(160) Intention to treat and on treatment exposure definition in observational safety studies: a simulation study using real world data

Background: Different exposure definitions have been applied to observational safety studies using register data, like intention to treat (ITT), on treatment (OT), and time varying (TV). Although ITT is a commonly used approach it is controversial since it may bias the results towards the null.

Objectives: To explore the potential bias introduced by using the ITT and OT exposure definitions under variation of effect size, treatment adherence and latency period in simulated datasets based on real-world data.

Methods: We used data from the Swedish National Registers. Individuals exposed to biologic immunosuppressive treatments for any indication, from 2016 until 2022, were included with a new-user approach. All individuals with at least one filled prescription of tumour necrosis factor inhibitors (TNFi) or interleukin inhibitors (ILi) were eligible for inclusion. We considered 64 different scenarios varying the known effect size (true hazard ratios [HR]): 1.0, 1.3, 1.9, 2.8; the proportion of non-adherence: 2.5%, 5%, 10%, 20%; and the length of the latency period (time between exposure and disease onset/detection): 0, 30,180, 365 days. For each scenario, we used a modified plasmode approach to generate 100 simulated survival times for a hypothetical outcome for each of 50000 individuals.
In the ITT exposure definition, all individuals were assigned to the first treatment and followed from the beginning of treatment until the end of the study period (31 December 2022), emigration, the occurrence of the hypothetical outcome or death. In the OT exposure definition, individuals were censored at the end of medication supply or switch of treatment in addition to the ITT censoring events. HRs were estimated using Cox proportional hazards regression models for ILi versus TNFi adjusted for age, sex, diabetes and history of hypertension with 95% confidence intervals estimated by bootstrapping.
The TV approach which may be more suitable to estimate HR in safety studies will be added to the study.

Results: With the ITT exposure definition, average HR (min - max) were 1.0 (1.00 - 1.01), 1.3 (1.30 - 1.31), 1.9 (1.89 - 1.90), and 2.77 (2.76 - 2.79) among simulated datasets with true HR of 1.0, 1.3, 1.9, 2.8, respectively. With the OT approach, average HR (min - max) were 1.02 (1.01 - 1.03), 1.33 (1.32 - 1.34), 1.94 (1.93 - 1.96) and 2.85 (2.83 - 2.87), respectively. For both exposure definitions, we did not observe any variation in the HR with increasing non-adherence or increasing latency periods.

Conclusions: Estimates from the simulated datasets showed a small bias towards the null for larger true HR using the ITT exposure definition. With the OT exposure definition, a small bias away from the null was observed with increasing true HR, but with wider confidence intervals, as expected.