Last chance! 50% off unlimited learning
Sale ends in
Locate extrema of fungible regression weights.
fungibleExtrema(
R.X,
rxy,
r.yhata.yhatb,
Nstarts = 100,
MaxMin = "Max",
Seed = NULL,
maxGrad = 1e-05,
PrintLevel = 1
)
cosine between OLS and alternate weights.
extrema of fungible weights.
k weights.
z weights: a normalized random vector.
OLS weights.
p x 1 vector of u weights.
Correlation between yhata and yhatb.
Correlation between y and yhatb.
Gradient of converged solution.
p x p Predictor variable correlation matrix.
p x 1 Vector of predictor-criterion correlations.
Correlation between least squares (yhatb) and alternate-weight (yhata) composites.
Maximum number of (max) minimizations from random starting configurations.
Character: "Max" = maximize cos(a,b); "Min" = minimize cos(a,b).
Starting seed for the random number generator. If Seed = NULL then the program will sample a random integer in the (0, 100,000) interval. Default (Seed = NULL).
The optimization routine will end when the maximimum of the (absolute value of the ) function gradient falls below the value specified in maxGrad. Default (maxGrad = 1E-05).
(integer). If PrintLevel = 1 then the program will print additional output during function convergence. Default (PrintLevel = 1).
Niels Waller and Jeff Jones
Koopman, R. F. (1988). On the sensitivity of a composite to its weights. Psychometrika, 53(4), 547--552.
Waller, N. & Jones, J. (2009). Locating the extrema of fungible regression weights in multiple regression. Psychometrika, 74, 589--602.
if (FALSE) {
## Example
## This is Koopman's Table 2 Example
R.X <- matrix(c(1.00, .69, .49, .39,
.69, 1.00, .38, .19,
.49, .38, 1.00, .27,
.39, .19, .27, 1.00),4,4)
b <- c(.39, .22, .02, .43)
rxy <- R.X %*% b
OLSRSQ <- t(b) %*% R.X %*% b
theta <- .02
r.yhata.yhatb <- sqrt( 1 - (theta)/OLSRSQ)
Converged = FALSE
SEED = 1234
MaxTries = 100
iter = 1
while( iter <= MaxTries){
SEED <- SEED + 1
cat("\nCurrent Seed = ", SEED, "\n")
output <- fungibleExtrema(R.X, rxy,
r.yhata.yhatb,
Nstarts = 5,
MaxMin = "Min",
Seed = SEED,
maxGrad = 1E-05,
PrintLevel = 1)
Converged <- output$converged
if(Converged) break
iter = iter + 1
}
print( output )
## Scale to replicate Koopman
a <- output$a
a.old <- a
aRa <- t(a) %*% R.X %*% a
## Scale a such that a' R a = .68659
## vc = variance of composite
vc <- aRa
## sf = scale factor
sf <- .68659/vc
a <- as.numeric(sqrt(sf)) * a
cat("\nKoopman Scaling\n")
print(round(a,2))
}
Run the code above in your browser using DataLab