robust: Robust D-Optimal Designs

Description

Finds Robust designs or optimal in-average designs for nonlinear models. It is useful when a set of different vectors of initial estimates along with a discrete probability measure are available for the unknown model parameters. It is a discrete version of bayes.

Usage

robust(formula, predvars, parvars, family = gaussian(), lx, ux, iter, k,
  prob, parset, fimfunc = NULL, ICA.control = list(),
  sens.minimax.control = list(), initial = NULL,
  npar = dim(parset)[2], plot_3d = c("lattice", "rgl"), x = NULL)

Arguments

formula

A nonlinear model formula. A symbolic description of the model consists of predictors and the unknown model parameters. Will be coerced to a formula if necessary.

predvars

A vector of characters. Denotes the predictors in the formula.

parvars

A vector of characters. Denotes the unknown parameters in the formula.

family

A description of the response distribution and the link function to be used in the model. This can be a family function, a call to a family function or a character string naming the family. Every family function has a link argument allowing to specify the link function to be applied on the response variable. If not specified, default links are used. For details see family. By default, a linear gaussian model gaussian() is applied.

Vector of lower bounds for the predictors. Should be in the same order as predvars.

Vector of upper bounds for the predictors. Should be in the same order as predvars.

iter

Maximum number of iterations.

Number of design points. Must be at least equal to the number of model parameters to avoid singularity of the FIM.

prob

A vector of the probability measure $\pi$ associated with each row of parset.

parset

A matrix that provides the vector of initial estimates for the model parameters, i.e. support of $\pi$. Every row is one vector (nrow(parset) == length(prob)). See 'Details'.

fimfunc

A function. Returns the FIM as a matrix. Required when formula is missing. See 'Details' of minimax.

ICA.control

ICA control parameters. For details, see ICA.control.

sens.minimax.control

Control parameters to verify the general equivalence theorem. For details, see the function sens.minimax.control.

initial

A matrix of the initial design points and weights that will be inserted into the initial solutions (countries) of the algorithm. Every row is a design, i.e. a concatenation of x and w. Will be coerced to a matrix if necessary. See 'Details' of minimax.

npar

Number of model parameters. Used when fimfunc is given instead of formula to specify the number of model parameters. If not given, the sensitivity plot may be shifted below the y-axis. When NULL, it is set to dim(parset)[2].

plot_3d

Which package should be used to plot the sensitivity (derivative) function for two-dimensional design space. Defaults to "lattice".

A vector of candidate design (support) points. When is not set to NULL (default), the algorithm only finds the optimal weights for the candidate points in x. Should be set when the user has a finite number of candidate design points and the purpose is to find the optimal weight for each of them (when zero, they will be excluded from the design). For design points with more than one dimension, see 'Details' of sensminimax.

Value

an object of class minimax that is a list including three sub-lists:

arg

A list of design and algorithm parameters.

evol

A list of length equal to the number of iterations that stores the information about the best design (design with least criterion value) of each iteration. evol[[iter]] contains:

`iter`		Iteration number.
`x`		Design points.
`w`		Design weights.
`min_cost`		Value of the criterion for the best imperialist (design).
`mean_cost`		Mean of the criterion values of all the imperialists.
`sens`		An object of class `'sensminimax'`. See below.
`param`		Vector of parameters.

empires

A list of all the empires of the last iteration.

alg

A list with following information:

`nfeval`		Number of function evaluations. It does not count the function evaluations from checking the general equivalence theorem.
`nlocal`		Number of successful local searches.
`nrevol`		Number of successful revolutions.
`nimprove`		Number of successful movements toward the imperialists in the assimilation step.
`convergence`		Stopped by `'maxiter'` or `'equivalence'`?

The list sens contains information about the design verification by the general equivalence theorem. See sensminimax for more details. It is given every ICA.control$checkfreq iterations and also the last iteration if ICA.control$checkfreq >= 0. Otherwise, NULL.

param is a vector of parameters that is the global minimum of the minimax criterion or the global maximum of the standardized maximin criterion over the parameter space, given the current x, w.

Details

Let $\Theta$ be a set of initial estimates for the unknown parameters. A robust criterion is evaluated at the elements of $\Theta$ weighted by a probability measure $\pi$ as follows: $$B(\xi, \pi) = \int_{\Theta}|M(\xi, \theta)|\pi(\theta) d\theta.$$ A robust design $\xi^*$ maximizes $B(\xi, \pi)$ over the space of all designs.

When the model is given via formula, columns of parset must match the parameters introduced in parvars. Otherwise, when the model is introduced via fimfunc, columns of parset must match the argument param in fimfunc.

To verify the optimality of the output design by the general equivalence theorem, the user can either plot the results or set checkfreq in ICA.control to Inf. In either way, the function sensrobust is called for verification. One can also adjust the tuning parameters in ICA.control to set a stopping rule based on the general equivalence theorem. See 'Examples' below.

Examples

Run this code

# NOT RUN {
# Finding a robust design for the two-parameter logistic model
# See how we set a stopping rule.
# The ELB is computed every checkfreq = 30 iterations
# The optimization stops when the ELB is larger than stoptol = .95
res1 <- robust(formula = ~1/(1 + exp(-b *(x - a))),
              predvars = c("x"), parvars = c("a", "b"),
              family = binomial(),
              lx = -5, ux = 5, prob = rep(1/4, 4),
              parset = matrix(c(0.5, 1.5, 0.5, 1.5, 4.0, 4.0, 5.0, 5.0), 4, 2),
              iter = 1, k =3,
              ICA.control = list(stop_rule = "equivalence",
                                 stoptol = .95, checkfreq = 30))

# }
# NOT RUN {
res1 <- iterate(res1, 100)
# stops at iteration 51
# }
# NOT RUN {

# }
# NOT RUN {
res1.1 <- robust(formula = ~1/(1 + exp(-b *(x - a))),
              predvars = c("x"), parvars = c("a", "b"),
              family = binomial(),
              lx = -5, ux = 5, prob = rep(1/4, 4),
              parset = matrix(c(0.5, 1.5, 0.5, 1.5, 4.0, 4.0, 5.0, 5.0), 4, 2),
              x = c(-3, 0, 3),
              iter = 150, k =3)
plot(res1.1)
# not optimal
# }