Randomized Clinical Trials, Analysis Sets, Statistical Analysis, Reporting Issues. Principal Features Analyses, the Cochrane Risk-of-Bias-Tool
The current chapter will address the statistical analysis of randomized controlled trials and reporting issues. First, types of analysis sets will be discussed, including the intention to treat analysis (ITT), as well as the per protocol (PP) analysis, otherwise called completed protocol (CP) analysis . The ITT analysis includes patients who are lost, and the PP analysis only includes the patients who completed the study.
Second, statistical principles required for data analysis will be reviewed. They are based on statistical reasoning. Statistical reasoning uses three general approaches: (1) statistical estimation, (2) statistical hypothesis testing, and (3) statistical modeling. Special attention will be given to the issues:
- - stratification, baseline covariates,
- - missing values, withdrawals, drop-outs (often a PP analysis uses the last observation carried forward principle (LOCF)),
- - safety & tolerability issues: often analyzed in subgroups with special populations (age, gender, comedication groups).
Third, the CONSORT (consolidated standards of randomized trials, a statement of medical journal editors referring to homogeneity, standard terminologies, and uniform units) will be reviewed.
Fourth, the issue of reporting bias with bias defined as systematic errors that no one recognizes, and other reporting issues will be the subject of this chapter.
The principal features of the study’s statistical analysis must be in the protocol. A more technical and detailed elaboration of it should be in the SAP (statistical analysis plan), as recommended by the International Conference of Harmonization guidelines (in the sections ICH E9 & E3). The principal features should, of course, be finalized, before database lock/unblinding . And a separate description of the analysis of primary outcomes (the confirmatory analysis), and secondary outcomes (the exploratory analysis) is recommended.
© Springer International Publishing Switzerland 2017
TJ. Cleophas, A.H. Zwinderman, Understanding Clinical Data Analysis,
We should add, that a principal features (PF) analysis is, currently, increasingly important. For example, a PF analysis is vital, if your outcome is, for example, as complex as functional magnetic resonance imaging producing multivoxel patterns for analysis (voxels are pixels = picture elements with 3 instead of 2 dimensions). Even better in the given situation will probably be a blinded PF analysis of the data prior to database-lock, that may call for changes in the principal features of the study. These PFs should, then, be documented in a protocol amendment. Only the latter amendment can be regarded as confirmatory, and will be in the confirmatory data analysis .
In the current chapter, we will also address pretty novel, but relevant, subjects, like
- - blinded principal features analyses,
- - outcome adjustments for
subgroups, random effects and baseline characteristics,
- - routine use of check lists before data lock,
- - the handling of missing data with either
intention to treat population, imputation methods, or multiple imputations.
Publication bias, and reporting biases, including the Cochrane risk-of-bias tool will also be reviewed here.