Quantifying selection bias due to unobserved patients in pharmacoepidemiologic studies of severe COVID-19 cohorts

TL;DR

This study examines how missing data from unobserved severe COVID-19 patients may bias treatment effect estimates in COPD patients.

Contribution

The paper introduces a quantitative bias analysis to assess selection bias from unobserved patients in severe COVID-19 studies.

Findings

The odds ratio for ICS/LABA versus LABA/LAMA users remained near 1 across all scenarios.

Wide confidence intervals suggest uncertainty in the effect estimates due to potential selection bias.

Substantial differences in death rates among non-hospitalised patients would be needed to alter study conclusions.

Abstract

The COVID-19 pandemic caused hospital pressures resulting in some patients with severe COVID-19 not being admitted. Studies aiming to measure treatment effects in patients with severe COVID-19 might produce biased estimates if restricted to hospitalised cohorts as a subset of the target population remained unobserved.

To quantify the effects of potential selection bias due to deaths outside of hospital in a case study of inhaled corticosteroids (ICS) and COVID-19 death among people with chronic obstructive pulmonary disease (COPD) hospitalised with COVID-19.

Using Clinical Practice Research Datalink Aurum linked to hospitalisation and death registries, we defined a cohort with COPD on 01 Mar 2020, followed up until 31st August 2020. We assessed the odds of COVID-19 death (International Classification of Diseases, 10th Revision U07) among hospitalised COVID-19 patients, comparing current users of ICS/long-acting β-agonist (LABA) and LABA/long-acting muscarinic antagonist (LAMA)). Our target population was those with COPD and severe COVID-19. We evaluated potential selection bias due to non-admission of severe COVID-19 cases using quantitative bias analysis (QBA) in four plausible scenarios, varying assumed death rates among non-hospitalised patients. Selection probabilities for deaths due to COVID-19 were known. The assumptions were: (1) equal odds of death between non-hospitalised and hospitalised groups; (2) doubled odds of death in non-hospitalised ICS/LABA group compared to hospitalised; (3) halved odds of death in non-hospitalised ICS/LABA group; and (4) doubled odds of death in both treatment groups among non-hospitalised patients. We calculated bootstrapped 95% confidence intervals (CIs).

During the study period, 107 ICS/LABA users and 133 LABA/LAMA users were hospitalised with COVID-19. COVID-19 deaths occurred in 42 (39.3%) ICS/LABA users versus 50 (37.6%) LABA/LAMA users. The OR after inverse probability of treatment weighting was 1.01 (95% CI 0.59–1.72). In scenario 1, the OR was unchanged (OR 1.07, 95% CI 0.70–1.67). In scenario 2, the corrected OR was 1.28 (95% CI 0.83–2.00). In scenario 3, the corrected OR was 0.81 (95% CI 0.52–1.23). In scenario 4, the corrected OR was 1.08 (95% CI 0.69–1.71).

QBA facilitated an assessment of the sensitivity of study results to potential selection bias due to non-admission of a subset of patients with severe COVID-19. The results of the four scenarios presented are in line with the null hypothesis, but CIs were wide. Death rates in the non-hospitalised would have needed to be substantially different in the treatment groups to change the study conclusions.

The online version contains supplementary material available at 10.1186/s12874-025-02732-w.

Figures