Abstract
Responses are often correlated in clinical trials. A patient who is not evaluable for a response may still provide some information to the response through his or her status on other responses. Under the assumption of missing at random, we propose the utilization of a self-consistent estimator. We show that the proposed estimators are more efficient than the conventional estimators by asymptotic relative efficiency and simulation study. An example from a Phase II clinical trial on children with chronic myelogenous leukemia is provided.
Notes
a Results were obtained by simulation of 10,000 random samples with P EE = 0.50, P EI = P IE = 0.25, the specified p ij , and the specified sample size.
b Estimated bias of the proposed estimator for R 1 = p 11 + p 12.
c Estimated coverage probability of 90% confidence intervals for R 1 = p 11 + p 12.
d
The hypothesis test was performed for the null hypothesis H
0: R
1 ≤ p
11 + p
12 vs. the alternative hypothesis H
1: R
1 > p
11 + p
12. The significance level was computed under the specified p
ij
. The power was computed with p
11 increased by , p
12 and p
21 unchanged, and p
22 decreased by
.
e The relative efficiency is defined as the ratio of the MSE of the conventional estimator over that of the proposed estimator for R 1 = p 11 + p 12. When the sample size is specified as ∞, the relative efficiency is the asymptotic relative efficiency.