corCi: Bootstrapped and normal confidence intervals for raw and composite correlations

Description

Although normal theory provides confidence intervals for correlations, this is particularly problematic with Synthetic Aperture Personality Assessment (SAPA) data where the individual items are Massively Missing at Random. Bootstrapped confidence intervals are found for Pearson, Spearman, Kendall, tetrachoric, or polychoric correlations and for scales made from those correlations. If given a correlation matrix and sample size(s), normal theory confidence intervals are provided.

Usage

corCi(x, keys = NULL, n.iter = 100,  p = 0.05,overlap = FALSE, 
 poly = FALSE, method = "pearson", plot=TRUE,minlength=5,n=NULL,...)
 
cor.ci(x, keys = NULL, n.iter = 100,  p = 0.05,overlap = FALSE, 
 poly = FALSE, method = "pearson", plot=TRUE,minlength=5,n=NULL,...)

Value

rho: The original (composite) correlation matrix.
means: Mean (of Fisher transformed) correlation retransformed back to the r units
sds: Standard deviation of Fisher transformed correlations
ci: Mean +/- alpha/2 of the z scores as well as the alpha/2 and 1-alpha/2 quantiles. These are labeled as lower.emp(ircal), lower.norm(al), upper.norm and upper.emp.
replicates: The observed replication values so one can do one's own estimates

Arguments

x: The raw data, or a correlation matrix if not doing bootstrapping
keys: If NULL, then the confidence intervals of the raw correlations are found. Otherwise, composite scales are formed from the keys applied to the correlation matrix (in a logic similar to cluster.cor but without the bells and whistles) and the confidence of those composite scales intercorrelations.
n.iter: The number of iterations to bootstrap over. This will be very slow if using tetrachoric/or polychoric correlations.
p: The upper and lower confidence region will include 1-p of the distribution.
overlap: If true, the correlation between overlapping scales is corrected for item overlap.
poly: if FALSE, then find the correlations using the method specified (defaults to Pearson). If TRUE, the polychoric correlations will be found (slowly). Because the polychoric function uses multicores (if available), and corCi does as well, the number of cores used is options("mc.cores")^2.
method: "pearson","spearman", "kendall"
plot: Show the correlation plot with correlations scaled by the probability values. To show the matrix in terms of the confidence intervals, use cor.plot.upperLowerCi.
minlength: What is the minlength to use in abbreviations of the cis? Defaults to 5
n: If finding confidence intervals from a correlation matrix, specify the n
...: Other parameters for axis (e.g., cex.axis to change the font size, srt to rotate the numbers in the plot)

Author

William Revelle

Details

If given a correlation matrix, then confidence intervals are found based upon the sample sizes using the conventional r2z fisher transformation (fisherz and the normal distribution.

If given raw data, correlations are found. If keys are specified (the normal case), then composite scales based upon the correlations are found and reported. This is the same procedure as done using cluster.cor or scoreItems.

Then (with raw data) the data are recreated n.iter times by sampling subjects (rows) with replacement and the correlations (and composite scales) are found again (and again and again). Mean and standard deviations of these values are calculated based upon the Fisher Z transform of the correlations. Summary statistics include the original correlations and their confidence intervals. For those who want the complete set of replications, those are available as an object in the resulting output.

Although particularly useful for SAPA (https://www.sapa-project.org/) type data where we have lots of missing data, this will work for any normal data set as well.

Although the correlations are shown automatically as a cor.plot, it is possible to show the upper and lower confidence intervals by using cor.plot.upperLowerCi. This will also return, invisibly, a matrix for printing with the lower and upper bounds of the correlations shown below and above the diagonal (see the first example).

References

For SAPA type data, see Revelle, W., Wilt, J., and Rosenthal, A. (2010) Personality and Cognition: The Personality-Cognition Link. In Gruszka, A. and Matthews, G. and Szymura, B. (Eds.) Handbook of Individual Differences in Cognition: Attention, Memory and Executive Control, Springer.

Examples

Run this code

#find confidence intervals of a correlation matrix with specified sample size
ci <- corCi(Thurstone[1:6,1:6],n=213)
ci  #show them
R <- cor.plot.upperLowerCi(ci)  #show them graphically
R #show them as a matrix 


#confidence intervals by bootstrapping requires raw data
corCi(psychTools::bfi[1:200,1:10])  # just the first 10 variables
#The keys have overlapping scales
keys <- list(agree=c("-A1","A2","A3","A4","A5"), conscientious= c("C1", 
  "C2","C3","-C4","-C5"),extraversion=c("-E1","-E2","E3","E4","E5"), neuroticism= 
  c("N1", "N2", "N3","N4","N5"), openness = c("O1","-O2","O3","O4","-O5"), 
  alpha=c("-A1","A2","A3","A4","A5","C1","C2","C3","-C4","-C5","N1","N2","N3","N4","N5"),
beta = c("-E1","-E2","E3","E4","E5","O1","-O2","O3","O4","-O5") )

  
#do not correct for item overlap
rci <-  corCi(psychTools::bfi[1:200,],keys,n.iter=10,main="correlation with overlapping scales") 
#also shows the graphic -note the overlap
#correct for overlap
rci <-  cor.ci(psychTools::bfi[1:200,],keys,overlap=TRUE, n.iter=10,main="Correct for overlap") 
#show the confidence intervals
ci <- cor.plot.upperLowerCi(rci)  #to show the upper and lower confidence intervals
ci   #print the confidence intervals in matrix form

Run the code above in your browser using DataLab