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
- /
- Parameter Estimates for Mixing Probabilities in Pr...

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
- Permalink
- Email to a Friend
- Report Inappropriate Content

05-23-2013 12:22 AM

Greetings,

Does anyone know if it is possible to force SAS to show you all of the estimates for mixing probabilities in Proc FMM? It defaults to omitting the last estimate, and I'm interested in this because I would like to know the standard error for all of the estimates for mixing probabilities. Alternatively, does anyone know how SAS calculates the standard error for the parameter estimates for mixing probabilities?

Thank you in advance,

Lynn

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

Posted in reply to lynnlaurie

05-23-2013 06:00 AM

Since the sum of the mixing probabilities is 1, the last estimate is 1 - sum_of_other_estimates. If you fit a model with k components, there are only (k-1) parameters.

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

Posted in reply to Rick_SAS

05-23-2013 09:47 AM

Hi Rick. Thanks for the response. I understand how to find the mixing probability (p) of the omitted value, but I'm interested in the standard error for that value. Do you know how it is calculated or how I can force Proc FMM to show it?

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

Posted in reply to lynnlaurie

05-24-2013 11:06 AM

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

Posted in reply to Rick_SAS

05-30-2013 02:54 PM

Not sure, but since the parameterization is of a generalized logit, with the last predictor set to zero, try fitting with the NOINT option in the model statement. If there is a single independent variable, that should address the issue.

(I hope)

Steve Denham

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

Posted in reply to lynnlaurie

05-30-2013 06:09 PM

If you have only two mixing probabilities, p and 1-p, then you are luck. There is symmetry in the SEs because FMM models the mixing process using the symmetric logit link. For instance, if p = 0.2, then 1-p=0.8 (obviously). Logit(p) = -logit(1-p). With p=0.2, logit(.2) = -1.386, and logit(.8) = +1.386. The variances carry through in the same way. I just checked this by writing a NLMIXED program for a two-component mixture and getting p and 1-p from logit(p) and logit(1-p). The SEs come from the inverse Hessian matrix in FMM (the mixing probability is part of the likelihood).

You can get a hint of all of this by comparing FMM with GENMOD for a zero-inflated Poisson (a simple mixing problem). Both PROCs fit this model, but FMM represents the mixing probability by p, whileGENMOD represents the mixing probability by 1-p. But the estimates and SEs all agree.

If you have more than two mixing probabilities, then you can't take advantage of the symmetry. You would have to get 1-(p1+p2+...) by hand; the logit of this is easy, but use of the delta method to get the variance (and SE) would be tricky.

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

05-30-2013 06:12 PM

And I have not seen any way to force FMM to print 1-p or logit(1-p).

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

05-31-2013 09:03 AM

Agreed. I have to say that I think PROC FMM is a great program, but very frustrating. Hopefully the next update will hit on some of these issues.

Steve Denham