Analyzing quilt data

[mcgregorian1]

Hi @teixeirak

Sorry, I should've opened this as an issue to begin with.

From your email: "Species should be combined. In this analysis, we don’t care about identity of the species, except as a factor to pair canopy and sub canopy."

Based on this, I get the following result for the current growing season (this is also saved in this repo): Canopy_subcanopy_correlation.pdf

I think this is more what you were thinking of?

[teixeirak]

Yes, this is what I had in mind. Could you please run it for all months?

[teixeirak]

Which analysis time period are you using?

[mcgregorian1]

This analysis is based on 1901-2016 with the CRU climate data

Sure thing. Sorry for the delay, I was trying to get the graphs to be in order.

It is here.

[teixeirak]

Thanks for the plot! I'm not noticing any interesting trends outside the current growing season, but wanted to check.

Regarding the analysis, I think what we want here is a mixed effects model--one for each climate variable--with canopy position (canopy/subcanopy) as a fixed effect, species as a random effect, and month within current growing season (May-June-July-Aug) as a fixed effect. The motivation for combining months is that we're not interested in differences for individual months, but for the whole growing season. (Note: an alternate way to do this analysis would be to aggregate months for the tree ring correlation analysis.)

[mcgregorian1]

Ok, sounds good! I can start looking into this on Monday

[mcgregorian1]

Hi @teixeirak

I've run the mixed effect model by variable and I found the following table. Here, this shows the significance of both the canopy position and the month on the correlation coefficient from @ValentineHerr 's quilts.

• Canopy position is significant except for tmn and tmp.
• Month is insignificant except for pet, pet_sum, and wet.

AIC vs BIC

Some notes about AIC vs BIC. The table above was done with AICc, a variation of AIC. There's a lot of back and forth about which is better to use to determine the best model (AIC or BIC). The best indication I could find was from stack_overflow and PennState. As I understand it:

• AIC: a lower AIC means a model is considered to be closer to the truth
• BIC: a lower BIC means that a model is considered to be more likely to be the true model

In other words, AIC is better for showing what's true for current conditions based on the data as we have it, and BIC is better for showing predictions in the future.

Based on this, AIC tells us that the full model as you described it and whose results are seen in the above table is the majority best model for each variable.

For predicting the future, BIC says that using only the fixed effect for position is majority best.

Alternative

Depending on if we want to use AIC or BIC in analysis, we run a different model for each variable based on what these charts show. I'm not sure, though, if that's a sound practice.

What do you think?

[teixeirak]

Note: if we end up including this in the paper, we should re-run with response ("slope")--as opposed to correlation--values. This relates to this issue.

[teixeirak]

We've said that we probably don't want to include this analysis in the paper, so closing this issue.