test canopy position as continuous variable

[teixeirak]

@mcgregorian1,

One other modification: let's represent canopy position as a continuous variable. This will take advantage of the fact that we classified canopy position into 4 categories, not just canopy/subcanopy (categories created for dividing into chronologies, which we're no longer doing; issue #4).

[mcgregorian1]

Hi @teixeirak

I can definitely put in the other canopy positions but I'm not sure I can categorize them as continuous because I only have 4 categories that are character strings. Year, for example, I can make either discrete or continuous based on whether I classify it as character or numeric, but I don't think I can do that with canopy position.

[teixeirak]

You could number 1-4. You could also try crown illumination (1-5).

[mcgregorian1]

I've put in all canopy positions and there is a very large change.

In short, putting in position_all drastically improves the model.

• I updated the outline showing for our base models what happens when we put position_all in.
• I've also updated the results table. Now we're seeing negative coefficient across the drought years.
• I also ran the full model like we did yesterday. I'll show you the AIC table here, but for me the most important aspect is that the null model (the best model without position_all) has a dAIC of 75. The R2 is 0.16.

The coefficients of the best model here are all positive for canopy position, with the strength of each saying to me that co-dominant actually are most resistant to drought.

• The one table I haven't updated yet is full_models_combined_years.csv. I can do that after fixing the dendrobands.

[mcgregorian1]

You could number 1-4. You could also try crown illumination (1-5).

I can do that. Let me know what you think based on the results I've found.

[teixeirak]

Nice! Either way works; you can stick with this. It's probably more correct to treat as a categorical variable. It's reassuring that they come out in the expected order.

[teixeirak]

Let's definitely stick with the 4 categories of canopy position (unless crown illumination index comes out better). The canopy/subcanopy categorization is obviously not the best we can do.

[teixeirak]

Oh wait, I just noticed that co-dominant trees have twice the resistance of dominant trees. So we may be seeing two mechanisms at play-- dominant trees are stressed by exposure, intermediate and suppressed by competition.

[teixeirak]

We therefore shouldn't treat this as a continuous variable, and same goes for crown illumination index.

[mcgregorian1]

Adding in crown illumination index as a discrete variable seems to have an interaction with crown position. That is, illum is in the best model, but immediately not in the second best model (also notice how the top three models are all within AIC values of 0.1; in other words, I believe they're roughly equal). If you look at dAIC of the first model that doesn't include either position_all or illum, you see a 75 similar to what we see with just position_all.

Interestingly, the coefficients change. Dominant is now giving negative correlation. Technically illum3 has the strongest correlation out of the illums but it's pretty slight.

[teixeirak]

I agree that they're roughly equivalent. We'd only want one or the other-- whichever is best. It looks like position performs slightly better, so let's stick with that.

[mcgregorian1]

Ok that's fair. I agree

[mcgregorian1]

@teixeirak The full models above (with just position_all, no illum) show a 0.0106 dAIC between the first model (everything but elev) and the second (everything). Because those are so close, does it make sense to keep elev in there when making my table over the different hypotheses? I ask particularly from your comment in #22

[teixeirak]

I think we're done with this issue.