diazrenata / revisitbecs Goto Github PK

decrease p threshold accordingly?

• Data
  • NEON and other data
  • Subsample/restrict datasets to a comparable time scale & grain as NEON
• BSED 'uniformity'
  • Multiple questions in play:
    • EER/an equal amount of energy per body size. This requires a nonrandom division of energy units among individuals, with preference going to smaller individuals.
    • Uniform size-abundance distributions (this is what was tested in Ernest 2005). Results in a non-uniform BSED, as larger-bodied species control more energy per capita.
    • Multimodality: see Thibault paper for inspiration
  • Constraint-based generative algorithm: borrow from feasiblesads?
    • Take as constraints N and E, and $E_max$ and $E_min$. Allocate $E_min$ to all individuals and $E_max$ to one individual. Randomly distribute E - E_max - E_min(N-1) among (N-2) individuals.
    • Needs more thought. But, some extreme forms of these distributions are impossible depending on $E_min$, $E_max$, $E$, and $N$. For example, maybe the uniform size-abundance distribution can't be achieved, because you can't have $N$ individuals evenly distributed between $E_min$ and $E_max$ without using more energy than $E$.
    • None of these approaches consider the constraint imposed by the (presumed, by me) constraint on intraspecific size variation for mammals.
• BSDs and KS tests
  • Alternative to KS test.
    • At minimum use tests for grouped data.
    • Consider simply bootstrapping?
• Scaling to larger dataset
  • drake?
  • False discovery rate

to validate results

HJ Andrews Experimental Forest 1995-1999

Niwot Ridge small mammals 1981-1990

Sev - this one has 8 sites, so I will need to figure out which ones Morgan used.

https://cran.r-project.org/web/packages/neonUtilities/
https://github.com/NEONScience/NEON-utilities

see diazrenata/smammal-mete for some datasets

e.g. Portal, UHURU, Kelt, Hog Island...

test with simulated datasets including communities that are and are not different...

right now they are in the .Rmd which is not ideal

See Zar 1999 at Marston

From Zar 1999:

To find the test statistic for the K-S goodness of fit for continuous data, calculate both

• D_i = |rel F_i - rel Fhat_i| where rel F_i = cumulative relative frequency and Fhat_i is the cumulative expected relative frequency (from comparison/proposed distribution). (Cumulative relative frequency = for each measurement X_i, f_i = frequency of that measurement. For each X_i, rel F_i is the proportion of measurements in the sample <= X_i.)
• and D'_i = |rel F_i-1 - rel Fhat_i|
  for each i. F_0 = 0.
  The test statistic is
  D = max[max(D_i), max(D'_i)]; D = the largest D_i or the largest D'_i, whichever is larger.
  Compare values of D to critical values based on sample size and alpha; if D > the critical value, reject H_0.

For small sample sizes (n < 25ish), increase power by employing the correction expounded by Harter, Khamis, and Lamb (1984) and Khamis (1990, 1993) (look these up)
For each i we determine
rel F = F_i / (n + 1)
rel F'i = (F_i - 1) / (n - 1)
Then obtain differences
Ddelta0,i = |rel Fi - rel Fhat_i|
Ddelta1,i = |rel F'i - rel Fhat_i|
The test statistic is either max Ddelta0,i or max Ddelta1,i, whichever leads to the highest level of significance (i.e. the smallest probability).
Compare to critical values in appendix.