(i) When the three-parameter Weibull distribution is correct, the three discrimina
tion criteria and the estimated three-parameter Weibull distribution are identical in
terms of BIAS and RRMSE for both the maximum and 98% quantities for all values of
the shape parameter (namely, 0.5, 1, 2 and 3). Underfitting the correct Weibull distri
bution with its two-parameter variant yields considerably larger BIAS and RRMSE for
all values of the shape parameter, although the degree of error is reduced as the value
of the shape parameter is increased. Thus, estimating the correct three-parameter
distribution is not preferable in terms of BIAS and RRMSE as compared with using
the three discrimination criteria, although underfitting the three-parameter Weibull
distribution by setting the location parameter
7to zero yields much larger BIAS and
RRMSE values in all cases considered. This result is broadly similar in qualitative
terms to those obtained for the gamma distribution although, in the latter case, esti
mating the correct distribution is preferred in terms of BIAS for a large value of the
shape parameter.
(ii) As compared with the case of the gamma distribution, overfitting the correct
two-parameter Weibull distribution yields some surprising results in terms of BIAS,
especially for larger values of the shape parameter. When the shape parameter is 0.5,
the estimated two-parameter Weibull distribution has much lower BIAS values than
those obtained using the discrimination criteria and the estimated three-parameter
Weibull distribution for both the maximum and 98% quantities. As the shape param
eter is increased from 0.5 to 1, the correctly estimated two-parameter distribution is
still preferred for the maximum quantity but is inferior to SIC, which favours the lower
dimensioned model, for the 98% quantity. When the shape parameter is increased to
2 or 3, all three discrimination criteria are preferred to the estimated two-parameter
Weibull distribution in terms of BIAS. Indeed, when the shape parameter is set to
3, even the overfitted three-parameter distribution has lower BIAS than the correctly
fitted two-parameter variant for both the maximum and 98% quantities. In terms of
RRMSE, the correctly estimated two-parameter distribution is preferred in all cases,
although its superiority is diminished as the shape parameter is increased. While there
is not a noticeable difference between the overfitted th ree-p aram eter distribution and th e th ree discrim ination criteria, SIC is always second best to th e estim ated two- pa ram eter d istrib u tio n , and th e estim ated th ree-p aram eter distribution is generally the worst. Q ualitatively, th e results are reasonably sim ilar to those for th e gam m a distri bution in th a t overfitting does not generally increase th e BIAS and RRM SE values for the m axim um and 98% quantities for th e Weibull distribution. However, it is interest ing to note th a t estim atin g th e correct tw o-param eter W eibull distribution does not always yield th e sm allest BIAS value relative to th e th ree discrim ination criteria, or even to th e overfitted th ree-p aram eter distribution.
Finally, Table 7.3 contains th e results from experim ents for the lognorm al d istri bution. Since th e lognorm al d istribution has th e opposite behaviour to the gam m a and W eibull d istrib u tio n s as th e shape p aram eter is increased, it is useful for com par ative purposes to exam ine th e results as th e shape p aram eter is decreased rath er than increased. T he principal points to note from th e tab le are as follows:
(i) W hen th e th ree-p aram eter lognorm al distribution is correct (7 = 1) and the shape p a ra m ete r is 1.2 or 0.9, th e correctly fitted th ree-p aram eter d istribution and th e th ree discrim ination criteria have identical BIAS and RRM SE values for both the m axim um and 98% q u an tities, whereas the underfitted tw o-param eter lognorm al dis trib u tio n has su b stan tially higher values for BIAS and RRM SE. A lthough th e under fitted tw o-param eter d istrib u tio n still has the largest BIAS and RRM SE values when th e shape p a ra m ete r is reduced to 0.5 or 0.4, the oth er four m ethods do not rem ain identical. W hen th e shape p aram eter is 0.5, the LR m ethod has by far th e smallest BIAS for th e m axim um q u a n tity while SIC has th e largest; for th e 98% quantity, AIC and th e correctly fitted th ree-p aram eter distribution have th e sm allest BIAS, and SIC again has th e largest. These rankings are not m aintained when th e shape param eter is reduced to 0.4. T h e sm allest BIAS values for th e m axim um and 98% quantities are AIC and th e estim ated th ree-p aram eter lognorm al distribution, respectively, w ith SIC th e worst of th e four m ethods in each case. On th e basis of RRM SE, however, the th ree-p aram eter lognorm al distribution has the sm allest value, followed closely by