On the Detection of Subpopulations

T-test) was estimated as the number of times the null hypothesis of no subpopulation was rejected using a significance level of 0.01 divided by the total number of simulations. 

Under the basic design, the Type I error rate for the T-test and LRT was 2 and 1%, respectively, using a significance level of 1%. The power to detect the subpopulation increased for all methods when the proportion of subjects in the subpopulation increased. A total of 80% power was reached for the LRT when 20% of the population was in the subgroup and never reached 80% for the T-test. Power did not increase much for either method when the percent of subjects in the subpopulation was increased to 30%. The LRT had greater power at detecting a subpopulation than did T-test under all conditions studied. 

Many modifications to the basic scenario were also examined. In the first modification, no effect on overall power was observed when the total number of subjects was doubled from 100 to 200 (but keeping the percent of subjects in the subpopulation the same). In the second modification, no difference in power was observed when the sampling times were allowed to vary randomly by + 10% from their fixed values. In the original design, samples were collected at 1 and 11.5 h on Day 10. In the third modification, an additional sample was collected at 5-h postdose. Power to detect a subpopulation increased dramatically with 80% power being achieved with only 10 subjects in the subpopulation for both the LRT and T-test. Hence, three samples was much better than two samples. 

Lee s study clearly showed that experimental design factors can have an enormous impact at detecting subpopulations within a population. Increasing the number of subjects in a population from 100 to 200 did 

---