MaAsLin2 1.12.0

Hi,

I have two statistics questions with Maaslin2 (linear model), multi-wilcoxon with correlation test, and Tukey tests in ANOVA analysis. When I set a variable with 2 levels and let Maaslin2 (linear model) run it, does it have any statistical power difference between Maaslin2 (linear model) and multi-wilcoxon with correlation test? Another question is that when I set a variable with more than 2 levels (e.g., 5 levels) and let Maaslin2 (linear model) run it, does it have any statistical power difference between Maaslin2 (linear model with 5 levels) and Tukey test in ANOVA analysis (5 groups in total)?

Best wishes,

Shuyuan

Here we add premises that we have two different sample sizes. One is 10 samples in each group and another one is 100 samples in each group.

Hi Shuyuan,

I talked with both @andrewGhazi and @sma and these are the thoughts that they came up with:

If the linear model assumptions are satisfied, the linear model will have higher power both before and after multiple testing correction when compared to a wilcoxon test. However, in many cases the linear model assumptions are almost never strictly satisfied, so determining the real world power difference would likely require a simulation study.

Similarly, it seems both Tukey’s and Maaslin-ANOVA are parametric tests. The difference being Tukey’s test statistic focuses on the *max* difference between the groups, whereas ANOVA is a (weighted) sum of squares of *all* pairwise differences between groups. This is pure intuition, but I suspect Tukey’s test is more powerful when there is an “outlier” group whose mean is vastly different from other groups, contributing to the max difference. On the other hand, if all groups are only moderately different from each other, then ANOVA would have better power.

Hope that is helpful!

Cheers,

Jacob Nearing