In recent years, data collected in routine practice by General Practitioners (GP) have become more widely used to investigate the safety and effectiveness of drugs. However, in some cases, the results are in conflict with results from more traditional studies such as randomized trials. This highlights the need to explore potential issues and biases in the analysis of routinely collected health data.
Two key issues that may be responsible for these inconsistent results are missing data and treatment switching. GPs record only health information relevant to the care of the patient, thus information required to fully address the research question may not always be available, resulting in missing data. Treatment switching refers to a patient swapping prescriptions from one treatment to another, which complicates the comparison between patients on the different treatments.
Using a recent example in CPRD where the results were inconsistent with randomized trial results, we aim to apply novel statistical methods to handle the aforementioned problems. We can then better understand when it is relevant to take account of such characteristics of the data and provide guidance regarding the relative benefits of different methods of analysis for future studies.
A recent CPRD cohort study suggested between person confounding remained a problem when investigating a potential interaction between PPIs and clopidogrel, as biologically implausible harmful associations were observed. Results from a self-controlled case series (SCCS) on the same data showed no increased risk of MI with PPI exposure and were thought to be more reliable as, in this analysis, fixed between person confounding is removed by design. SCCS limitations mean a more general solution is needed.
We identified treatment (PPI) switching and exclusion of potentially important confounders due to missing data as key issues.
To investigate treatment switching, we would perform "intent-to-treat", "per-protocol" and "as-treated" analyses using Cox models, incorporating probability weights accounting for participant differences between those who did and didn't change exposure status during follow-up. We will extend this idea by using more complex approaches such as marginal structural models, splitting data into, for example, 3-month intervals.
For missing data, we would initially incorporate information from confounders, previously omitted using missing categories approaches and then proceed to multiple imputation based analyses in the different analysis settings outlined above.
Finally, we would investigate methods to incorporate both issues.
Through this work we seek only to improve methodological approaches in future studies, not answer additional clinical questions.
Myocardial Infarction
- All-cause mortality
- Myocardial Infarction or all-cause mortality
Ian Douglas - Chief Investigator - London School of Hygiene & Tropical Medicine ( LSHTM )
John Tazare - Corresponding Applicant - GSK
Elizabeth Williamson - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
John Tazare - Collaborator - GSK
Liam Smeeth - Collaborator - London School of Hygiene & Tropical Medicine ( LSHTM )
MINAP