Can instrumental variable analysis overcome confounding by indication? Validation of physicians' prescribing preferences as instrumental variables.

Aims:
We will provide novel empirical and simulated evidence about the validity of physicians' prescribing preferences as instrumental variables and guidance about their interpretation and use to patients, clinicians, regulators and researchers.

In recent studies we have found evidence that instrumental variable analysis of electronic medical records can produce results comparable to a meta-analysis of randomised controlled trials (RCTs) for a fraction of the cost [13]. This suggests that instrumental variable analyses of observational studies could provide a useful additional source of reliable evidence.

Instrumental variable analysis may therefore be an extremely useful tool for pharmacoepidemiologists to investigate the effects of medications using observational data. Our results to date suggest that it is possible to obtain evidence as precise as all published experimental studies combined using an observational dataset. The combined investment in all 44 randomised studies included in our meta-analysis is likely to be orders of magnitude greater than the cost of conducting the analysis reported in Thomas et al. (2013). Furthermore, resources like the CPRD are continually accruing new data, both due to existing General Practices adding further years of records and because additional General Practices are added. Therefore, over time more observational data are becoming available which will greatly increase the precision of observational estimates. Eventually, this will allow researchers to address even more refined hypotheses such as the effect of treatments within pre-specified sub-groups of the population.

The results of this project will be written up as two papers. The first will present the results of the empirical negative control studies and the second will describe the simulation study. We anticipate both papers will merit publication in high impact medical or statistical journals, such as Statistics in Medicine and Epidemiology. We will present the results of these studies at two international conferences (specifically the International Society for Pharmacoepidemiology in Dublin 2016 and the Society for Epidemiologic Research in Miami in 2016). We have requested resources for these conference presentations and open access publication fees.