Closed mcgregorian1 closed 5 years ago
Looks good overall. I do wonder though-- Should we have the null model for height*elev be one without elev at all? This would tell us how much elev contributes alone and in interaction with height, which makes the term as a whole more significant. For now, let's run it against both null models, and we can then pick which to present in the table.
Good point. I added a different model run (combining 1.2c2 and 1.3b1 where I used the same model with the interaction, but took out elev). The new table is here(for the record I didn't assign NAs yet so we can see the absolute dAIC).
My impressions are it seems height still dominates, both from when it's taken out and how it's the biggest contributor to variance across most models. Elev apparently contributes to variance a degree more when you're looking at an interaction of elev and height compared to additive.
Just so I understand this. In this table, we are saying that...
1.1- We predict that resistance decreases with height. This should come out true when we account for the interaction term (as calculated this AM). Correct? (You might run a model with just height, no interaction.)
1.2b- Correct. So much for the canopy exposure hypothesis... I suppose the more crowded/ suppressed trees are suffering more (after one accounts for height).
1.2c- Correct. Again, so much for the canopy exposure hypothesis... Rather, higher elevation trees have higher resistance (less sensitive).
1.2c2 & 1.3b1: This one is more complex. 1.2c2 and 1.3b1 (sorry... horrible numbering system is my fault!) predicts directions in opposite interactions.
1.2c2 predicts that tall height + high elevation (i.e., exposure) are both problems (--> lower resistance; - coefficients), and the effect is compounded (- coefficient; original predicted coefficient was wrong). 1.2c2 is rejected.
1.3b1 predicts that higher elevation (i.e., vertical distance to water) leads to lower resistance (1.2c, - corr of elevation, rejected), and this effect is especially strong in small trees (+ coefficient). There is a + coefficient, so elevation increases resistance, particularly in big trees. This still means that small trees do less well at higher elevations.
1.2c21.3b1: we predict that when heightelev, elev is positively correlated with resistance values. This is true. - incorrect: there are two coefficients here, predictions are as above. This test simply tells us the overall significance of elev
2.1- Correct.
2.2- Correct.
There's potential for some confounding effects (as we see in Cocoli!). However, looking at this file, it looks like the coefficients all go in the same direction regardless of whether we use the full model or just single variables.
For 1.1, we need a model with no interaction in order to get the dAIC for height alone.
For 1.1, we need a model with no interaction in order to get the dAIC for height alone.
Isn't that already done in this file?
I mean using the full model, but putting height + elev, not height * elev.
Gotcha. Not including position_all, height gives a dAIC (compared to no height) of 54.81.
I'm going to redo the table of full models using position_all
@teixeirak does it matter that in taking out illum, we'd be having the coefficient of dominant be positive again?
@teixeirak does it matter that in taking out illum, we'd be having the coefficient of dominant be positive again?
No, we're concerned about relative differences, not coefficient values (absolute coefficient values are pretty meaningless outside the context of the full model).
I think we're done with this issue (can reopen if not).
Hi @teixeirak
I've recreated the table as I think you were indicating, where the model variables has all the variables, and the null model is the one lacking. The table can be found here. To make it more clear, I've included a column saying which variable the coefficient is representing (aka, the variable not included in the null model).