Date: Mon, 13 Apr 2009 12:16:20 -0700
Reply-To: "Yaacov P." <ympetscher@GMAIL.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: "Yaacov P." <ympetscher@GMAIL.COM>
Organization: http://groups.google.com
Subject: PROC GLIMMIX and HLM comparison for Cross-classified
Content-Type: text/plain; charset=ISO-8859-1
Greetings,
I was wondering if anyone out there has looked at some comparisons
between GLIMMIX and HLM6 for cross-classified models. From the models
i have been running, it appears that when using method=mspl (as
opposed to default rspl) the variance components are a near identical
match (to .001 level). What is somewhat inconsistently different
between SAS and HLM6; however, are the estimates of the fixed effects.
Consider the code:
proc glimmix data=a.lvl noclprint method=mspl;
class a1_student_id letter;
model ls (event='1')= ln pa_total ln*pa_total /dist=binary solution
oddsratio ddfm=bw;
nloptions maxiter=10000;
random intercept /subject=a1_student_id;
random intercept /subject=letter;
covtest zerog /cl;
where ls ne 9;
run;
SAS output shows:
Solutions for Fixed Effects
Standard
Effect Estimate Error DF
t Value Pr > |t|
Intercept -3.6585 0.3052 25
-11.99 <.0001
ln 3.1963 0.2035
7272 15.71 <.0001
PA_TOTAL 0.03683 0.01716 7272
2.15 0.0319
ln*PA_TOTAL 0.04671 0.01579 7272
2.96 0.0031
HLM output shows:
Final estimation of fixed effects: (Unit-specific model)
------------------------------------------------------------------------------
Standard Approx.
Fixed Effect Coefficient Error T-ratio
d.f. P-value
------------------------------------------------------------------------------
For INTRCPT1, P0
INTERCEPT,theta0 -3.225108 0.274924 -11.731 7587
0.000
PA_TOTAL, G01 0.036828 0.017156 2.147 7587
0.032
For A1_LN, P1
INTERCEPT,theta1 3.745992 0.173212 21.627 7587
0.000
PA_TOTAL, G11 0.046714 0.015792 2.958 7587
0.004
------------------------------------------------------------------------------
In a more complex model that includes 2-way and 3-way interactions,
the results become more discrepant, varying by more than 1.0 log-odds.
Interestingly, the different in df for the intercept are quite
different between SAS (df = 25) and HLM (df = 7587).
Has anyone else encountered such findings? I haven't seen much in the
way of comparisons between programs, and while HLM software doesn't
appear to be designed to handle other 2-way and 3-way interactions
easily, I wonder the extent to which results are valid or meaningful.
Thanks for any insights!
Yaacov
