Conference on Applied Statistics in Agriculture

Simulation Comparison of Methods to Estimate Confidence Intervals of the Mitigated Fraction

In the area of veterinary medicine, efficacy studies are conducted to support licensure of vaccines. Such studies are typically designed to assess a vaccine's ability to prevent or mitigate clinical disease. For example, reduction of duration/severity of clinical signs or the severity of lung lesions are often considered as primary or secondary criteria of success. Studies designed to measure efficacy typically utilize two or more treatment groups and often use blocking structures to accommodate animal housing or litter related effects. When the criteria of interest are continuous or ordinal variables, as is the case with the above measurements, the mitigated fraction (MF) is often used to quantify a vaccine effect. One common approach involves determining the confidence interval for the MF using a bootstrap procedure. For data arising from studies with a blocking structure, there are two bootstrap procedures that are often used. The first resamples the blocks with replacement (randomized cluster bootstrap). The second resamples the blocks and the subjects within blocks, both with replacement (two-stage bootstrap). In addition to the bootstrap procedures, an asymptotic estimate of the variance of the MF can be calculated and used to construct a confidence interval. With three potential methods, it is of interest to determine coverage related to the associated intervals using study designs commonly used for in efficacy studies. In addition, coverage was assessed in situations with and without a treatment effect. Using parameter estimates obtained from the data in which lung lesions were measured, we conducted a simulation experiment estimating the MF and confidence interval using each method described above. The results from this simulation study suggest the bootstrap procedures perform poorly when no treatment effect is present, while the confidence interval estimated using the asymptotic variance performs well. However, none of the methods perform particularly well in the presence of a treatment effect.