Auxiliary function as user interface for fine-tuning RPM model fitting algorithm, which computes the MLPLE of the Revealed Preferences Model via optimization.
control.rpm(
init_theta = NULL,
algorithm = "NLOPT_LD_SLSQP",
print_level = 0,
xtol_rel = 1e-08,
ftol_rel = 1e-08,
ftol_abs = 1e-06,
lower.bound = -10,
upper.bound = 10,
maxeval = 2000,
bs.maxeval = 2000,
bs.xtol_rel = 1e-08,
bs.save.data = FALSE,
check_derivatives = FALSE,
bootstrap = TRUE,
hessian = FALSE,
seed = NULL,
parallel.type = "PSOCK",
parallel.ncores = 1,
ncores = 1,
constraints = c("none", "M_single"),
logodds_single = FALSE,
save.data = TRUE,
robust.cov = FALSE,
local_opts = list(algorithm = "NLOPT_LD_SLSQP", xtol_rel = 1e-07, maxeval = maxeval),
nbootstrap = 50,
nbootstrap.SD = 20,
large.population.bootstrap = 5000,
alpha = 0.05
)A list with arguments as components.
vector; numeric vector of starting parameter values. This value and other possible starting values are applied to find a good optimizer. This can either have length the number of parameters corresponding to the terms in the formula or in addition the equilibrium constraints.
string; The optimization algorithm to use. See nloptr::nloptr.print.options()
and the NLopt website for a description of the algorithms.
integer; possible values: 0, 1, 2, or 3. This controls how much output is shown during the optimization process. Possible values: 0 (default): no output; 1: show iteration number and value of objective function; 2: 1 + show value of equalities/constraints; 3: 2 + show value of controls.
scalar; Stop when an optimization step (or an estimate of the optimum) changes every parameter by less than xtol_rel multiplied by the absolute value of the parameter. If there is any chance that an optimal parameter is close to zero, you might want to set an absolute tolerance with xtol_abs as well. Criterion is disabled if xtol_rel is non-positive. Possible values: xtol_rel > 0. Default value: 1.0e-08.
scalar; Stop when an optimization step (or an estimate of the optimum) changes the log-likelihood by less than ftol_rel multiplied by the absolute value of the log-likelihood.
scalar; Stop when an optimization step (or an estimate of the optimum) changes the log-likelihood by less than ftol_abs. tolerance with xtol_abs as well. Criterion is disabled if ftol_abs is non-positive. Possible values: ftol_abs > 0. Default value: 1.0e-06.
numeric; lower bounds on the parameter estimates (that is, the beta and gamma parameters in the model). Can be a vector of the same size as the coefficient vector or a single number which is used for all bounds.
numeric; upper bounds on the parameter estimates (that is, the beta and gamma parameters in the model). Can be a vector of the same size as the coefficient vector or a single number which is used for all bounds.
integer; Stop when the number of function evaluations exceeds maxeval. This is not a strict maximum: the number of function evaluations may exceed maxeval slightly, depending upon the algorithm. Criterion is disabled if maxeval is non-positive. Default value: 1000.
integer; Stop the bootstrap optimization when the number of function evaluations exceeds bs.maxeval. This is not a strict maximum: the number of function evaluations may exceed bs.maxeval slightly, depending upon the algorithm. Criterion is disabled if bs.maxeval is non-positive. Default value:50
scalar; Stop the bootstrap optimization when an optimization step (or an estimate of the optimum) changes every parameter by less than bs.xtol_rel multiplied by the absolute value of the parameter. See the parameter xtol_rel for details.
logical; Should the bootstrapped data be saved in the bootstrap return list (as components
Xdata and Zdata).
logical; Compare the user-supplied analytic gradients with the finite difference approximations.
logical; If `TRUE` use a bootstrap to compute the standard errors and associated
covariance matrices. If `FALSE` base the standard errors and associated
covariance matrices on the Hessian of the
(constrained) log-likelihood.
In all cases the extended covariance matrix is returned in ext.covar.hessian.
This is the matrix of parameters, log-odds of being single and the Lagrange multipliers.
logical; Depreciated. The negation of the `bootstrap` argument.
Seed value (integer) for the random number generator. See
set.seed
The type of cluster to run. The typical choices are "MPI" and "PSOCK", where you must have "MPI" installed to use the former. The default is "PSOCK".
count; Depreciated. The renamed `ncores` argument.
Number of processors to use in the bootstrap computations. The default is 1, that is no parallel processing.
string; Additional constraints to force the proportions of singles to match the (weighted) population estimates? This should not be required, but does stabilize the estimates in cases where there is much uncertainty. The possible values are "none" and "M_single" (the numbers of male singles of each type are reproduced). Note that adding constraints leads to over-constrained optimization which may fail.
logical; Should the log-odds ratio of being single relative to a randomly chosen person of the same sex from the the population be returned. If FALSE the log-odds of being single relative is returned. This is a pure preference parameter.
logical; Should the data be saved in the return list (as components
Xdata and Zdata).
logical; Should the covariance matrix of the estimates be computed using a robust method (MASS::cov.mcd)? Only use if the bootstrap is unstable.
list; list of options for nloptr sub-algorithm. See the nloptr package, but these are rarely changed.
integer; Number of bootstrap resamples to take in the estimation of the covariance matrix of the parameter estimates.
integer; Number of bootstrap resamples to take in the estimation of the variances used in the studentized bootstrap. This is run for each nbootstrap sample and so is expensive.
integer; If the population size exceeds large.population.bootstrap then
the large population approximation is used to simulate the matchings in the bootstrap. Otherwise the
small population simulation is used (including the Gale-Shapley algorithm).
The small population method is more accurate in smaller populations, with the default cutoff being 5000 people.
proportion; Type I error rate for the confidence intervals produced by the bootstrap.
This function is only used within a call to the rpm
function.
Some of the arguments are not yet fully implemented. It will evolve slower to incorporate more arguments as the package develops.
rpm