SIMPER vs. MRPP + ISA

[Allie Gardner]

Hello,

I am working on an experiment where I am comparing the microbial communities associated with four groups with five replicates per group. Initially my approach with these data was to use MRPP to test for pairwise differences between groups. I then used ISA to identify the taxa that most strongly differentiate groups. This was all straightforward to do using PC-ORD and your helpful companion book. It has been suggested to me that I use SIMPER instead for this analysis (like what you could do using PAST or PRIMER-E). My questions are: 1) Is it possible to run SIMPER using PC-ORD? and 2) Do you see an advantage to using SIMPER over the MRPP+ISA method and/or would you expect to see a qualitative difference between the results of these approaches? To me they look fairly similar mathematically, both distance-based methods using different metrics for differences in individual taxa between groups. MRPP and ISA struck me as nice complementary methods to get a test of significance for the null hypothesis of no difference between groups and then identify key taxa. But if SIMPER is somehow more "valid," then SIMPER it is....

Thanks,

Allie Gardner

[Bruce McCune]

Allie, here are my views.

1. No SIMPER on PC-ORD (interestingly we have hardly had any requests for it).

2. SIMPER and ISA have pretty different fundamentals, in that ISA has two specific and attractive criteria for defining indicators (faithfulness and exclusiveness), while SIMPER is looking at individual species contributions to a specific distance measure. ISA is quite popular, I think in part because it is both effective and incredibly intuitive, so that even people who know nothing about community analysis can get the idea. But having said that, I know of no in-depth comparison between SIMPER and ISA (which doesn't mean there hasn't been one), and it is quite possible that they would often lead to the same qualitative conclusions. One other thing: it is standard practice with ISA to use a randomization test to evaluate statistical significance. I'm not sure that is part of SIMPER, but maybe a SIMPER user can chime in on that.

I've never heard anyone say that SIMPER is more valid than ISA -- they are just different.

Popularity is not a good measure of merit, but fyi, googling "Indicator species analysis" gives 2.1 million hits, while googling "SIMPER" gives 472,000 hits.

MRPP is very similar to ANOSIM, and in comparisons I have seen their performance (i.e. power) is quite similar. One nice thing about MRPP is that it is well grounded in the statistical literature (see papers and books by Paul Mielke).

Hope this helps,

Bruce McCune

--
You received this message because you are subscribed to the Google Groups "PC-ORD" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pc-ord+un...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Allie Gardner]

Dr. McCune,

Thanks for your quick reply and your insights. I've seen a couple papers that have used both SIMPER and ISA, though not specifically for the purpose of comparing the two approaches. The impression I have gotten from these few papers is that SIMPER does not seem to pick up any important taxa that ISA misses, but only a subset of the taxa picked up by SIMPER are statistically significant for ISA. I am not familiar with a permutation test for SIMPER, but I am new to all this so there well may be one that I don't know about. That is a good point you make that ISA is intuitive when it comes to interpreting the result. I'll think about this more and talk with my advisors about the best approach with our research questions.

If you don't mind, I have one other question regarding MRPP. In a separate experiment I also am comparing microbial communities under different treatments, but this time the comparisons are being done over time and also at different locations so there is a blocking variable. Am I correct in thinking that MRPP cannot handle this experimental design? I've gotten the sense from your book that if only the location block existed or only the time effect you could use blocked MRPP, but PC-ORD (and/or MRPP in general) could not handle the design with both block and location. In this case, would the best solution be to slice the experiment so I'm just comparing among the treatments within single time points and within each treatment across multiple time points?

Thanks again,

Allie

[Bruce McCune]

Allie, You can handle one or  two factors with MRPP or the purely nonparametric form of perMANOVA used in PC-ORD. But in your second paragraph you have three factors: blocks, treatments, and date. So yes, slicing it by date and comparing the treatment effect through time (e.g. graph test statistic over time) is one way to deal with that.

- Bruce McCune

[scott miller]

---

From: Allie Gardner <alli...@gmail.com>
To: PC-ORD <pc-...@googlegroups.com>
Sent: Sunday, March 13, 2016 7:41 PM
Subject: Re: SIMPER vs. MRPP + ISA