Asymmetric Finite-Sample Confidence Intervals for Inverse-Probability Estimation under Informative Missingness
Abstract
Unequal response rates can bias mean estimates when units respond infrequently. We consider bounded outcomes observed through independent Bernoulli indicators with known response probabilities. For a clipped inverse-weighted mean, we get exact bias and variance and a sharp worst-case exponential-moment bound. This gives asymmetric finite sample confidence intervals for the mean that directly adjust for clipping bias. We show that a clipping level chosen based solely on the response design maintains coverage without a selection penalty. We also permit each response chance to be in a fixed range and give a rule for its worst case. These intervals maintain coverage in simulations of three designs at two sample sizes but can be wide. Severe imbalance can cause normal weighted intervals to undercover, while complete-case intervals can miss the target. The results require fixed bounded outcomes, independent responses, and known or bounded response probabilities. These bounds can serve as an orientation in cases where normal approximations might not be valid.
Abstract
PDF
Full Text HTML