Solitary bee analysis steps.txt

Steps for solitary bee analysis (June 19, 2017)

-------------------------------------

Using traps that were in the same location both years:

1) Choose bee spp.

	Some have very low abundance in 1 or both years
	Top 10 candidates, filtering out ones with large year-to-year differences
                     Species y2015 y2016
       Anthophora terminalis   589   346
         Melissodes confusus   349   335
    Lasioglossum leucozonium   146   213
              Dufourea maura   130    62
         Megachile perihirta    87    55
  Lasioglossum dialictus_sp1    81    32
         Hylaeus hylaeus_sp9    35    63
     Anthophora occidentalis    50    29
          Melissodes rivalis    23    50
            Andrena peckhami    28    43


2) Truncate data to meaningful temporal range

	Surveys were done during different times of year. To compare year-to-year data, we need to use the same temporal range.

	Ideas:
	Use day of year (simplest)
	Use GDD (harder to explain, possibly more biologically relevant)
	Fit abundance-time curves to each site/year, and integrate over relevant ranges

3) Calculate overlap in 2015

	Ideas:
	Days of overlap in canola bloom (simplest)
	Fit curves to flower abundance, abundance, and integrate
	Standardize canola bloom (?) or populations (?)
	What landscape radius? (probably closest one)

4) Measure population responses in 2016 (Popn2016 ~ Popn2015 * Overlap) for each spp
	
	Ideas:
	Mean count per trapping effort (easy)
	Integrate over similar period from 2016

5) What else?? 

	Potential things to include:
	Other crops (e.g. sunflower, clover) or some measure of landscape flower availability (a la Jenn)?
	Different radii (use autocorrelogram)
	Honeybee abundance as a covariate (assumed to have a negative effect)		
	Crops in 2014 as a predictor of 2015 population? (What do we use as the "starting" population?)
	


Paul's comments:

Could do some kind of integral approach (as I suggested), or could do:

Nbee2016 ~ Nbee2015 * Ncanola2015 + s(GDD2016) + (1|spp/site)

	With this model, we would focus less on accept/reject model, but look at model coefficient b/w Nbee:NCanola



Other things to remember 
	- deployedHours must be used as an offset
	- consider the brms package in R for STAN-based Bayesian hierarchical models