Question

I am trying to adapt some (to me) very complicated code to work with my data.

I think the crux of my problem is that some of my variables lose dimension when I begin with a two-dimensional matrix, and I need to know how to make variables retain their dimensions.

I am starting with two variables, e (a data.frame), a portion of which looks like this:

e <- 
structure(list(X2hr = c(0.106, 0, 0, 0, 0.01, 0.042), X6hr = c(1, 
0.083, 0.006, 0, 1, 0.967), X12hr = c(0.049, 0.057, 0.098, 0.405, 
0.046, 0.029), X24hr = c(0.264, 0.301, 0.025, 0.15, 0.58, 0.487
), X36hr = c(0.284, 1, 0.114, 1, 0.671, 1), X48hr = c(0.274, 
0.235, 0.299, 0.253, 0.617, 0.636), X72hr = c(0.098, 0.021, 1, 
0.325, 0.283, 0.35)), .Names = c("X2hr", "X6hr", "X12hr", "X24hr", 
"X36hr", "X48hr", "X72hr"), row.names = c("cgd1_10", "cgd1_100", 
"cgd1_1000", "cgd1_1010", "cgd1_1020", "cgd1_1030"), class = "data.frame")

and m (a 2-dimensional matrix, with one column and 2913 rows), a portion of which looks like this:

m <- 
structure(c(0, 0, 1.174805088, 1.174805088, 0, 0), .Dim = c(6L, 
1L), .Dimnames = list(c("cgd1_10", "cgd1_100", "cgd1_1000", "cgd1_1010", 
"cgd1_1020", "cgd1_1030"), "X4_1110_2.motif2"))

I load the glmnet package define two functions, IDC.glmnet and PBM.glmnet.getCoefs:

library(glmnet)
IDC.glmnet <- function(e, m, mode="coef", randomize=F, alpha=0.5) {
  nona  <- !is.na(e)
  enona <- e[nona]
  mnona <- m[nona,]
  if(ncol(m)==1)
    dim(mnona) <- c(sum(nona),ncol=1)
  e.cv <- cv.glmnet( mnona, enona, nfolds=10)
  l <- e.cv$lambda.min
  #print(l)
  if (randomize == TRUE) {
    enona <- sample(enona)
  }
  e.fits <- glmnet( mnona, enona, family="gaussian", alpha=alpha, nlambda=100)
  if (mode == "predict") {
    cor.test(predict(e.fits, mnona, type="response", s=l), enona)$estimate
  } else {
    as.matrix(predict(e.fits, s=l, type="coefficients")[-1,])
  }
}

PBM.glmnet.getCoefs <- function(e, m, alpha=0.05, randomize=F, center=FALSE) {
  e.coef <<- apply(e, 2, IDC.glmnet, m, mode="coefficients",
                   alpha=alpha, randomize=randomize)
  if (dim(e)[2] > 1) { 
    e.coef.s <- t(apply(e.coef, 1, scale, center=center))
  } else {
    e.coef.s <- e.coef
  } 
  rownames(e.coef.s) <- colnames(m)
  colnames(e.coef.s) <- colnames(e)
  e.coef.s
}

Then I try to execute PBM.glmnet.getCoefs on my variables:

coefs <- PBM.glmnet.getCoefs(e, m)

And I get the following error message:

Error in t(apply(e.coef, 1, scale, center = center)) : 
  error in evaluating the argument 'x' in selecting a method for function 't':
  Error in apply(e.coef, 1, scale, center = center) : 
    dim(X) must have a positive length

The problem occurs when I use a single-column matrix for m. If I have multiple columns, it works fine. But I can’t use multiple columns because it skews the results, and I really need to be able to use a single-column m. From my limited troubleshooting abilities, I think that this line in the PBM.glmnet.getCoefs function is where the trouble begins:

e.coef <<- apply(e, 2, IDC.glmnet, m, mode="coefficients",
                 alpha=alpha, randomize=randomize)

e.coef is a vector when I use a single-column m. Then since e.coef is dimensionless, I get the error in t(apply) listed above.

e.coef looks like this:

> e.coef   
        X2hr         X6hr        X12hr        X24hr        X36hr        X48hr 
 0.025701875  0.004066947  0.043836383  0.020151361  0.003512643 -0.035211133 
       X72hr 
-0.034503722 

How can I make sure that e.coef retains the proper dimensions (1 row and 7 columns, column headings taken from top row of e, row values determined somewhere in the IDC.glmnet function)?

Answer 1

You correctly identified the line causing the issue. The problem is described in the Value section of ?apply: “‘apply’ returns a vector if ‘MARGIN’ has length 1”.

So make this small change to ensure the dimensions are correct:

PBM.glmnet.getCoefs <-
function(e, m, alpha=0.05, randomize=F, center=FALSE ) {
   e.coef <<- apply(e, 2, IDC.glmnet, m, mode="coefficients",
                    alpha=alpha, randomize=randomize)
   dim(e.coef) <<- c(ncol(m), ncol(e))
   if (dim(e)[2] > 1) { 
     e.coef.s <- t(apply(e.coef, 1, scale, center=center))
   } else {
     e.coef.s <- e.coef
   }    
   rownames(e.coef.s) <- colnames(m)
   colnames(e.coef.s) <- colnames(e)
   e.coef.s
}