OptimalDesign (version 0.2)

od.MISOCP: Efficient exact design using mixed integer second-order cone programming

Description

Computes an efficient exact experimental design under general linear constraints using the approach of mixed integer second-order cone programming.

Usage

od.MISOCP(F, b, A=NULL, w0=NULL, crit="D", R=NULL, kappa=1e-9, 
            tab=NULL, graph=NULL, t.max=120)

Arguments

F

The n times m matrix of real numbers. The rows of F represent the m-dimensional regressors corresponding to n design points. It is assumed that n>=m>=2. Use od.m1 for models with 1-dimensional regressors. For D-optimality, the current implementation supports the models with m<=10.

b, A

The real vector of length k and the k times n matrix of reals numbers. The linear constraints A%*%w<=b, w0<=w define the set of permissible designs w (where w0 is a described below.) The argument A can also be NULL; in that case b must be a non-negative number and A is set to the 1 times n matrix of ones.

w0

The non-negative vector of length n representing the design to be augmented. This argument can also be NULL; in that case, w0 is set to the vector of zeros.

crit

The optimality criterion. Possible values are "D", "A", "IV".

R

The region of summation for the IV-optimality criterion. The argument R must be a subvector of 1:n, or NULL. If R=NULL, the procedure uses R=1:n. Argument R is ignored if crit="D", or if crit="A".

kappa

A small non-negative perturbation parameter.

tab

A vector determining the regressor components to be printed with the resulting design. This argument should be a subvector of 1:n, or a subvector of colnames(F), or it can be NULL. If tab=NULL, the design is not printed.

graph

A vector determining the regressor components to be plotted with the resulting design. This argument should be a subvector of 1:n, or a subvector of colnames(F), or it can be NULL. If graph=NULL, the resulting design is not visualized.

t.max

The time limit for the computation.

Value

A list with the following components:

method

The method used for computing the design w.best.

w.best

the best permissible design found, or NULL. The value of w.best will be NULL if the computation fails. This can happen, if no permissible solution is found within the time limit, no permissible solution exists, or the problem is unbounded; see the status variable for more details. Note that even if w.best is a permissible design, then it still can have a singular information matrix; cf. the Phi.best variable.

Phi.best

The value of the criterion of optimality of the design w.best. If w.best has a singular information matrix or if the computation fails, the value of Phi.best will be 0.

status

The status variable of the gurobi optimization procedure; see the gurobi solver documentation for details.

t.act

The actual time taken by the computation.

Details

The procedure computes an efficient exact design by converting the optimal design problem to a specific problem of mixed integer second-order cone programming; see the reference for details. The advantage of this approach is the possibility to construct exact designs under a general system of linear constraints.

The model should be non-singular in the sense that there exists an exact design w satisfying the constraints 0<=w0<=w and A%*%w<=b, with a non-singular information matrix, preferably with the reciprocal condition number of at least 1e-5. If this requirement is not satisfied, the computation may fail, or it may produce a deficient design.

If the criterion of IV-optimality is selected, the region R should be chosen such that the associated matrix L (see the help page of the function od.crit) is non-singular, preferably with a reciprocal condition number of at least 1e-5. If this requirement is not satisfied, the computation may fail, or it may produce a deficient design.

The perturbation parameter kappa can be used to add n*m iid random numbers from the uniform distribution in [-kappa,kappa] to the elements of F before the optimization is executed. This can be helpful for increasing the numerical stability of the computation or for generating a random design from the potentially large set of optimal or nearly-optimal designs.

The performance strongly depends on the problem and on the hardware used, but in most cases the function can compute an optimal or nearly-optimal exact design for a problem with a hundred design points within minutes of computing time. We advise the user to verify the quality of the resulting design by comparing it to the result of an alternative method (such as od.IQP and od.RC) and/or by computing its efficiency relative to the corresponding optimal approximate design (computed using od.SOCP). In the special case of the single constraint on the size, it is generally more efficient to use the function od.KL, or the function od.RCs.

References

Sagnol G, Harman R (2015): Computing exact D-optimal designs by mixed integer second-order cone programming. The Annals of Statistics, Volume 43, Number 5, pp. 2198-2224.

See Also

od.IQP, od.RC, od.SOCP, od.KL, od.RCs

Examples

# NOT RUN {
if(require("gurobi")){
# Consider a dose-response study where both efficacy and toxicity are 
# observed as 0/1 outcomes for each patient, where the probability of 
# the outcome 1 under the dose x is modeled by the logistic function: 
# exp(ae+be*x)/(1+exp(ae+be*x)) for the efficacy, and 
# exp(at+bt*x)/(1+exp(at+bt*x)) for the toxicity. We can choose the 
# doses x in the range 1 mg to 150 mgs. 
# The aim is to estimate the parameters ae,be using the D-optimal
# design, or using the A-optimal design. 

# Because this is a non-linear model, the optimal designs will depend 
# on the unknown values of the parameters. We will use the approach of 
# local optimality with the following nominal values of the parameters: 
tle <- c(-10, 0.2)
tlt <- c(-20, 0.2)

# It is simple to show that the localized information matrix for 
# (ae,be) is the information matrix of the standard model with 
# the following regressors:
F.logistic <- matrix(0, nrow=150, ncol=2)
for (i in 1:150) 
  F.logistic[i, ] <- 
        c(sqrt(exp(tle[1]+tle[2]*i))/(1+exp(tle[1]+tle[2]*i)), 
        i*sqrt(exp(tle[1]+tle[2]*i))/(1+exp(tle[1]+tle[2]*i)))
                       
# The constraints on the experiment are twofold: We can have at most 
# N=100 subjects and we also require that the expected number of 
# "failed" trials is at most 10. A trial is considered to be a failure 
# if it leads to either a toxic response, or if it is not efficacious. 
# These constraints can be expressed as A*w<=b:
efficacy.prob <- function(x) 
    exp(tle[1]+tle[2]*x)/(1+exp(tle[1]+tle[2]*x))
toxicity.prob <- function(x) 
    exp(tlt[1]+tlt[2]*x)/(1+exp(tlt[1]+tlt[2]*x))
failure.prob <- function(x) 
    1 - (1 - toxicity.prob(x)) * efficacy.prob(x)
b <- c(100, 10); A <- rbind(rep(1,150), failure.prob(1:150))

# Now we can compute the designs:
res.D <- od.MISOCP(F.logistic, b, A, crit="D")
res.A <- od.MISOCP(F.logistic, b, A, crit="A")

# Let us verify the quality of the designs by computing their efficiency 
# relative to the approximate optimal designs:
res.D$Phi.best / od.SOCP(F.logistic, b, A, crit="D")$Phi.best
res.A$Phi.best / od.SOCP(F.logistic, b, A, crit="A")$Phi.best

# We can plot the failure probability curve (red), the toxicity 
# probability curve (black), the efficacy probability curve (green), 
# the D-optimal design (orange) and the A-optimal design (blue):
plot(failure.prob(1:150), type="l", 
     ylab="probability / proportion of subjects", 
     lwd=3, col="red")
lines(toxicity.prob(1:150), col="black")
lines(efficacy.prob(1:150), col="green")
lines(res.D$w.best/100, type="h", col="orange")
lines(res.D$w.best/100, type="h", col="orange", lwd=2)
lines(res.A$w.best/100, type="h", col="blue", lwd=2)

# Note that both designs perform the observations generally at two 
# different levels, one of which is a dose which leads to a 50-percent 
# failure of efficacy. Based on these designs, none of the patients 
# are put on dangerously high levels of doses.
}
# }