For anybody that deals with SAS-grad students or researchers.
I ‘m looking for an online "Dummies’-type guide to SAS statistical analysis. (Version 6.12 is what I got)
I’ve got repeated measurements data for which I need to program a covariance matrix structure. Toeplitz or et al blah blah blah…
Right now I do this by a ‘best fit’ analysis-i.e. trial and error. Anybody know of a SAS website that offers helpful programming tips?
With as aggravating as SAS can be, hopefully there’s a site to help the non-statistician with it. It would be nice to have an easier way.
Best fit analysis is trial by error?
Are you asking for help with a particular routine? Or with an explanation of statistical subjects??
Well, maybe both, RM.
I want to analyze treatment effects using repeated measures. I’m measuring weight gain over time. As I understand it, since I was asking about SAS, the variance of the initial measurements is included in the subsequent measurements and must be accounted for in the model.
Since there is no measurement before treatment, it is not a covariance structure in the sense that it not regression. This is important.
SAS, for repeated measurements, asks that you include the covariance structure in the mixed model (I’m using PROC MIXED).
My question is: Instead of programming the individual covariance matrixes (in SAS : SIM-simple; UN-unstructured; CS- compound symmetry; CHS-heterogeneous compound symmetry; …etc…)
Which gives me my trial and error results-I run the PROC MIXED with each different covariance structure and I see how ‘good they fit’. Trial and error.
PROC MIXED or (GLM for those not up to date) will compute Model Fitting Information for [whatever variable].
Akaike’s information criterion
Scharawtz’s Bayesians criterion
Etc…
Fluctuations of these determines my “best fit” to my data.
Instead of the witch’s brew of OUTPUT to determine what works best–
I’m looking for a web site or (God-Forbid I pay for It) a stat program that will run this shit for me.
Because now (in SAS 6…) I have to run each co-matrix on its’ own and compare to what’s the ‘best fit’ ergo-trial and error.
That is what I meant by T&E.
If you can explain the covariance matrix in plain English, I would appreciate it.
If you can explain covariance matrix to me in English, I’d be appreciative.
Thanks.
Doh! same last line! (phone call). But what’s the deal?
You almost lost me there in the jargon, but I can see a starting point here.
I guess you know what variance is, statistically–the square of the standard deviation. And if you have N variables, your covariance matrix is going to be N by N.
Take all possible correlation coefficients between your variables and arrange them in a matrix, also N by N. Since each variable will be exactly correlated with itself, the diagonal of the matrix will be all 1’s (perfect correlation). The off-diagonal elements will be symmetrical, and will be some value between -1 (perfect negative correlation) and 1. That is the correlation matrix.
Multiply the rows by the N corresponding standard deviations, and then the columns too, and you have the covariance matrix. The diagonal will be the variances of each variable (sd squared), and the off-diagonal elements will be the covariances between pairs of variables.
Large off-diagonal elements mean that the variables are highly correlated somehow, and may be measuring the same effect. Small off-diagonal elements mean that the variables are measuring relatively unique features (there is little correlation between them, anyway–they may be just nonlinearly related).
That’s probably not what you wanted, is it?
If your server supports it, try the usenet group COMP.SOFT-SYS.SAS (maybe DejaNews has it). You can also subscribe to the listserve SAS-L by sending a subscribe command to listserv@uga.cc.uga.edu (this may be your best bet). You can check SAS’s white papers. Finally, there is a statistical-specific user group listed on the SAS website here. I don’t know what their requirements for joining are.
Or you could use SPSS. 
::runs out of firing range::
Thanks for the feedback RM
But, I’ve got the possible structures for the covariance model–i.e. N v N (time v time) lies on the diagonal, but-as far as I understand , this only works for the simple model. Or does it?
What’s the relationship between varying variances on the main diagonal
Specifically, the CSH (heterogeneous compound symmetry) option (and I’m quoting for a SAS manual here) allows for unequal variances on the main diagonal. And other programming options deal all the variations, diagonal or off diagonal-equal or unequal.
I admit I don’t get all this shit. How do unequal variances correlate within a particular structure? I guess they can, since there is an unstructered option and I guess that all the other options are structured.
My question should be: what’s the easiest way to figure out, with a given data set, which structure to choose?
Brachy-thanks for the sites, I’ll be checking them out.