the omicron cases for which a positive test result was recorded might have included a relatively higher proportion of infected people who were prone to severe disease than the analogous delta cases—for example, because a higher proportion of people infected with the omicron variant who sought testing had comorbidity.

However, available data do not suggest a change in the proportion of infections being detected in England by community PCR testing during the study period (although the extent of community testing was reduced later²). We believe that the UK is unique in having conducted large-scale, population-based COVID-19 prevalence surveys,³ alongside its mass testing programmes. To assess the hypothesis of Yek and colleagues, we compared estimates of infection prevalence in the population with estimates of the corresponding prevalence of infections detected through community testing (appendix pp 1–4). Contrary to the hypothesis, we found that community testing detected similar proportions of people infected with the virus during the delta-dominant and omicron-dominant periods in England (appendix pp 2–3).

Yek and colleagues further argue that the relative risks of all-cause outcomes might be closer to the null than those of COVID-19-specific outcomes. They suggest that in the absence of direct measurement of COVID-19-specific outcomes, other data could indirectly discriminate probable COVID-19-related and COVID-19-unrelated events. In principle, we agree that cause-specific event data are desirable. However, assuming a constant background rate of unrelated hospitalisations and deaths, differential misclassification of outcome events by variant is unlikely, and non-differential misclassification is more likely to result in bias towards than away from the null. Further, we note that, during the study period, all individuals admitted to hospital in England were tested for COVID-19 at admission, so missed hospitalisation events in individuals with undetected COVID-19 is unlikely. Several studies that reported relative risks of COVID-19-specific hospitalisation have estimated relative risks consistent with those from our study.4, 5

We acknowledge that our dataset did not include comorbidity data. However, recent studies in other European countries with comorbidity data available reported only minor differences in comorbidity between delta and omicron cases, and provided comorbidity-adjusted relative risks consistent with those from our study.4, 5, 6 One of these studies explored the effect of adjusting versus not adjusting for comorbidity and found only marginal differences.⁵

Taken together, we believe the available data indicate that it is unlikely that the proposed mechanisms have strongly biased the results of our analysis.