Investigating biases in observational studies of inhaled corticosteroids and the risk of COVID-19-related outcomes

During the COVID-19 pandemic, there was an urgent need to understand whether existing medications could influence infection risk and COVID-19 disease severity. Many pharmacoepidemiologic studies were rapidly conducted but there were concerns about biases impacting the findings. Inhaled corticosteroids (ICS) are widely used maintenance medications in chronic obstructive pulmonary disease (COPD) which were investigated as potential disease-modifying drugs in COVID-19. Randomised controlled trials (RCTs) suggested a protective effect of one ICS, budesonide, in patients with mild COVID-19, whereas observational studies found conflicting results. It was unclear whether these results were due to differing study questions and populations, or whether results were affected by biases. Little was known about how biases could explain these inconsistencies or how bias analysis methods could be applied to improve pandemic pharmacoepidemiologic studies. The aim of this thesis is to investigate biases affecting pharmacoepidemiologic studies of COVID-19 using an applied example of ICS and COVID-19 outcomes among people with COPD. I estimated effects of ICS among patients with COPD at different stages of the COVID-19 disease pathway (i.e., ranging from susceptible individuals to those with severe COVID-19) and accounted for potential biases arising in these effect estimates during the first pandemic wave. All analyses were based on a cohort of patients with COPD defined from UK electronic health records (EHRs) in the Clinical Practice Research Datalink. I compare ICS/long-acting β-agonist (LABA) users to LABA/long-acting muscarinic antagonist (LAMA) users. Given results from RCTs and observational studies, I did not expect to see a causal harm of ICS in patients with COPD in the studies conducted in this thesis. I investigated the potential impact of outcome misclassification on the association between ICS and COVID-19 hospitalisation and death using probabilistic bias analysis (PBA). Models not accounting for outcome misclassification showed increased odds of COVID-19 hospitalisation and death among ICS users. Accounting for outcome misclassification modestly shifted ORs away from the null and increased uncertainty, but was unlikely to change study conclusions. However, results raised concerns about the potential impact of residual confounding. To account for this, I applied high-dimensional propensity scores (HDPS). HDPS aim to control for residual confounding by considering the data contained in the EHR as proxy variables. Using the same COPD cohort, I estimated propensity scores combining prespecified covariates with up to 1,000 HDPSderived covariates derived from primary care observations, prescriptions and hospitalisations. Conventional propensity score-weighted analyses excluding triple therapy users showed weak evidence of increased COVID-19 hospitalisation and death risks among ICS users. For COVID-19 deaths, HDPS-weighted estimates were consistently closer to the null, while for hospitalisations, results varied slightly based on the number of covariates included in HDPS. Finally, I used simple bias analysis to correct for differential selection in an analysis of COVID-19 death among hospitalised patients, aiming to make inference about patients with severe COVID-19. Due to hospital pressures, not everyone with severe COVID-19 was hospitalised, potentially causing selection bias when restricting to hospitalised patients. In those hospitalised for COVID-19, there was no evidence of an increased risk of COVID-19 deaths among ICS users. QBA across four plausible scenarios varying assumed death rates among non-hospitalised patients yielded corrected ORs between 0.81 and 1.28. All four results remained consistent with the null hypothesis, though confidence intervals were wide. In summary, I applied several approaches to deal with biases that threaten the validity of pharmacoepidemiologic studies. Misclassification and selection bias were unlikely to have changed study conclusions, and confounding was identified as the primary concern in this case. These findings demonstrate the value of systematic and comprehensive bias assessment in pharmacoepidemiologic research and provide methodological guidance for future studies. Careful consideration of mechanisms causing bias is important to appropriately address biases. Methods to account for biases should be integrated into workflows now to improve real-world evidence generation in preparation for future pandemics.