is compared with the corresponding
theoretical cumulative distribution function (CDF) given the sample of survival days. We
choose a probability distribution for which the EDF and CDF are the closest. By making
use of Mathematica v7.0 software (Wolfram Research [24]), or any advanced symbolic
computation package, one can calculate the mean squared and absolute differences
between the EDF and CDF, and select the model of best fit by visual inspection. The
question may arise as to whether the selection of the probability model providing the best
fit will be dependent on a given difference measure. Accordingly, two measures of
discrepancy are being used to determine the best fit model. The mean squared
differences i.e., ∑ ( ( ) ( ))
, and the absolute mean differences
i.e.,∑ | (
) (
)|
are considered, where isthe empirical distribution function, G
is the functional representation of the cumulativedistribution function, and n is the
observed sample size. Thus, it is obtained that the half-normal model best fits the data
considering the lowest differences (mean squared differences= 0.0010135; absolute
mean differences = 0.021602). The gamma model has the next lowest differences (mean
squared differences = 0.0015963; absolute mean differences = 0.0352381).The
exponential model has the highest differences (mean squared differences =
0.0094878;absolute mean differences = 0.086806). Furthermore, to confirm the best
model half-normal, the calculated Anderson-Darling Test Statistics is obtained as 0.4964
with its corresponding p-value = 0.7834. Eqn. (2) is used to the data, and the predictive
moments of a single future response are estimated with respect to some values of the
hyperparameters. The results of95% HPD intervals with respect to certain combination of
the hyperparameters are reported in Table 5.