The Inefficacy of Chauvenet's Criterion for Elimination of Data Points

Abstract

Chauvenet's criterion is commonly used for rejection of outliers from sample datasets in engineering and physical science research. Measurement and uncertainty textbooks provide conflicting information on how the criterion should be applied and generally do not refer to the original work. This study was undertaken to evaluate the efficacy of Chauvenet's criterion for improving the estimate of the standard deviation of a sample, evaluate the various interpretations on how it is to be applied, and evaluate the impact of removing detected outliers. Monte Carlo simulations using normally distributed random numbers were performed with sample sizes of 5–100,000. The results show that discarding outliers based on Chauvenet's criterion is more likely to have a negative effect on estimates of mean and standard deviation than to have a positive effect. At best, the probability of improving the estimates is around 50%, which only occurs for large sample sizes.