Clinical Trials

Clinical Trials: A Practical Guide ■❚❙❘carried out at each of the time points (see Chapter 28). This also leads to theproblem of multiple comparisons and therefore increases the possibility ofa false-positive result.Subgroup analysesA clinical trial is usually concerned with the overall impact of a treatment on thetrial population. However, individuals within a trial population vary in theircharacteristics. Common secondary analyses involve the investigation of whetherdifferences between treatments vary between different subgroups of the studypopulation. For example, we might test whether males or females benefit morefrom a certain treatment, or specifically diabetic patients, or patients above acertain age. Such analyses are called subgroup analyses. The statistical issuesrelated to these are complex and prone to confusion and often involve multipletesting, in addition to the problem that trials are often not powered to detecttreatment differences amongst subgroups (see Chapter 23) [4].Interim analysesInterim analyses are usually undertaken during the conduct of a trial for ethicaland economical reasons, with the possibility that the trial might be terminatedearly if significant treatment differences in efficacy or safety outcomes are found.It is worth noting that successive analyses conducted on the growing body of trialdata will also lead to an increase in the overall Type I error rate at the finalanalysis of the trial, unless adjustments are made (as discussed in the next section)(see also Chapter 31).Strategies for dealing with multiplicityMultiplicity needs to be considered at the design stage of the trial and whenwriting a statistical analysis plan – before any analysis is undertaken. We discussthree possible strategies for handling multiplicity in trials.Change limits to P-values of single testThe first strategy is to plan a predefined correction for the inflated Type I error.If the significance level for each individual test is reduced, then the overall levelof significance can be kept at 0.05 for the entire series of tests. There are differenttypes of statistical corrections that are based on this approach.A commonly used approach is the Bonferroni correction to the nominal significancelevel [5]. The significance level of each subtest is set to be the overall significancelevel divided by the total number of tests performed – eg, with five tests, the criticallevel of significance for each of the five subtests is set at 0.01 (= 0.05 / 5) instead333