Methods/Statistics
Comparison of immortal time bias across four time-zero settings Hiroya Morita* Hiroya Morita Masataka Taguri
Introduction
In epidemiological research, time-related biases can have a substantial impact on the results. One prominent example of such a bias is immortal time bias, for which various strategies have been proposed to address. One such approach is the prevalent new user design (PNUD), which can be used when non-users are selected as the comparison group. In this method, matching is performed between the non-user group and the treatment group based on the time elapsed since the cohort entry date (CED).
Purpose
In the PNUD, individuals who will eventually receive treatment could be matched as non-users. An inappropriate method is to exclude those subjects from being matched as the non-user, as this approach cannot eliminate bias. We call this method inappropriate PNUD. The aim of this study is to compare scenarios through simulation: inappropriate PNUD and another in which immortal time bias arises without using PNUD, in order to assess the degree of bias generated when PNUD is used incorrectly.
Method
For simplicity, our simulation is based on the situation that there is no confounding and no treatment effect. We compare four different settings of time-zero for treatment and control groups: (1) CED vs CED (misclassified immortal time bias), (2) treatment initiation date (TID) vs CED (excluded immortal time bias), (3) using an inappropriate PNUD, and (4) using appropriate PNUD.
Results:
Our simulation shows that both methods (1) and (2) exhibit large apparent preventive effects of the treatment due to immortal time bias. The magnitude of the bias is greater for method (1) than for method (2). In method (3), a bias similar in magnitude to that of method (2) was observed. Conversely, method (4) shows almost no bias. (Figure)
Discussion/Conclusion:
When PNUD is used inappropriately, it fails to eliminate the bias, highlighting the importance of using the method correctly.
