Get bootstrapSR to work with Blim definition

[marchtaylor]

srrTMB allows for the definition of Blim when fitting a segmented regression with fixed inflection point. Could this argument also be used in bootstrapSR ?

Cheers, Marc

[iagomosqueira]

Could you please test the latest commit? Blim is now ignored in non-segreg models

[marchtaylor]

Thanks for the quick response.

bootstrapSR works with Blim argument, but not when only segreg model is defined alone:

library(FLCore)
library(FLSRTMB)

data(ple4)

Blim <- 3e5


# works when 2 models are specified ------------
mods <- c("bevholt", "segreg")

fits <- srrTMB(as.FLSRs(ple4, models=mods), 
  spr0=mean(spr0y(ple4)), Blim = Blim)

plotsrs(fits)

srpars <- bootstrapSR(ple4, iters = 20,
  models=mods, method="best", Blim = Blim)

srpars@.Data["m",]


# doesn't work when only segreg is specified --------------
mods <- c("segreg")

fits <- srrTMB(as.FLSRs(ple4, models=mods), 
  spr0=mean(spr0y(ple4)), Blim = Blim)

plotsrs(fits)

srpars <- bootstrapSR(ple4, iters = 20,
  models=mods, method="best", Blim = Blim)
# Error in fits[[best]] : 
#   attempt to select less than one element in get1index

srpars@.Data["m",]

[iagomosqueira]

As far as I can see it is just that with that Blim value the bevholt provide a better fit and get selected. I tried

set.seed(887)
srpars1 <- bootstrapSR(ple4, iters = 20,
  models=mods, method="best", Blim = Blim)
c(srpars1$m)

set.seed(887)
srpars2 <- bootstrapSR(ple4, iters = 20,
  models=mods, method="best")
c(srpars2$m)

so the bootstrapped pairs should be the same. I can try outputting the actual pairs to check it.

[marchtaylor]

Hi there. In the second example, I changed mods to only be segreg, and then get the error with bootstrapSR.

[iagomosqueira]

OK, checking it again

[iagomosqueira]

The fit with fixed(b=Blim) returns no llkhd, so the selection in bootstrapSR fails. Will look in to it.

[iagomosqueira]

FLSRTMB uses FLCore::fmle to handle the segreg with Blim option. This call returned an FLPar with different dimensions so the call to combine at the end of boostrapSR was incorrect. Also, the bootstrapped year dimnames were confusing the FLQuant creator sometimes.

Both issues appear solved now, but I would appreciate some test from your side.

srpars <- bootstrapSR(ple4, iters = 50,
  models=mods, method="best", Blim = 2.5e5)

> c(srpars$m)

 [1] 1 1 1 1 3 1 1 1 1 1 1 1 1 1 1 1 1 3 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

[marchtaylor]

Looks good! Thanks Iago