BACKGROUND: Observing numerous events on day 0 of follow-up is common in studies of the association between diagnosed coronavirus disease 2019 (COVID-19) and adverse events (AE). Starting follow-up on day 0 may induce bias when evaluating some diseases or time periods, such as the COVID-19 pandemic.
OBJECTIVE: To highlight potential selection bias on day 0 in a cohort study that examined the hazard of AEs following COVID-19 diagnosis.
METHODS: This cohort study assessed the hazard of AEs after COVID-19 diagnosis (study period, 1 Apr 2020-10 Dec 2020) in 2 United States claims data sources—Merative TM MarketScan® Commercial Database (18-64 years), and Medicare Fee-For-Service (65+ years). Individuals with a COVID-19 diagnosis were matched 1:1 to a comparator on several factors, including hospitalization status on day 0. Inverse probability of treatment-weighted hazard ratios (HRs) and 95% confidence intervals (CI) were calculated both including and not including day 0 in follow-up. Multiple AEs were evaluated; results for nonhemorrhagic stroke (NHS) and myocarditis/pericarditis (MYO/PER) are presented as examples. Subgroup analyses of individuals hospitalized on day 0 were used to illustrate the effect of matching on hospitalization.
RESULTS: Starting follow-up on day 0 or day 1 resulted in HRs for NHS (Medicare) of 0.89 (95% CI, 0.86-0.93) and 1.35 (95% CI, 1.28-1.43), respectively; 60% of all comparator NHS events occurred on day 0 compared with 39% of the COVID-19 group. In subgroup analyses starting follow-up on day 0, the HR for NHS was 0.54 (95% CI, 0.51-0.57) for participants hospitalized on day 0 and 1.67 (95% CI, 1.56-1.78) among those not hospitalized on day 0. In contrast, the HRs for MYO/PER (MarketScan) starting follow-up on day 0 or day 1 were 6.20 (95% CI, 4.46-8.62) and 4.51 (95% CI, 3.15-6.45), respectively, with more MYO/PER events in the COVID-19 group occurring on day 0 (34% vs. 11%).
CONCLUSIONS: For AEs typically treated in hospitals (e.g., NHS), most AEs in the comparator group occurred on day 0, suggesting that matching on hospitalization during the pandemic selected comparators more likely to be admitted due to emergent, non-COVID-19 conditions (e.g., NHS) than the COVID-19 group, resulting in paradoxically protective HRs when day 0 was included in follow-up. Additionally, the recording of all hospital diagnoses on the admission date may obscure the timing of events. Conversely, for AEs not required to be hospitalized (e.g., MYO/PER), more AEs in the COVID-19 group occurred on day 0, a potential form of reverse causality where care seeking for the AE prompted COVID-19 testing, resulting in inflated HRs when day 0 was included. Considering day 1 as the first day of follow-up in cohort studies may reduce induced bias but at the cost of complete case ascertainment.
