We thank Lai et al. (1) for the attention given to our article on the effect of metformin on lung cancer incidence (2). We take this opportunity to address an important methodological concern with many of the observational studies published on this topic to date, namely time-related bias.
The authors’ own study used the Taiwanese National Health Insurance databases to identify a cohort of 19,624 patients treated with antidiabetic drugs (3). It found that the use of metformin was associated with a significant 45% (hazard ratio 0.55; 95% CI 0.37–0.82) reduction in the incidence of lung cancer and an identical reduction with the use of thiazolidinediones. The study design is regretfully subject to immortal time bias (4). This bias was introduced in the study by misclassifying the time during which patients were not yet exposed to metformin as exposed, time during which they could not have lung cancer. To account for this immortal time, the analysis should consider the users of metformin as unexposed until they actually receive their first metformin prescription and use statistical methods for time-varying exposures. In a review article of this question, we in fact performed a crude calculation to illustrate the bias of this study and found that with an average delay of only 9.5 months between cohort entry and the start of metformin therapy, the proper classification of exposure changes the hazard ratio of 0.55 to a hazard ratio of 1.0, indicating no reduction of any kind in the incidence of lung cancer (5). The authors are invited to conduct this corrected analysis and report back.
With respect to the meta-analysis by Noto et al. (6), the pooled data suggesting a reduction in lung cancer with metformin is dominated by the study by Libby et al. (7) (63% of the weight), which is affected by time-lag bias. This bias was caused by the fact that nonusers of metformin who should have been excluded because of a previous cancer were in fact used and counted as outcome events, thus artificially increasing the rate of cancer in the nonuser group (5). It is noteworthy in this meta-analysis that the three randomized trials that do not have such biases, namely the UK Prospective Diabetes Study (UKPDS), A Diabetes Outcome Progression Trial (ADOPT), and the Rosiglitazone Evaluated for Cardiac Outcomes and Regulation of Glycemia in Diabetes (RECORD) study, find no effect of metformin on either cancer incidence or mortality (6).
In all, time-related biases, which are straightforwardly avoidable with a proper design and analysis, are regrettably frequent in observational studies of diabetes research (5). They have led to substantial, yet misleading, underestimates of the effect of metformin on cancer incidence and mortality.
Acknowledgments
No potential conflicts of interest relevant to this article were reported.
