HR: 1330h
AN: H12B-0967    [PDF]
TI: The Application of Censored Regression Models in Low Streamflow Analyses
AU: * Kroll, C
EM: cnkroll@esf.edu
AF: Environmental Resources Engineering, SUNY College of Environmental Science and Forestry, 1 Forestry Drive, Syracuse, NY 13210 United States
AU: Luz, J
EM: jgluz@ufba.br
AF: Instituto de Geociencias, Universidade Federal da Bahia, Rua Barao de Geremoabo, Salvador, 40170-290 Brazil
AB: Estimation of low streamflow statistics at gauged and ungauged river sites is often a daunting task. This process is further confounded by the presence of intermittent streamflows, where streamflow is sometimes reported as zero, within a region. Streamflows recorded as zero may be zero, or may be less than the measurement detection limit. Such data is often referred to as censored data. Numerous methods have been developed to characterize intermittent streamflow series. Logit regression has been proposed to develop regional models of the probability annual lowflows series (such as 7-day lowflows) are zero. In addition, Tobit regression, a method of regression that allows for censored dependent variables, has been proposed for lowflow regional regression models in regions where the lowflow statistic of interest estimated as zero at some sites in the region. While these methods have been proposed, their use in practice has been limited. Here a delete-one jackknife simulation is presented to examine the performance of Logit and Tobit models of 7-day annual minimum flows in 6 USGS water resource regions in the United States. For the Logit model, an assessment is made of whether sites are correctly classified as having at least 10% of 7-day annual lowflows equal to zero. In such a situation, the 7-day, 10-year lowflow (Q710), a commonly employed low streamflow statistic, would be reported as zero. For the Tobit model, a comparison is made between results from the Tobit model, and from performing either ordinary least squares (OLS) or principal component regression (PCR) after the zero sites are dropped from the analysis. Initial results for the Logit model indicate this method to have a high probability of correctly classifying sites into groups with Q710s as zero and non-zero. Initial results also indicate the Tobit model produces better results than PCR and OLS when more than 5% of the sites in the region have Q710 values calculated as zero.
DE: 1812 Drought
DE: 1869 Stochastic processes
SC: Hydrology [H]
MN: 2003 Fall Meeting