The randomization has a strong impact on the bounds of the confidence intervals. As an example, when n = 10 and X = 2
the upper bound of the externally randomized split sample Wilson interval ranges from 0.476 to 0.558, meaning that the
range of the upper bound is 0.082% or 18% of the interval’s conditional expected length 0.45. The impact of the randomization
can be substantial even for relatively large n. Consider for instance the case n = 200, X = 40. The conditional mean length
of the split sample Wilson interval is 0.11, whereas the range of the upper bound of the interval is 0.016, or 14.5% of the
conditional expected length of the interval.
The U(−1/2, 1/2) Wilson and Stevens intervals are similarly affected by the randomization for small n: when n = 10
and X = 2 the conditional expected length of the Stevens interval is 0.48 and the range of the upper bound is 0.111% or 23%
of the conditional expected length. However, the impact of the randomization on these interval decreases quicker than it
does for the split sample Wilson interval. When n = 200 and X = 40 the conditional expected length of the Stevens interval
is 0.11. The range of the upper bound is 0.005, which is no more than 4.5% of the conditional expected length.
Fig. 4 shows the range of the upper bounds of the three intervals for different combinations of X and n. The intervals are
about as bad for n = 20, but the U(−1/2, 1/2) Wilson and Stevens intervals are much less affected by the randomization
for n = 100 and n = 500. Because of the equivariance of the intervals, the figures for the lower bound are identical if X is
replaced by n − X.
The randomization has a strong impact on the bounds of the confidence intervals. As an example, when n = 10 and X = 2
the upper bound of the externally randomized split sample Wilson interval ranges from 0.476 to 0.558, meaning that the
range of the upper bound is 0.082% or 18% of the interval’s conditional expected length 0.45. The impact of the randomization
can be substantial even for relatively large n. Consider for instance the case n = 200, X = 40. The conditional mean length
of the split sample Wilson interval is 0.11, whereas the range of the upper bound of the interval is 0.016, or 14.5% of the
conditional expected length of the interval.
The U(−1/2, 1/2) Wilson and Stevens intervals are similarly affected by the randomization for small n: when n = 10
and X = 2 the conditional expected length of the Stevens interval is 0.48 and the range of the upper bound is 0.111% or 23%
of the conditional expected length. However, the impact of the randomization on these interval decreases quicker than it
does for the split sample Wilson interval. When n = 200 and X = 40 the conditional expected length of the Stevens interval
is 0.11. The range of the upper bound is 0.005, which is no more than 4.5% of the conditional expected length.
Fig. 4 shows the range of the upper bounds of the three intervals for different combinations of X and n. The intervals are
about as bad for n = 20, but the U(−1/2, 1/2) Wilson and Stevens intervals are much less affected by the randomization
for n = 100 and n = 500. Because of the equivariance of the intervals, the figures for the lower bound are identical if X is
replaced by n − X.
การแปล กรุณารอสักครู่..
