turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

Find a Community

- Home
- /
- SAS Programming
- /
- SAS Procedures
- /
- Pooling results from multiple GLIMMIX output

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

10-09-2015 08:47 PM

Hi everyone,

I need your help to analyze a GLIMMIX parameter estimates output using PROC MIANALYZE. The output is created by group, and for each group treatment effect is estimated with its standard error. Basically I want to pool estimates from multiple simulated datasets. Unlike PROC MI procedure imputed values are not variables but they are estimates. Thanks.

Accepted Solutions

Solution

10-26-2015
09:27 PM

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

Posted in reply to MetinBulus

10-10-2015 11:22 PM

Usually, simulations are used to assess the effect of sources of variation/uncertainty that are not well accounted for by available models. The variation of parameter estimates from simulation runs is assumed to reflect the variability built into the simulation process. Thus, the estimated standard errors are usually ignored, unless you are interested in the distribution of the standard error itself. The properties (location, variance, covariance, etc.) of your set of parameter estimates can normally be investigated and summarized with SAS base procedures (univariate, corr, sgplot).

I think however that simulations should involve more than 10 repetitions. I usually go for 100 to 2000.

PG

All Replies

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

Posted in reply to MetinBulus

10-09-2015 11:42 PM

Why not look at the distribution of your parameter estimates with standard tools like PROC UNIVARIATE and PROC CORR?

PG

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

Posted in reply to PGStats

10-10-2015 09:20 AM - edited 10-10-2015 09:22 AM

I have the regression coefficient estimated ten times for example, each estimation has a standard error associated with it. Using Proc means or Univariate won't take into account standard errors, as far as I know. I don't know how to come up with a single value for ten coefficients, taking into account their standard errors.

Solution

10-26-2015
09:27 PM

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

Posted in reply to MetinBulus

10-10-2015 11:22 PM

Usually, simulations are used to assess the effect of sources of variation/uncertainty that are not well accounted for by available models. The variation of parameter estimates from simulation runs is assumed to reflect the variability built into the simulation process. Thus, the estimated standard errors are usually ignored, unless you are interested in the distribution of the standard error itself. The properties (location, variance, covariance, etc.) of your set of parameter estimates can normally be investigated and summarized with SAS base procedures (univariate, corr, sgplot).

I think however that simulations should involve more than 10 repetitions. I usually go for 100 to 2000.

PG