Modified intention-to-treat analysis did not bias trial results

Abstract

Objective

To investigate whether analysis of the modified intention-to-treat (mITT) population with postrandomization exclusion of patients from analysis is associated with biased estimates of treatment effect compared to the conservative intention-to-treat (ITT) population.

Study Design and Setting

Placebo-controlled, blinded randomized trials on biological or targeted interventions for rheumatoid arthritis were identified through a systematic search. Two authors independently extracted data. A random-effects meta-analysis was used to combine odds ratios as an expression of treatment effect and stratify according to the different analysis populations.

Results

Seventy-two randomized trials were included and analyzed (23,842 patients). Thirty trials analyzed the ITT population, 37 analyzed an mITT population, and 5 trials had an unclear analysis population. The treatment effect of active intervention compared to control, when based on mITT, was comparable to ITT (odds ratio 3.76 [95% confidence interval 3.09, 4.57], and 3.47 [2.77, 4.34]; comparison P = 0.60).

Conclusion

We found no difference in the treatment effect between randomized trials using ITT and mITT analyses populations. This suggests that the mITT approach in rheumatoid arthritis trials investigating biological or targeted interventions does not introduce bias compared to ITT.

Introduction

The intention-to-treat (ITT) principle is recommended for analysis as it protects against attrition bias by providing an unbiased estimate of treatment effect when reporting randomized controlled trials [1], [2], [3]. The ITT principle requires all patients in a randomized trial to be analyzed according to their original randomized allocation and requires all patients to be included in the analysis [4]. This conserves randomization and avoids attrition bias when evaluating a treatment assignment [5], [6]. Frequently, a less strict ITT approach is used [7], [8], [9], [10], [11]—commonly referred to as modified ITT (mITT). Generally mITT is based on exclusion of patients—for different reason(s)—postrandomization [12], which may introduce bias [13], [14].

Both ITT and mITT analyses are affected by missing data which also can bias estimates of treatment effect [15], [16], [17]. Methods to handle missing data in randomized trials are as follows: last observation carried forward [10], [11], [18], [19], [20], [21], nonresponder imputation, and multiple imputation [11], [22]. Last observation carried forward is widely criticized and advised against [16], [18], [19], [20], [23], [24], [25]. Nonresponder imputation attributes withdrawal to lack of efficacy—and missing data are imputed as “failure” [18], [26]—and represents a more conservative method than last observation carried forward for dichotomous or categorical outcomes. Preferably, multiple imputation should be used to handle missing data [18], [23].

Estimates of treatment effect can be biased by the use of mITT and common different imputations methods, but the impact and direction of bias has not been assessed previously within rheumatology. To examine whether ITT treatment effect is different from mITT treatment effect and to examine the impact of different imputation methods, we carried out a meta-epidemiological study focusing on trials assessing US Food and Drug Administration or European Medicines Agency–approved biological or targeted interventions (see Methods in the following for definition) in patients with rheumatoid arthritis. These interventions are relatively recent (first approved in 1998) and yield a more modern approaches to trial design, that are believed to have low risk of bias [27]. They therefore represent a good case for examining the nuances of attrition bias.