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
- /
- Non-estimable lsmeans from glm

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

02-11-2013 06:02 AM

Hi everyone.

Could I please ask how can I resolve this?

I wish to get the lsmeans of Intervation but SAS gives me a warning saying 'WARNING: The ADJUST=DUNNETT p-values and confidence limits cannot be computed because the Dunnett-Hsu approximation did not converge. Try ADJUST=SIMULATE.' after I run the below glm model. I tried the ADJUST=SIMULATE but it gives me an error saying 'ERROR: Cannot compute simulated p-values with inestimable LSMEANS.'. I have about 60000 data points with up to 6 categories of Intervention and up to 7 categories of experimentNum. Their interaraction term is significant <0.0001. There are zero observations for some combination groups of Intervention*experimentNum. However, the second lsmeans statement runs without problem.

Your help/insight is greatly appreciated!!

proc glm data=tmp order=internal plots=none;

class Intervention experimentNum;

model logconc=Intervention|experimentNum / solution e;

lsmeans Intervention / cl stderr e pdiff=control('No control');

lsmeans Intervention*experimentNum / cl pdiff stderr e om adjust=tukey;

run; quit;

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

02-11-2013 07:58 AM

The main effect marginal means for intervention cannot be calculated, for as you point out, there are "empty cells". The interaction LSmeans are presented, because that is the level of observation, essentially. Note that certain combinations (those with N=0) do not appear. The best way around this is to fit a means model (see Milliken and Johnson's Analysis of Messy Data). To get what you need in GLM, you would have to construct ESTIMATE statements that combine the non-zero values to get main effect means, and then take differences. Unfortunately, the ESTIMATE statement does not allow for multiple comparisons.

Enter PROC MIXED (or GLIMMIX), and the LSMESTIMATE statement. This will enable calculation of main effect means and differences, in a way that control for multiple comparisons can be handled. The statements may be complex, getting 42 possible interaction means combined as needed.

Steve Denham

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

02-11-2013 06:31 PM

Thanks Steve for your reply! Much appreciated! I'll give it a go as you suggested.