The safety of angiotensin-converting enzyme inhibitors (ACEI), a class of antihypertensive drugs, regarding the risk of lung cancer has been recently questioned. Studies investigating the association between ACEI and the risk of lung cancer have found mixed results. A recent study by Hicks et al. (2018) has found an increase of the risk of lung cancer. As the ACEI is a commonly prescribed drug class, thorough investigation of its safety is essential.
We can compare the risk among patients which have been prescribed the drug of interest to a comparison group. Differences in risk of the event between the two groups may indicate an effect of the drug on that event.
However, the two groups (the one prescribed the drug of interest and the comparison group) might be different. Not considering this difference can lead to bias in estimating the association between ACEI and the risk of lung cancer.
We will therefore consider these potential differences to estimate the association between ACEI and risk of lung cancer.
The main challenge of such a study comes from the fact that electronic health record (EHR) contains a huge quantity of variables that record the medical history of patients. A comparison of existing approaches to analyse such data will be done, with the optimal method subsequently applied in the analysis.
A recent population-based cohort study by Hicks et al. (2018) using CPRD data has found an association between ACEI prescription and the risk of lung cancer. However, confounding bias is a major problem when estimating the association between drug prescription and the outcome of interest. We propose a number of analyses to explore, quantify and where possible, remove, that bias:
(i) Applying standard PS analyses and High-Dimensional Propensity Score (HDPS) or modified versions, as described below, to account for confounding
(ii) Negative controls outcomes
(iii) Sensitivity analyses to explore confounding by smoking and detection bias (“treatment” effects mediated through early diagnosis due to ACEI effects on cough).
To achieve this, a subsidiary goal is to assess the optimal statistical methodology for conducting propensity score analyses in EHR using high-dimensional confounder adjustment. While the HDPS is gaining in popularity, a number of other approaches exist and may perform better in certain settings.
The HDPS consists of four steps: (1) creating a set of candidate confounder variables from patient codes, (2) selecting relevant confounder variables from this pool, (3) fitting a model relating the drug prescription to the selected confounder variables to estimate the PS, (4) estimating the treatment effect using the PS. We will evaluate different statistical methods for achieving the second and third steps and comparing them both in these data and in a plasmode simulation study based on these data (Franklin et al. 2014). For the second step, approaches based on the least absolute shrinkage and selection operator (lasso) and the elastic net (Franklin et al. 2015) and others based on dimensionality reduction (Schneeweiss et al. 2017) exist. For the third step, techniques proposed in the literature include neural networks, generalized boosting, ridge regression, random forest, CART (Franklin et al. 2015, Schneeweiss et al. 2018, Karim et al. 2018).
We are interested in evaluating the association between prescription of ACEI and risk of lung cancer.
Primary outcome: incident lung cancer (diagnosis or death with lung cancer listed as a cause if no prior diagnosis)
Secondary outcomes (negative control outcomes):
- Lung related: chronic obstructive pulmonary disease (COPD); sarcoidosis; asbestosis
- Not lung related: fracture (any); herpes zoster; incident colon cancer
It is likely that the primary question of interest will be heavily confounded, in particular by inadequate adjustment for smoking, thus a number of negative and positive controls will be used to help identify residual confounding. For the lung-related negative controls, we expect these to be strongly associated with smoking status but there is less clear rationale for believing a causal effect of ACEI prescription exists (particularly asbestosis). Thus, apparent treatment effects would suggest inadequate adjustment for smoking. Fracture is included as a general marker of frailty; colon cancer is unlikely to be affected by ACEI prescription but smoking and other lifestyle factors are risk factors, again allowing us to pick up signals of residual confounding by these characteristics.
Elizabeth Williamson - Chief Investigator - London School of Hygiene & Tropical Medicine ( LSHTM )
Elizabeth Williamson - Corresponding Applicant - London School of Hygiene & Tropical Medicine ( LSHTM )
Angel Chang - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Caroline Morton - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Ian Douglas - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
James Carpenter - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Jessica Caterson - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
John Tazare - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Krishnan Bhaskaran - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Laurie Tomlinson - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Liam Smeeth - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Paris Baptiste - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Patrick Bidulka - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Sean Maguire - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Yi Ting Nikki Yu - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Caroline Morton - Collaborator - Queen Mary University of London
David Turner - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
Fatima Batool - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
HES Admitted Patient Care;ONS Death Registration Data;Patient Level Index of Multiple Deprivation;Practice Level Index of Multiple Deprivation