This is a test version of Biostars. For the public version, visit https://www.biostars.org.
Some 'NA's in plink logistic regression results

Hi All,

When I perform logistic regression on my data using PLINK, some SNPs (not all SNPs) have NA in OR,SE,L95,U95,STAT and P columns

What does it mean for that SNP? Is it associated?

Thanks in advance

gwas logistic regression plink --logistic
• 9,477 views • link
updated by chrchang523 chrchang523 • written by Mel • Mel

1 answer

This can happen if all your genotypes at a locus are identical (no association analysis is possible then); have you checked whether that's the case?

• 0 views • link
written by chrchang523 chrchang523

Hi chrchang523,

Thanks for the reply!

Yes I've check this, I'm just looking at a group of SNPs in particular, when I perform a straight forward -assoc test all these SNPs have minor allele frequencies in either cases, controls or both. The SNPs that have minor allele frequency of 0 in one either case or control, then I'd expect the logistic regression to be NA for these (see table SNP 1). But for some SNPs there is a MAF in cases and controls, and yet the logistic regression is NA (see table SNP 2 and 3)

snpName    MAFcase    MAFcontrol    LogisticRegressionPvalue    OR        L95       U95
1          4.014      0             NA                          NA        NA        NA
2          4.119      0.6536        NA                          NA        NA        NA
3          0.1144     2.288         NA                          NA        NA        NA
4          17.73      30.59         1.71E-02                    0.5943    0.3875    0.9115

Could it be that the MAF is to small for SNP 2 and 3? Like with a Chi squared test needs to see a minimum of 4 observations for the test to work.

• 0 views • link
updated by Ram Ram • written by Mel • Mel

The other two most likely issues are:

  1. Multicollinearity. If you have a covariate (or a linear combination of them) which behaves almost identically to the genotype, the regression doesn't converge to a unique solution.
  2. Random convergence failure, even though mathematically there shouldn't be a problem; the logistic regression algorithm employed by PLINK isn't perfect. It was recently updated, though; you might want to check whether PLINK 1.9 also gives you NAs.
• 0 views • link
updated by Ram Ram • written by chrchang523 chrchang523

Log in to answer this question.

More posts like this