Page 50 - Demo

Chapter 248behaviors, settings, or disorders that may be either closely or distally related to the target behavior.15One main shared shortcoming was the lack of statistical analyses. None of the 12 studies included a justification for the sample size. Sample size calculations are important to ensure that clinically relevant effects can be detected while not including, and hence burdening, too many patients. In N-of-1 trials, a power analysis can help decide on the number of periods required to detect a clinically relevant treatment effect within a patient and, in case of a series of N-of-1 trials, for the number of participants required to determine an average treatment effect in the study sample. Formulas and methods for calculation of the required sample size for these different objectives are available for N-of-1 studies.53The majority of the studies only described results using graphical or tabular methods, whereas (non)parametric statistical analyses are now considered the standard for testing for an intervention effect in N-of-1 studies.54 (Non)parametric and ancillary analyses should be performed to evaluate period effects, intrasubject correlations, and subgroup and adjusted effects. Rather than attempting to adjust for carryover effects, it is preferred to choose the (washout) periods long enough for carryover not to occur.Both mixed-effects models and Bayesian models can properly address the inter- and intrapatient variability in series of N-of-1 trials.41 A clear overview is given of the various frequentist analyses proposed for N-of-1 trials that may serve different purposes.40 Most importantly, the statistical methods should properly account for the method of randomization used. Simple analyses such as a paired t test and a summary measure approach can be acceptable for testing the hypothesis of a difference between treatments. For assessing heterogeneity of the treatment effects between individuals, a mixed model approach is required40 with an ANOVA type test for hypothesis testing. The latter can also be done in a Bayesian framework using hierarchical modeling. In a Bayesian framework, it is quite natural to update an estimation when data from new N-of-1 trials become available. If one wishes to produce shrunken estimates or predict the effects for future Annelieke Muller sHL.indd 48 14-11-2023 09:07