Nonparametric bootstrap confidence intervals for the Log-Pearson type III distribution

Estimating the frequency of floods is an important problem in hydrology, commonly solved by fitting a probability distribution to observed maximum annual floods. An essential step which must follow the estimation of a quantile is a quantification of its precision. First-order parametric approximations are commonly used to obtain confidence intervals (CIs) for flood flow quantiles. Nonparametric computer-intensive Bootstrap CIs are compared with parametric CIs for simulated samples, drawn from a log-Pearson type III (LP) distribution. Using this methodology, biased in favour of parametric CIs since the parent distribution is known, Bootstrap CIs are shown to be more accurate for small to moderate confidence level (80%), when parameters are estimated by the indirect method of moment (WRC). However, the actual level of Bootstrap CIs is almost always lower than the target level. It is expected that, compared to parametric CIs, Bootstrap CIs perform even better when applied to actual series of maximum annual floods, since they need not come from a LP distribution.