Hello,
I'm currently trying to model the movement of a slope (v.obs) with a
regression model.
The data can be found following the given links:
either
http://www.sendspace.com/file/dnugwc
or
http://rapidshare.com/files/420569660/sel.day.txt
I want to use the Box-Cox transformation to normalize the response as
well as the predictor variables.
The scatterplot looks like this:
library(zoo)
library(alr3)
load("sel.day.txt")
sel.p1<-window(sel, start=as.POSIXct("2008-04-05"),
end=as.POSIXct("2009-04-01"))
pairs(~v.obs+ snow+ HH6.1+ Q.Enz+ pcpt+
qd,data=sel.p1,gap=0.4,cex.labels=1.5)
In Sheather: "A Modern Approach to Regression with R" the function
bctrans is used to calculate lambda for the variables. I use
"yeo.johnson" since there are values=0 in the data.
Doing this creates following output:
2> summary(bctrans(~v.obs+ snow+ pcpt+ Q.Enz+ qd+ HH6.1, data=sel.p1,
family="yeo.johnson"))
yeo.johnson Transformations to Multinormality
Est.Power Std.Err. Wald(Power=0) Wald(Power=1)
v.obs -49.9674 5.5747 -8.9632 -9.1426
snow -4.1130 0.3326 -12.3655 -15.3719
pcpt 0.6111 0.0811 7.5341 -4.7950
Q.Enz -0.8584 0.0904 -9.4967 -20.5601
qd -26.1100 2.3432 -11.1427 -11.5695
HH6.1 -6.0205 0.0023 -2653.7643 -3094.5528
LRT df p.value
LR test, all lambda equal 0 549.4523 6 0
LR test, all lambda equal 1 1414.1770 6 0
So what to do with that. I tried transforming my variables with the
Est.Power given in the output. I rounded the values more or less
arbitrarily for the first try:
v.obs<-(sel.p1$v.obs^(-0.5)-1)/-0.5
snow<-(sel.p1$snow^(-4)-1)/-4
pcpt<-(sel.p1$pcpt^(0.5)-1)/0.5
Q.Enz<-(sel.p1$Q.Enz^(-0.9)-1)/-0.9
qd<-(sel.p1$qd^(-26)-1)/-26
HH6.1<-(sel.p1$HH6.1^(-6)-1)/-6
trans<-merge(v.obs,qd,pcpt,snow,HH6.1,Q.Enz)
This gives me a lot of -Inf's which I d'ont like too much. I thought
about transforming the data first, e.g v.obs<-v.obs*10^5. But that
doesn't seem the right way, and doing that i often get errors from bctrans:
2> summary(bctrans(~ v.obs+ snow+ pcpt+ Q.Enz+ qd+ HH6.1, data=sel.p1,
family="yeo.johnson"))
Error in optim(start, neg.kernel.profile.logL, hessian = TRUE, method =
"L-BFGS-B", :
L-BFGS-B needs finite values of 'fn'
These errors also happen when i try another formula without the response
variable:
2> summary(bctrans(~ snow+ pcpt+ Q.Enz+ qd+ HH6.1, data=sel.p1,
family="yeo.johnson"))
Error in optim(start, neg.kernel.profile.logL, hessian = TRUE, method =
"L-BFGS-B", :
L-BFGS-B needs finite values of 'fn'
Does anybody have an idea how to cope with the data to get proper
parameters for the transformation?
Thanks a lot
Axel Kasparek
TU M?nchen
