Information Technology Reference
In-Depth Information
1
B
*
Â
()
=
()
ˆ
ˆ
.
r
=
RMB
,
MSE
r
B
b
1
B
B
b
=
1
Using a component-of-variance calculation (Gong 1982), for Simulation
1.1
()
=
(
)
12
12
M
0 1070
.
~
0 1078
.
=
M
100
;
so if we are interested in comparing root mean squared errors about
the excess error, we need not perform more than
B
= 100 bootstrap
replications.
In each simulation, I included 400 experiments and therefore used the
approximation
400
1
400
2
Â
2
()
∫-
ˆ
[
ˆ
]
[
ˆ
]
MSE
1
rErR
~
rR
e
-
,
e
e
=
11
r
where
e
and
R
e
are the estimate and true excess of the
e
th experiment.
Figure 2 and 3 show 95% nonsimultaneous confidence intervals for
RMSE
1
's and RMSE
2
's. Shorter intervals for RMSE
1
's would be prefer-
able, but obtaining them would be time-consuming. Four hundred experi-
ments of simulation 1.1 with
p
= 4,
n
= 20, and
B
= 100 took 16
computer hours on the PDP-11/34 minicomputer, whereas 400 experi-
ments of simulation 2.3 with
p
= 6,
n
= 60, and
B
= 100 took 72 hours.
Halving the length of the confidence intervals in Figures 2 and 3 would
require four times the number of experiments and four times the com-
puter time. On the other hand, for each simulation in Figure 3, the confi-
dence interval for RMSE
2
(
ideal
) is disjoint from that of RMSE
2
(
boot
),
and both and disjoint from the confidence intervals for RMSE
2
(
jack
),
RMSE
2
(
cross
), and RMSE
2
(
app
). Thus, for RMSE
2
, we can convincingly
argue that the number of experiments is sufficient.
r
r
r
r
r
5. THE RELATIONSHIP BETWEEN CROSS-VALIDATION
AND THE JACKKNIFE
Efron (1982) conjectured that the cross-validation and jackknife estimates
of excess error are asymptotically close. Gong (1982) proved Efron's con-
jecture. Unfortunately, the regularity conditions stated there do not hold
for Gregory's rule. The conjecture seems to hold for Gregory's rule,
however, as evidenced in Figure 4, a scatterplot of the jackknife and cross-
validation estimates of the first 100 experiments of simulation 1.1. The
plot shows points hugging the 45° line, whereas a scatterplot of the boot-
strap and cross-validation exhibits no such behavior.