cjs_tmb: Fitting function for CJS models

Description

A function for computing MLEs for a specified Cormack-Jolly-Seber open population capture-recapture model for processed dataframe x with user specified formulas in parameters that create list of design matrices dml. This function can be called directly but is most easily called from crm that sets up needed arguments.

Usage

cjs_tmb(
  x,
  ddl,
  dml,
  model_data = NULL,
  parameters,
  accumulate = TRUE,
  initial = NULL,
  method,
  hessian = FALSE,
  debug = FALSE,
  chunk_size = 1e+07,
  refit,
  itnmax = NULL,
  control = NULL,
  scale,
  crossed = TRUE,
  compile = TRUE,
  extra.args = NULL,
  reml,
  clean = FALSE,
  getreals = FALSE,
  prior = FALSE,
  prior.list = NULL,
  tmbfct = "f1",
  useHess = FALSE,
  ...
)

Value

The resulting value of the function is a list with the class of crm,cjs such that the generic functions print and coef can be used. Elements are 1) beta: named vector of parameter estimatesm 2) lnl: -2*log likelihood, 3) AIC: lnl + 2* number of parameters, 4) convergence: result from optim; if 0 optim thinks it converged, 5) count:optim results of number of function evaluations, 6) reals: dataframe of data and real Phi and p estimates for each animal-occasion excluding those that occurred before release, 7) vcv:var-cov matrix of betas if hessian=TRUE was set.

Arguments

x: processed dataframe created by process.data
ddl: list of dataframes for design data; created by call to make.design.data
dml: list of design matrices created by create.dm from formula and design data
model_data: a list of all the relevant data for fitting the model including imat, Phi.dm,p.dm,Phi.fixed,p.fixed, and time.intervals. It is used to save values and avoid accumulation again if the model was re-rerun with an additional call to cjs when using autoscale or re-starting with initial values. It is stored with returned model object.
parameters: equivalent to model.parameters in crm
accumulate: if TRUE will accumulate capture histories with common value and with a common design matrix for Phi and p to speed up execution
initial: list of initial values for parameters if desired; if each is a named vector from previous run it will match to columns with same name
method: method to use for optimization; see optim
hessian: if TRUE will compute and return the hessian
debug: if TRUE will print out information for each iteration
chunk_size: specifies amount of memory to use in accumulating capture histories; amount used is 8*chunk_size/1e6 MB (default 80MB)
refit: non-zero entry to refit
itnmax: maximum number of iterations
control: control string for optimization functions
scale: vector of scale values for parameters
crossed: if TRUE it uses cjs.tpl or cjs_reml.tpl if reml=FALSE or TRUE respectively; if FALSE, then it uses cjsre which can use Gauss-Hermite integration
compile: if TRUE forces re-compilation of tpl file
extra.args: optional character string that is passed to admb if use.admb==TRUE
reml: if set to TRUE uses cjs_reml if crossed
clean: if TRUE, deletes the tpl and executable files for amdb if use.admb=T
getreals: if TRUE, will compute real Phi and p values and std errors
prior: if TRUE will expect vectors of prior values in list prior.list
prior.list: which contains for normal distributions 1) mu_phi_prior: vector of mu values for phi_beta, 2) sigma_phi_prior: vector of sigma values for phi_beta, 3) mu_p_prior: vector of mu values for p_beta, 4) sigma_p_prior: vector of sigma values for p_beta, 5) random_mu_phi_prior: vector of mu values for ln sigma of random effects, 6) random_sigma_phi_prior: vector of sigma values for ln sigma_phi, 7) random_mu_p_prior: vector of mu values for ln sigma_p, 8) random_sigma_p_prior: vector of sigma values for ln sigma_p.
tmbfct: either "f1" - default or "f2" - any random effects treated as fixed effects or "f3" fixed effects fixed at mode and no random effects.
useHess: if TRUE, the TMB hessian function is used for optimization; using hessian is typically slower with many parameters but can result in a better solution
...: any remaining arguments are passed to additional parameters passed to optim or cjs.lnl

Author

Jeff Laake

Details

It is easiest to call cjs through the function crm. Details are explained there.

Be cautious with this function at present. It does not include many checks to make sure values like fixed values will remain in the specified range of the data. Normally this would not be a big problem but because cjs.lnl calls an external FORTRAN subroutine, if it gets a subscript out of bounds, it will cause R to terminate. So make sure to save your workspace frequently if you use this function in its current implementation.

References

Pledger, S., K. H. Pollock, et al. (2003). Open capture-recapture models with heterogeneity: I. Cormack-Jolly-Seber model. Biometrics 59(4):786-794.