Inclusion of top variables

[mcgregorian1]

This is related to #95 - we decided that we would drop RP from the top models list as a covariate because its coefficient was not consistent across years.

That's fine for resistance. I'm testing recovery now (haven't gotten to resilience yet), and I'm getting the same top variables as candidate traits (RP, TLP, PLA). However, now RP is consistent across years but neither TLP nor PLA are. Based on our rationale for resistance, does this mean for recovery I should only test between RP and the other covars (height and TWI)?

I'm assuming resilience will be different as well.

[teixeirak]

I suppose its best to go with them. I don't love the xylem porosity distinction because we only have 2 diffuse porous species (+1 semi-ring), but we should stick with the "rules" and go with what we get. That probably makes sense; I wouldn't hypothesize that TLP and PLA would be important, but I think ring porous species are generally thought to have higher resilience. I'm not sure though-- will have to find some lit on that.

[teixeirak]

I see that ring porous actually have lower recovery, and pretty strongly... interesting. I'll try to dig up some literature on this. What I think I remember reading several year ago is that they're better at recovering xylem function after cavitation, which isn't the same as growth.

[mcgregorian1]

Oops please ignore all these commits referencing this. I meant for these to be in #131

[teixeirak]

@mcgregorian1, a few points here:

1- I'm confused... tested_traits_best_coeff_resilience_CPout.csv and tested_traits_best_coeff_recovery_CPout.csv both show ring-porous species with consistently lower Rs or Rc, but the current plot shows mixed responses. What's going on?

But wait... never mind. I think we want to drop the ring/ diffuse distinction. Fagus and L. tulipifera are on opposite ends of the spectrum in terms of Rs:

With just 2 diffuse porous species (+JUNI), xylem porosity just isn't a good variable.

[teixeirak]

2- Because Rt, Rc, and Rs are all related, I think we want to test the same set of candidate variables across all variables. (I'll come back to this.)

[mcgregorian1]

@mcgregorian1, a few points here:

1- I'm confused... tested_traits_best_coeff_resilience_CPout.csv and tested_traits_best_coeff_recovery_CPout.csv both show ring-porous species with consistently lower Rs or Rc, but the current plot shows mixed responses. What's going on?

I believe this may be due to how I took out height and TWI from these models (see my response from last night for #131)

[teixeirak]

@mcgregorian1, a few points here: 1- I'm confused... tested_traits_best_coeff_resilience_CPout.csv and tested_traits_best_coeff_recovery_CPout.csv both show ring-porous species with consistently lower Rs or Rc, but the current plot shows mixed responses. What's going on?

I believe this may be due to how I took out height and TWI from these models (see my response from last night for #131)

Be sure to see my edit to the comment above.

[mcgregorian1]

I believe this may be due to how I took out height and TWI from these models (see my response from last night for #131)

Be sure to see my edit to the comment above.

Ah yes this is why you were wanting to not include RP from the get-go when we were doing just Rt.

Sounds good then. You mentioned we want the same candidate variables across metrics, so I'll hold off on further thoughts (ie taking out RP for Rc would leave only one variable in the model - I assume what you're thinking will alleviate this).

[teixeirak]

2- Because Rt, Rc, and Rs are all related, I think we want to test the same set of candidate variables across all variables. (I'll come back to this.)

Back to this...

• [ ] I think we want xylem porosity removed altogether (from this figure)
• [ ] We should remove the H x TWI term, at least from the set of models plotted here. That is, if the best model includes Hx TWI, we should plot the second-best.
• [ ] all terms in the top model should be plotted, regardless of whether they are significant (i.e., included in all top models). We could make the lines dashed for non-significant terms.

[teixeirak]

Fortunately, it looks like we get the same set of candidate traits for Rt, Rs, and Rc. That is, all 3 give us TLP and PLA (once RP is removed).

[mcgregorian1]

To make sure I understand, my workflow will be the following for Rt, Rc, and Rs:

• look at the best models for each drought year
• select the top model from each year that does not include height*TWI
  • does this mean only take out the interaction? So I can still choose a model that has height or TWI alone, or together but with no interaction?
  • one possible outcome is we have no top models that fit this category under dAICc<1. However, I'm saying this without looking at the files so maybe we're fine. I'm going to have rerun top_models to not include RP at all, so the overall results will be slightly different.
• each Rt/Rc/Rs will have TLP and PLA, and then it will vary whether or not they have height and/or TWI.

[teixeirak]

select the top model from each year that does not include height*TWI

* does this mean only take out the interaction? So I can still choose a model that has height or TWI alone, or together but with no interaction?

* one possible outcome is we have no top models that fit this category under dAICc<1. However, I'm saying this without looking at the files so maybe we're fine. I'm going to have rerun `top_models` to not include RP at all, so the overall results will be slightly different.

This is the one question I'm still thinking about. Do we want to completely take this out for Rs and Rc, or just not plot it. For instance, the top 2 models for Rs have the interactive term. I'd initially thought of removing it for ease of interpretation, but now I'm leaning towards keeping it. There's no real reason to include for Rt but not the other metrics. For Rs, the term is negative, combined with a positive response to TWI, which means that resilience is higher in wetter locations, particularly for smaller trees.

[teixeirak]

So... let's include height x TWI as candidate variables, but skip plotting them.

[teixeirak]

So, the process will be:

1. For all 3 metrics, candidate variables will include ln[H], ln[TWI], ln[H]xln[TWI], TLP, PLA.

2. Plot the best models for each drought year UNLESS it includes ln[H]xln[TWI]. If it includes ln[H]xln[TWI], go down to the best model that doesn't include this term. ln[H]xln[TWI] is never significant, so there is never a case where no top model lacks the term.

3. Plot significant terms (those in all top models) with solid lines, non-sig terms with dashed lines. Note that models with ln[H]xln[TWI] will always include an ln[H] term, but its coefficient may be opposite. This does not mean that the effect of ln[H] is inconsistent, as you need to account for the interaction in order to understand its effect within the range of the data.

[mcgregorian1]

So, the process will be:

1. For all 3 metrics, candidate variables will include ln[H], ln[TWI], ln[H]xln[TWI], TLP, PLA.
2. Plot the best models for each drought year UNLESS it includes ln[H]xln[TWI]. If it includes ln[H]xln[TWI], go down to the best model that doesn't include this term. ln[H]xln[TWI] is never significant, so there is never a case where no top model lacks the term.
3. Plot significant terms (those in all top models) with solid lines, non-sig terms with dashed lines. Note that models with ln[H]xln[TWI] will always include an ln[H] term, but its coefficient may be opposite. This does not mean that the effect of ln[H] is inconsistent, as you need to account for the interaction in order to understand its effect within the range of the data.

Based on # 2, I shouldn't have any models with the height*TWI interaction anyways, so # 3 is not referring to that specific term, right? Or am I missing something? As I see it, you mention # 3 to make sure I know that the top models can still have height as insignificant?

[teixeirak]

Based on # 2, I shouldn't have any models with the height*TWI interaction anyways, so # 3 is not referring to that specific term, right? Or am I missing something? As I see it, you mention # 3 to make sure I know that the top models can still have height as insignificant?

That comment was just to clarify any potential confusion regarding the fact that the interaction term can change the sign of the ln[H] term, making it look like it's direction is not consistent across models (and therefore wouldn't be significant). For example, the top all-droughts model for Rc includes the interaction term, and there the effect of ln[H] is positive. However, in the top model without the interaction (which is the one you'll plot), the term is negative.

Here, ln[H] would be counted as significant because its in all top models, and we don't worry about the fact that the sign changes.

[mcgregorian1]

Ahh ok gotcha. Thanks for clarifying! I should hopefully be able to get to some of this tonight

[mcgregorian1]

1. For all 3 metrics, candidate variables will include ln[H], ln[TWI], ln[H]xln[TWI], TLP, PLA.

To clarify here, I'm going to re-run all models without RP and only with these candidate traits. Then I'll redo the associated figures as we discussed.

[teixeirak]

Right.

[mcgregorian1]

After running through the code and doing these updates I can confirm that no variables are in all top models (thus I did not change the line types). Here's what the figure looks like - note how Rc has no 1966 or 1999 line - this is because the top model for both of those contained no variables (only the random term).

Now I more clearly see what you were talking about with height.

[teixeirak]

Great!

After running through the code and doing these updates I can confirm that no variables are in all top models (thus I did not change the line types).

Regarding dashed lines, I meant in all top models for the metric and year(s) in question. I know that there are some (e.g., offhand, ln[H] for resistance for all-droughts model).

[mcgregorian1]

Ah...so you mean for this plot it would only be the black line for height and the green line for TLP should be dashed?

[teixeirak]

It will be different for each panel (i.e., each term in each model). It should be solid if the term is included in all models within dAICc of the top model, dashed if not.

[teixeirak]

For the all-years combined models, lines should match table 1– significant terms (solid lines) as “yes”, non-sig (dashed lines) as “(yes)”. This is at least true of Rt. The results for Rs and Rc are preliminary based on the results in the repo this AM.

[teixeirak]

@mcgregorian1 , a few minor things here to finalize the figure:

• [x] regarding the dashed lines, please disregard what I said above. What we'd want is a comparison of the dAICc relative to a model without the term in question, but that may be complex to code. You can just leave them all solid.
• [x] Please reorder as Rt, Rc, Rs (and let's use that as the standard order throughout)
• [x] fix error bar on Rt for PLA_dry
• [x] you only need the variable name in the bottom row (keep the Rt labels for TLP and PLA)
• [x] you can save space by moving the legend into the plot space