randomness in lasso

[Brian Holtemeyer]

I noticed that my lasso results (using "cv" selction method) change when the sortedby property changes.

Here's an example:

set seed 1234
clear
set obs 1000

gen y = rnormal()

forv z = 1/500 {
    gen x`z' = rnormal() * y * .1
}

qui foreach k in x1 x2 {
    sort `k'    
    isid `k'
    lasso     linear y x*, selection("cv")     
    noi di " names of all selected variables when sorted by `:sortedby' :  `e(allvars_sel)' "
}

here's the output from that code above:

 names of all selected variables when sorted by x1 :  x57 x94 x160 x176 x198 x230 x300 x305
 names of all selected variables when sorted by x2 :  x57 x84 x94 x122 x157 x160 x176 x198 x206 x230 x287 x300 x305 x434 x491

I hadn't thought enough about it and I naively assumed that by setting the seed I'd get the same results regardless of -sortedby-. After having thought about it, I now assume that the -sortedby- property is affecting the "k-fold" groups that are used by the cross validation?

Finally, I assume that the seed choice determines the k-fold groups?

Thanks for helping me think through this.