Evaluating methodological approaches to assess the severity of infection with SARS-CoV-2 variants: scoping review and applications on Belgian COVID-19 data

Abstract

Background Differences in the genetic material of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) variants may result in altered virulence characteristics. Assessing the disease severity caused by newly emerging variants is essential to estimate their impact on public health. However, causally inferring the intrinsic severity of infection with variants using observational data is a challenging process on which guidance is still limited. We describe potential limitations and biases that researchers are confronted with and evaluate different methodological approaches to study the severity of infection with SARS-CoV-2 variants. Methods We reviewed the literature to identify limitations and potential biases in methods used to study the severity of infection with a particular variant. The impact of different methodological choices is illustrated by using real-world data of Belgian hospitalized COVID-19 patients. Results We observed different ways of defining coronavirus disease 2019 (COVID-19) disease severity (e.g., admission to the hospital or intensive care unit versus the occurrence of severe complications or death) and exposure to a variant (e.g., linkage of the sequencing or genotyping result with the patient data through a unique identifier versus categorization of patients based on time periods). Different potential selection biases (e.g., overcontrol bias, endogenous selection bias, sample truncation bias) and factors fluctuating over time (e.g., medical expertise and therapeutic strategies, vaccination coverage and natural immunity, pressure on the healthcare system, affected population groups) according to the successive waves of COVID-19, dominated by different variants, were identified. Using data of Belgian hospitalized COVID-19 patients, we were able to document (i) the robustness of the analyses when using different variant exposure ascertainment methods, (ii) indications of the presence of selection bias and (iii) how important confounding variables are fluctuating over time. Conclusions When estimating the unbiased marginal effect of SARS-CoV-2 variants on the severity of infection, different strategies can be used and different assumptions can be made, potentially leading to different conclusions. We propose four best practices to identify and reduce potential bias introduced by the study design, the data analysis approach, and the features of the underlying surveillance strategies and data infrastructure.

Surfing the Waves: Differences in Hospitalised COVID-19 Patients across 4 Variant Waves in a Belgian University Hospital

“…We are confident that the Belgian national surveillance data reflects our catchment area’s epidemiology, as shown in another project published by our group recently [ 29 ]. We were also reassured to read a recent paper by Meurisse et al [ 11 ], who found that exposure ascertainment based on time periods leads to the most precise results when studying severity per variant wave. We, therefore, chose periods in which each variant was circulating at levels of at least 80%, based on national COVID-19 surveillance data.…”

“…The inclusion periods were defined as the periods in which one particular strain of the SARS-CoV-2 virus accounted for 80% or more of the circulating viruses in Belgium [ 11 ]. We identified four periods of interest, shown below, based on the Belgian epidemiology; the Beta VOC did not reach 80% dominance at any point and was therefore not included.…”

Homologous and Heterologous Prime-Boost Vaccination: Impact on Clinical Severity of SARS-CoV-2 Omicron Infection among Hospitalized COVID-19 Patients in Belgium

“…First, in the main analysis, the variant of infection was assigned based on linkage with WGS-confirmed lineage data from the COVID-19 TestResult database. This might result in selection bias and drastically limits the sample size [ 43 ]. The small sample sizes reduce the statistical power and the reliability of significant findings [ 53 ].…”

“…However, the underlying missing data mechanism is unknown. Sixth, different selection steps were performed to obtain data of our study population (i.e., selection of patients tested for COVID-19, treated in the hospital for COVID-19 symptoms, registered in the CHS, and with an Omicron BA.1, BA.2, or BA.4/5 infection identified by linking WGS-confirmed lineage data) [ 43 ]. Hence, patients in our study population might have specific characteristics distinctive from the population of all hospitalized patients in Belgium and the general population, thereby limiting generalization.…”

“…Patients without a linked WGS-confirmed lineage or with a linked WGS-confirmed lineage different from Omicron BA.1, BA.2, BA.4 or BA.5 were excluded from the study population in the main analysis. As linkage with the non-exhaustive WGS test results limited the sample size significantly (only for 11% of the eligible patients a WGS-confirmed lineage could be linked) and might have resulted in selection bias [ 43 ], a sensitivity analysis was performed where patients were categorized based on their COVID-19 diagnosis date within time periods with known circulation of the Omicron sub-lineages [ 33 ]. Patients diagnosed with COVID-19 during a restricted time period with an approximated circulation of BA.1, BA.2 or BA.4/5 of more than 80%, were assigned to the BA.1, BA.2 or BA.4/5 Omicron sub-lineage, respectively.…”