help interpreting output during rjunglesparse imputation and RJungle (GWAS data)

[Allison Cox]

Hi,

I am using Random Jungle to look for epistasis in GWAS.  

I had some missing SNP values so I imputed the values in rjunglesparse.  The imputation outputs a confusion matrix and gini importance values.  What is this output for - is it for the imputation or for the case-control random-forest analysis?  

The reason I ask is that the confusion matrix from the imputation output is MUCH BETTER that what I'm getting in the regular analysis.  In the regular analysis, I'm getting a confusion matrix that looks like this:

Number of variables: 506177

Test/OOB set: 

(real outcome == rows / predicted outcome == columns )

0 1 error

0 4 90 0.957447

1 6 112 0.0508475

0.45283

And the highest gini importance index is 0.1.

The output from the imputation looks like this:

Number of variables: 506177

Test/OOB set: 

(real outcome == rows / predicted outcome == columns )

0 1 error

0 15 18 0.545455

1 22 27 0.44898

0.487805

And the highest gini importance index is over 10.

Can you clarify what I'm seeing here?  I'm running the epistasis analysis because there weren't any good hits via regular GWAS - is the uneven confusion matrix just evidence of no strong association with the phenotype?

Thanks for your help!

Allison

[Jochen Kruppa]

Dear Allison,

the random jungle algorithm uses the original imputation method by Breiman. Breiman stated that the results of the imputation algotithm might be over optimistic. We do not recommend to use this approach with SNPs because it ignores the given LD structure. Please have a closer look at PLINK or IMPUTE.

A PLINK tutorial is given here: http://pngu.mgh.harvard.edu/~purcell/plink/pimputation.shtml

Best

Jochen

2013/5/6 Allison Cox <aj...@yahoo.com>

Allison

--
You received this message because you are subscribed to the Google Groups "randomjungle" group.
To unsubscribe from this group and stop receiving emails from it, send an email to randomjungle...@googlegroups.com.
To post to this group, send email to random...@googlegroups.com.
Visit this group at http://groups.google.com/group/randomjungle?hl=en.
For more options, visit https://groups.google.com/groups/opt_out.