The R checks for linear regression fail when using Eigen > 3.2.1

[lckarssen]

When running make check on e.g. an Ubuntu 15.10 machine, the R tests for linear regression fail. The Jenkins tests (http://jenkins.genabel.org) are OK, since that machine runs Ubuntu 14.04 LTS.

Some digging later it turns out that the Eigen library is the source.

[lckarssen]

Some more detailed test results:

• 3.2.0 OK (in Ubuntu 14.04)
• 3.2.1 OK
• 3.2.2 BROKEN
• 3.2.3 BROKEN
• 3.2.5 BROKEN (in Ubuntu 15.10)
• 3.3-alpha1 BROKEN
• 3.2 "current" (e059accf541a) BROKEN
• 3.3-beta2 BROKEN (in Ubuntu 16.10)
• 3.3.1 BROKEN

See also the discussion on the KDE forum: https://forum.kde.org/viewtopic.php?f=74&t=129151&sid=f20413e93d1953f6f1762f6548aa20dc

Edit: updated with new data points (3.3-beta2, 3.3.1)

[lckarssen]

Strangely enough with Eigen 3.2.8, the linear regression test script passes, but the logistic one fails for the 2df model. One value is one off in the 6th digit when comparing text input with filevector input.

[lckarssen]

Hmm, it seems that using Eigen 3.2.8 results in failure of the palinear R test script as well.

Output of the 2DF model for test SNP 6 from R:

 beta_SNP_A1A2 sebeta_SNP_A1A2 beta_SNP_A1A1 sebeta_SNP_A1A1 chi2_SNP_2df
 6.734722        6.490411            NA              NA              1.0975797 

Output ProbABEL with Eigen 3.2.1:

rs3174230 G C 0.8123 1.2e-05 0.9991 0.0185 182 0.997253 19 7845238 0 0 -6.73472 6.50872 1.09758

Output ProbABEL with Eigen 3.2.8:

rs3174230 G C 0.8123 1.2e-05 0.9991 0.0185 182 0.997253 19 7845238 352 7.683e+07 345.265 7.683e+07 1.09758

[lckarssen]

After some discussion on the genabel-devel mailing list we have decided to include Eigen 3.2.1 with the ProbABEL source. This way, this issue won't delay the release of ProbABEL v0.5.0.

[lckarssen]

One more data point: Eigen version 3.3-beta2 distributed with Ubuntu 16.10 also fails the R check, as does v3.3.1.