Statistical methods for elimination of guarantee-time bias in cohort studies: a simulation study

Aspirin has been considered to be beneficial in preventing cardiovascular diseases and cancer. Several pharmaco-epidemiology cohort studies have shown protective effects of aspirin on diseases using various statistical methods, with the Cox regression model being the most commonly used approach. However, there are some inherent limitations to the conventional Cox regression approach such as guarantee-time bias, resulting in an overestimation of the drug effect. To overcome such limitations, alternative approaches, such as the time-dependent Cox model and landmark methods have been proposed. This study aimed to compare the performance of three methods: Cox regression, time-dependent Cox model and landmark method with different landmark times in order to address the problem of guarantee-time bias.

Through statistical modeling and simulation studies, the performance of the above three methods were assessed in terms of type I error, bias, power, and mean squared error (MSE). In addition, the three statistical approaches were applied to a real data example from the Korean National Health Insurance Database. Effect of cumulative rosiglitazone dose on the risk of hepatocellular carcinoma was used as an example for illustration.

In the simulated data, time-dependent Cox regression outperformed the landmark method in terms of bias and mean squared error but the type I error rates were similar. The results from real-data example showed the same patterns as the simulation findings.

While both time-dependent Cox regression model and landmark analysis are useful in resolving the problem of guarantee-time bias, time-dependent Cox regression is the most appropriate method for analyzing cumulative dose effects in pharmaco-epidemiological studies.