Sa. Lawal et We. Watt, FREQUENCY-ANALYSIS OF LOW FLOWS USING THE AKAIKE INFORMATION CRITERION, Canadian journal of civil engineering, 23(6), 1996, pp. 1180-1189
It is the current practice in frequency analysis of low flows to consi
der only three-parameter distributions in which one of the parameters
represents a nonzero lower bound. When applied to the small samples ty
pically available, this practice results in incorrect low flow estimat
es. These errors are related to errors in the estimated lower bound. T
o preclude this possibility, it is proposed that the current practice
be changed to include the selection of a two-parameter distribution in
certain situations. To assess this proposal, the Akaike information c
riterion (AIC) is used to compare the suitability of the most commonly
used three-parameter distribution (three-parameter Weibull) and three
two-parameter distributions (two-parameter Weibull, Gumbel, and logno
rmal) to low flow data for 51 long-term hydrometric stations across Ca
nada. For 75% of the stations, a two-parameter distribution is selecte
d over the three-parameter distribution if the selection criterion is
minimum AIC. In about one third of the remaining 25% of the stations w
here the three-parameter Weibull distribution gave the minimum AIC, th
e estimated lower bound is sufficiently close to the minimum observed
low flow to indicate overfitting and hence unreliable quantile estimat
es. When the AIC is supplemented with visual examination of goodness o
f fit on probability plots, it is found that the lognormal distributio
n could very well fit those cases where the AIC selected the three-par
ameter Weibull distribution.