# RE: st: Polychoric and missing values in correlation matrix

 From jverkuilen To Subject RE: st: Polychoric and missing values in correlation matrix Date Thu, 9 Apr 2009 15:13:51 -0400

```Maarten already addressed why you have missing values.

Tetra- and poly-choric correlations assume underlying *bivariate* normality With some binary and other continuous variables, you really want some tetrachoric (for binay/binary pairs), Pearson correlations (for continuous/continuous pairs), and biserial correlations (for binary/continuous pairs).

I am guessing the tetrachorics on skewed binary variables are the trouble spot, and you can get them using -tetrachoric- on only the binary variables and checking the bivariate margins with -tab-.

There is no guarantee the resulting "correlation" matrix is consistent with multivariate normality, however, but checking for PSD requires you compute the eigenvalues, which is easy enough using -pcamat-.

You can often approximate tetrachoric and biserial correlations quite well using a common correction factor based on the ordinate of the Gaussian, a result found in most test theory books, e.g.,

R. P. McDonald. 1999. Test Theor: A Unified Treatment. LEA.

-----Original Message-----
From: "Juan Julio Gutierrez" <juanju@yahoo.com>
To: statalist@hsphsun2.harvard.edu
Sent: 4/9/2009 7:50 AM
Subject: st: Polychoric and missing values in correlation matrix

Dear all

I am using the polychroric command to calculate the correlation matrix of a set of 17 variables ( 8 binary 0-/1 and 9 metric).

After lenghty calculations STATA shows a matrix  with missing values in the matrix, mostly between metric and binary variables. 4 metric variables have missing values correlations with binary variables.

The results window show these messages. In red:

could not calculate numerical derivatives

missing values encountered

initial values not feasible

too few quotes

no variables defined

And in green:

numerical derivatives are approximate

nearby values are missing

Do you have an idea why am I facing these results?

How can I assure that the corr. matrix is positive defined?

Do you know if using R poly command would dela with the missing values? .

Thank you

Juan Julio

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/
```