ovaskain
commented
3 years ago The values in the association matrix are not (and cannot be) independent of each other, as the matrix needs to be positive definite. This reflects (partially) biology as well, not just model structure: if A and B have high positive association, and B and C as well, then necessarily A and C must be positively associated. On top of this, the latent variable structure "looks for" correlated associations. If you have 100 species, you have ca. 5000 associations. If you wish to get very accurate estimates of each of those, you should have a huge amount of data. The latent variable structure tries to get you those associations as accurately as possible with limited amount of data. Can you trust that the associations are real rather than noise or something forced by the statistical model? Test this e.g. by conditional cross-validation, see e.g. the book for more details.
O2
On 11.12.2020 15.40, Ryan Burner wrote:
Hi HMSC team,
I had a question about the patterns I typically observe in species association scores. In the plots below, I’ve ordered species by their occupancy and plotted association score with >95% support. (More common species have more associations, which is what I typically see.)
One observation I commonly make, which is the focus of my question, is that most species that have any highly supported associations have associations with [almost] all other species that have any associations at all. This essentially means that association scores between a given species and all other species (i.e. the scores in a single row, or single column, in the association matrix) do not appear to be independent of each other. This is most apparent in the second plot below, but is also somewhat true in the first.
This doesn’t seem likely to reflect a biological reality, but seems more likely to be an artefact of something about the model structure. Maybe the latent variables that are used to estimate each species’ associations are few enough that the estimates are in this case gross oversimplifications that result in these simple patterns?
My main question then is whether I should 'believe' these association scores, in which case it could be meaningful for me to e.g. try to test whether differences in trait values between the two members of given species pair is related to their probability of having an association? Or is it (as it appears to me) clear from the plots below that these particular models at least are resulting in association score value estimates that I shouldn't trust?
Thanks!!
Ryan
cor1 https://user-images.githubusercontent.com/36172156/101908695-c2061100-3bbc-11eb-8c24-220257a54a20.jpg cor2 https://user-images.githubusercontent.com/36172156/101908706-c5010180-3bbc-11eb-98df-4ef9e7c872a9.jpg
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/hmsc-r/HMSC/issues/74, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEIYMZWRYZNJ6UTONZEVL6LSUIOLLANCNFSM4UWUWR3A.
rburner
Hi HMSC team,
I had a question about the patterns I typically observe in species association scores. In the plots below, I’ve ordered species by their occupancy and plotted association score with >95% support. (More common species have more associations, which is what I typically see.)
One observation I commonly make, which is the focus of my question, is that most species that have any highly supported associations have associations with [almost] all other species that have any associations at all. This essentially means that association scores between a given species and all other species (i.e. the scores in a single row, or single column, in the association matrix) do not appear to be independent of each other. This is most apparent in the second plot below, but is also somewhat true in the first.
This doesn’t seem likely to reflect a biological reality, but seems more likely to be an artefact of something about the model structure. Maybe the latent variables that are used to estimate each species’ associations are few enough that the estimates are in this case gross oversimplifications that result in these simple patterns?
My main question then is whether I should 'believe' these association scores, in which case it could be meaningful for me to e.g. try to test whether differences in trait values between the two members of given species pair is related to their probability of having an association? Or is it (as it appears to me) clear from the plots below that these particular models at least are resulting in association score value estimates that I shouldn't trust?
Thanks!!
Ryan