The two one-sided tests (TOST) procedure for average bioequivalence is operationally equivalent to an ordinary \(100 (1 - 2 \alpha)\%\) or an expanded \(100 (1 - \alpha)\%\) confidence interval.
OrdCI <- c(est - sd * qt(1 - alpha, n - 1) / sqrt(n),
est + sd * qt(1 - alpha, n - 1) / sqrt(n))
ExpCI <- c(min(0, est - sd * qt(1 - alpha, n - 1) / sqrt(n)),
max(0, est + sd * qt(1 - alpha, n - 1) / sqrt(n)))
We simulate 10,000 normal datasets with \(\sigma=0.1\), and \(n=20\) and calculate both the 90% ordinary and the 95% expanded CI using a simulation function Sim(mu, sd, n, alpha, nsim)
(not shown here).
For \(\mu=log(1.25)\) we get the expected coverage probabilities (CP) of 90% and 95%; the implied tests, however, have a “power” of 5% for both CIs. The expanded CI is substantially wider on average, but this involves (by construction) no power disadvantage whatsoever.
Sim(mu=log(1.25), sd=0.1, n=20, alpha=0.05, nsim=10000)
## ordinary expanded
## Coverage 0.902 0.952
## Power 0.048 0.048
## Width 0.076 0.261
For \(\mu=0\) the CP of the expanded CI is 100% because it always contains zero (again by construction). The discrepancy in average width is now much smaller because few CIs are actually being “expanded”.
Sim(mu=0, sd=0.1, n=20, alpha=0.05, nsim=10000)
## ordinary expanded
## Coverage 0.901 1.000
## Power 1.000 1.000
## Width 0.076 0.077