Randomised clinical trials: statistical investigations of bias, inefficiency and misinterpretation
thesisposted on 28.03.2022, 02:08 by Ishani Manjula Schou
Randomised clinical trials (RCTs) compare treatment interventions using the health out-comes of individuals assigned to their treatment at random. RCTs are the gold standard for comparing treatment efficacy, but many factors in their design, conduct and analysis can lead to bias, inefficiency or misinterpretation. This thesis by publication presents statistical investigations of three such areas. The first area relates to potential bias from early stopping of RCTs. Some researchers have claimed that early stopping of RCTs based on interim analyses leads to overestimation of the treatment effect and that this is particularly problematic for meta-analyses that synthesise the results of multiple studies. This thesis presents extensive theoretical and simulation studies of this potential source of bias. It is concluded that early stopping is not a substantive sourceof bias for meta-analyses of RCTs. The second area relates to the potential for misinterpretation of RCT subgroup analyses, particularly subgroups defined by geographical region in global studies. Subgroup analysis principles require a significant test of interaction to conclude heterogeneity of subgroup treatment effects. However, overly optimistic expectations of treatment effect homogeneity often lead to over-interpretation of apparent differences between subgroups. This thesis proposes a suite of graphical analyses that supplement a test of interaction with a visual assessment of the extent to which chance can explain the observed differences between subgroups. An open-source software package for the R computing environment is presented. The third area relates to efficient design of RCTs having several treatments compared to a common control. Standard balanced designs have equal numbers of individuals on each treatment, but are inefficient in this context. This thesis considers efficient unbalanced designs that minimise variance or maximise power. New results in optimal design theory, and some guidelines for the efficient planning of RCTs having several treatments, are presented.