Thank you so much for this. Extremely grateful for your clear explanations, pertinent choice of content, provision of the TH-cam recording and patience answering so many questions.
Thank you, Gavin, for making available your excellent teaching material. I have been following your content on both gams and mutivariate methods and I find it extremely helpful for my doctoral thesis.
Although late I have a question about restricted permutations. Say I got 4 habitats sampled seasonally. I guess I would call habitats=blocks. However, I cant figure out what I would do with seasons. I read an example below with pigs which seemed kinda similiar and made sense but could you confirm my train of thoughts?
If you are testing the effect of a covariate within habitats, then yes, you want to block on habitat. As for season, it depends on whether you are testing effect of a covariate within or across seasons, and whether season is the covariate you are wanting to test. You could include season in your model (if you want to estimate variation due to seasons. You could use Condition(season) if you want to exclude the seasonal effects (partial them out), which would be if season was a nuisance variable. In the pattern case you probably want to use season as the plot strata too as variation between seasons shouldn't be included in the permutation test (`plot = Plot(strata = season)`). With these constrained permutation tests, you don't have one set-up for the whole model or even a single model; you might need to fit several model with different permutation designs to answer the questions you have. But you'll need to explain more about your set-up/experiment/observations/hypotheses for me to be able to say more.
@@ftboth I actually didn't want to waste your time but perhaps this might also help someone else. So this is the setup: I sampled for habitat types seasonally (3 replicates each). I wanted to examine whether the community varies among habitat and among season within habitat. Thank you for your response!
No worries @tasvak3452, your description is pretty clear and my answering your question will likely help others too. To assess if the community varies by habitat you will need to permute the habitats. As we never permute blocks, you will need to set habitat as the plot-level strata. You will also need to set the type of within-plot permutation to be "none". You will also need a model with habitat as a covariate, so something like ord1
@@ftboth Ive tried the method you proposed and I get this error. "Error in check(sn, control = control, quietly = quietly) : Permutation 'type' is "none" for both 'plots' & 'within'. Nothing to permute." This is what I used habitatperm
Thanks so much for this video and others! I've been seeking out all information for writing how() arguments for my experiment's design and I'm still coming up a bit short. We have 20 pigs total, 10 pigs on a control diet and the other 10 on an experimental diet. Each pig is sampled 3 times (baseline, day 7, and day 15). So for permuting, I am wanting to say that it is how(within=Within(type="series", constant = TRUE), plots = Plots(strata = Pig, type="free"). Then testing with adonis2(BrayMatrix ~ Diet*Time+Point) with by = "margin". I am a bit confused about the type arguments in the within and plot functions. Any advice would be greatly appreciated!
The thing with permutation tests is that you need to meet the assumption of exchangability of samples under the null hypothesis. In your case you have dependence within individual pigs and the time effect. You modelled the time effect (from the looks of your model - how you actually include Time is a missing detail [an ordered factor would seem appropriate]) so that really all you need to deal with is the dependence among observations within pigs. So, from the information you provide I would just do `how(blocks = Pig)` as we don't want to shuffle the data at the Pig (plot) level and we don't want to restrict permutations within Pigs as we modelled the Time effect. There may still be some unmodelled temporal dependence as the Time effect is only at the population level. As such you might need to include something for Time. If you want that, you can't use `constant = TRUE` as there are only 3 cyclic permutations of the time points 0, 7, 15. With `constant` your minimum p value is >1/3. So you would need to relax that restriction and just use `how(blocks = Pig, within = Within(type = "series"))` and make sure the observations within the levels of Pig are arranged in Time order.
@@ftboth Hello! Do you cover ANOSIM in this webinar? I'm not sure it's worth watching the entire 3+ hours yet (although I hope to eventually since I find this very interesting) if it's not covered since I need to do ANOSIM sooner, if possible. Thank you so much!
@@ftboth thank you so much for the reply! But I thought ANOSIM and Permanovas were different (one looking at similarities and the latter looking at differences). And ehen I ran both on my data, I definitely ended up with different results. Anyway...I've been struggling with my PhD data for forever (my advisor isn't very adept with stats) and any advice would be greatly appreciated! Eg, is there a reason you might not recommend ANOSIM?
Sir, thank you for such information-rich sessions. I am working on data from plant (92 species) x soil chemistry(16 parameters). The p-value is significant, and the r2 is low at 0.15. How can I explain these results? Can you please suggest any references?
With so many response variables (species) you are unlikely to explain a high percentage of the variance in the data, so small variance explained is the typical situation. But the permutation test is telling you that even explaining the small amount that you do is unlikely if there was no relationship between the species and your environmentAl variables. You have to recalibrate your brain a bit with multivariate data and these measures of variance explained and what is large or small. I would look to see if there is structure in the unexplained variance of your data; if there is (I.e. large eigenvalues in the unconstrained axes) then that suggests while the things you measured explain more variance than you’d expect if they had no effect, you are missing some/a lot of what structures your plant community.
Thank you so much for this. Extremely grateful for your clear explanations, pertinent choice of content, provision of the TH-cam recording and patience answering so many questions.
Thank you very much for this video, it's really interesting. Can I have your email?
You can find my email pretty easily online
Thank you, Gavin, for making available your excellent teaching material. I have been following your content on both gams and mutivariate methods and I find it extremely helpful for my doctoral thesis.
Great to hear!
Although late I have a question about restricted permutations. Say I got 4 habitats sampled seasonally. I guess I would call habitats=blocks. However, I cant figure out what I would do with seasons. I read an example below with pigs which seemed kinda similiar and made sense but could you confirm my train of thoughts?
If you are testing the effect of a covariate within habitats, then yes, you want to block on habitat. As for season, it depends on whether you are testing effect of a covariate within or across seasons, and whether season is the covariate you are wanting to test. You could include season in your model (if you want to estimate variation due to seasons. You could use Condition(season) if you want to exclude the seasonal effects (partial them out), which would be if season was a nuisance variable. In the pattern case you probably want to use season as the plot strata too as variation between seasons shouldn't be included in the permutation test (`plot = Plot(strata = season)`). With these constrained permutation tests, you don't have one set-up for the whole model or even a single model; you might need to fit several model with different permutation designs to answer the questions you have. But you'll need to explain more about your set-up/experiment/observations/hypotheses for me to be able to say more.
@@ftboth I actually didn't want to waste your time but perhaps this might also help someone else. So this is the setup: I sampled for habitat types seasonally (3 replicates each). I wanted to examine whether the community varies among habitat and among season within habitat. Thank you for your response!
No worries @tasvak3452, your description is pretty clear and my answering your question will likely help others too.
To assess if the community varies by habitat you will need to permute the habitats. As we never permute blocks, you will need to set habitat as the plot-level strata. You will also need to set the type of within-plot permutation to be "none". You will also need a model with habitat as a covariate, so something like
ord1
@@ftboth amazing explanation!! I had it all wrong in my head. Its pretty clear now. Thank you and keep it up!!! Greetings from Greece
@@ftboth Ive tried the method you proposed and I get this error. "Error in check(sn, control = control, quietly = quietly) :
Permutation 'type' is "none" for both 'plots' & 'within'.
Nothing to permute."
This is what I used
habitatperm
Thanks, Gavin, for making all this material available. Great stuff.
Thank you very much for the valuable big amount of informations.
Best wishes.
I think a webinar dedicated to plotting would be useful.
Thank god I found your tutorials.....literally saving me right now.
Thanks so much for this video and others! I've been seeking out all information for writing how() arguments for my experiment's design and I'm still coming up a bit short. We have 20 pigs total, 10 pigs on a control diet and the other 10 on an experimental diet. Each pig is sampled 3 times (baseline, day 7, and day 15). So for permuting, I am wanting to say that it is how(within=Within(type="series", constant = TRUE), plots = Plots(strata = Pig, type="free"). Then testing with adonis2(BrayMatrix ~ Diet*Time+Point) with by = "margin".
I am a bit confused about the type arguments in the within and plot functions. Any advice would be greatly appreciated!
The thing with permutation tests is that you need to meet the assumption of exchangability of samples under the null hypothesis. In your case you have dependence within individual pigs and the time effect. You modelled the time effect (from the looks of your model - how you actually include Time is a missing detail [an ordered factor would seem appropriate]) so that really all you need to deal with is the dependence among observations within pigs. So, from the information you provide I would just do `how(blocks = Pig)` as we don't want to shuffle the data at the Pig (plot) level and we don't want to restrict permutations within Pigs as we modelled the Time effect.
There may still be some unmodelled temporal dependence as the Time effect is only at the population level. As such you might need to include something for Time. If you want that, you can't use `constant = TRUE` as there are only 3 cyclic permutations of the time points 0, 7, 15. With `constant` your minimum p value is >1/3. So you would need to relax that restriction and just use `how(blocks = Pig, within = Within(type = "series"))` and make sure the observations within the levels of Pig are arranged in Time order.
Thank you for the video. Please can the PERMANOVA involve interactions between categorical and continuous variables?
Yes, it can
Thanks for the effort. Very informative
@@ftboth Hello! Do you cover ANOSIM in this webinar? I'm not sure it's worth watching the entire 3+ hours yet (although I hope to eventually since I find this very interesting) if it's not covered since I need to do ANOSIM sooner, if possible. Thank you so much!
@@ftboth thank you so much for the reply! But I thought ANOSIM and Permanovas were different (one looking at similarities and the latter looking at differences). And ehen I ran both on my data, I definitely ended up with different results. Anyway...I've been struggling with my PhD data for forever (my advisor isn't very adept with stats) and any advice would be greatly appreciated! Eg, is there a reason you might not recommend ANOSIM?
Sir, thank you for such information-rich sessions. I am working on data from plant (92 species) x soil chemistry(16 parameters). The p-value is significant, and the r2 is low at 0.15. How can I explain these results? Can you please suggest any references?
With so many response variables (species) you are unlikely to explain a high percentage of the variance in the data, so small variance explained is the typical situation. But the permutation test is telling you that even explaining the small amount that you do is unlikely if there was no relationship between the species and your environmentAl variables. You have to recalibrate your brain a bit with multivariate data and these measures of variance explained and what is large or small.
I would look to see if there is structure in the unexplained variance of your data; if there is (I.e. large eigenvalues in the unconstrained axes) then that suggests while the things you measured explain more variance than you’d expect if they had no effect, you are missing some/a lot of what structures your plant community.
Thanks!
Brilliant work. Thanks for making the lecture available for everyone!
It's my pleasure
Hi Gavin, please can you confirm how to retrieve the pairwise tests between groups for PERMANOVA? Thank you
Perhaps the pairwiseAdonis package by Pedro Martinez Arbizu would suit your needs : github.com/pmartinezarbizu/pairwiseAdonis
The two sessions were so incredibly helpful thank you!
Glad it was helpful!
Thank you!!
You're welcome!