scoreOverlap: Find correlations of composite variables (corrected for overlap) from a larger matrix.

Description

Given a n x c cluster definition matrix of -1s, 0s, and 1s (the keys) , and a n x n correlation matrix, or an N x n data matrix, find the correlations of the composite clusters. The keys matrix can be entered by hand, copied from the clipboard (read.clipboard), or taken as output from the factor2cluster or make.keys functions. Similar functionality to scoreItems which also gives item by cluster correlations. scoreBy does this for individual subjects after a call to statsBy.

Usage

scoreOverlap(keys, r, correct = TRUE, SMC = TRUE, av.r = TRUE, item.smc = NULL, 
     impute = TRUE,select=TRUE, scores=FALSE,  min=NULL,max=NULL)
scoreBy(keys,stats, correct = TRUE, SMC = TRUE, av.r = TRUE, item.smc = NULL, 
     impute = TRUE,select=TRUE,min.n=3,smooth=FALSE)
cluster.cor(keys, r.mat, correct = TRUE,SMC=TRUE,item.smc=NULL,impute=TRUE)

Value

cor: the (raw) correlation matrix of the clusters
sd: standard deviation of the cluster scores
corrected: raw correlations below the diagonal, alphas on diagonal, disattenuated above diagonal
alpha: The (standardized) alpha reliability of each scale.
G6: Guttman's Lambda 6 reliability estimate is based upon the smcs for each item in a scale. G6 uses the smc based upon the entire item domain.
av.r: The average inter item correlation within a scale
size: How many items are in each cluster?
MIMS: A matrix of the average correlations, corrected for overlap, between the various scales.

Arguments

keys: A list of scale/cluster keys, or a matrix of cluster keys
r.mat: A correlation matrix
r: Either a correlation matrix or a raw data matrix
stats: The output from statsBy. Needed for scoreBy.
correct: TRUE shows both raw and corrected for attenuation correlations
SMC: Should squared multiple correlations be used as communality estimates for the correlation matrix?
item.smc: the smcs of the items may be passed into the function for speed, or calculated if SMC=TRUE
impute: if TRUE and scores=FALSE, impute missing scale correlations based upon the average interitem correlation, otherwise return NA. If scores=TRUE, impute missing data by means, medians, or none.
av.r: Should the average r be used in correcting for overlap? smcs otherwise.
select: By default, just find statistics for items included in the scoring keys. This allows for finding scores from matrices with bad items if they are not included in the set of scoring keys.
min.n: The minimum number of pairwise observations needed to define a correlation (in scoreBy)
smooth: If the matrices used in scoreBy are not positive semi-definite, should we smooth them?
scores: Find the raw average scores for each subject. Same functionality as scoreItems
min: May be specified as minimum item score allowed, else will be calculated from data. min and max should be specified if items differ in their possible minima or maxima. See notes for details.
max: May be specified as maximum item score allowed, else will be calculated from data.

Author

Maintainer: William Revelle revelle@northwestern.edu

Details

These are three of the functions used in the SAPA (https://www.sapa-project.org/) procedures to form synthetic correlation matrices. Given any correlation matrix of items, it is easy to find the correlation matrix of scales made up of those items. This can also be done from the original data matrix or from the correlation matrix using scoreItems which is probably preferred unless the keys are overlapping. It is important to remember with SAPA data, that scale correlations should be found from the item correlations, not the raw data.

In the case of overlapping keys, (items being scored on multiple scales), scoreOverlap will adjust for this overlap by replacing the overlapping covariances (which are variances when overlapping) with the corresponding best estimate of an item's ``true" variance using either the average correlation or the smc estimate for that item. This parallels the operation done when finding alpha reliability. This is similar to ideas suggested by Cureton (1966) and Bashaw and Anderson (1966) but uses the smc or the average interitem correlation (default).

A typical use in the SAPA project is to form item composites by clustering or factoring (see fa, ICLUST, principal), extract the clusters from these results (factor2cluster), and then form the composite correlation matrix using cluster.cor. The variables in this reduced matrix may then be used in multiple correlatin procedures using mat.regress.

The original correlation is pre and post multiplied by the (transpose) of the keys matrix.

If some correlations are missing from the original matrix this will lead to missing values (NA) for scale intercorrelations based upon those lower level correlations. If impute=TRUE (the default), a warning is issued and the correlations are imputed based upon the average correlations of the non-missing elements of each scale.

Because the alpha estimate of reliability is based upon the correlations of the items rather than upon the covariances, this estimate of alpha is sometimes called ``standardized alpha". If the raw items are available, it is useful to compare standardized alpha with the raw alpha found using scoreItems. They will differ substantially only if the items differ a great deal in their variances.

The MIMS matrix reflects the average correlations of the Multiple Items within each of the Multiple Scales. This has been suggested as a measure of scale validity. These correlations are adjusted for item overlap.

scoreOverlap answers an important question when developing scales and related subscales, or when comparing alternative versions of scales. For by removing the effect of item overlap, it gives a better estimate the relationship between the latent variables estimated by the observed sum (mean) scores.

scoreBy finds the within subject correlations after preprocessing with statsBy. This is useful if doing an ESM study with multiple occasions for each subject. It also makes it possible to find the correlations for subsets of subjects. See the example. Note that it likely that for ESM data with a high level of missingness that the correlation matrices will not be positive-semi-definite. This can lead to composite score correlations that exceed 1. Smoothing will resolve this problem.

scoreBy is useful when examining multi-level models where we want to examine the correlations within subjects (e.g., for ESM data) or within groups of subjects (when examining the stability of correlational structures by subgroups). For both cases the data must be processed first by statsBy. To find the variances of the scales it is necessary to use the cor="cov" option in statsBy.

Starting in version 2.4.1, the scores option has been added. This will score the data (as does scoreItems) if it is not a correlation matrix but does not have all the options of scoreItems. Obviously, while the correlations can be corrected for item ovelap, the raw scores can not be so corrected.

References

Bashaw, W. and Anderson Jr, H. E. (1967). A correction for replicated error in correlation coefficients. Psychometrika, 32(4):435-441.

Cureton, E. (1966). Corrected item-test correlations. Psychometrika, 31(1):93-96.

Examples

Run this code

if(require(psychTools)) {
#use the msq data set that shows the structure of energetic and tense arousal
small.msq <- psychTools::msq[ c("active", "energetic", "vigorous", "wakeful", 
"wide.awake", "full.of.pep", "lively", "sleepy", "tired", "drowsy","intense", 
"jittery", "fearful", "tense", "clutched.up", "quiet", "still",    "placid",
 "calm", "at.rest") ]
small.R <- cor(small.msq,use="pairwise")
keys.list <- list(
EA = c("active", "energetic", "vigorous", "wakeful", "wide.awake", "full.of.pep",
       "lively", "-sleepy", "-tired", "-drowsy"),
TA =c("intense", "jittery", "fearful", "tense", "clutched.up", "-quiet", "-still", 
       "-placid", "-calm", "-at.rest") ,

high.EA = c("active", "energetic", "vigorous", "wakeful", "wide.awake", "full.of.pep",
       "lively"),
low.EA =c("sleepy", "tired", "drowsy"),
lowTA= c("quiet", "still", "placid", "calm", "at.rest"),
highTA = c("intense", "jittery", "fearful", "tense", "clutched.up")
   ) 
   
keys <- make.keys(small.R,keys.list)
      
adjusted.scales <- scoreOverlap(keys.list,small.R)
#compare with unadjusted
confounded.scales <- cluster.cor(keys,small.R)
summary(adjusted.scales)
#note that the EA and high and low EA and TA and high and low TA 
# scale correlations are confounded
summary(confounded.scales) 
}
bfi.stats <- statsBy(bfi,group="education",cors=TRUE ,cor="cov")
 #specify to find covariances
 bfi.plus.keys <- c(bfi.keys,gender="gender",age ="age")
bfi.by <- scoreBy(bfi.plus.keys,bfi.stats)
bfi.by$var   #to show the variances of each scale by groupl
round(bfi.by$cor.mat,2)  #the correlations by group
bfi.by$alpha

Run the code above in your browser using DataLab