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 mixed effect size: Does a solution exist?

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

11-22-2016 11:51 PM

I have spent so much time trying to find out how to calculate an effect size for linear mixed models, but still success evades me. I am an avid SPSS user. Given the mixed model limitation in SPSS and lack of my success, I finally decided to transition to SAS in hopes of possibly overcoming this problem with the effect size. Is there anyone who has any clue how to calculate an effect size for lognitudinal linear mixed models with complex covariance structures like ARH1, ante(1), AR(1), etc?

Here is a sample code:

```
proc mixed data = sas.aLTP covtest scoring=5;
class dose Mouse Time;
model ln_aLTP = dose cTime cTime*cTime / solution ddfm=sat influence;
repeated Time /type=ante(1) subject=mouse;
random int cTime / subject = mouse type = arh(1) solution g v;
lsmeans dose /adjust=dunnett pdiff=control('4') CL;
run;
```

I am aware of this thread for Cohen's d in proc mixed, which did not generate a solid resolution.

I'm also aware of this paper: A practical guide to calculating Cohen’s f2, a measure of local effect size, from PROC MIXED, but that covers only the simplest covariance structures. I have emailed various authors and still have not found a practical solution. So I'm wondering if any SAS users have somehow overcome this issue, since the community is large.

Some more specific questions to guide the discussion:

- If I want to at least report Cohen's d for the difference between two means in lsmeans, does it sound resonable to run a
*t*-test for those two means of interesy using the predicted values from proc mixed and then calculate Cohen's d? - As far as converting SE to SD, are there actual sources that validate such an approach? The thread referenced above mentions
**SD=SE*SQRT(DF+1)**. Is this a justifyable approach? - One of the authors who didn't have much time to explain his solution to me, said that he would compute the effect size as the ratio of the estimated fixed effect parameter to the sqrt of total marginal outcome (at a given timepoint if marginal variance isn't constant). It seems that SAS can produce the marginal variance output. Can anyone elaborate on this and perhaps give some tips about syntax? Do I need to add "marginal" or "var" somewhere in the syntax to get that output? I am not sure what would such an effect size be called in terms of the standardized variants, like Cohen's f sq, pseudo-r sq, or something else.

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

12-09-2016 11:45 PM

The reference to marginal variance likely refers to proc glimmix not proc mixed, but that's ok, both will estimate your models.

See

http://support.sas.com/resources/papers/proceedings16/11663-2016.pdf

for an example of specifying, exporting and calculating other statistics from proc glimmix covariance estimate output.

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

12-09-2016 11:59 PM

I think this posting is also relevant to your inquiry:

https://communities.sas.com/t5/SAS-Statistical-Procedures/Proc-Mixed/m-p/296000#U296000