Search code examples
rmodel-validationmodel-fitting

L1 penalized regression fails to predict from model

This question may be too package specific, but I would value input on what can be wrong in my use of the predict function on my data set.

The procedure I'm using is the following:

require(penalized)
# neg contains negative data
# pos contains positive data

Now, the procedure below aims to construct comparable (balanced in terms os positive and negative cases) training and validation data sets. 

# 50% negative training set
negSamp <- neg %>% sample_frac(0.5) %>% as.data.frame()
# Negative validation set
negCompl <- neg[setdiff(row.names(neg),row.names(negSamp)),]
# 50% positive training set 
posSamp <- pos %>% sample_frac(0.5) %>% as.data.frame()
# Positive validation set
posCompl <- pos[setdiff(row.names(pos),row.names(posSamp)),]
# Combine sets
validat <- rbind(negSamp,posSamp)
training <- rbind(negCompl,posCompl)

Ok, so here we now have two comparable sets.

[1] FALSE  TRUE
> dim(training)
[1] 1061  381
> dim(validat)
[1] 1060  381
> identical(names(training),names(validat))
[1] TRUE

I fit the model to the training set without a problem (and I've tried using a range of Lambda1 values here). But, fitting the model to the validation data set fails, with a just odd error description.

> fit <- penalized(VoiceTremor,training[-1],data=training,lambda1=40,standardize=TRUE)
# nonzero coefficients: 13
> fit2 <- predict(fit, penalized=validat[-1], data=validat)
Error in .local(object, ...) : 
  row counts of "penalized", "unpenalized" and/or "data" do not match

Just to make sure that this is not due to some NA's in the data set:

> identical(validat,na.omit(validat))
[1] TRUE

Oddly enough, I may generate some new data that is comparable to the proper data set:

data.frame(VoiceTremor="NVT",matrix(rnorm(380000),nrow=1000,ncol=380) ) -> neg
data.frame(VoiceTremor="VT",matrix(rnorm(380000),nrow=1000,ncol=380) ) -> pos
> dim(pos)
[1] 1000  381
> dim(neg)
[1] 1000  381

and run the procedure above, and then the second fit works! How come? What could be wrong with my second (not training) data set?

Solution

  • Ok,

    I found the solution to this problem. The problem was in my finding of complementary data sets. 

    neg[setdiff(row.names(neg),row.names(negSamp)),]

    does not do the right thing, but

    neg %>% 
rownames_to_column() %>% 
filter(! rowname %in% row.names(negSamp)) %>% 
column_to_rownames() %>% data.frame()

    does. With this change, along with using data.frame instead of as.data.frame then it all works.