Re: Differing results with PROC GENMOD and PROC MIXED

annekolean@YAHOO.COM wrote:
>Hi, I fitted a model for a continuous outcome (range
>0-7) using both proc genmod and proc mixed and the
>results (ie. parameter estimates and signifcance
>values) are completely different. I was told that the
>two methods should give similar results, is that not
>the case?
>
>My data comes from a longitudinal study (12 weeks)
>with 135 subjects and two treatment arms. There were 4
>factors that I was interested in: tx (A or B), week
>(continuous), baseline value of outcome (continuous),
>and a categorical improvement measure cov1
>(improved/not improved). Here's my proc mixed
>specification:
>
>
>proc mixed data=midata;
>class id tx wave cov1;
>model y = tx week ybl cov1
> tx*ybl tx*cov1 week*ybl ybl*cov1
> ybl*cov1*tx/solution outpred=allpred;
>repeated wave/ subject=id(tx) type=ar(1) rcorr;
>run;
>
>
>This model gives me the following type3:
>
>Effect DF DF F Value Pr > F
>tx 1 121 0.01 0.9104
>Week 1 930 0.20 0.6530
>ybl 1 121 79.29 <.0001
>cov1 1 121 0.05 0.8212
>ybl*tx 1 121 0.50 0.4823
>tx*cov1 1 121 4.50 0.0360
>Week*ybl 1 930 5.75 0.0166
>ybl*cov1 1 121 11.61 0.0009
>ybl*tx*cov1 1 121 4.14 0.0441
>
>
>Here's my genmod specification:
>
>
>proc genmod data=midata;
>class id tx cov1;
>model y = tx week ybl cov1
> tx*ybl tx*cov1 week*ybl ybl*cov1
> ybl*cov1*tx/dist=normal type3;
>repeated subject=id(tx)/type=ar(1) covb corrw;
>output out=genout pred=preds reschi=resid;
>run;
>
>And here are the type3 score statistics:
>
>Source DF Chi-Square Pr > ChiSq
>
>tx 1 0.01 0.9155
>Week 1 0.54 0.4611
>ybl 1 24.45 <.0001
>cov1 1 0.09 0.7701
>ybl*tx 1 0.16 0.6879
>tx*cov1 1 3.74 0.0532
>Week*ybl 1 4.50 0.0340
>ybl*cov1 1 3.68 0.0551
>ybl*tx*cov1 1 1.48 0.2234
>
>
>For one, the 3-way interaction is no longer
>significant....

I'll disagree here. I think your results are very consistent. Your 3-way
interaction may be driven more by errors and noise than by real effects.
Everything else looks surprisingly close. However, an artificial
less-or-more-
than-0.05 call will point out what look to me to be really minor differences
in places. If you call p-values of .053 and .036 'different', then you are
making an arbitrary judgment.

>Is there a problem with the specification? Are the two
>models not equivalent? If they are, which result is
>more valid??

[1] Is there a problem with the model specifications? I can't tell. You
have a
complex model, with no residual diagnostics and no residual plots and no
evaluation
of the underlying assumptions. In the PROC MIXED formulation, you are
assuming
a simple autoregressive time series structure. In the PROC GENMOD
formulation,
you are using GEE with an analogous covariance structure. Perhaps neither
is
reaosnable, given the data. I *do* suspect that your DIST=NORMAL option
is not correct. Is Y *really* continuous, or does it just have a number of
different
possible values? You have stated that it is bound between 0 and 7, so it
doesn't
meet the specs I would expect to see in a real normally distributed
variable.

[2] Are the two methods not equivalent? They are not. See my above
discussion. And be apprised that the underlying methods used to fit the
models
are not exactly the same, either.

[3] If they are, which result is more valid?? I'm going to guess here that
NEITHER is valid. I'm going to guess that the normal errors are not
correct.
I'm going to guess that the AR(1) model may not be reasonable. I'm going
to guess that when you do your diagnostics, you'll find at least one outlier
or leverage point which could distort your results.

HTH,
David
--
David L. Cassell
mathematical statistician
Design Pathways
3115 NW Norwood Pl.
Corvallis OR 97330

__________________________________________________ _______________
FREE pop-up blocking with the new MSN Toolbar – get it now!
http://toolbar.msn.click-url.com/go/...ave/direct/01/