Lasso with SL.glmnet

[PFMB]

Hi,

I tried to estimate the AverageTreatmentEffect with 'ltmle' using solely
SL.library="SL.glmnet" for the LASSO for variable selection which results in:

Error in lognet(x,is.sparse, ix, jx, y, weights, offset, alpha, nobs, : one
multinomial or binomial class has 1 or 0 observations; not allowed
Additional warning:
In FUN(X[[i]], ...) : Error in algorithm SL.glmnet
The Algorithm will be removed from the Super Learner (i.e. given weight0)

Error occured during call to SuperLearner:
Q.kplus1 ~ A.1 + L.1 + Y.1 + A.2 + L.2 + Y.2 + A.3
Note that some SuperLeaner libraries crash when called with continuous dependent variables, as in the case of initial Q regressions with continuous Y or subsequent Q regressions even if Y is binary.
The error reported is:
Error in system.time({ : All algorithms dropped from library

Used R-Code:

set.seed(123)
n <- 36 # no. of obs are particularly low
t <- 3 # points in time

A <- data.frame(matrix(rbinom( n*t ,1,0.6), n, t))
L <- A+data.frame(matrix(rnorm( n*t ,0,1), n, t))
Y <- L*data.frame(matrix(rgamma( n*t,2,4), n, t))
df <- data.frame(A[,1],L[,1],Y[,1],A[,2],L[,2],Y[,2],A[,3],L[,3],Y[,3]) # assume A->L->Y
colnames(df) <- c("A.1","L.1","Y.1","A.2","L.2","Y.2","A.3","L.3","Y.3")
YRANGE <- c(min(Y),max(Y))

SL.lib1 <- c("SL.glmnet")
SL.lib2 <- c("SL.stepAIC","SL.knn","SL.gam","SL.glm.interaction")

ltmle_est <- ltmle(df, Lnodes = c(2,5,8), Anodes=c(1,4,7), Cnodes = NULL,
Ynodes=c(3,6,9), Yrange= YRANGE, estimate.time = FALSE,
gbounds=c(0.05,1),
abar = list(treament = rep(1L,t),
control = rep(0L,t)),
SL.library=SL.lib1)

I assume that the internal transformation of the continuous response variable
according to Yrange to [0,1] is the origin of the problem. I guess a change in the family argument is needed? I tried to define my own wrapper by modifying https://github.com/ecpolley/SuperLearner/blob/master/R/SL.glmnet.R without success. What do I need to change to fix this?

Many thanks and regards!

@ck37

[joshuaschwab]

Hi,
I don't know glmnet especially well, but I think the problem is that glmnet does not work with family="binomial" and continuous Y.
n <- 100
p <- 5
X <- matrix(rnorm(n * p), n, p)
beta <- rnorm(p)
Y.continuous <- plogis(X %*% beta + rnorm(n))
Y.binary <- rbinom(n, 1, Y.continuous)
m1 <- SL.glmnet(Y.binary, X, X, family=binomial(), obsWeights = rep(1, n)) #ok
m2 <- SL.glmnet(Y.continuous, X, X, family=binomial(), obsWeights = rep(1, n)) #error

You could use a different SL.library, but if you only have 36 observations, you may be better off using glm instead of SuperLearner (SL.library = "glm", the default).

Josh