In search of the perfect model

Administrative data has obvious advantages: it is inexpensive, time-efficient, representative of real-world results and patients, and there are massive amounts of it available in various countries around the world; however, it can have potentially significant limitations when it is used for clinical research. Nayan et al investigated a common inadequacy — lack of adjustment for smoking status and obesity.1 Numerous studies have identified the importance of these two variables as risk factors for poor outcomes across various disease states. Often, administrative data does not have access to these kind of lifestyle variables; however, surrogates (such as a physician billing code for smoking cessation counselling or operative interventions on morbidly obese patients) or linked household survey results can be used to attempt to address some of this known residual confounding. The thought of prospective data collection to supplement an administrative data study is quite daunting, and thus often avoided if possible. In Nayan’s study, the addition of these two covariates made little difference to their multivariable models predicting mortality after kidney cancer