[S] differences SAS PROCMIXED and lme

Siem Heisterkamp (S.H.HEISTERKAMP@amc.uva.nl)
Thu, 20 Aug 1998 16:58:01 +0200


dear all,
I have allarmig differnces between my interpretaion of SAS output and
S-PLUS lme concerning standard errors of the predicted means. The means
themselves were slightly different.
=====
this is the SPLUS output

:\necosad\_data > summary(basiscat.lme,re=F)
Call:
Fixed: PCS ~ tijd * CAT + PCSBAS
Random: ~ 1
Cluster: ~ NECNUM
Data: rmspl3m

Estimation Method: RML
Convergence at iteration: 3
Restricted Loglikelihood: -1171.087
Restricted AIC: 2360.174
Restricted BIC: 2395.049

Variance/Covariance Components Estimate(s):
Structure: compound symmetry
Standard Deviation(s) of Random Effect(s)
(Intercept)
4.484634

Cluster Residual Variance: 29.97762

Fixed Effects Estimate(s):
Value Approx. Std.Error z ratio(C)
(Intercept) 8.5752692 2.68019729 3.1994918
tijd12 -0.6220664 0.92257753 -0.6742701
tijd18 -1.4514370 0.93184751 -1.5575906
CAT -2.0162583 1.28861974 -1.5646651
PCSBAS 0.8127410 0.06218279 13.0701923
tijd12CAT 0.1776739 1.44085896 0.1233111
tijd18CAT -1.1188441 1.46182566 -0.7653745

Conditional Correlation(s) of Fixed Effects Estimates
(Intercept) tijd12 tijd18 CAT
PCSBAS tijd12CAT
tijd12 -0.165213876

tijd18 -0.152608131 0.468059448

CAT -0.070255654 0.331028476 0.329190675

PCSBAS -0.953022168 0.005974993 -0.005587017 -0.126515797

tijd12CAT 0.110817754 -0.640328452 -0.299667517 -0.532445421
-0.009105626
tijd18CAT 0.086962704 -0.298301965 -0.637515103 -0.525330893
0.014388136 0.475987198

cluster means at 42.11 of the covariate PCSBAS
by using tapply() on predict.lme( ) output

1 2
6 42.79979 40.65197
12 42.31673 40.22127
18 41.62779 37.98649

"se" by using tapply() etc
sqrt(var.basis/n.basis)
1 2
6 0.3906523 0.5298230
12 0.4239170 0.5609732
18 0.4212500 0.5854842

numbers:
n.basis
[1] 76 68 66 51 49 46

Using SAS PROCMIXED we have exactly the same fit, same se for the
effects but small differences in LSMEANS
(we checked that the means were computed at the same level of PCSBAS)
and LARGE differences in the se's

to be precise: The SAS estimates for the SE are almost 2 times as large.
What do I wrong in S-PLUS?
(or is my interpretation of the LSMEANS SAS-output wrong)

The syntax in SAS was:

proc sort data=a.rmspl3m; by descending cat descending tijd; run;

proc mixed data=a.rmspl3m order=data method=reml;
class cat necnum tijd;
title2 "ANOVA - MIXED - CS - Serial meas.";
model pcs = cat|tijd pcsbas / solution ;
repeated tijd/ type=cs sub=necnum(cat) r;
lsmeans cat|tijd / e;
run;

basiscat.lme<-lme(fixed=PCS~tijd*CAT+PCSBAS,cluster=~NECNUM,random=~1,da
ta=rmspl3m,est.method="RML",na.ac

Dr. S.H. Heisterkamp
University of Amsterdam
Department of Clinical Epidemiology and Biostatistics
room J2-220
PO Box 22700 1100 DE Amsterdam
tel: +31-(0)-20-5668520
fax:+31-(0)-20-6912683
s.h.heisterkamp@amc.uva.nl

-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news