Notes from meeting with Adam June 8

revise Balsamroot vol, brix and mass models to include date random effect
BRIX ~ treatment * year +(1|plot/plant) + (1|year:date), data = balsboth
redo ls means analysis for these new models
redo model diagnostics for these models

in exploration, plot date by BRIX
plot(as.factor(balsam$date, balsam$BRIX) - shows high variablility among dates so need to add as random effect

Be careful about interpretation : when flowers are producing nectar, this is what we see...

revise Buckwheat models to incorporate quadrant random effect
BRIX ~ treatment * year +(1|plot/quadrant) + (1|year:date), data = buckboth
check to see if high variability among dates and if we need the year:date random effect
redo ls means for these new models
redo model diagnostics for these models

For presence/absence
data exploration
do ls means analysis
run model diagnostics
back-transform effects from logit scale
Think about: remove data points from end of season?

Update model results spreadsheet

June 10
Updated buckwheat munging files to include quad as a separate variable.
Plotted BRIX, mass and volume against quadrant, but almost no variability among quadrants for all 3 variables
Added plots to data exploration files
Ran models with and without quadrant nested random effect, and difference is negligable (see comment in sugar code)
Decided not to include nested quadrant effect in model

Plotted BRIX, mass and volume against date, added plots to data exploration file
June 30 2015 is a major outlier for both volume (high) and BRIX (low) - maybe too wet that day? check field notes
Question: remove 6/30/15 from dataset?
Variablility in volume and sugar mass is much lower in 2016 than 2015
ran models both with and without +(1|year:date), and negligible difference in estimates and p-value for both mass and BRIX
p-value for volume went down from 0.0485 without year:date effect to 0.0063 with year:date effect
  estimates are only negligibly different
decided not to include year:date effect in model

June 21
Updated balsamroot analysis (for all 3 response variables): modified model to include year:date random effect
Updated model analysis (including ls means) and model diagnostics to reflect new model
Update data exploration files to plot response variables by date and treatment plot

Re-looked at new date exploration plots from June 10 for buckwheat - shows fair amount of variation by date for all 3 response variable
Decided to update buckwhat models to also include year:date random effect
Updated model analysis (including ls means) and model diagnostics to reflext new model

Ran analysis on presence/absence of nectar for both balsamroot and buckwheat (using same models as 3 response variables)
Plotted presence/ansence versus date and treatment plot for both species
Ran model diagnostics for models - CHECK RESIDUALS PLOT (ESP FOR BUCKWHEAT) WITH ADAM

need to update model results summary spreadsheet
need to re-run pdfs with new files
need to update graphs?

Notes from meeting with Adam June 22

For volume and sugar mass

Trumpet residuals are a problem.  Try log transform, then maybe square or cube root to randomize residuals.
	See BalsamModsDiag file: line 20 plot of residuals
Don't need quadrant random effect (i.e., no (1|plot/quad)) in buckwheat models

For nectar presence/absence

2015 Buckwheat is a problem: it's hard to analyze binary data when it's all 1's or all 0's. 2015 buckwheat is almost all 1's
Cut 2015 buckwheat from analysis 

Diagnostics: Don't worry about residuals.  Adam says because all the variables are categorical the model will fit
(he used the following call, and it didn't tell him anything useful: plot(fitted(modbals), abs(residuals(modbals, type="deviance"))), or try without the absolute value (abs)
	This uses deviance residuals

For number of flowers

Can't use buckwheat data - don't have absolute numbers
For balsamroot data - need to get a number for each plant (i.e., one number that captures the total number of flowers that plant produced)
Can work through data - # of buds to flowers vs aborted buds, stalks - to get a number for each plant, or, grab a time slice
To do: plot # of fully flowering over time.  If all at the same time, then maybe take a time slice

Notes June 26

Updated balsamroot models to log transform volume and sugar mass.  Residuals look much better now.
Balsamroot BRIX model was fine, no changes.
Updated buckwheat models to log transform volume and sugar mass.  Residuals look much better now.
Buckwheat BRIX model has odd residuals - might be okay but check with Adam on this one.

Added ln(volume) and ln(mass) to balsamroot and buckwheat data diagnostics

deleted old pdfs of model and data explore runs
need to make and upload new pdfs

Notes June 27

Updated presence/absence model to remove 2015 buckwheat (because all 1's)
Moved data exploration into this file
Moved obsolete pres/abs model diagnostics file to "old analysis"
ran pdfs for all volume, mass, BRIX and presence/absence models  (the knitter doesn't like the "influence" function)
Updated model results file with new model runs

Notes July 1
wrote munging for balsamroot phenology 2016.  Did some data exploration.
Need to think about how best to represent # of flowers each plant produces

Notes July 5
Spoke with John Pleasants on July 3 about how best to represent # of flowers each plant produces.  He suggested that the specific strategy (i.e., taking max flowers, taking max scenesed, etc.) wasn't as important as being consistent with the method for assigning a number across both treatments
Take home: use whatever method seems best, just be consistent across both control and treated plants

Finished munging for 2015 and 2016 Balsamroot 
Need now to sit down with the by-plant dfs and assign a total number of flowers to each plant
Then put in a new spreadsheet, import into R, and run analysis.

Notes July 6

Assigned total flowers to each plant:
Looked primarily at the number of senesced flowers and TF at the end of the season.  Most plants were straightforward - it was obvious how many flowers the plant had produced.  However, some plants had inconsistently recorded flower numbers, in this case we made a judgement based on our knowledge of field techniques.  Again, most plants were straighforward.
Plants where we had to make a judgement call: (2015): CHSR8-4, CHSR8-6, EHSR2-8, EH4-6, WHSR9-3 (5/47 or 11%)
(2016): CHSR8-4, EH4-6, EHSR1-1, EHSR1-2, WHSR9-1, WHSR9-2, CC6-1, CSR7-3, CSR7-9, ESR2-2, ESR2-3, ESR2-4, WSR10-1 (13/59 or 22%)
In all of these cases, a different person might have assigned a different number, but most of the time it won't be off my more than one.

Many plants produced flowers in one year but not the other (2 plants produced in 2015 that didn't in 2016, 20 produced in 2016 that didn't produce in 2015)
Double-checked original data for these plants, and this is accurate.

Look at some plots of the data - it looks like there's something going on there
Need to ask Adam about appropriate model (i.e., random effect by plant, year?, plot?)

Notes July 8

Added tentative models for flower numbers, glmer(#offlowers ~ treatment + (1|plant), family = poisson), for 2015, 2016, and both
checked disperson and it looks good (very close to 1 for all 3 models)
checked model fit for all three models - also looks good, but residuals plot is a little funky - not sure how to interpret
Ask Adam = poisson distribution correct?  How to interpret residuals plot?

Notes from meeting with Adam July 10
Plotted 2015 vs 2016 total number of flowers, colored by treatment
> plot(flowers$year15, flowers$year16, xlim=c(0,22), ylim=c(0,22)) #drops the 1 outlier off the graph so we can see the rest clearly
> with(subset(flowers, treatment=="H"), points(year15, year16, pch=16, col="blue")) #color heat treatment data blue
No pattern of treatment by year in total number of flowers (i.e., no clustering of treatment data)
Add plot-level random effect to # of flowers models for 2015 and 2016, so that model matches experimental design
Get rid of model that combines both years.  The model I had was incorrect, and by the time you correct it you may as well have just run the two different years separately.
Models for nectar charactersitics and presence/absence look good - they're done.
Saved R history from this session in Documents folder

Notes July 11th
Corrected models for total number of flowers per notes from Adam above
Calling the analysis done - moving on to manuscript
Need to reorganize files and develop DMP, then archive

Notes August 2
Updated figures files:
boxplots etc only for those models that show a significant difference between treatments
volume graph - data from Opler 1983, which includes raw data for volumes for plants that fall into a "pollination syndrome"
	data based on measured volumes in plants that fall into "pollination syndromes"
	Need to be careful about this - defining pollination syndrome based on some nectar characteristics
concentration graph - based on Willmer 2011 p. 206 (only those values for "optimum concentration", i.e., optimal concentration for fast intake as recorded in labratory work), as well as Nicolson 2007 p. 303 - many of the same sources as Willmer 2011
	Values for concentrations are based on empirical measurements of sucrose intake