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
- /
- Analytics
- /
- Stat Procs
- /
- PROC MI FCS - how to monitor convergence for class...

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

05-04-2017 06:30 AM

In PROC MI, when using the FCS method, the OUTITER and PLOTS=TRACE options can be specified to monitor convergence. However, it only seems possible to output the mean and standard deviation of continuous variables which are being imputed (or are fully observed but included as variables in the call to PROC MI). If (for example) logistic regression is used to impute a binary variable, there appears to be no way of checking or assessing convergence.

For example, using code such as:

proc mi data=miinput seed=279813 nimpute=10 out=imps; var y1 y2 y3 ; class y1 y2 y3 ; fcs nbiter=100 OUTITER=trace logistic(y1 = y2 y3 / DETAIL ); fcs logistic(y2 = y1 y3 / DETAIL ); fcs logistic(y3 = y1 y2 / DETAIL );

run;

one obtains the warning message "WARNING: The OUTITER= option is only applicable with continuous variables in the FCS model.".

The DETAIL option means you can see what parameter values were used in each logistic regression for each imputation, but to assess convergence properly one really needs to see the parameter values by iteration for each imputation, as one can do with the MCMC statement. Is this possible somehow?