SCM

[#5726] ProbABEL Cox regression gives (too) many NaNs, does not match with R output

View Trackers | Bugs | Download .csv | Monitor

Date:
2014-05-26 13:43
Priority:
4
State:
Closed
Submitted by:
Lennart Karssen (lckarssen)
Assigned to:
Lennart Karssen (lckarssen)
Resolution:
Fixed
Operating System:
All
Severity:
major
Hardware:
All
Version:
PA v0.4.3
Component:
ProbABEL
URL:
http://forum.genabel.org/viewtopic.php?f=10&t=866
 
Summary:
ProbABEL Cox regression gives (too) many NaNs, does not match with R output

Detailed description
The URL associated with this report points to the forum discussion. Aside from the people in the forum discussion I also got a report via e-mail about the same bug.

Comparing CoxPH regression output between R and pacoxph gives different results for many (all?) SNPs. pacoxph gives nan for those SNPs, whereas R does give output.

The reports are inconclusive about whether this behaviour is present in v0.4.3 only or also in v0.4.2.

In v0.4.2 I implemented additional convergence checks in the estimate() function of pacoxph, based on the way R checks the output.

As user wuttke points out in the forum I made a mistake there. R's survival package checks each component in the regression coefficients (beta vector) for anomalies. I misread R's code when implementing the checks. Maybe this is the cause of the problems...

Followup

Message
Date: 2014-08-19 06:43
Sender: Lennart Karssen

According to reports on the forum this bug is fixed in the coxfix branch in r1751.

Attached Files:

Changes:

Field Old Value Date By
ResolutionAccepted As Bug2014-10-28 08:48lckarssen
close_dateNone2014-10-28 08:48lckarssen
status_idOpen2014-10-28 08:48lckarssen
Thanks to:
Vienna University of Economics and Business University of Wisconsin - Madison Powered By FusionForge