sampling and survivorship biases

[teixeirak]

Tree-ring analyses of change through time are subject to a number of sampling and survivorship biases (see Box 3 of Walker et al. 2020 and refs therein).

I need to dig into understanding these and how they should be handled. At a minimum, we need to discuss the reliability of the trends. It may be necessary to limit the analysis to species that meet certain sampling criteria, although I'm not yet sure exactly what those would be. One may be limiting the analysis to instances where trees were recruited more-or-less continuously over a broad range of years, as opposed to in cohorts (see Brienan et al. 2017). Another may be focusing on sites sampling the whole population (SC, and ZOF may be close) or a random subset (SCBI), as opposed to just canopy trees (Nehrbass-Ahles et al. 2014).

@ValentineHerr , if you can do so easily, it would be helpful if you could generate a file like cores_vs_dbh_included_in_analysis.pdf, but with year instead of DBH on the x-axis.

[teixeirak]

I need to turn my attention to other projects for awhile and come back to this later.

[ValentineHerr]

I added cores_vs_Year_included_in_analysis.pdf but I think it may be hard to see what you want... example:

I am guess these plots would be more informative. They are density plot of the first year of each core. When there are spikes (like in at LD for FRAM), it means that most of the cores from that species started within few years of each other (not that there are few cores from FRAM at Lilley Dickey, and they may come from the same tree). No spike, like at BCI, would mean that recruitment happened evenly across time.

[teixeirak]

Oh, that's awesome; thank you!

[ValentineHerr]

(note: I just updated the pdf, I think they may have been some cores where I wasn't taking the first year.)

[teixeirak]

These are useful. I'll at least incorporate this info into the discussion.

Note: for CB, they truncated what they sent us at 1850 (although apparently not for quite all), so we want to start the analysis then and not count 1850 as a recruitment year.

[teixeirak]

I still need to read up on the potential biases, but one that particularly concerns me is that if most of the trees recruited at the same time, fast growing trees will reach a set size sooner, resulting in an artificially negative trend (Brienan et al. 2017). (Although to what extent would the random effect of tree take care of that?)

From a first look, I don't see a ton of evidence for that dynamic, but we might remove from the analysis species that had, say, >50% of their recruitment within a 20 year period.

But.... I really need to set this project aside for now and turn attention to the ERL review! (Sad, as I have good momentum with it at the moment.)

[teixeirak]

@ValentineHerr , could you please make the density plots separately by species, with one small panel per species, ideally taking < 1/2 page, except at sites with many species? Because the analysis is separate for each species, and relative abundance of species varies, its more helpful to see them separately. I plan to probably include these in the SI, alongside Cameron's decadal plots:

e.g.,

My basic plan is to acknowledge and provide info relevant to some of the potential biases. Fully controlling for them would require an analysis that's beyond our scope here, so my goal is to acknowledge and discuss the fact that stand dynamics can influence these patterns. Fortunately, I think that our analysis does a number of things right--probably as much as possible without a big sensitivity analysis.

[teixeirak]

Eventually, I'll want to annotate the year trends plot with which potential biases may apply at each site. I'm still figuring out exactly how that will work.

[teixeirak]

Here are some notes on the various biases:

slow-grower survivorship bias , a.ka. modern sample bias (Brienan et al. 2012)

problem

• assumes size-dependent mortality (trees die when they hit a certain size), and that individual trees are persistently fast or slow growers
• persistently slow growers are over-represented in older records, so mean growth rates are underestimated for those times
• problem is modest, and dependent upon assumptions about mortality

solution

• This bias is modest
• I believe random effect of tree should help quite a bit with this, and perhaps solve it altogether

big tree selection bias (Brienan et al. 2012)

problem

• see fig above
• assumes age-dependent mortality, and that individual trees are persistently fast or slow growers
• sampling often focuses on the largest trees, which would tend to be fast growers, so it will appear that growth rate increases over time
• problem can be severe, particularly when the minimum size threshold for sampling is high

solution

• Sample small minimum diameters. Ideally, sample all trees, or sample sizes in proportion to their abundance
• This would be a problem for some of our sites, but not all. I want to annotate the plots to note where this bias is likely to occur
• Again, I believe that the random effect of tree should help quite a bit to overcome this bias

clustered age distribution bias (Brienan et al. 2017)

problem

• If there are clustered age distributions/ little recruitment in recent years, bias can arise because "persistent growth differences result in fast-growing individuals reaching the size class in which growth rates are calculated at a younger age compared to slow growers."
• This problem does apply in regional curve standardization approach
• This effect was first described by Vlam (2014), and its possible influence on growth trend evaluations was discussed by Groe- nendijk et al. (2015).

solution

• Again, I believe that the random effect of tree should help quite a bit to overcome this bias
• This would be a problem for some of our sites/species (assessed using these plots), but not all. I want to annotate the plots to note where this bias is likely to occur

‘predeath bias’ (Bowman et al. 2013)

problem

• Due to the marked slowing in tree growth rates in the several years [37] to several decades [38] prior to tree death, widths of recent growth rings, which include those from individuals that are dying (but currently still alive) and have lower growth rates, may be smaller than those from the more distant past [11]

solution

• we remove 20 years prior to death. This doesn't fully solve the problem though.

‘juvenile selection effect’, see Bowman et al. 2013 , Groenendijk et al., 2015).

problem

• "The juvenile selection bias occurs if large canopy trees of today were relatively fast growing in the past, that is slow-growing juvenile trees have a lower chance to reach maturity than fast growers. This bias may lead to the detection of growth decreases over time, or may mask growth increases." (Groenendijk et al 2015)

solution

[ValentineHerr]

These figures are great!

[ValentineHerr]

@ValentineHerr , could you please make the density plots separately by species, with one small panel per species, ideally taking < 1/2 page, except at sites with many species?

I pushed the updated pdf. Please have a look and let me know if each page look good. If yes, I'll try to save each site separately in a jpeg or png file and adapt the size of figure according to n species...

[teixeirak]

Thanks! If it's easy to do, it would be nice to have these stacked vertically, with max y-value independent for each species. However, these are SI figures, so no need to waste time making them too beautiful.

I plan to use these to identify and probably skip analysis of species with a potentially large clustered age distribution bias (explained above), but need to read more to figure that out.

For the big tree bias above, I want to set a minimum DBH sampling threshold, which should eliminate ABAL at ZOF but probably no others (which have already been eliminated through the other criteria)

[ValentineHerr]

I pushed individual plots for each sites... Let me know if they look good.

One issue is the color coding and species codes.. I would need to bring our standardizing over from this project... I probably can do it (but not tonight), if you think it is worth the effort

[ValentineHerr]

Note: for CB, they truncated what they sent us at 1850 (although apparently not for quite all), so we want to start the analysis then and not count 1850 as a recruitment year.

I still need to do that

[teixeirak]

These are great! No need to worry about species color-coding. I'll look at them carefully later.

[teixeirak]

Some minor things on these density figures:

• [x] please save without spaces in the name (those mess up knitting into document)
• [x] Remove site name
• [x] For SCBI, y-axes are too scrunched up to read. A two-column plot would be fine
• [x] Standardize x-axes, and include up to the coring year (or 2020 if easier) even if there’s zero recruitment.
• [x] for CB, it’s probably okay (best?) to plot 1850 as the recruitment year for trees established before then, and I’ll just note that anomaly in the caption.
• [x] If easy, replace species codes with scientific names
• [x] also, if easy, it's a bit easier to keep the species names as plot titles (as you had and removed), rather than as a legend

[ValentineHerr]

I think I addressed everything above.

[teixeirak]

Thanks; looks great! I'm realizing, though, that my idea for CB isn't working, especially because it seems that not all records are cut at 1850:

Probably the easiest is to adjust that x-axis to start at 1875 (without adjusting y-axis). Then I'll note in the caption that for species with low establishment post-1875, establishment was prior.

[teixeirak]

Also, did you standardize the x-axis max? Differs across species here:

If easiest, it would be okay to just set them all to 2020.

[ValentineHerr]

Are you sure you have the latest version? There should be only one x-axis, on the lower panel.

[teixeirak]

Looks like the Zofin fig wasn't updated (but all others were):

[ValentineHerr]

I pushed the new version for cedar breaks. That removed Pinus longaeva

[teixeirak]

(oops, wrong screenshot! That was a family zoom event. :-) Fixed now.)

[teixeirak]

I pushed the new version for cedar breaks. That removed Pinus longaeva

Thanks! Could you please either hold the y-axes where they were (keep the same analysis, just adjust the x-min on plotting) or annotate with % of cores starting before 1875?

[ValentineHerr]

Sorry, I did not realize that moved the y limits:

Does something like this work? (there is only 3 cores starting after 1875 for PILO, not enough for a density plot I suppose).

[teixeirak]

Yeah, that works; thanks!

[ValentineHerr]

ok, I pushed it.

[teixeirak]

Thanks! And please don't miss the Zofin plot, which for some reason didn't get updated in the last round (see above)

[ValentineHerr]

Not sure what happened there, it is not in the data anymore!! I will investigate where/when it disappeared, this afternoon. Probably a minor issue somewhere.

[ValentineHerr]

Ok, updated the dendro repo, so Zofin is there...

I am not sure what happened, but re-running everything changed some of the outputs files... I think it is mostly just the same data but in a different order. Except for Harvard, where things look a little different.... I don't think it is super serious changes but I'll try to investigate when I have a little more time.

[teixeirak]

Thanks! You're just talking about differences to these figures (not the full results), right?

[teixeirak]

minor thing: just noticed that PCAB got species name lable of "NA"

[ValentineHerr]

I fixed PCAB (to PIAB) so that it would match properly with latin name.

[ValentineHerr]

Thanks! You're just talking about differences to these figures (not the full results), right?

unclear... I am going to do a test run to see if the results change.

[ValentineHerr]

Thanks! You're just talking about differences to these figures (not the full results), right?

unclear... I am going to do a test run to see if the results change.

So the results (response curves, interactions etc....) have not changed for HF, but there are a few instances where the curves are drawn with a different line type, implying the significance of some of the terms have changed....

Example:

I really don't have a good understanding of what happened... maybe I had not updated HF after the last change in the dendro repo???... All sites are run together, so I don't thing that could have happened...

I'll re-run everything to update all figures and tables

[teixeirak]

Allo-db changed a bit, right?

[ValentineHerr]

yes, that could be it