I'm interested in ranking 9 domains/attributes in terms of derived
importance methods.
I'm using the Standardised Beta in Linear Regression [SPSS], and the 2
importance measures offered in Random Forest [R] .
I'm experimenting with Random Forest as I getting wary of the high
inter-domain correlations violating linear regression assumptions.
Average inter-domain correlation is about 0.75.
My data consists of mean ratings [0-100%].
Used same data & dependent variable for both methods.
I get different rankings for each method. I would expect this but I
would like comments on -
# Would it make any sense in averaging the 3 ranking and using this as
a "best" estimate ?
# If averaging is not acceptable - which one is preferred ?
# Are correlations of 0.75 high in this context - should I be worried
about regression assumptions ?
# Of the 2 RF importance measures, which is the most commonly used /
preferred.
|
|
Rank
|
Rank
|
Rank
|
|
Domains
|
Regression
|
RF-%IncMSE
|
RF-IncNodePurity
|
|
Dom1 Integrity
|
5
|
2
|
5
|
|
Dom2 Service
|
9
|
6
|
7
|
|
Dom3 SelfDevelopment
|
3
|
9
|
8
|
|
Dom4 SelfExpression
|
2
|
1
|
1
|
|
Dom5 Expertise
|
8
|
7
|
9
|
|
Dom6 Teamwork
|
6
|
3
|
3
|
|
Dom7 Enthusiam
|
4
|
8
|
4
|
|
Dom8 Action
|
1
|
4
|
2
|
|
Dom9 Stability
|
7
|
5
|
6
|
--
Mark Webb
Line +27 (21) 786 4379
Cell +27 (72) 199 1000
Fax to email +27 (86) 5513075
Skype webbmark
Email [hidden email]
=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD
Locked
