Impact of "time zero" of Follow-Up Settings in a Comparative Effectiveness Study Using Real-World Data with a Non-user Comparator: Comparison of Six Different Settings

Abstract

Background: Time-related bias can lead to misleading conclusions. Properly setting the "time zero" of follow-up is crucial for avoiding these biases. However, the time-zero setting is challenging when comparing users and non-users of a study drug because the latter do not have a time point for starting treatment.

Objective: This methodological study aimed to illustrate the impact of different time-zero settings on effect estimates in a comparative effectiveness study using real-world data with a non-user comparator.

Methods: Data for type 2 diabetes patients were extracted from an administrative claims database, and the onset of diabetic retinopathy (study outcome) was compared between users (treatment group) and non-users (non-use group) of lipid-lowering agents. We applied six time-zero settings to the same dataset. The adjusted hazard ratio (HR) for the outcome was estimated using a Cox regression model in each time-zero setting, and the obtained results were compared among the settings.

Results: Of the six settings, three (study entry date [SED] vs SED [naïve approach], treatment initiation [TI] vs SED, TI vs Matched [random order]) showed that the treatment had a reduced risk of the outcome (HR [95% CI]: 0.65 [0.61-0.69], 0.92 [0.86-0.97], and 0.76 [0.71-0.82], respectively), one (TI vs Random) had an increased risk (HR [95% CI]: 1.52 [1.40-1.64]) , and two (SED vs SED [cloning method], and TI vs Matched [systematic order]) had neither increased nor decreased risk (HR [95% CI]: 0.95 [0.93-1.13], and 0.99 [0.93-1.07], respectively).

Conclusions: This study demonstrates that different time-zero settings can lead to different conclusions, even if the same dataset is analyzed for the same research question, probably because improper settings can introduce bias. To minimize such biases, researchers should carefully define time zero, particularly when designing a non-user comparator study using real-world data.

Fig. 1

Study time frame for each time-zero setting method. The follow-up period started at time zero and ended when censored on the incidence of an outcome event, exit from the database, or the end of the study period, whichever occurred first. The time before time zero was the washout period for the outcomes; patients with an outcome event during the washout period were excluded from the analysis. SED study entry date (the date of the first prescription of a glucose-lowering agent plus the diagnostic record of type 2 diabetes in the same month), TI treatment initiation (first prescription date of lipid-lowering agents)

Fig. 2

Hazard ratio in each time-zero setting method. SED versus SED (naïve approach): time zero set at SED for both groups (naïve approach). SED versus SED (cloning method): time zero set at SED for both groups (cloning method). TI versus SED: time zero set at treatment initiation for the treatment group and SED for the non-use group. TI versus Random: time zero set at treatment initiation for the treatment group and at a randomly sampled hospital visit date for the non-use group. TI versus Matched (random order): time zero set at treatment initiation for the treatment group and at the matched date for the non-use group (matching in random order). TI versus Matched (systematic order): time zero set at treatment initiation for the treatment group and at the matched date for the non-use group (matching in systematic order). For SED versus SED (cloning method), only the “group” was included as an explanatory variable in the Cox regression model because covariate adjustment was already performed in the estimation of weights for the inverse-probability-of-censoring weighting. CI confidence interval, HR hazard ratio, SED study entry date (the date of the first prescription of a glucose-lowering agent plus the diagnostic record of type 2 diabetes in the same month), TI treatment initiation (first prescription date of lipid-lowering agents)