I'm running a comparison of fish species communities between 6 sites at 3 locations (North, East and West) where the sites either have a reef at them (Yes, No). I found a significant effect through an ANOSIM and PERMANOVA so I did a follow up analysis via a two-way crossed analysis in SIMPER to try to determine which species were contributing to the differences. My SIMPER analysis identified 6 species which explained >70% of the variation between the sites with a reef versus the sites without a reef. The number one species (F0301) contributed 19.55% to the dissimilarity between the two groups. To illustrate this point visually I tried to build some simple bar graphs showing the average number of F0301 at the reef sites versus the non-reef sites but these values were almost identical (mean = 4.15, and 4.16; SE = 0.9, 1.1).

What is happening here? I have tried controlling for location, and running a paired T-test but there's still no difference between the raw values. However the SIMPER analyses appear legitimate-- an NMDS plot appears to show a higher concentration of species F0301 in the no reef group, and F0301 is the only species that has a correlation factor > 0.6.

All tests were run in PRIMER V7

  • $\begingroup$ You've compared the number of F0301, but have you tried comparing the proportional abundance? i.e. the number of F0301 might be similar between different sites, but if total diversity is lower in one type of site then the proportion of species that are F0301 will be higher. Remember that the NMDS is based on distances between sites, where as your t-tests are looking at the raw numbers. What distance measure did you use - Bray-Curtis? $\endgroup$
    – rw2
    Commented Jan 18, 2021 at 11:38
  • $\begingroup$ @rw2 I reran it with the proportional abundances and the values are as close to identical as you can get (mean = 0.21214, 0.21214; SE = 0.034, 0.029), and yes I am using a Bray-Curtis dissimilarity index. $\endgroup$
    – Dugan
    Commented Jan 18, 2021 at 15:47
  • $\begingroup$ @rw2 I also get similar results if I determine the rank of the species abundances and compare those between the reef versus non-reef group: (mean = 2.99, 2.91; SE = 0.38, 0.41) $\endgroup$
    – Dugan
    Commented Jan 18, 2021 at 16:05
  • $\begingroup$ I'm coming to the conclusion that SIMPER is just a garbage analysis as was highlighted by Warton et al. in their paper: "Distance-based multivariate analyses confound location and dispersion effects" $\endgroup$
    – Dugan
    Commented Jan 18, 2021 at 19:35
  • $\begingroup$ Yes, it could be that - it's difficult to be sure without seeing the data. Maybe you could try using Joint Species Distribution Modelling, if your number of species/sites is not too high. $\endgroup$
    – rw2
    Commented Jan 20, 2021 at 15:01

1 Answer 1


After extensive research I figured out what the issue was-- essentially there's an interaction effect that's taking place where species F0301 is more common at island sites at the North and East locations, but more common at non-island sites in the Western locations-- this is partially explained by a lack of nearby alternative habitat at the North and East locations which forces a tighter association with the island sites. Unfortunately SIMPER doesn't flag interaction effects well, so it was easy to miss. Also, there are lots of issues with SIMPER overall in that it can only detect significant species when they're relatively abundant, and there's a significant amount of variation in the number of individuals sampled between sites. This stems from the fact that SIMPER, like PERMANOVA attempts to condense species interactions to a single string of numbers to facilitate computation of all of the possible permutations-- this is because this method was developed in the 80's when computing power was significantly lower. The latest approach uses a permutation of the whole matrix using a general linear mixed model for multivariate data and analyzes all possible species/species correlations. It's quick and easy to set up in R through the mvabund package, and if you run a follow up multivariate ANOVA you can analyze the effect of each factor (e.g., location, presence of a reef and the interaction effect between the two) on each species, which prevents possible confusion about how each factor is affecting the species.

Check out this video for a more detailed explanation of why general linear multivariate models are a better option https://www.youtube.com/watch?v=KnPkH6d89l4


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.