Evaluation of time-related bias with non-user control

Abstract

Background: In observational studies estimating the association between treatment and time-to-event outcomes, time-related biases can substantially impact results. Immortal time bias is one of such biases, and two types are known: misclassified immortal time bias and excluded immortal time bias. These biases often arise from incorrect time-zero definition, especially with non-user controls. This study aims to illustrate immortal time bias in non-user controls using formulas, simulations, and real-world data.

Methods: For our simulations, we considered two scenarios: one with no confounding and no treatment effect, and the other with time-dependent confounding. We compare three different settings of time-zero for treatment and control groups. Method 1: Both groups were followed from cohort entry date (CED). Method 2: The treatment group was followed from treatment initiation date (TID), while the non-user group was followed from CED. Method 3: The treatment group was followed from TID, and non-users were matched to treatment patients, followed from the corresponding TID of their matched patient.

Results: Our simulation showed that both Method 1 and Method 2 can exhibit large biases in the estimated treatment effect due to immortal time bias. The magnitude of the bias is greater for Method 1 than for Method 2. On the other hand, Method 3 showed almost no bias. Even in the presence of time-dependent confounding, Method 3 did not introduce bias.

Conclusions: To reduce time-related biases, it is crucial for researchers to carefully define an appropriate time-zero, especially when using a non-user control group.