Motivated by recent reports on associations between diabetes and cancer, many researchers have used administrative databases to examine risk association of cancer with drug use in patients with diabetes. Many of these studies suffered from major biases in study design and data analysis, which can lead to erroneous conclusions if these biases are not adjusted. This article discusses the sources and impacts of these biases and methods for correction of these biases. To avoid erroneous results, this article suggests performing sensitivity and specificity analysis as well as using a drug with a known effect on an outcome to ascertain the validity of the proposed methods. Using the Hong Kong Diabetes Registry, we illustrated the impacts of biases of drug use indication and prevalent user by examining the effects of statins on cardiovascular disease. We further showed that 'immortal time bias' may have a neutral impact on the estimated drug effect if the hazard is assumed to be constant over time. On the contrary, adjustment for 'immortal time bias' using time-dependent models may lead to misleading results biased towards against the treatment. However, artificial inclusion of immortal time in non-drug users to correct for immortal time bias may bias the result in favour of the therapy. In conclusion, drug use indication bias and prevalent user bias but not immortal time bias are major biases in the design and analysis of drug use effects among patients with diabetes in non-clinical trial settings.
© 2011 Blackwell Publishing Ltd.