Error Using OEM with Categorical Variables

[chandrabanerjee]

Hello,
I am trying to use OEM for grouped LASSO on a tall dataset which contains many categorical variables (and as a result, lots of binary variables in the model matrix). While running cv.oem, I keep getting the following error:

Error in oemfit.binomial(is.sparse, x, y, family, penalty, weights, groups, :
TridiagEigen: failed to compute all the eigenvalues

This does not just extend to my use case, but also to smaller datasets like the birthwt data from the MASS package. I am pasting a reproducible example below:

`
library(MASS)
library(splines)
library(oem)

#Load and create Model Matrix
data("birthwt")
view(birthwt)

birthwt$race = as.factor(birthwt$race)
birthwt$smoke = as.factor(birthwt$smoke)
birthwt$low = as.factor(birthwt$low)

X = model.matrix(low~ns(age,3)+ns(lwt,3)+race+smoke+ptl, birthwt)[,-1]
Y = birthwt$low

#Define Groups
grouping = c(1,1,1,2,2,2,3,3,4,5,5,5)

#Run cv.oem for Logistic Regression with Group LASSO penalty:
cvoem = cv.oem(X, Y, family = "binomial", penalty = "grp.lasso", groups = grouping, nfolds = 10)
`

Any help regarding: 1) an explanation of the issue and 2) a workaround would be much appreciated. cv.gglasso is just too slow!

Thanks.

[jaredhuling]

I am not able to reproduce your error. This is a known error, however, and should be fixed with the current development version on github. If you install the current version of oem on github do you still have this error?

I'm confused by your example, since the number of columns in X in your code is 10, but you have specified a group structure that has length 12. When I fix your code above, there is no error:

`
library(MASS)
library(splines)
library(oem)

#Load and create Model Matrix
data("birthwt")
view(birthwt)

birthwt$race = as.factor(birthwt$race)
birthwt$smoke = as.factor(birthwt$smoke)
birthwt$low = as.factor(birthwt$low)

X = model.matrix(low~ns(age,3)+ns(lwt,3)+race+smoke+ptl, birthwt)[,-1]
Y = birthwt$low

#Define Groups
grouping = c(1,1,1,2,2,2,3,3,4,5)

#Run cv.oem for Logistic Regression with Group LASSO penalty:
cvoem = cv.oem(X, Y, family = "binomial", penalty = "grp.lasso", groups = grouping, nfolds = 10)
`

[chandrabanerjee]

Hi Jared,
Thanks for the quick response. I think I deleted one categorical variable by mistake. Attached is the updated code.

`
library(MASS)
library(splines)
library(oem)
#Load and create Model Matrix
data("birthwt")
view(birthwt)

birthwt$race = as.factor(birthwt$race)
birthwt$smoke = as.factor(birthwt$smoke)
birthwt$ptl = as.factor(birthwt$ptl)
birthwt$low = as.factor(birthwt$low)

X = model.matrix(low~ns(age,3)+ns(lwt,3)+race+smoke+ptl, birthwt)[,-1]

Y = birthwt$low

grouping = c(1,1,1,2,2,2,3,3,4,5,5,5)

cvoem = cv.oem(X, Y, family = "binomial", penalty = "grp.lasso", groups = grouping, nfolds = 10)

This returns the error. I will try and test out the development version of the package.