Mixed models

[BlueNose]

Without the noint statement I get exactly the tables above.

An intercept for each level is expected, since the model is of the form: log(Pr(y<=j)/Pr(y>j)) = aj+bx

I think it makes sense to get 4 intercepts (the fifth has no meaning). In the table above there are 2 slopes. I have 3 treatments, which are 2 dummy variables, so that also makes sense in a way. What makes you suspect that the results are not reasonable ?

The results are in the same direction of the descriptive statistics, however a few things concern me:

a. Odds ratios saying that the new treatment is 100 times better, well, 100 sounds a little bit too much

b. standard errors of the covariance parameters vary strongly when I change models. If I choose only animalsID I get S.E of approximately 1, when I take AnimalID, location and technique I get 6 and if only animalID and technique I get 14.

c. How to choose the right model. Both from the S.E point of view, and from any IC point of view, the model with only AnimalID is best...which means ignoring completely the technique and location. However if I want to choose from the other two options, AIC and BIC give opposite answers.

May I run the model as if my response is Gaussian just to get LSMEANS for descriptive purposes ?

[SteveDenham]

Looks like my response got eaten by the Flying Spaghetti Monster.

I'll try again.

Estimate statements should give the odds ratios you are looking for.

For example:

estimate 'Score=0 v Score=1, Control 1 v New treatment' intercept -1 1 0 0 0 treatment -1 0 1 /exp cl;

estimate 'Score=0 v Score=(1 or 2), Control 1 v New treatment' intercept -1 0 1 0 0 treatnent -1 0 1 /exp cl;

estimate 'Score=(1 or 2) v Score=(1,2 or 3), Control 2 v New treatment' intercept 0 -1 1 0 0 treatment 0 -1 1 /exp cl;

Whatever odds ratios of interest you may have can be calculated this way.  These are only some of the possibilities.

I think the size of the standard errors really depends on the number of experimental units as well as the variance estimates.  My general feeling on random effects is that if they are design factors AND they contribute a non-negligible variance component, they should be included, no matter what the IC might say. I guess this is the "Design factors trump all others" school of thought, and I think it is critical when denominator degrees of freedom can't be adjusted, when the method is adaptive quadrature or Laplace.

Lastly, regarding LSMEANS.  If the scores are relatively evenly spaced, then the lsmeans would have some descriptive ability.  But what if the scores are ways of summarizing health, with 0=no problems and 5=dead.  Is a 2.5 that you get from averaging a 0 and a 5 really the same as the 2.5 you get from averaging a 2 and a 3?  Probably not, or we would never have wandered down this road of ordinal responses.

Good luck.

Steve Denham

[BlueNose]

Thank you for keep replying, this has become a Flying Spaghetti Monster.

I have used the estimate statement, but only now I see the difference, you specified also the intercepts while I specified only treatments.

I did what you suggested, and the last question I have for this monster thread is how do I interpret such a thing ? I thought I figured out how to interpret it until you added the intercept into the picture 🙂

My output is:

Label	Estimate	S.E	DF	t	Pr>t	Alpha	Lower	Upper	exp(estimate)	exp lower	exp upper
Score=0 vs Score=1, Control #1-New Treatment	-2.9764	0.9793	25	-3.04	0.0055	0.05	-4.9933	-0.9595	0.05098	0.006783	0.3831
Score=0 vs Score=(1 or 2), Control #1-New Treatment	-2.2188	0.9109	25	-2.44	0.0223	0.05	-4.0948	-0.3428	0.1087	0.01666	0.7098

Cov Parameter	Subject	estimate	S.E
Intercept	Animal_ID*Location	1.2534	6.8323
Intercept	Animal*Technique*Locati	0.06052	6.7719

Treatment	Treatment	estimate	DF	lower	upper
Control 1	New Treatment	0.010	25	<0.001	0.103
Control 2	New Treatment	0.049	25	0.007	0.335

As for the LSMEANS, you are right of course, my response is numeric with meaning, that's why I even thought of presenting averages, I wouldn't average a nominal or ordinal with no numeric value.

Thanks a lot Steve, I learned a lot from this thread !!

[SteveDenham]

Well, exp(estimate) is the odds ratio for the given comparison.  You can set up as many as needed, for the comparisons of interest.  The first says that the odds ratio for a score of 1 compared to a score of 0 for the contrast of control 1 to new treatment is 0.05.  Other odds ratios can be derived.  It all depends on the comparisons you need to make.

Steve Denham

[BlueNose]

Thank you Steve for all your assistance, I think I understand now how to make interpretation of this kind of output.

The only thing I don't understand is why you wrote:

estimate 'Score=0 vs Score=(1 or 2), Control 1- New Treatment' intercept -1 0 1 0 0 Treatment 1 0 -1/cl or;

and not

estimate 'Score=0 vs Score=(1 or 2), Control 1- New Treatment' intercept -1 0.5 0.5 0 0 Treatment 1 0 -1/cl or;

since these are contrasts. my guess is that it is related to the cumulative log, but the rational still isn't entirely clear.

Thanks for everything !

P.S Regarding the first OR you gave me, the simple one, I actually think I managed to reconstruct what SAS did manually.

[SteveDenham]

Your guess is correct--because the link is the cumulative log, the intercepts are for cumulative comparisons.  Not real sure, but if you were fitting a generalized logit link, your latter estimate might be more appropriate.  Perhaps someone ( or 1zmm) could help out here.

Steve Denham