4G-06 - Quantifying selection bias due to hospital pressures in pharmacoepidemiological studies of COVID-19

Background: During the first wave of the COVID-19 pandemic there were significant pressures on hospitals, leading to some patients with severe COVID-19 never being admitted and dying outside of hospitals. This might result in selection bias in studies of hospitalised patients.

Objectives: To quantify the effects of potential selection bias in a study of inhaled corticosteroids (ICS) and COVID-19 death among people with COPD hospitalised with COVID-19.

Methods: Using data from Clinical Practice Research Datalink Aurum linked to hospitalisation and death registries, we defined a cohort with chronic obstructive pulmonary disease (COPD) on 01 Mar 2020. We assessed the odds of COVID-19 death (ICD-10 U07.1/U07.2) among those hospitalised with a primary diagnosis of COVID-19, comparing current users of ICS/long-acting β-agonist (LABA) and users of LABA/long-acting muscarinic antagonist (LAMA)). The target population is people with COVID-19 severe enough to warrant hospitalisation. Because some people in the target population were never hospitalised, the estimated treatment effect is at risk of selection bias which can be evaluated using quantitative bias analysis (QBA). We explored the impact of this bias under various selection probabilities by exposure and outcome, estimated by assuming death rates among the unobserved. Due to death registry linkage, selection probabilities for cases (deaths due to COVID-19) are known. In scenario 1, we assume equal death rates among the non-hospitalised and the hospitalised. In scenario 2, the death rate in the ICS group is doubled in the non-hospitalised compared to the hospitalised (i.e., the non-hospitalised were sicker than the hospitalised). In scenario 3, the rate in the ICS group is halved in the non-hospitalised compared to the hospitalised (i.e., the non-hospitalised were healthier than the hospitalised).

Results: 107 ICS users and 133 LABA/LAMA users were hospitalised with COVID-19 during the study period. In the ICS group, 42 (39.3%, odds = 0.65) died of COVID-19, compared with 50 (37.6%, odds = 0.60) in the LABA/LAMA group. The crude OR was 1.07 (95% CI 0.63 – 1.81), which remained similar after inverse probability of treatment weighting (OR 1.01 (95% CI 0.59-1.72)).
In scenario 1, as expected, the OR is unchanged (OR = 1.07). In scenario 2, the corrected OR is 1.28. In scenario 3, the corrected OR is 0.81. Results for a range of other assumptions will also be presented.

Conclusions: QBA is a simple method to assess sensitivity of study results to potential selection bias. Our findings indicate that selection probabilities would have needed to be substantially different in the treatment groups to change the study conclusions, which is considered unlikely given assumed similar selection pressures in the treatment groups.