aceglm {acebayes}

Approximate Coordinate Exchange (ACE) Algorithm for Generalised Linear Models

Description

Functions implementing the approximate coordinate exchange (ACE) algorithm (Overstall & Woods, 2017) for finding Bayesian optimal experimental designs for generalised linear models (GLMs).

Usage

aceglm(formula, start.d, family, prior, B, 
criterion = c("D", "A", "E", "SIG", "NSEL", "SIG-Norm", "NSEL-Norm"),
method = c("quadrature", "MC"), Q = 20, N1 = 20, N2 = 100, lower = -1, 
upper = 1, progress = FALSE, limits = NULL)

paceglm(formula, start.d, family, prior, B, 
criterion = c("D", "A", "E", "SIG", "NSEL", "SIG-Norm", "NSEL-Norm"),
method = c("quadrature", "MC"), Q = 20, N1 = 20, N2 = 100, lower = -1, 
upper = 1, limits = NULL, mc.cores = 1, n.assess = 20)

Arguments

formula

An object of class "formula": a symbolic description of the model. The terms should correspond to the column names of the argument start.d.

start.d

For aceglm, an n by k matrix, with column names used by the argument formula, specifying the initial design for the ACE algorithm.

For paceglm, a list with each element being an n by k matrix, with column names used by the argument formula, specifying the initial design for each repetition of the ACE algorithm.

family

A description of the error distribution and link function to be used in the model. This can be a character string naming a family function, a family function or the result of a call to a family function. (See family for details of family functions.)

prior

An argument specifying the prior distribution.

For method = "MC", a function with one argument: B; a scalar integer. This function should return a B by p matrix, with p the number of model parameters, containing a random sample from the prior distribution of the parameters. The value of p should correspond to the number of terms specified by the formula argument.

For method = "quadrature", a list specifying a normal or uniform prior for the model parameters. For a normal prior distribution, the list should have named entries mu and sigma2 specifying the prior mean and variance-covariance matrix. The prior mean may be specified as a scalar, which will then be replicated to form an vector of common prior means, or a vector of length p. The prior variance-covariance matrix may be specified as either a scalar common variance or a vector of length p of variances (for independent prior distributions) or as a p by p matrix. For a uniform prior distribution, the list should have a named entry support, a 2 by p matrix with each column giving the lower and upper limits of the support of the independent continuous uniform distribution for the corresponding parameter.

B

An optional argument for controlling the approximation to the expected utility. It should be a vector of length two.

For method = "MC", it specifies the size of the Monte Carlo samples, generated from the joint distribution of unknown quantities. The first sample size, B[1], gives the sample size to use in the comparison procedures, and the second sample size, B[2], gives the sample size to use for the evaluations of Monte Carlo integration that are used to fit the Gaussian process emulator. If left unspecified, the default value is c(20000,1000).

For method = "quadrature", it specifies the tuning parameters (numbers of radial abscissas and random rotations) for the implemented quadrature method; see Details for more information. If left unspecified, the default value is c(2, 8).

criterion

An optional character argument specifying the utility function. There are currently seven utility functions implemented as follows:

1. pseudo-Bayesian D-optimality (criterion = "D");

2. pseudo-Bayesian A-optimality (criterion = "A");

3. pseudo-Bayesian E-optimality (criterion = "E").

4. Shannon information gain with Monte Carlo (MC) approximation to marginal likelihood (criterion = "SIG");

5. Shannon information gain with normal-based Laplace approximation to marginal likelihood (criterion = "SIG-Norm");

6. negative squared error loss with importance sampling approximation to posterior mean (criterion = "NSEL");

7. negative squared error loss with normal-based approximation to posterior mean (criterion = "NSEL-Norm");

If left unspecified, the default is "D" denoting pseudo-Bayesian D-optimality. See Details for more information.

method

An optional character argument specifying the method of approximating the expected utility function. Current choices are method = "quadrature" for a deterministic quadrature approximation and method = "MC" for a stochastic Monte Carlo approximation. The first of these choices is only available when the argument criterion = "A", "D" or "E". The second choice is available for all possible values of the argument criterion. If left unspecified, the argument defaults to "quadrature" for criterion = "A", "D" or "E" and to "MC" otherwise. See Details for more information.

Q

An integer specifying the number of evaluations of the approximate expected utility that are used to fit the Gaussian process emulator. The default value is 20.

N1

An integer specifying the number of iterations of Phase I of the ACE algorithm (the coordinate exchange phase). The default value is 20.

N2

An integer specifying the number of iterations of Phase II of the ACE algorithm (the point exchange phase). The default value is 100.

lower

An argument specifying the design space. This argument can either be a scalar or a matrix of the same dimension as the argument start.d which specifies the lower limits of all coordinates of the design space. The default value is -1.

upper

An argument specifying the design space. This argument can either be a scalar or a matrix of the same dimension as the argument start.d which specifies the upper limits of all coordinates of the design space. The default value is 1.

progress

A logical argument indicating whether the iteration number and other information detailing the progress of the algorithm should be printed. The default value is FALSE.

limits

An argument specifying the grid over which to maximise the Gaussian process emulator for the expected utility function. It should be a function with three arguments: i, j and d which generates a one-dimensional grid for the ijth coordinate of the design when the current design is d. The default value is NULL which generates values uniformly on the interval (lower[i,j],upper[i,j]) or (lower,upper) depending on whether the arguments lower and upper are matrices or scalars, respectively.

mc.cores

The number of cores to use, i.e. at most how many child processes will be run simultaneously. Must be at least one (the default), and parallelisation requires at least two cores. See mclapply for more information and warnings for mc.cores > 1.

n.assess

If method = "MC", the approximate expected utility for the design from each repetition of the ACE algorithm will be calculated n.assess times. The terminal design returned will be the design with the largest mean approximate expected utility calculated over the n.assess approximations.

Details

The aceglm function implements the ACE algorithm to find designs for the class of generalised linear models (GLMs) for certain cases of utility function meaning the user does not have to write their own utility function.

Two utility functions are implemented.

1. Shannon information gain (SIG)

   The utility function is

   u^{SIG}(d) = \pi(\theta|y,d) - \pi(\theta),

   where \pi(\theta|y,d) and \pi(\theta) denote the posterior and prior densities of the parameters \theta, respectively.

2. Negative squared error loss (NSEL)

   The utility function is

   u^{NSEL}(d) = - \left(\theta - E(\theta |y,d)\right)^T \left(\theta - E(\theta |y,d)\right),

   where E(\theta | y,d) denotes the posterior mean of \theta.

In both cases the utility function is not available in closed form due to the analytical intractability of either the posterior distribution (for SIG) or the posterior mean (for NSEL). The acebayes package implements two approximations to both utility functions. If criterion = "SIG" or criterion = "NSEL" then sampling-based Monte Carlo or importance sampling approximations will be employed. This was the original approach used by Overstall & Woods (2017). If criterion = "SIG-Norm" or criterion = "NSEL-Norm" then approximations based on approximate normality of the posterior (Overstall et al., 2017) will be used.

The normal approximation to the posterior can be taken further leading to the approximation by some scalar function of the Fisher information matrix, \mathcal{I} (\theta;d), which only depends on \theta (Chaloner & Verdinelli, 1995). In the case of SIG, the approximate utility is given by

u^{D}(d) = \log \vert \mathcal{I} (\theta;d) \vert,

and the resulting design is typically called pseudo-Bayesian D-optimal. For NSEL, the approximate utility is given by

u^A(d) = - \mathrm{tr} \left\{ \mathcal{I} (\theta;d)^{-1} \right\}

with the resulting design termed pseudo-Bayesian A-optimal. These designs are often used under the frequentist approach to optimal experimental design and so to complete the usual set, the following utility for finding a pseudo-Bayesian E-optimal design is also implemented:

u^E(d) = \mathrm{min} \mbox{ } e\left(\mathcal{I} (\theta;d) \right),

where e() denotes the function that calculates the eigenvalues of its argument.

The expected utilities can be approximated using Monte Carlo methods (method = "MC" for all criteria) or using a deterministic quadrature method (method = "quadrature", implemented for the D, A and E criteria). The former approach approximates the expected utility via sampling from the prior. The latter approach uses a radial-spherical integration rule (Monahan and Genz, 1997) and B[1] specifies the number, n_r, of radial abscissas and B[2] specifies the number, n_q, of random rotations. Larger values of n_r will produce more accurate, but also more computationally expensive, approximations. See Gotwalt et al. (2009) for further details.

Note that the utility functions for SIG and NSEL are currently only implemented for logistic regression, i.e. family = binomial, or Poisson regression, i.e. family = poisson(link="log"), whereas the utility functions for pseudo-Bayesian designs are implemented for generic GLM families.

Similar to all coordinate exchange algorithms, ACE should be repeated from different initial designs. The function paceglm will implement this where the initial designs are given by a list via the argument start.d. On the completion of the repetitions of ACE, paceglm will approximate the expected utility for all final designs and return the design (the terminal design) with the largest approximate expected utility.

For more details on the ACE algorithm, see Overstall & Woods (2017).

Value

The function will return an object of class "ace" (for aceglm) or "pace" (for paceglm) which is a list with the following components:

utility

The utility function resulting from the choice of arguments.

start.d

The argument start.d.

phase1.d

The design found from Phase I of the ACE algorithm.

phase2.d

The design found from Phase II of the ACE algorithm.

phase1.trace

A vector containing the evaluations of the approximate expected utility of the current design at each stage of Phase I of the ACE algorithm. This can be used to assess convergence.

phase2.trace

A vector containing the evaluations of the approximate expected utility of the current design at each stage of Phase II of the ACE algorithm. This can be used to assess convergence.

B

The argument B.

Q

The argument Q.

N1

The argument N1.

N2

The argument N2.

glm

If the object is a result of a direct call to aceglm then this is TRUE.

nlm

This will be FALSE.

criterion

If the object is a result of a direct call to aceglm then this is the argument criterion.

method

If the object is a result of a direct call to aceglm then this is the argument method.

prior

If the object is a result of a direct call to aceglm then this is the argument prior.

family

If the object is a result of a direct call to aceglm then this is the argument family.

formula

If the object is a result of a direct call to acenlm then this is the argument formula.

time

Computational time (in seconds) to run the ACE algorithm.

binary

The argument binary. Will be FALSE for the utility functions currently implemented.

d

The terminal design (paceglm only).

eval

If deterministic = "MC", a vector containing n.assess approximations to the expected utility for the terminal design (paceglm only).

If deterministic = "quadrature", a scalar giving the approximate expected utility for the terminal design (paceglm only).

final.d

A list of the same length as the argument start.d, where each element is the final design (i.e. phase2.d) for each repetition of the ACE algorithm (paceglm only).

besti

A scalar indicating which repetition of the ACE algorithm resulted in the terminal design (paceglm only).

Note

These are wrapper functions for ace and pace.

Author(s)

Antony M. Overstall [e-mail], David C. Woods, Maria Adamou & Damianos Michaelides

References

Chaloner, K. & Verdinelli, I. (1995). Bayesian experimental design: a review. Statistical Science, 10, 273-304.

Gotwalt, C. M., Jones, B. A. & Steinberg, D. M. (2009). Fast computation of designs robust to parameter uncertainty for nonlinear settings. Technometrics, 51, 88-95.

Meyer, R. & Nachtsheim, C. (1995). The coordinate exchange algorithm for constructing exact optimal experimental designs. Technometrics, 37, 60-69.

Monahan, J. and Genz, A. (1997). Spherical-radial integration rules for Bayesian computation,” Journal of the American Statistical Association, 92, 664–674.

Overstall, A.M. & Woods, D.C. (2017). Bayesian design of experiments using approximate coordinate exchange. Technometrics, 59, 458-470.

Overstall, A.M., McGree, J.M. & Drovandi, C.C. (2018). An approach for finding fully Bayesian optimal designs using normal-based approximations to loss functions. Statistics and Computing, 28(2), 343-358.

See Also

ace, acenlm, pace, pacenlm.

Examples

## This example uses aceglm to find a Bayesian D-optimal design for a 
## first order logistic regression model with 6 runs 4 factors. The priors are 
## those used by Overstall & Woods (2017), with each of the five
## parameters having a uniform prior. The design space for each coordinate is [-1, 1].

set.seed(1)
## Set seed for reproducibility.

n<-6
## Specify the sample size (number of runs).

start.d<-matrix(2 * randomLHS(n = n,k = 4) - 1,nrow = n,ncol = 4,
dimnames = list(as.character(1:n), c("x1", "x2", "x3", "x4")))
## Generate an initial design of appropriate dimension. The initial design is a 
## Latin hypercube sample.

low<-c(-3, 4, 5, -6, -2.5)
upp<-c(3, 10, 11, 0, 3.5)
## Lower and upper limits of the uniform prior distributions.

prior<-function(B){
t(t(6*matrix(runif(n = 5 * B),ncol = 5)) + low)}
## Create a function which specifies the prior. This function will return a 
## B by 5 matrix where each row gives a value generated from the prior 
## distribution for the model parameters.

example1<-aceglm(formula=~x1+x2+x3+x4, start.d = start.d, family = binomial, 
prior = prior, method = "MC", N1 = 1, N2 = 0, B = c(1000, 1000))
## Call the aceglm function which implements the ACE algorithm requesting 
## only one iteration of Phase I and zero iterations of Phase II. The Monte
## Carlo sample size for the comparison procedure (B[1]) is set to 100.

example1
## Print out a short summary.

#Generalised Linear Model 
#Criterion = Bayesian D-optimality 
#Formula: ~x1 + x2 + x3 + x4
#
#Family: binomial 
#Link function: logit 
#
#Method:  MC 
#
#B:  1000 1000 
#
#Number of runs = 6
#
#Number of factors = 4
#
#Number of Phase I iterations = 1
#
#Number of Phase II iterations = 0
#
#Computer time = 00:00:01

example1$phase2.d
## Look at the final design.

#          x1          x2          x3         x4
#1 -0.4735783  0.12870470 -0.75064318  1.0000000
#2 -0.7546841  0.78864527  0.58689270  0.2946728
#3 -0.7463834  0.33548985 -0.93497463 -0.9573198
#4  0.4446617 -0.29735212  0.74040030  0.2182800
#5  0.8459424 -0.41734194 -0.07235575 -0.4823212
#6  0.6731941  0.05742842  1.00000000 -0.1742566

prior2 <- list(support = rbind(low, upp))
## A list specifying the parameters of the uniform prior distribution

example2<-aceglm(formula = ~ x1 +x2 + x3 + x4, start.d = start.d, family = binomial, 
prior = prior2, N1 = 1, N2 = 0)
## Call the aceglm function with the default method of "quadrature"

example2$phase2.d
## Final design

#          x1          x2          x3         x4
#1 -0.4647271  0.07880018 -0.94648750  1.0000000
#2 -0.7102715  0.79827332  0.59848578  0.5564422
#3 -0.7645090  0.39778176 -0.74342036 -1.0000000
#4  0.4514632 -0.33687477  0.55066110  0.3994593
#5  0.7913559 -0.41856994  0.01321035 -0.8848135
#6  0.6337306  0.11578522  1.00000000  1.0000000

mean(example1$utility(d = example1$phase2.d, B = 20000))
#[1] -11.61105
mean(example2$utility(d = example2$phase2.d, B = 20000))
#[1] -11.19737
## Compare the two designs using the Monte Carlo approximation