Presentation (CTI)

OBJECTIVES: In small-sample settings or when data are skewed—common situations in health economics—the assumptions required for parametric methods to perform well are often violated. As a result, confidence intervals derived from these methods may be inaccurate and exhibit poor coverage probabilities. This simulation study aims to examine the limitations of such parametric approaches and explore non-parametric bootstrap techniques as alternatives for confidence interval estimation, given small samples, in the context of Health Technology Assessment (HTA).
METHODS: We compare standard parametric confidence intervals (95% confidence level) with four non-parametric bootstrap methods: percentile, basic, studentized (bootstrap-t), and bias-corrected accelerated (BCa). A Monte Carlo simulation study assesses these methods in terms of coverage probability and interval width, across different underlying distributions (normal, Student’s t, exponential, and chi-squared) and small sample sizes (n = 5, 10, and 20).
RESULTS: Only the bootstrap-t method achieves satisfactory performance, producing relatively narrow intervals with coverage probabilities close to the nominal 95%, when the sample size is at least 20. In the best-case scenario, it yields a coverage probability of 94.1% with a mean interval width only 1.1 times greater than that of the parametric approach. For smaller samples, approximately 70.1% of the coverage probabilities fall below 90%, and all methods produce overly wide intervals, particularly for skewed distributions. To illustrate the practical implications, we applied both the standard parametric method and the best-performing non-parametric method (bootstrap-t) to estimate confidence intervals for measures of effectiveness in the context of a meta-analysis.
CONCLUSIONS: In small-sample settings or skewed data scenarios—common in health economic evaluations—parametric confidence intervals often underperform. The bootstrap-t method provides a more reliable alternative when the sample size is at least 20, offering improved coverage without excessively widening the interval.