Internal function: MARSS parameter estimation using TMB

Source: R/estimate_marss2.R

This model is in the general "marss" vectorized form. In estimate_marss2(), I use the approach in MARSS::MARSSoptim() where I just use the chol() of these matrices. Technically there are 2 equal solutions since the diagonals appear as the square so -a and a are the same. But I have not observed that this affects the LL of optim() but it definitely seems to slow things down. In estimate_marss(), the diagonals and offdiagonals of R and Q are split apart like Tim Cline did. I had some problems with the splitting approach for some models with Q unconstrained, though now it seems fixed.

estimate_marss2(
  MLEobj,
  method = c("TMB", "nlminb_TMB", "BFGS_TMB"),
  opt.control = NULL,
  ...
)

Arguments

MLEobj

A properly formatted MARSS model as output by MARSS()

method

Normally passed in as MLEobj$method, but allows user to pass in a new method if they want to use MLEobj with another method. Allowed values are "TMB", "nlminb.TMB", "BFGS.TMB".

opt.control

Normally this is passed in as MLEobj$control, but if the MLEobj was set up using a different method, then you will need to set the opt.control options. See details.

...

not used

Value

A list with the objective and optimization objects.

  • obj is the raw output from the TMB::MakeADFun() call.

  • op is the raw output from the optimization call (optim or nlminb). Note that the function is minimizing the negative log-likelihood so the sign will be opposite of the log-likelihood reported by MARSS()

  • opt.control is the controls sent to the optimization function.

  • method method used for optimization

Details

Minimal error checking is done in this function. Normal calling function is MARSS::MARSS() with method="TMB".

The main thing this does is

  • collapse the 3D fixed and free matrices into 2D

  • separate out the diag and offdiag parameter elements of R and Q

Restrictions

  • V0 fixed (not estimated)

  • Q and R cannot be time-varying (at the moment)

opt.control is what is passed to the control argument in nlminb() or optim(). If you use MARSS(x, method="TMB"), this will be set to appropriate defaults which you can see via MLEobj$control. But if you call estimate_marss() with a MLEobj from a call such as MARSS(x, method="kem") (so not a TMB method), you will need to set opt.control if you want values different from the base defaults for those functions. Note as a shortcut for nlminb(), you can set both eval.max, iter.max to the same value with opt.control=list(maxit=1000). Note, if you pass in opt.control, this will replace all values currently in MLEobj$control that are associated with the optimizer function.

The defaults set in MARSS::MARSS() are

  • nlminb: eval.max = 5000, iter.max = 5000 and trace = 0.

  • optim: maxit = 5000 and trace = 0

All other controls for the optimization function are left at NULL.

Author

Eli Holmes.

Examples

library(MARSS)
data(lakeWAplankton, package = "MARSS")
phytoplankton <- c("Cryptomonas", "Diatoms", "Greens", "Unicells", "Other.algae")
dat <- as.data.frame(lakeWAplanktonTrans) |>
  subset(Year >= 1980 & Year <= 1989) |>
  subset(select=phytoplankton) |>
  t() |>
  MARSS::zscore()

# set-up the model
mod <- MARSS(dat, model=list(m=3, tinitx=1), form="dfa", fit=FALSE, silent=TRUE)
# fit
fit <- estimate_marss(mod)
#> MARSSfit.TMB: The method in the marssMLE object is not TMB. Setting to nlminb_TMB for fitting.