Central limit theorems for doubly adaptive biased coin designs (Q1895356): Difference between revisions

Subjects are assumed to arrive sequentially at an experimental site and are immediately assigned to one of two treatments \(A\) or \(B\). Responses from the subjects for the treatments are assumed to be independent random variables from exponential families. Let \(\rho\) denote the desired allocation proportion under Eiseles biased coin design [\textit{J. R. Eisele}, J. Stat. Plann. Inference 38, No. 2, 249-261 (1994; Zbl 0795.62066)]. For the first \(n\) arrivals, let \(S_n\) denote the difference between the number of subjects assigned to treatment \(A\) and the desired number to be assigned proportional to \(\rho\). Let \(\widehat {\rho}_n\) denote the maximum likelihood estimate of \(\rho\). The authors establish that \[ V_n= (S_n, n(\widehat {\rho}_n- \rho)), \] properly normalized, converges to a normal distribution. In the proof, \(D_n V_n\) is written as the sum of a martingale and a remainder term, for a specified \((D_n)\) of square matrices. The martingale part is shown to converge to normality and the remainder to 0. As a consequence \((V_n)\), properly normalized, is shown to converge to normality. A functional form of the central limit theorem is obtained. An example of the Behrens-Fisher problem in the normal case is discussed.