Estimation of the treatment effect in the presence of non‐compliance and missing data

Abstract

Treatment non-compliance and missing data are common problems in clinical trials. Non-compliance is a broad term including any kind of deviation from the assigned treatment protocol, such as dose modification, treatment discontinuation or switch, often resulting in missing values. Missing values and treatment non-compliance may bias study results. Follow-up of all patients until the planned end of treatment period irrespective of their protocol adherence may provide useful information on the effectiveness of a study drug, taking the actual compliance into account. In this paper, we consider non-compliance as discontinuation of treatment and assume that the endpoint of interest is recorded for some non-complying patients after treatment discontinuation. As a result, the patient's longitudinal profile is dividable into on- and off-treatment observations. Within the framework of depression trials, which usually show a considerably high amount of dropouts, we compare different analysis strategies including both on- and off-treatment observations to gain insight into how the additional use of off-treatment data may affect the trial's outcome. We compare naïve strategies, which simply ignore off-treatment data or treat on- and off-treatment data in the same way, with more complex strategies based on piecewise linear mixed models, which assume different treatment effects for on- and off-treatment data. We show that naïve strategies may considerably overestimate treatment effects. Therefore, it is worthwhile to follow up as many patients as possible until the end of their planned treatment period irrespective of compliance, including all available data in an analysis that accounts for the different treatment conditions.

Addressing missing participant outcome data in dental clinical trials

“…21 However, this assumption is rarely plausible in clinical trials and is generally not recommended. 11,12,18,20 3.2.…”

“…12,17 However, distinguishing between ignorable and non-ignorable missingness is rarely possible. 11,18,20 In this case, a versatile strategy could involve starting the analysis by assuming MAR and then evaluating the robustness of these results under various scenarios of the outcome of the missing participants using a sensitivity analysis. Any inconsistency in the conclusions between the different analysis scenarios is a strong indication that missing outcome data may be MNAR.…”

“…The first group is known as exclusions or non-adherent participants and several analytic methods have been proposed to address these exclusions. 1,11 The second group may be termed as last observations carried forward (LOCF) because the missing responses of the participants are usually replaced with their last observed values. 12 Finally, the third group comprises the missing participants or missing outcome data; the latter will be considered in detail in this article.…”

“…15,38 However, the use of LOCF has been heavily criticized for yielding biased treatment estimates in terms of safety and efficacy either in favour of or against the intervention, especially when early dropouts occur and when the response variable is expected to change over time. 4,11,12,27,28,[34][35][36][39][40][41][42] For instance, if an intervention improves a health condition over time, then LOCF will underestimate the efficacy of this intervention, while the opposite will occur if an intervention is linked with deterioration of a health condition over time. The weaknesses of this approach lie on its unrealistic assumptions; LOCF assumes that the responses remain constant after dropping out early from the trial and hence, ignores the natural history of the disease.…”

“…18 The starting point of many statistical techniques for missing outcome data is the assumption that the data are MAR. 11 Similar to MCAR, MAR can be assessed by exploring possible relationships between missingness and observed covariates. 12,17 3.3.…”