Hi Julie,

sometimes it is helpful to find out why the observations are outliers - which from your variables makes it to be outlier (it could be one, more or all of variables). Then, when you, for example, realize that some of your variables are responsible for that (e.g. beacause of instable or unreproducible variable), you could try to remove that variables instead of sample removing.

This is one (but not only) possibility what you can try.

Dear Julie,

If you use e.g. metaboanalyst.ca, there is a test for outlier detection under random forest. In the faq there is a Link for detailed information. Then you can define cut-offs which is a bit more objective than removal by eye.

Hope this helps! Best,

Simon

In addition to Simons answer I always use a trias of PCA and hierarchical cluster analysis in combination with random forrest outlier testing with the MetaboAnalyst tool for detection of outliers.

Good look :)

Hi, don't remove the samples just use a robust method for data analysis like Manuel mentioned, there are already some available. Here is another one: 10.1093/biostatistics/kxl005

best

Hi, This is a very interesting and relevant question for the entire research community working in metabolomics. Please feel free to register (for free to create a login with your Email ID!) and join the International Metabolomics Society’s Forum: http://www.metabolomics-forum.com/index.php?action=forum;?action=forum#c12 to ask more specific questions any software, database, approach and platform (mass spectrometry, NMR) issues where you can get very specific responses from expert users and researchers! Thanks again, Biswa

Dear all,

thank you for your input.

I just read some paper where they removed outliers by PCA "based on a tolerance of 95% beyond the Hotelling's T2 ellipse". Any opinion on this method?

Thanks

Best

Julie