Originally, a confidence interval was considered of level 1α only if it achieved at least this level for all values in the
parameter space; and this condition is still desirable if one is going to use the interval to carry out level-α tests for a null
hypothesis. However, confidence intervals are not always used for testing purposes, which has led many researchers to
allow for the empirical level to fall somewhat below the nominal level at some parameter values. In Agresti and Coull (1998)
it was argued that this was a good practice.
Let cpðp1; p2Þ denote the simulated coverage probability. Let pi ði ¼ 1; 2Þ denote the vectors containing the possible values
for p1 and p2 and let M be the number of all possible combinations of p1 and p2. For our simulation studies we report various
assessment measures for each of the methods across a range of values for p1 and p2. The first is the average coverage given
as cp ¼ f∑p1 Ap1∑p2 Ap2 cpðp1; p2Þg=M.