R: Auto-constructing Frechet derivative of D-criterion based on...
cfderiv
R Documentation
Auto-constructing Frechet derivative of D-criterion based on general equivalence theorem
Description
Auto-constructs Frechet derivative of D-criterion at M(ξ, β) and in direction M(ξ_x, β) where M is Fisher information matrix, β is vector of parameters, ξ is the interested design and ξ_x is a unique design which has only a point x. The constructed Frechet derivative is an R function with argument x.
Usage
cfderiv(ymean, yvar, param, points, weights)
Arguments
ymean
a character string,
formula of E(y) with specific satndard: characters b1, b2, b3, ... symbolize model parameters and x1, x2, x3, ... symbolize explanatory variables. See 'Examples'.
yvar
a character string, formula of Var(y) with specific standard as ymean. See 'Details' and 'Examples'.
param
a vector of values of parameters which must correspond to b1, b2, b3, ... in ymean.
points
a vector of points which belong to design ξ . See 'Details'.
weights
a vector of ξ points weights. The sum of weights should be 1; otherwise they will be normalized.
Details
If response variables have the same constant variance, for example σ^2, then yvar must be 1.
Consider design ξ with nm-dimensional points. Then, the vector of ξ points is
(x_1, x_2, …, x_i, …, x_n),
where x_i = (x_{i1}, x_{i2}, …, x_{im}). Hence the length of vector points is mn.
Value
fderiv
a function in which its argument is a vector x, an m-dimentional design point, and its output is the value of Frechet derivative at M(ξ, β) and in direction M(ξ_x, β).
Note
A design ξ is D-optimal if and only if Frechet derivative at M(ξ, β) and in direction M(ξ_x, β)is greater than or equal to 0 on the design space. The equality must be achieved just at ξ points. Here, x is an arbitrary point on design space.
This function is applicable for models that can be written as E(Y_i) = f(x_i,β)
where y_i is the ith response variable, x_i is the observation vector of the ith explanatory variables, β is the vector of parameters and f is a continuous and differentiable function with respect to β.
In addition, response variables must be independent with distributions that belong to the Natural exponential family. Logistic,Poisson, Negative Binomial, Exponential, Richards, Weibull, Log-linear, Inverse Quadratic and Michaelis-Menten are examples of these models.
Author(s)
Ehsan Masoudi, Majid Sarmad and Hooshang Talebi
References
Masoudi, E., Sarmad, M. and Talebi, H. 2012, An Almost General Code in R to Find Optimal Design, In Proceedings of the 1st ISM International Statistical Conference 2012, 292-297.
Kiefer, J. C. 1974, General equivalence theory for optimum designs (approximate theory), Ann. Statist., 2, 849-879.7.