On Sequential Treatment Allocations in Clinical Trials
This dissertation treats baseline-dependent sequential designs of two-treatment parallel-group clinical trials. The treatment assignments are chosen in order to minimize the variance, in a linear model, of the treatment effect. This is done for each new allocation using a generalized biased coin design, or a non-randomized minimization' method.
For the minimization' method, the balance process is shown to be tight. It follows that the loss, defined roughly as the number of patients lost d ue to imbalance, is of order N 1 (where N is the trial size).
The ANCOVA statistic is used in both parametric and randomization tests when the design is randomized. Deficiency (or second-order efficiency), of design and analysis combin ed, is defined in terms of expected p-value. The asymptotic deficiency of the randomization analysis following a biased coin design is obtained when prognostic factors are ignored. It can be arbitrarily close to zero re lative the balanced t-test when assuming a normal model. Similar results, when prognostic variables are used, are indicated by simulations.
As a comparison, the expected loss and asymptotic expected p-value deficiency, relative a balanced parametric test, equal the number of prognostic variables when using independent randomizations.
AMS 1991 subject classification: 62L05, 60K30, 62G10, 62P10
biased coin designs
balance of prognostic factors