Primary Submission Category: Applications in Health and Biology
Positive and negative control outcomes to inform target trial emulations with observational data: an application to diabetes medications in the BESTMED Consortium
Authors: Lucia Petito, Emma Hegermiller, Indhumathy Chelliah, Golnaz Loftian, Ryan Carnahan, Satyender Goel, Alan Kaul, Andrea DeVries, Cecilia Lansang, Marie McDonnell, Vinit Nair, Elisa Priest, Vincent Willey, Alexander Turchin, Miguel Hernan,
Presenting Author: Lucia Petito*
Although emulating a target trial precludes many design-induced biases, confounding due to noncomparability of treatment groups on unmeasured factors remains a concern. Employing control outcomes—outcomes for which the magnitude of the treatment effect is known and the confounding structure is similar to the one for the main outcome of interest—may help identify whether residual confounding exists after adjustment for measured factors. Here, we describe the use of control outcomes for confounding assessment in a study of the effect of second-line treatments for type 2 diabetes (dipeptidyl peptidase 4 inhibitors; glucagon-like peptide-1 receptor agonists; sodium-glucose cotransporter-2 inhibitors; sulfonylureas [SU]; basal insulin) on the 3-year risk of cardiovascular events using data on 57,910 individuals from the BESTMED Consortium (http://www.bestmed.org). We study 2 control outcomes: 3-month risk of cardiovascular events (negative control outcome) and 12-month change in hemoglobin A1c (positive control outcome). An outcome that does not share the same confounding structure, 3-year risk of herpes zoster, is considered as it has been used in previous falsification analyses. We conclude that selection of control outcomes with similar confounding structures is critically important to identify the presence of substantial confounding after adjustment for the information available in the observational database.