[R-sig-eco] Adonis output with or without strata
Gaëlle Quere
g@e||equere @end|ng |rom hotm@||@com
Wed Apr 15 09:06:08 CEST 2020
Dear list members,
Thanks Jakob Russel for your last email. This discussion is really useful to me in conducting the right analyses on my data. I�m hoping you won�t mind pursuing it.
I ran adonis again:
ad4 <- adonis(dist ~ Island + Dis + Island:Dis, data = meta)
Df SumsOfSqs MeanSqs F.Model R2 Pr(>F)
Island 6 4.6265 0.77108 2.0056 0.21504 0.001 ***
Dis_status 1 0.7927 0.79266 2.0618 0.03684 0.001 ***
Island:Dis_status 6 3.0240 0.50400 1.3109 0.14055 0.001 ***
Residuals 34 13.0716 0.38446 0.60756
Total 47 21.5148 1.00000
---
The interaction factor is significant and now, to be honest, I am a bit confused on how to interpret it. I don't have the hypothesis that the difference between Healthy and Diseased is different on the 7 islands but I would be interested in testing whether the 7 islands are more/less similar in the Healthy versus the Diseased.
The ordination plot clearly separates the healthy and diseased groups but I can't find a way to show it statistically. Would running pairwise comparisons following adonis do it? I tried, and the combinations are numerous and it's hard to see a clear pattern.
To sum up, if I run:
ad <- adonis(dist ~ Island + Dis, data = meta)
Df SumsOfSqs MeanSqs F.Model R2 Pr(>F)
Island 6 4.6265 0.77108 1.9163 0.21504 0.001 ***
Dis. 1 0.7927 0.79266 1.9699 0.03684 0.001 ***
Residuals 40 16.0956 0.40239 0.74812
Total 47 21.5148 1.00000
Am I allowed to say that Healthy and Dis are significantly different regardless of the island?
How do I test whether the 7 islands are more/less similar in the Healthy versus the Diseased?
Thank you all,
Gaelle
De : Jakob Russel <russel2620 using gmail.com>
Envoy� : vendredi 10 avril 2020 14:16
� : Ga�lle Quere <gaellequere using hotmail.com>
Objet : Re: Adonis output with or without strata
Hi Ga�lle,
It looks like you have 7 islands, but in your adonis output it says that island only has 1 degree of freedom, this is expected to be 7-1 = 6. Are you sure that the Island factor is coded as a factor and not a numeric?
Your model tests the overall effect of island and disease. If your Island factor is significant you can say that your islands have different microbiomes. If your Disease factor is significant you can say that Healthy and Diseased have different microbiomes, independent of islands. From your plot it is expected that both are significant. If you include the interaction between Island and Disease you test the hypothesis that the difference between Healthy and Diseased is different on the 7 islands, or conversely that the 7 islands are more/less similar in the Healthy versus the Diseased.
If you want to know which has the strongest effect, you can look at the R2, which is the effect size. You can also calculate an unbiased effect size with one of my R-packages (https://github.com/Russel88/MicEco#adonis_omegasq).
De : Ga�lle Quere <gaellequere using hotmail.com>
Envoy� : vendredi 10 avril 2020 13:38
� : r-sig-ecology using r-project.org <r-sig-ecology using r-project.org>; Jakob Russel <russel2620 using gmail.com>
Objet : RE: Adonis output with or without strata
Hi Jakob,
Thank you for your time and useful answers. I had tried to include Island as a fixed factor earlier and had a significant result for both factors. I got confused on how to interpret this output.
I was interested in the difference between Healthy and Diseased but Island had an effect too so I thought: should I test the interaction factor then? Also, what does it tell me about which of Island or Disease has the most influence etc. this sort of questions. Since we are here and if you have time, maybe we could go through it.
I ran this model:
adonis <- adonis(dist ~ Island + Dis, data = meta)
and obtained this output:
Df SumsOfSqs MeanSqs F.Model R2 Pr(>F)
Island 1 0.9302 0.93024 2.6688 0.18269 0.001 ***
Dis 1 0.6759 0.67592 1.9392 0.13275 0.002 **
Residuals 10 3.4856 0.34856 0.68456
Total 12 5.0918 1.00000
Q1: Does the model test the effect of Island within the Healthy group and the Diseased group separately or does it test an Island effect on the whole data set?
Q2: If I visualize my data using an NMDS, there is a clear separation of the Healthy and Diseased groups (see pdf enclosed) meaning that healthy samples are �closer� to each other than diseased samples from the same island? but now the adonis seems to say that Island has a strong effect too. How do I distinguish between the two factors, how do I know which factor has the �strongest effect� on the microbiomes?
Cheers,
Ga�lle
________________________________
De : Jakob Russel <russel2620 using gmail.com>
Envoy� : vendredi 10 avril 2020 10:04
� : Ga�lle Quere <gaellequere using hotmail.com>
Objet : Re: Adonis output with or without strata
Dear Ga�lle,
I have just received your email through r-sig-ecology, so it seems to have passed through.
Q1:
Yes, your model tests whether the microbiomes of Healthy and Diseased algae are different, regardless of island. (Under the assumption that the Bray-Curtis dissimilarity captures this difference in microbiome composition)
Q2:
If you are not interested in the islands, I would say your model is correct. You could try to include islands as a fixed effect to see the effect size, but if you have a balanced experimental design it will not affect the Dis factor. If your design is unbalanced the order matters and you should add the island as the first fixed effect adonis <- adonis(dist ~ Island + Dis, data = meta).
Regards,
Jakob
Den fre. 10. apr. 2020 kl. 11.28 skrev <gaellequere using hotmail.com<mailto:gaellequere using hotmail.com>>:
Dear Jakob,
Sorry for sending a private email. I tried to reply directly on the forum but it keeps getting rejected and I don't understand why.
Here was my answer:
Dear Jakob,
Thank you for your prompt reply. May I take the opportunity to ask two related additional questions?
As mentioned in my first post, my research question is to compare microbiomes depending on health status of the host algae (Healthy vs Diseased). Healthy and diseased samples were collected in different islands.
So far, this is the model I have been using to answer my research question:
Variable: Bray-Curtis dissimilarity matrix (dist)
Fixed factor: Disease (Dis)
Random factor: Island
adonis<- adonis(dist ~ Dis, strata = Island, data = meta)
Q1: If I understood correctly, this model compares Healthy and Diseased samples within each island, is that correct? The p-value being significant, does it mean that Healthy and Diseased samples are different within each island or can I make a general statement like: overall Healthy and Diseased microbiomes are different regardless of the island. In other words, microbiomes associated with diseased samples share significant similarities among island and are distinct from healthy samples?
Q2: Is my model correct or should I consider Island as a fixed factor and consider potential interactions as well in order to test the effect of both Island and Dis on the microbiomes?
Thanks once more,
Ga�lle
<quote author='Jakob Russel'>
Dear Ga�lle,
The p-values in PERMANOVA are generated by permutations. These permutations
are by default free, but they can be constrained by the strata factor.
The strata argument only affects the permutations, and hence will only
affect the p-value.
The Sum of Squares (and R2) are only based on the fixed effects.
Regards,
Jakob
Den fre. 10. apr. 2020 kl. 09.42 skrev Ga�lle Quere <GaelleQUERE using hotmail.com<mailto:GaelleQUERE using hotmail.com>
>:
> Dear list members,
>
> I am having trouble understanding the output of an adonis model I am
> running first with then without the argument �strata� in the formula. I
> have seen this question asked before but no answer were provided.
>
> I would like to compare microbial communities associated with healthy and
> diseased reef algae. Healthy and diseased algae have been collected in
> different islands. I am more interested in the difference between the
> Healthy and Diseased groups, I am not really interested in comparing the
> islands but still would like to take the islands into account. From what I
> read it looks like a way to do it is to include Island as a random factor
> in my model.
>
> Variable: Bray-Curtis dissimilarity matrix (dist)
>
> Fixed factor: Disease (Dis)
>
> Random factor: Island
>
> adonis1<- adonis(dist ~ Dis, strata = Island, data = meta)
>
> adonis2<- adonis(dist ~ Dis, data = meta)
>
>
>
> However, I get the exact same output for both models:
>
> Df SumsOfSqs MeanSqs F.Model R2 Pr(>F)
>
> Dis_status 1 0.8934 0.89345 1.9845 0.0397 0.001 ***
> Residuals 48 21.6103 0.45021 0.9603
> Total 49 22.5038 1.0000
> ---
>
>
> I tried with a subset and I had the same output besides a different
> p-value.
>
>
>
> Could someone please explain to me why the outputs are identical?
>
>
>
> I am sorry if that sounds trivial but but I don�t really understand why
> R2, SumsOfSqs and MeanSqs are exactly identical with or without the strata
> argument. I�d like to understand better what�s behind it.
>
>
> Thanks a lot,
>
> Gaelle
>
>
>
>
>
>
>
>
> [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-ecology mailing list
> R-sig-ecology using r-project.org<mailto:R-sig-ecology using r-project.org>
> https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
>
[[alternative HTML version deleted]]
_______________________________________________
R-sig-ecology mailing list
R-sig-ecology using r-project.org<mailto:R-sig-ecology using r-project.org>
https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
</quote>
Quoted from:
http://r-sig-ecology.471788.n2.nabble.com/Adonis-output-with-or-without-strata-tp7580521p7580522.html
_____________________________________
Sent from http://r-sig-ecology.471788.n2.nabble.com
[[alternative HTML version deleted]]
More information about the R-sig-ecology
mailing list