What is the "Effect Size" in LefSe?

Question

[dpc]

Hi - I have a basic question regarding the tool LefSe which is used for biomarker discovery. What I understand so far is, this tool at first does a Kruskal-Wallis test between groups followed by Wilcoxon rank-sum test among the subgroups. So far, so good. At last, it calculates the Effect Size. To my knowledge, the effect size may be of different types: Standardised mean difference, correlation co-efficient, Odd ratio. Can you please tell me which one is calculated in LefSe? And, how it is calculated?

Thanks, dpc

Answer 1

[zorbax]

According to the original paper the effect sizes "are obtained by averaging the differences between class means (using unmodified feature values) with the differences between class means along the first linear discriminant axis, which equally weights features' variability and discriminatory power.". The implementation in R through Python is in the L219-L220:

robjects.r('effect.size <- abs(mean(LD[sub_d[,"class"]=="'+p[0]+'"]) - mean(LD[sub_d[,"class"]=="'+p[1]+'"]))')
scal = robjects.r('wfinal <- w.unit * effect.size')

L224 takes absolute value like the coefficient.

[dpc]

Thanks zorbax. The default LDA value has kept as 2.00 in LefSe. Can you please suggest the minimum LDA score I can consider to find biomarker by LefSe analysis?

[stro0070]

I would also be interested in knowing what effect size can be considered significant. For example, if no biomarkers have an effect size of 2.0, could one consider those with effect sizes greater than 1.0 as being marginally significant? They’re reported on a log10 scale, so does an effect size of 1.0 mean 10x difference?