psych (version 2.1.9)

corr.test: Find the correlations, sample sizes, and probability values between elements of a matrix or data.frame.

Description

Although the cor function finds the correlations for a matrix, it does not report probability values. cor.test does, but for only one pair of variables at a time. corr.test uses cor to find the correlations for either complete or pairwise data and reports the sample sizes and probability values as well. For symmetric matrices, raw probabilites are reported below the diagonal and correlations adjusted for multiple comparisons above the diagonal. In the case of different x and ys, the default is to adjust the probabilities for multiple tests. Both corr.test and corr.p return raw and adjusted confidence intervals for each correlation.

Usage

corr.test(x, y = NULL, use = "pairwise",method="pearson",adjust="holm",
alpha=.05,ci=TRUE,minlength=5,normal=TRUE)
corr.p(r,n,adjust="holm",alpha=.05,minlength=5,ci=TRUE)

Arguments

x

A matrix or dataframe

y

A second matrix or dataframe with the same number of rows as x

use

use="pairwise" is the default value and will do pairwise deletion of cases. use="complete" will select just complete cases.

method

method="pearson" is the default value. The alternatives to be passed to cor are "spearman" and "kendall". These last two are much slower, particularly for big data sets.

What adjustment for multiple tests should be used? ("holm", "hochberg", "hommel", "bonferroni", "BH", "BY", "fdr", "none"). See p.adjust for details about why to use "holm" rather than "bonferroni").

alpha

alpha level of confidence intervals

r

A correlation matrix

n

Number of observations if using corr.p. May be either a matrix (as returned from corr.test, or a scaler. Set to n - np if finding the significance of partial correlations. (See below).

ci

By default, confidence intervals are found. However, this leads to a noticable slowdown of speed, particularly for large problems. So, for just the rs, ts and ps, set ci=FALSE

minlength

What is the minimum length for abbreviations. Defaults to 5.

normal

By default, probabilities for method="spearman" and method="kendall" are found by normal theory. If normal=="FALSE", then repetitive calls are made to cor.test. This is much slower, but gives more accurate p values. exact is set to be FALSE which means that exact p values for small samples are not found given the problem of ties.

Value

r

The matrix of correlations

n

Number of cases per correlation

t

value of t-test for each correlation

p

two tailed probability of t for each correlation. For symmetric matrices, p values adjusted for multiple tests are reported above the diagonal.

se

standard error of the correlation

ci

the alpha/2 lower and upper values.

ci2

ci but with the adjusted pvalues as well. This was added after tests showed we were breaking some packages that were calling the ci object without bothering to check for its dimensions.

These are the adjusted ((Holm or Bonferroni) confidence intervals. If asking to not adjust, the Holm adjustments for the confidence intervals are shown anyway, but the probability values are not adjusted and the appropriate confidence intervals are shown in the ci object.

stars

For those people who like to show magic asterisks denoting statistical significance" the stars object flags those correlation values that are unlikely given normal theory. See the last example for how to print these neatly.

Details

corr.test uses the cor function to find the correlations, and then applies a t-test to the individual correlations using the formula $$t = \frac{r * \sqrt(n-2)}{\sqrt(1-r^2)}$$

$$se = \sqrt(\frac{1-r^2}{n-2})$$

The t and Standard Errors are returned as objects in the result, but are not normally displayed. Confidence intervals are found and printed if using the print(short=FALSE) option. These are found by using the fisher z transform of the correlation and then taking the range r +/- qnorm(alpha/2) * se and the standard error of the z transforms is $$se = \sqrt(\frac {1}{n-3})$$. These values are then back transformed to be in correlation units. They are returned in the ci object.

Note that in the case of method=="kendall" since these are the normal theory confidence intervals they are slightly too big.

The probability values may be adjusted using the Holm (or other) correction. If the matrix is symmetric (no y data), then the original p values are reported below the diagonal and the adjusted above the diagonal. Otherwise, all probabilities are adjusted (unless adjust="none"). This is made explicit in the output. Confidence intervals are shown for raw and adjusted probabilities in the ci object.

corr.p may be applied to the results of partial.r if n is set to n - s (where s is the number of variables partialed out) Fisher, 1924.

cor.test for tests of a single correlation, Hmisc::rcorr for an equivalant function, r.test to test the difference between correlations, and cortest.mat to test for equality of two correlation matrices.

Also see cor.ci for bootstrapped confidence intervals of Pearson, Spearman, Kendall, tetrachoric or polychoric correlations. In addition cor.ci will find bootstrapped estimates of composite scales based upon a set of correlations (ala cluster.cor).

In particular, see p.adjust for a discussion of p values associated with multiple tests.

Other useful functions related to finding and displaying correlations include lowerCor for finding the correlations and then displaying the lower off diagonal using the lowerMat function. lowerUpper to compare two correlation matrices.

Examples

Run this code
# NOT RUN {
ct <- corr.test(attitude)  #find the correlations and give the probabilities
ct #show the results
cts <- corr.test(attitude[1:3],attitude[4:6]) #reports all values corrected for multiple tests

print(corr.p(cts$r,n=30),short=FALSE) #don't adjust the probabilities print(corr.test(sat.act[1:3],sat.act[4:6],adjust="none"),short=FALSE) #print out the stars object without showing quotes print(corr.test(attitude)$stars,quote=FALSE)  #note that the adjusted ps are given as well