search

EcoClimLab / ForestGEO-tree-rings

Repository for analysis of tree-ring data from 10 globally distributed forests (Anderson-Teixeira et al., in press, Global Change Biology)
2 stars 2 forks source link

Try ClimWin for tree-ring analysis #1

Closed teixeirak closed 4 years ago

teixeirak commented 4 years ago

@ValentineHerr,

As discussed in person, I'd like to try the climwin R package, described in this paper, on our tree cores from SCBI.

Note that Sean McMahon is using climwin and help with some questions: 


Hi Krista, 

Funny you should write this now. I’ve been working with climwin the past few weeks to look at the phenology of bole expansion and cessation. I have code you can use that might help you set up the dates (POSiXct stuff which can always be a headache), interpret output, and a bunch of recommendations, which we should talk about. 

Helene’s right about overfitting, and climwin has some randomization algorithm that is supposed to account for that after you get a ‘best’ model. I haven’t really focused on that aspect, as I’m not so much trying to get a precise hypothesis test (i.e., rejecting nulls), and not trying to build a predictive model (i.e., a x increase in climate variable will lead to a y advance of spring expansion), but instead trying to get any realistic potential climate signal of so many possible variables and time lags which can be pointed to as suggestions of causes, and guides for future physiological work. 

There are problems with climwin. Some have to do with the package, but really it is pretty straightforward. More important, there are just some things that correlations of patterns with climate drivers just fall prey to. E.g., you should use ‘absolute’ vs. ‘relative’ searches, or you will just tend to get a correlated trend (you will get significant results that are counter-intuitive, such as warmer weather is associated with later spring growth. But with relative searches, this is because weather is warmer later in the spring, not because it causes later growth!). Lots of little things like this you should not have to re-discover!

Again, let’s chat about it. Not a magic bullet, but a pretty solid search tool.

Cheers,

Sean
ValentineHerr commented 4 years ago

Posting here a couple things to think about:

  1. multiple signals. See this thread on a climwin issue. I our case QUAL, for example shows 2 signals. I implemented recommendation and the second signal persists somewhat, meaning it is not due to an auto-correlation issue. I suppose we should always check for that in the future.

    QUAL best model output with true null model:

    image

    QUAL best model output with quadratic term of tmx of the strongest signal in previous figure as a fixed effect (as the "null model"):

    image

  2. this thread also shows how a non detrended response variable could be an issue in the analysis... Sounds like we can add Year in our baseling model to account for the trends...

teixeirak commented 4 years ago

Thanks, @ValentineHerr !

  1. Makes sense.
  2. Thanks for finding this. It sounds like it will make sense and may be the solution, but I want to dig into it more deeply.
ValentineHerr commented 4 years ago

@teixeirak , I uploaded the figure for the mixed model with dbh (roughly calculated for now) as fixed effects and species as random effect. Best model is with dtr of the year before but \it doesn't look like a striking pattern to me.