Change in LDA score

[Duy Dinh]

Hi Nicola,

I have another question I want to ask.  Have you change how the LDA score is calculated recently?

I'm asking because I have used the same data to look for significant features.  I noticed that on my old LDA score, all the features have about  .5 more in LDA score.  I took notice because initially our Barnesiella had an LDA score of over 2 but no Barnesiella was reported the second time.  

I lowered the LDA threshold to 1.5 and Barnesiella again showed as significant.  Can there be any other reason for this discrepancy? 

I have attached the two files here.  The first one is the "HIV- and HIV+ no subclass...".  The second is the older one which is "12_12_13 LefSe ..".

Thank you,

Duy

PS sorry about the multiple questions.

[Nicola Segata]

Hi Duy,

 thanks for getting in touch. The LEfSe algorithm has not been changed recently. The last change was a small bug fix for rare cases in which very few samples were present. Some updates might have occurred to some of the libraries LEfSe internally use, but this should not affect the values considerably. Notice also that the LDA computation is based on boostrapping which has a randomization component, but the differences between runs are minimal. When very few samples are present the random variability can be a little higher. So unless you even slightly changed the input it's hard for me to find an explanation.

Notice, however, that the LDA score and the threshold at 2.0 are not an absolute "significance" estimate. The most important aspect of the LDA score is the ranking it induces and the relative differences among biomarkers and both these aspects seem identical in the two plots.

I hope this helps

thanks

Nicola

[Duy Dinh]

Thanks Nicola!

[lina....@gmail.com]

Hello Nicola,

I ran into the same issue that Duy had. My collaborator wanted me to change the colors and orders of the two groups, and I figured I could just change the labels and rerun the Lefse analysis, and now I am getting different orders of the organisms within the groups.

I attached below two figures. 

Can this still be attributed to the random effect in the bootstrapping procedure?

Thanks for any insight you might have!

Best,

~Lina

[Nicola Segata]

Hi Lina,

 it's a bit unusual but yes, I think that is due to the random effect in the bootstrapping procedure.

cheers

Nicola

[lina....@gmail.com]

Thanks for the clarification, I appreciate it!
~Lina

[capricy gao]

Hello Nicola,

Could you please explain how the lda scores are calculated. The paper only mentioned the bootstrap steps, but was not clear how it is really calculated. I recently start to use it and wonder how those scores are related to the coefficients of linear discriminants?

Thank you very much for your attention!

C.

[capricy gao]

Hello Nicola,

Could you please explain how the lda scores are calculated. The paper only mentioned the bootstrap steps, but was not clear how it is really calculated. I recently start to use it and wonder how those scores are related to the coefficients of linear discriminants?

Thank you very much for your attention!

C.