Assessing the internal and external validity of confidence interval calibration in observation studies

Date of ISAC Approval: 
21/04/2017
Lay Summary: 
Studies using existing data such as electronic health records could give erroneous results because people undergo different treatments for a reason, and so when comparing treatment groups we could be comparing apples to oranges. This limitation of using the data is often acknowledged in studies with a single sentence, but not considered further. We think this is one of the causes for the many conflicting study results in the field, and propose to measure how accurate studies are. For this we use controls, which are questions where we know the answer, and see whether the study setup produces the correct answer. We want to go even further and calibrate the result of the study to take into account the deviation that we observe for the controls. To demonstrate this approach we want to reproduce two previous studies that found conflicting results, and use controls to evaluate and calibrate these studies. We first want to show that our calibration procedure corrects the results for our controls. Next, we would like to show that the calibration makes the two studies more in agreement.
Technical Summary: 
This study will include two analyses investigating the risk of upper Gastrointestinal (GI) bleeding during exposure to selective serotonin reuptake inhibitors (SSRIs), intended to replicate the analyses performed by Tata et al.1 as closely as possible. The first analysis is a matched case-control study; the second analysis is a self-controlled case series (SCCS). The same analyses used to create estimates for upper GI bleeding will also be used to produce estimates for a set of 50 negative control outcomes which are not believed to be caused by SSRIs and where the true relative risk is therefore believed to be one. Additional estimates will be generated for 3x50 synthetic positive controls derived from the negative controls. Positive controls will be generated by fitting outcome models for each negative control, and inserting additional outcomes based on the predicted probabilites of patients for the outcomes, thus preserving any observed confounding stucture while increasing the true relative risk to three predefined levels: 1.5, 2, and 4. Based on the estimates for negative and positive controls a model of systematic error will be fitted, and confidence intervals will be calibrated. This study will use the CPRD in Common Data Model as described elsewhere
Health Outcomes to be Measured: 
- Upper Gastrointestinal bleeding
Collaborators: 

Dr Martijn Schuemie - Chief Investigator - Janssen US
Dr Martijn Schuemie - Corresponding Applicant - Janssen US
Dr Patrick Ryan - Collaborator - Johnson & Johnson (JnJ - USA)