Using probabilistic bias analysis to account for outcome misclassification of COVID-19 death in pharmacoepidemiologic studies

Background: Early in the pandemic, there was concern about underascertainment of COVID-19 deaths and its impact on treatment effects in pharmacoepidemiologic studies.

Objectives: We assessed the impact of outcome misclassification in an analysis of inhaled corticosteroids (ICS) on the risk of COVID-19 death in the UK during the first pandemic wave using probabilistic bias analysis (PBA).

Methods: Using data from Clinical Practice Research Datalink Aurum linked to Hospital Episode Statistics data and Office for National Statistics death registry, we defined a cohort with chronic obstructive pulmonary disease (COPD) on 01 Mar 2020 (i.e., baseline). We compared the risk of COVID-19 death (ICD-10 codes U07.1/U07.2 as a cause of death on the death certificate) among current users of ICS/ long-acting β-agonist (LABA) (+/- long-acting muscarinic antagonist (LAMA)) and users of LABA/LAMA using inverse-probability of treatment weighted (IPTW) logistic regression. Follow-up continued until death, deregistration, or 31 Aug 2020, whichever came first. We used PBA to assess the impact of outcome misclassification, assuming non-differential misclassification. Based on the literature, we assigned beta distributions to reflect our assumptions about sensitivity and specificity and sampled from these 20,000 times for record-level PBA. Using these values, we simulated a new outcome variable and applied IPTW logistic regression models to adjust for both confounding and misclassification. In a sensitivity analysis, we excluded ICS+LABA+LAMA (triple therapy) users as they tend to be sicker than dual therapy users.

Results: Of 161,411 patients with COPD, 35.4% were prescribed ICS/LABA and 13.8% LABA/LAMA at baseline. 366 deaths with COVID-19 occurred in the study period, 294 among ICS users. ICS users had increased odds of COVID-19 death compared with LABA/LAMA users (OR 1.63, 95% CI 1.26 – 2.11). After IPTW, the OR moved towards the null (OR 1.42, 95% CI 1.11 – 1.81), as it did after exclusion of people using triple therapy (OR 1.24, 95% CI 0.88 – 1.74).
After record-level PBA, the OR modestly increased in the primary model (IPTW-OR 1.55 (95% simulation interval (SI) 1.09 – 2.54) and in the model excluding ICS users on triple therapy (IPTW-OR 1.26, 95% SI 0.78 – 2.10) relative to models without PBA.

Conclusions: PBA shifted the OR away from the null while uncertainty increased, as expected given the heuristic that non-differential misclassification biases effect estimates towards the null on average, and the adjustment for more sources of error. Exclusion of triple therapy users suggests there may have been residual confounding by indication. These results provide some reassurance that outcome misclassification was unlikely to change the conclusions of the study but confounding by indication remains a concern.