Biology Reference
In-Depth Information
44 PIP3 PIP2 1.00 0.5
56 p44.42 pakts473 1.00 0.5
66 pakts473 p44.42 1.00 0.5
67 pakts473 PKA 1.00 0.5
77 PKA pakts473 1.00 0.5
89 PKC P38 1.00 0.5
90 PKC pjnk 1.00 0.5
99 P38 PKC 1.00 0.5
109 pjnk PKC 1.00 0.5
> avg.catnet = averaged.network(sa, threshold = 0.85)
Again,
avg.catnet
presents some small differences from both
avg.boot
and
avg.start
. Such differences can be attributed to the different scores and
structure learning algorithms used to build the sets of high-scoring networks. In
particular, it is very common for arc directions to change between different learning
algorithms as a result of score equivalence.
2.5.2 Choosing the Significance Threshold
The value of the threshold beyond which an arc is considered significant, which is
often called the
significance threshold
, does not seem to have a huge influence on
the analysis of the data analyzed in
Sachs et al.
(
2005
). In fact, any value between
0
.
5and0
85 yields exactly the same results. So, for instance,
> all.equal(averaged.network(boot, threshold = 0.50),
+
.
averaged.network(boot, threshold = 0.70))
[1] TRUE
The same holds for
avg.catnet
and
avg.start
. However, this is often not
the case. Therefore, it is important to use a statistically motivated algorithm for
choosing a suitable threshold instead of relying on ad hoc values.
A solution to this problem is presented in
Scutari and Nagarajan
(
2012
)and
implemented in
bnlearn
as the default value for the
threshold
argument in
averaged.network
.
> averaged.network(boot)
Random/Generated Bayesian network
model:
[praf][plcg][p44.42][PKC][pmek|praf][PIP2|plcg]
[pakts473|p44.42][P38|PKC][PIP3|plcg:PIP2]
[PKA|p44.42:pakts473][pjnk|PKC:P38]
nodes:
11
Search WWH ::
Custom Search