do a true out of sample comparison

[sdtaylor]

Use some hold out data in the NPN and LTS datasets to actual out of sample comparisons at the end

[sdtaylor]

years used for testing data within each dataset

dataset_test_years = list(c(2012,2013,2014), #harvard
                          c(2015,2016),      #npn
                          c(2013,2014,2015), #hubbard
                          c(2014,2015),      #hjandrews
                          c(NA))             #jornada, no formal test data

[sdtaylor]

leaving out specific years for each dataset leads to some species/phenophases with many test samples, and some with very few. Instead I'll do, for each species/phenophases, a random 20% of observations

[sdtaylor]

done 9e5e436