Clinical Trials

Clinical Trials: A Practical Guide ■❚❙❘known efficacious drug in order to look for evidence of drug A’s superiority overan established standard therapy.AssumptionsFour key assumptions are required in the two-sample t-test. Firstly, we assumethat the two treatment group populations from which the samples are drawn aredistributed normally [2–4].Secondly, we assume that the variances (or standard deviations) of the twopopulations are equal [2–4], ie, σ 12= σ 22= σ 2 . The equality of variancesassumption can be formally verified with an F-test [2–4]. We can also do aninformal check by looking at the relative magnitude of the two-sample variancesS 12and S 22. For example, if S 12/ S 22is considerably different from 1 then theassumption that σ 12= σ 22= σ 2 will be in doubt. In cases where σ 12≠σ 22, we needto use a modified t-test or nonparametric method [3–5].Thirdly, we assume that the observations in the two treatment groups areindependent of each other, ie, no observation in one group is influenced byanother observation in the second group [3–5]. Taking the posttreatment FEV 1as an example, the value of posttreatment FEV 1in the active treatment group isnot affected by that in the placebo group. Therefore, the values of the two sets ofposttreatment FEV 1measurements constitute two independent samples.Finally, we assume that the two populations are homogeneous in terms of theobserved and unobserved characteristics of patients (ie, free from confounding).These characteristics might be demographics (eg, age), prognosis (eg, clinicalhistory, disease severity), or baseline measurements of outcome variables (eg,pretreatment FEV 1in the CAL trial).Although we might never know the unobservable heterogeneity (differences)between two populations, we can assess whether the two populations arecomparable by looking at the observed summary statistics, such as means orproportions at baseline by treatment. This is why a table that summarizes thebaseline information in a clinical trial by treatment group is always provided ina clinical report. If the two treatment groups are not balanced with regard to someof the predictors of outcome, covariate adjustment by means of stratification orregression modeling can be employed (see Chapters 24–26) [1,4,5].209