I'm running a correlation analysis between a set of variables in SPSS.
In a first analysis I ran a crosstab analysis entering variables as rows and columns and performed both Spearman, Kendall's Tau-b and Tau-c correlation tests and annotated the results.
In a second moment I performed the same analysis, but going trough the bivariate correlation option in SPSS to obtain a correlation matrix and to facilitate the view of significant correlations.
It turned out that for Kendall's correlation, the p-values were different between analyses, despite the correlation coefficient and the number of valid samples being identical.
E.g.: For a given correlation I've got Tau-b = 0.175; p = 0.032; N = 75 in the first analysis, but obtained Tau-b = 175; p = 0.067; N = 75. For the Spearman's correlation the results were consistent between analyses, Rho = 0.213; p = 0.067; N = 75.
Can somebody grasp why this occurs? There are differences in p-value estimations when running correlation analyses in these two ways or it is probably be an error?
I forgot two say that in both cases the analyses were two-tailed.
Hello Glauco,
A search of IBM SPSS 24 Algorithms outlines the approximate SE method used to evaluate Tau-b in the Crosstabs subprogram, but no mention is made of what goes on in the Correlations subprogram.
I tried a sample data set with each in SPSS 26: Both yielded the same Tau-b value, but differed in the computed probability under the null hypothesis. Therefore, it is safe to conclude that these two subprograms evaluate significance level differently (and were likely implemented by different developers). Caveat emptor!
Good luck with your work!
Hello, David Morse !
Thank you so much for the answer!
So, do you think that using the values from crosstab would be safer, since it is more well documented?
Running a "validation" correlation analysis in another program would also be nice to evaluate the reliability of results, right?
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics