In short, I am running bootstrapping on an association analysis data set in
R (using the
Arules package) and collecting the fit measures (support, confidence and lift) for the full data as well as the bootstrap samples. The number of bootstrap samples is equal 1000.
The final resultant data set I have is the set of 1000 confidence measures for each rule as found by the Arules package. From this, I am calculating the 95% confidence intervals for the confidence measure of each rule.
I noticed that for about 20% of the rules, the full data estimate of confidence is actually less than the lower 0.025th percentile of the set of bootstrapped confidence measures. This is a little counter-intuitive, since I would expect the full data estimate to be somewhere in the middle of the bootstrapped estimates (the bootstrapped sample being to the sample, as the sample is to the population etc.). What am I missing here?