add in sequence for a max of 3 (e.g: for age 5 test: age 4, 4+3, 4+3+2 )
pick best based on AIC
current plan
put in option to select pooled 2 or pooled 3
age 6 ~ p54 or p543
age 5 ~ p43 or p432
age 4 ~ p32 or p321
match based on brood year
-> now have sibreg.prep() fn that generates a compact object for the simple and pooled sibregs
-> have revised fitModel() and/or sibreg.simple.est() to use this rather than the current computation ("run year offset by 1")
for youngest: use naive
for second youngest: use simple sib reg
for third youngest: always use pooled 2
for older: selection of pooled 2 vs pooled 3 has an effect
-> verified
to discuss
[ ] NA handling end years: put sample table and note into wiki
[ ] NA handling: what if one of the age classes is NA -> make the pooled sum NA, and lose that data point from the sib reg fit? -> NAs NOT ALLOWED! -> make sure it is in the wiki and the app help!
[x] covariates for pooled: NO
[ ] need to discuss: which years used in FC step for pooled sib reg models? BY for forecasted age,right?
Multiple
need new stepwise code
possible combinations constrained as follows:
max of 3 predictors
always use previous age class
add up to more 2 age classes (in sequence, don't skip)
any covariates up to 3 total predictors
need to decide: use stepwise pkg that does all the combinations or write code generate valid model variations (text strings for lm input)
option for choosing max predictors in stepwise pkg? -> some yes, some no -> need to test
Pooled
original plan
current plan
put in option to select pooled 2 or pooled 3
age 6 ~ p54 or p543 age 5 ~ p43 or p432 age 4 ~ p32 or p321
match based on brood year
-> now have sibreg.prep() fn that generates a compact object for the simple and pooled sibregs -> have revised fitModel() and/or sibreg.simple.est() to use this rather than the current computation ("run year offset by 1")
for youngest: use naive for second youngest: use simple sib reg for third youngest: always use pooled 2 for older: selection of pooled 2 vs pooled 3 has an effect -> verified
to discuss
Multiple
need to decide: use stepwise pkg that does all the combinations or write code generate valid model variations (text strings for lm input)
Ideas:
consider alternative versions:
Basic Questions
transformation / standardized -> for now, hardwire a default normalize/centralize, in the future plan for transformation options.
retrospective: run a new model selection for each retrospective step ("dynamic") -> but depends on the speed. Maybe do 2 options.
selection criteria: r.sq vs AIC vs. BIC vs. KIC (e.g. this discussion