Re: next version of PSPP

[Alan Mead]

William Simpson wrote:

I'm honestly shocked that no one else has mentioned correlations in this

    

I have 0.7.2 and it has
analyse/linear regression
I tested it just now and it gives r-squared. Take the square root and
use the sign of the slope -- voila, Pearson r.

As for reliability...
Isn't that just Pearson r?
If so, see above.
  

Right (about the Pearson r).  Or I could use Calc. Or just use my own code.  Calculating a single correlation is trivial and PSPP already does it (and GSL already does it), which is why it's puzzling that CORRELATIONS is missing from PSPP.

But let's say that you were developing a new assessment and you wanted to look at the correlation between each item and some criterion, to ensure that you included all the most predictive items. 

Or let's say that you were doing a study involving regression and you measured a dozen variables and you wanted to publish your work, so you followed the APA guidelines for data presentation, which require that you present basic descriptive statistics for all the measures used in your study--that would mainly be a table showing the mean, standard deviation, and intercorrelation of each variable used in your study (perhaps with some measure of the reliability of your measures on the diagonal of that matrix).  [And those basic stats aren't merely window-dressing; simple colinearity is a major problem in many applications of regression and a correlation matrix is the standard way to assess this.  So, implementing regression/GLM without the correlation procedure is bass ackwards.]

Or you have a survey and you wanted to know the correlation of several questions.

Or, in short, doing almost any applied project in the behavioral sciences.  Then, it would be pretty inconvenient to use linear regression to calculate each correlation. 

Also, SPSS correlations has some other features.  It allows you to pick a portion of the matrix of correlations (the syntax is something like "CORR /VAR A B C WITH X Y Z" and it gives you just the lower left quarter of the matrix--which can be fairly important if you are working with a lot of variables), optionally it produces Spearman correlations, you can cause it to export a correlation or covariance matrix, etc.  And, although (like many) I'm not a big fan of blind NHST, SPSS correlations sprinkles the stars into the matrix for you.

Not to be argumentative, but why do you think SPSS included the correlations procedure, if so many other procedures could be tortured to produce correlations?

So, my point is that this is a huge gaping hole in the functionality of PSPP and I could never use PSPP in a class, nor can I recommend PSPP to anyone as a replacement for SPSS.

I said as much to Ben a good 9-10 years ago and he said something like "Well, look, I'm doing this in my spare time and, listen, the code is GPL and you could always jump in if you want it that badly" and that's cool. I completely understand!  I just wanted to take advantage of this thread to add a voice for current SPSS/PASW users like me... A long-time, extremely dissatisfied SPSS user who uses SPSS/PASW in teaching and research.  I think there's probably quite a few and many of them probably aren't on this list.

As for reliabilities, they are estimated in different ways and the RELIABILITIES command produces Cronbach's alpha (and a couple different forms of the intra-class correlation coefficient; ICC).  And it optionally produces a table, very useful for developing measures, showing the corrected item-total and "alpha-if-item-deleted" values for each item in a scale.  Corrected item-total correlations are a royal PITA to calculate because the "correction" is that you remove the target item from the total when you calculate the correlation for that item.  So, to do this in SPSS syntax, you need to create a total for each item separately:

compute total = sum(x1, x2, x3, x4).
compute total1 = sum(x2, x3, x4).
compute total2 = sum(x1, x3, x4).
compute total3 = sum(x1, x2, x4).
compute total4 = sum(x1, x2, x3).
execute.

corr / var x1 total1.
corr / var x2 total2.
corr / var x3 total3.
corr / var x4 total4.

This would get prohibitively tedious for a long measure or 15 10-item tests (which was a project I helped analyze) and most SPSS users these days aren't interested in using SPSS like a programming language (or else they would be using SAS or R).  And, of course, alpha, ICCs, etc. are not available this way (and, of course, this wouldn't work anyway because CORR is not available).  If PSPP optionally allowed one to calculate biserial, rather than point-biserial/Pearson correlations, for dichotomous variables that would be even better.

-Alan

-- 
Alan D. Mead, Ph.D. 
Assistant Professor, Institute of Psychology
Scientific Adviser, Center for Research and Service
Illinois Institute of Technology
3101 South Dearborn, 2nd floor
Chicago IL 60616

Skype: alandmead
+312.567.5933 (Campus)
+815.588.3846 (Home Office)
+312.567.3493 (Fax)

http://mypages.iit.edu/~mead
http://www.center.iit.edu
http://www.alanmead.org

I slept and dreamt that life was joy.
I awoke and saw that life was service.
I acted and behold, service was joy.
 -- attributed to Rabindranath Tagore 

amead.vcf
Description: Vcard