Promax: Perform Procustes,bifactor, promax or targeted rotations and return the inter factor angles.

Description

The bifactor rotation implements the rotation introduced by Jennrich and Bentler (2011) by calling GPForth in the GPArotation package. promax is an oblique rotation function introduced by Hendrickson and White (1964) and implemented in the promax function in the stats package. Unfortunately, promax does not report the inter factor correlations. Promax does. TargetQ does a target rotation with elements that can be missing (NA), or numeric (e.g., 0, 1). It uses the GPArotation package. target.rot does general target rotations to an arbitrary target matrix. The default target rotation is for an independent cluster solution. equamax facilitates the call to GPArotation to do an equamax rotation. Equamax, although available as a specific option within GPArotation is easier to call by name if using equamax. The varimin rotation suggested by Ertl (2013) is implemented by appropriate calls to GPArotation.

Usage

faRotate(loadings,rotate="oblimin",...)
bifactor(L, Tmat=diag(ncol(L)), normalize=FALSE, eps=1e-5, maxit=1000)
biquartimin(L, Tmat=diag(ncol(L)), normalize=FALSE, eps=1e-5, maxit=1000)
TargetQ(L, Tmat=diag(ncol(L)), normalize=FALSE, eps=1e-5, maxit=1000,Target=NULL)
TargetT(L, Tmat=diag(ncol(L)), normalize=FALSE, eps=1e-5, maxit=1000,Target=NULL)
Promax(x,m=4, normalize=FALSE, pro.m = 4) 
Procrustes(L,Target) #adapted from Niels Waler
target.rot(x,keys=NULL)
varimin(L, Tmat = diag(ncol(L)), normalize = FALSE, eps = 1e-05, maxit = 1000)
vgQ.bimin(L)   #called by bifactor
vgQ.targetQ(L,Target=NULL)  #called by TargetQ
vgQ.varimin(L)  #called by varimin
equamax(L, Tmat=diag(ncol(L)), eps=1e-5, maxit=1000)

Value

loadings: Oblique factor loadings
rotmat: The rotation matrix applied to the original loadings to produce the promax solution or the targeted matrix
Phi: The interfactor correlation matrix

Arguments

x: A loadings matrix
L: A loadings matrix
loadings: A loadings matrix
rotate: Which rotation should be used?
m: the power to which to raise the varimax loadings (for Promax)
pro.m: the power to which to raise the various loadings in Promax.
keys: An arbitrary target matrix, can be composed of any weights, but probably -1,0, 1 weights. If missing, the target is the independent cluster structure determined by assigning every item to it's highest loaded factor.
Target: A matrix of values (mainly 0s, some 1s, some NAs) to which the matrix is transformed.
Tmat: An initial rotation matrix
normalize: parameter passed to optimization routine (GPForth in the GPArotation package and Promax)
eps: parameter passed to optimization routine (GPForth in the GPArotation package)
maxit: parameter passed to optimization routine (GPForth in the GPArotation package)
...: Other parameters to pass (e.g. to faRotate) include a Target list or matrix

Author

William Revelle

Details

The two most useful of these functions is probably biquartimin which implements the oblique bifactor rotation introduced by Jennrich and Bentler (2011). The second is TargetQ which allows for missing NA values in the target. Next best is the orthogonal case, bifactor. None of these seem to be implemented in GPArotation (yet).

TargetT is an orthogonal target rotation function which allows for missing NA values in the target.

faRotate is merely a convenient way to call the various GPArotation functions as well as the additional ones added here.

The difference between biquartimin and bifactor is just that the latter is the orthogonal case which is documented in Jennrich and Bentler (2011). It seems as if these two functions are sensitive to the starting values and random restarts (modifying T) might be called for.

bifactor output for the 24 cognitive variable of Holzinger matches that of Jennrich and Bentler as does output for the Chen et al. problem when fm="mle" is used and the Jennrich and Bentler solution is rescaled from covariances to correlations.

Promax is a very direct adaptation of the stats::promax function. The addition is that it will return the interfactor correlations as well as the loadings and rotation matrix.

varimin implements the varimin criterion proposed by Suitbert Ertl (2013). Rather than maximize the varimax criterion, it minimizes it. For a discussion of the benefits of this procedure, consult Ertel (2013).

In addition, these functions will take output from either the factanal, fa or earlier (factor.pa, factor.minres or principal) functions and select just the loadings matrix for analysis.

equamax is just a call to GPArotation's cFT function (for the Crawford Ferguson family of rotations.

TargetQ implements Michael Browne's algorithm and allows specification of NA values. The Target input is a list (see examples). It is interesting to note how powerful specifying what a factor isn't works in defining a factor. That is, by specifying the pattern of 0s and letting most other elements be NA, the factor structure is still clearly defined.

The target.rot function is an adaptation of a function of Michael Browne's to do rotations to arbitrary target matrices. Suggested by Pat Shrout.

The default for target.rot is to rotate to an independent cluster structure (every items is assigned to a group with its highest loading.)

target.rot will not handle targets that have linear dependencies (e.g., a pure bifactor model where there is a g loading and a group factor for all variables).

References

Ertel, S. (2013). Factor analysis: healing an ailing model. Universitatsverlag Gottingen.

Hendrickson, A. E. and White, P. O, 1964, British Journal of Statistical Psychology, 17, 65-70.

Jennrich, Robert and Bentler, Peter (2011) Exploratory Bi-Factor Analysis. Psychometrika, 1-13

Examples

Run this code

jen <- sim.hierarchical()
f3 <- fa(jen,3,rotate="varimax")
f3   #not a very clean solution
Promax(f3)  #this obliquely rotates, but from the varimax target
target.rot(f3)  #this obliquely rotates to wards a simple structure target

#compare this rotation with the solution from a targeted rotation aimed for 
#an independent cluster solution
#now try a bifactor solution
fb <-fa(jen,3,rotate="bifactor")
fq <- fa(jen,3,rotate="biquartimin")
#Suitbert Ertel has suggested varimin
fm <-  fa(jen,3,rotate="varimin") #the Ertel varimin
fn <- fa(jen,3,rotate="none")  #just the unrotated factors
#compare them
factor.congruence(list(f3,fb,fq,fm,fn))
# compare an oblimin with a target rotation using the Browne algorithm
 #note that we are changing the factor #order (this is for demonstration only)
 Targ <- make.keys(9,list(f1=1:3,f2=7:9,f3=4:6)) 
 Targ <- scrub(Targ,isvalue=1)  #fix the 0s, allow the NAs to be estimated
 Targ <- list(Targ)  #input must be a list
#show the target
 Targ
 fa(Thurstone,3,rotate="TargetQ",Target=Targ)  #targeted oblique rotation
#compare with oblimin
f3 <- fa(Thurstone,3)
#now try a targeted orthogonal rotation
Targ <- make.keys(9,list(f1=1:3,f2=7:9,f3=4:6)) 
faRotate(f3$loadings,rotate="TargetT",Target=list(Targ)) #orthogonal

Run the code above in your browser using DataLab