Coefficient value greater than 1

Downstream analysis and statistics MaAsLin
1

Hello,

I am having results from running Maaslin2 and one feature reports a coefficient of 1.26. Is this normal? It seems weird.

Thank you,
Giorgio

2

Hi Giorgio - one quick way to verify this would be to look at the corresponding scatter/box plot and see if this effect size translates to the strong association your are seeing.

3

Hi Himel,

Thank you for your answer. So you are telling that it is possible to have a value greater/smaller than +/- 1 and it is only a matter of verifying the strength. Should maximum strength max out at |1| ?

4

Hi @Giorgio_Casaburi - if you standardize (z-score) the metadata (which is only done for continuous variables when prompted), you will indeed get standardized beta coefficients ranging between -1 and 1. Otherwise, you can get coefficients greater than 1.

1 Like
5

Sir @himel.mallick , I also have faced the same thing as I run with everything default. Can you please tell me what is the minimum co-efficient I should consider to call the features as differentially abundant? Or, should I consider all the features in the significant_results.tsv file?

Thanks,
DC7

6

Hi,

I have a follow-up question on this topic. I have recently run Maaslin2 version 1.12.0 looking at group differences (so the sole predictor in my model is a binary categorical variable). I have included the parameter standardize=TRUE, but I am still getting coefficients larger than 1 (range = -0.03 to 1.57). Is that expected behavior?

Thanks so much!
Fran

7

Standardizing continuous metadata puts the regression coefficients in units of standard deviations instead of the variable’s units. For instance, if you used height in meters as a covariate, the units for a coefficient in the Maaslin2 outputs would be “log rel abundance units / meter”. If you re-ran the analysis with height in centimeters, the coefficients would all be smaller by a factor of 100. But if you standardized height, the units will be “log rel abundance units / standard deviation of height” and you’ll get the same coefficient values regardless of whether you use meters or centimeters.

With the default log transform, there’s no guarantee that the coefficients will fall in a set (-1, 1) range. I think Himel’s original comment to that effect was in the context of raw relative abundance proportions, which will have a restricted range since you can’t get a bigger effect than 0% → 100%.