2

Hi I am trying perform a variable selection using the lasso approach, but my model includes a random effect.

I have run the model without the random effect using both the cv.glmnet and optL1 functions in glmnet and penalized (and got results); however, when I try to run the model with the random effect using glmmLasso I get NAs for much of the output.

Here is a sample of what my data looks like. I have standardized all the predictor variable (x1-x5). The random factor is plot (there are two observations per plot)

head(data)
     Plot     y       x1         x2            x3       x4         x5
[1,]    1  33.38637 -0.3778876 -0.6493490 -1.9874184  0.13397776 -0.70846326
[2,]    4 26.75993 -1.9377779 -0.6932049  0.6747408  1.28276574  0.04818805
[3,]    5 29.25769 -0.1512598  0.7053084  1.0550493 -0.66133701 -0.72988495
[4,]    6 26.47273 -0.3711473  1.5288232  0.2944324 -0.83807362 -0.58073941
[5,]    6 24.32503 -0.9079206 -1.0732886  0.6747408  0.04560945  2.48661737
[6,]    7 32.93299 -0.7321292  0.1546743 -0.0858761 -0.13112716 -0.50712361

Here is the model I ran:

mod<-glmmLasso(y~x1+x2+x3+x4+x5, rnd=list(Plot=~1), data=data,lambda=10)

And the output:

    summary(x)
Call:
glmmLasso(fix = y~ x1+x2+X3+X4+x5, rnd = list(Plot = ~1), data = data, lambda = 10)


Fixed Effects:

Coefficients:
               Estimate StdErr z.value p.value
(Intercept)  53.4562879     NA      NA      NA
x1          -32.2591971     NA      NA      NA
x2           -0.7791392     NA      NA      NA
x3           -0.0023207     NA      NA      NA
x4           -0.0148370     NA      NA      NA
x5           -0.1593585     NA      NA      NA

Random Effects:

StdDev:
         Plot
Plot 3.294686

I originally thought it was not working due to correlations between the predictor variables based on this: Getting p-values for all included parameters using glmmLasso

But when looking at correlations, nothing is too high. I even tried removing x2 as it had the highest correlations but I still got NAs in the model output.

            x1           x2             x3          x4            x5
   x1    1.00000000    0.1632633 -0.08125663   -0.15017421    0.11155284
   x2    0.16326326    1.0000000 -0.31222700    0.30135509    0.14519762
   x3    -0.08125663   -0.3122270  1.00000000   -0.09517447   -0.05023544
   x4    -0.15017421    0.3013551 -0.09517447    1.00000000    0.19224986
   x5    0.11155284    0.1451976 -0.05023544    0.19224986    1.00000000

I am wondering why I am getting NAs. Is it due to my sample size (either only 2 observations per random effect grouping or only having a total of 31 observations?) Any thoughts would be appreciated. Thanks!

broch
  • 97
  • 6

1 Answers1

2

Set final.re=TRUE in your glmmLasso() call.

This performs the final Fisher scoring re-estimation.

user11307953
  • 21
  • 2