Abstract
Change in the area of premalignant lesions is an end point in estimating the efficacy of chemopreventive agents. When examiners round measurements of lesion length and width, they introduce variability, which perturbs the relative percent change in lesion area and, consequently, the percent of subjects showing a clinical response. We use simulations to illustrate the resulting bias when the agent under test is effective in reducing lesion area. We simulated 500 oral leukoplakia lesions per run, with 2,500 runs at each of five levels of agent effectiveness, namely, true relative percent reduction in area of 25%, 45%, 50%, 55%, and 75%. Realistic values of lesion lengths and widths were generated randomly and then rounded to the nearest multiple of five. The product is the distribution of mean relative percent change in lesion area and the corresponding percent of subjects showing a clinical response. Even the fifth percentile of the distribution of mean relative percent change in lesion area consistently underestimated the true value, by about 6 percentage points. The percent showing a clinical response was underestimated by 50%, 37%, and 11% for true values of reduction in lesion area of 50%, 55%, and 75%, respectively. This could easily double the required sample size for a modest phase II study. We suggest that it is cost-effective to train observers of lesion length and width to eschew rounding of measurements in the chemoprevention setting.