topic Re: combining HPMIXED and MIXED for individual cross-classified growth models in SAS Procedures

combining HPMIXED and MIXED for individual cross-classified growth models

huanaca — Sun, 03 Mar 2013 07:15:34 GMT

I have a large, longitudinal data set -- seven years of data for approximately 115,000 students (approximately 500,000 observations), attending approximately 600 schools. I am doing cross-classified individual growth modeling (cross-classified because students change schools). Since the data set is so large, I am estimating models with small (2 percent) subsamples of the data set. However, with the 2 percent subsamples, the data are sparse; and so PROC MIXED cannot estimate the covariance parameters. Therefore, I am using a combination of HPMIXED and MIXED. This way I can estimate the covariance parameters in HPMIXED, and then pass them along to MIXED. Then I can compute sequential F-tests and make adjustments for multiple comparisons in testing differences between LSMEANS.

The grouping variable of interest is ProFnc: 1=proficient, 0=non-proficient. When I put the grouping variable in the model with no interactions, the fixed effects produced by HPMIXED and MIXED are the same. However, HPMIXED estimates a fixed effect for ProFnc=1, where as MIXED estimates an effect for ProFnc=0.

Then, when I add to the model interactions with the grouping variable, the fixed effects seem to behave strangely. I don't know if I should ignore the fixed effects that HPMIXED produces, and just go with the fixed effects that MIXED produces. And if that is the case, can I trust the covariance parameters coming out of HPMIXED? Can I trust the fixed effects that MIXED is producing? If there is a problem, is there a change I can make in the code to fix this?

Below I have pasted the code and output for the two models I described above. Again, I first have the model with the grouping variable but no interactions; and second, I have the model with the grouping variable and its interactions. (Note: I realize these particular interactions are not significant, but I am trying to understand what SAS is doing here and which numbers I can trust.)

Thank you in advance for any help with this.

/* MODEL 1 - the grouping variable (ProFnc) and NO interactions */

/* HPMIXED */

PROC HPMIXED DATA=sub2pct_long noclprint;

CLASS stdpseudoid schcode ProFnc;

MODEL zELA = Timec|Timec|Timec|Timec

Female

PrEdClGrd

PrEdSmClHS

FRL

ProFnc

zCeldt /solution ;

RANDOM intercept Timec /subject=stdpseudoid type=un;

RANDOM intercept /subject=schcode type=un;

ODS OUTPUT covparms=hpmcov2pct;

RUN;

/* MIXED */

PROC MIXED DATA=sub2pct_long noclprint covtest lognote method=reml;

CLASS stdpseudoid schcode ProFnc;

MODEL zELA = Timec|Timec|Timec|Timec

Female

PrEdClGrd

PrEdSmClHS

FRL

zCeldt

ProFnc /solution htype=1;

RANDOM intercept Timec /subject=stdpseudoid type=un;

RANDOM intercept /subject=schcode type=un;

PARMS/PDATA=hpmcov2pct hold=1,2,3,4,5 noiter;

RUN;

THE HPMIXED PROCEDURE
Data Set	WORK.SUB2PCT_LONG
Response Variable	zELA
Estimation Method	Restricted Maximum Likelihood (REML)
Degrees of Freedom Method	Residual

Number of Observations Read	10175
Number of Observations Used	8898

Dimensions
G-side Cov. Parameters	4
R-side Cov. Parameters	1
Columns in X	12
Columns in Z	4969
Subjects (Blocks in V)	1

Optimization Information
Optimization Technique	Dual Quasi-Newton
Parameters in Optimization	4
Lower Boundaries	3
Upper Boundaries	0
Residual Variance	Profiled

Iteration History
Iterations	Evaluations	Objective Function	Change	Max Gradient
0	4	18310.276802	. 3333	425.1763
1	4	18306.028084	4.24871820	56.80605
2	3	18305.679162	0.34892203	74.52385
3	3	18305.05507	0.62409210	39.65128
4	4	18305.032432	0.02263798	38.72127
5	4	18304.982517	0.04991456	4.723682
6	3	18304.981792	0.00072531	0.027503
7	3	18304.981792	0.00000002	0.000091

Convergence criterion (GCONV=1E-8) satisfied.

Covariance Parameter Estimates
Cov Parm	Subject	Estimate
UN(1,1)	StdPseudoId	0.6356
UN(2,1)	StdPseudoId	-0.00929
UN(2,2)	StdPseudoId	0.01124
UN(1,1)	SchCode	0.009595
Residual		0.2287

Fit Statistics
-2 Res Log Likelihood	18305
AIC (smaller is better)	18315
AICC (smaller is better)	18315
BIC (smaller is better)	18305
CAIC (smaller is better)	18310
HQIC (smaller is better)	18305

Solution for Fixed Effects
Effect	ProFnc	Estimate	Standard Error	DF	t Value	Pr > \|t\|
Intercept		-0.1662	0.04025	8887	-4.13	<.0001
Timec		-0.02220	0.03404	8887	-0.65	0.5144
*TimecTimec**		-0.00221	0.02929	8887	-0.08	0.9398
*TimecTimecTimec*		0.002277	0.008339	8887	0.27	0.7848
*TimeTimeTimeTimec**		-0.00020	0.000738	8887	-0.26	OC0.7913
Female		0.05291	0.03625	8887	1.46	0.1444
PrEdClGrd		0.1948	0.09722	8887	2.00	0.0452
PrEdSmClHS		0.1982	0.04100	8887	4.83	<.0001
FRL		-0.04520	0.03694	8887	-1.22	0.2212
ProFnc	0	0	.	.	.	.
ProFnc	1	0.2171	0.03673	8887	5.91	<.0001
zCeldt		0.2361	0.01803	8887	13.10	<.0001

THE MIXED PROCEDURE
Data Set	WORK.SUB2PCT_LONG
Dependent Variable	zELA
Covariance Structures	Unstructured, Variance Components
Subject Effects	StdPseudoId, SchCode
Estimation Method	REML
Residual Variance Method	Parameter
Fixed Effects SE Method	Model-Based
Degrees of Freedom Method	Containment

Covariance Parameters	5
Columns in X	20
Columns in Z	4969
Subjects	1
Max Obs Per Subject	10175

Number of Observations
Number of Observations Read	10175
Number of Observations Used	8898
Number of Observations Not Used	1277

Parameter Search
CovP1	CovP2	CovP3	CovP4	CovP5	Res Log Like	-2 Res Log Like
0.6356	-0.00929	0.01124	0.009595	0.2287	-9152.4909	18304.9818

Covariance Parameter Estimates
Cov Parm	Subject	Estimate	Standard Error	Z Value	Pr Z
UN(1,1)	StdPseudoId	0.6356	0	.	.
UN(2,1)	StdPseudoId	-0.00929	0	.	.
UN(2,2)	StdPseudoId	0.01124	0	.	.
UN(1,1)	SchCode	0.009595	0	.	.
Residual		0.2287	0	.	.

Fit Statistics
-2 Res Log Likelihood	18305.0
AIC (smaller is better)	18305.0
AICC (smaller is better)	18305.0
BIC (smaller is better)	18305.0

Solution for Fixed Effects
Effect	ProFnc	Estimate	Standard Error	DF	t Value	Pr > \|t\|
Intercept		0.05087	0.04413	425	1.15	0.2497
Timec		-0.02220	0.03404	2040	-0.65	0.5144
*TimecTimec**		-0.00221	0.02929	4190	-0.08	0.9398
*TimecTimecTimec*		0.002277	0.008339	4190	0.27	0.7848
*TimeTimeTimeTimec**		-0.00020	0.000738	4190	-0.26	0.7913
Female		0.05291	0.03625	4190	1.46	0.1444
PrEdClGrd		0.1948	0.09722	4190	2.00	0.0452
PrEdSmClHS		0.1982	0.04100	4190	4.83	<.0001
FRL		-0.04520	0.03694	4190	-1.22	0.2212
ProFnc	0	-0.2171	0.03673	4190	-5.91	<.0001
ProFnc	1	0	.	.	.	.
zCeldt		0.2361	0.01803	4190	13.10	<.0001

Type I Tests of Fixed Effects
Effect	Num DF	Den DF	F Value	Pr > F
Timec	1	2040	0.66	0.4156
*TimecTimec**	1	4190	10.33	0.0013
*TimecTimecTimec*	1	4190	0.03	0.8713
*TimeTimeTimeTimec**	1	4190	0.01	0.9256
Female	1	4190	5.38	0.0204
PrEdClGrd	1	4190	3.38	0.0659
PrEdSmClHS	1	4190	31.73	<.0001
FRL	1	4190	0.56	0.4555
ProFnc	1	4190	33.85	<.0001
zCeldt	1	4190	171.49	<.0001

/* MODEL 2 - the grouping variable (ProFnc), AND its interactions */

/* HPMIXED */

PROC HPMIXED DATA=sub2pct_long noclprint;

CLASS stdpseudoid schcode ProFnc;

MODEL zELA = Timec|Timec|Timec|Timec

Female

PrEdClGrd

PrEdSmClHS

FRL

zCeldt

ProFnc

ProFnc*Female

ProFnc*PrEdClGrd

ProFnc*PrEdSmClHS

ProFnc*zCeldt /solution ;

RANDOM intercept Timec /subject=stdpseudoid type=un;

RANDOM intercept /subject=schcode type=un;

ODS OUTPUT covparms=hpmcov2pct;

RUN;

/* MIXED */

PROC MIXED DATA=sub2pct_long noclprint covtest lognote method=reml;

CLASS stdpseudoid schcode ProFnc;

MODEL zELA = Timec|Timec|Timec|Timec

Female

PrEdClGrd

PrEdSmClHS

FRL

ProFnc

zCeldt

ProFnc*Female

ProFnc*PrEdClGrd

ProFnc*PrEdSmClHS

ProFnc*zCeldt /solution htype=1;

RANDOM intercept Timec /subject=stdpseudoid type=un;

RANDOM intercept /subject=schcode type=un;

PARMS/PDATA=hpmcov2pct hold=1,2,3,4,5 noiter;

RUN;

THE HPMIXED PROCEDURE
Data Set	WORK.SUB2PCT_LONG
Response Variable	zELA
Estimation Method	Restricted Maximum Likelihood (REML)
Degrees of Freedom Method	Residual

Number of Observations Read	10175
Number of Observations Used	8898

Dimensions
G-side Cov. Parameters	4
R-side Cov. Parameters	1
Columns in X	20
Columns in Z	4969
Subjects (Blocks in V)	1

Optimization Information
Optimization Technique	Dual Quasi-Newton
Parameters in Optimization	4
Lower Boundaries	3
Upper Boundaries	0
Residual Variance	Profiled

Iteration History
Iteration	Evaluations	Objective Function	Change	Max Gradient
0	4	18320.233278	.	430.4562
1	4	18315.885135	4.34814326	57.74109
2	3	18315.533274	0.35186101	75.8847
3	3	18314.895565	0.63770890	40.89073
4	4	18314.872387	0.02317834	39.85029
5	4	18314.819313	0.05307395	3.502647
6	3	18314.818918	0.00039410	0.020809
7	3	18314.818918	0.00000001	0.000075

Convergence criterion (GCONV=1E-8) satisfied.

Covariance Parameter Estimates
CovParm	Subject	Estimate
UN(1,1)	StdPseudoId	0.6363
UN(2,1)	StdPseudoId	-0.00940
UN(2,2)	StdPseudoId	0.01123
UN(1,1)	SchCode	0.009609
Residual		0.2287

Fit Statistics
-2 Res Log Likelihood	18315
AIC (smaller is better)	18325
AICC (smaller is better)	18325
BIC (smaller is better)	18315
CAIC (smaller is better)	18320
HQIC (smaller is better)	18315

Solution for Fixed Effects
Effects	ProFnc	Estimate	Standard Error	DF	t Value	Pr > \|t\|
Intercept		-0.1545	0.04406	8883	-3.51	0.0005
Timec		-0.02224	0.03404	8883	-0.65	0.5136
*TimecTimec**		-0.00219	0.02929	8883	-0.07	0.9404
*TimecTimecTimec*		0.002277	0.008339	8883	0.27	0.7848
*TimeTimeTimeTimec**		-0.00020	0.000738	8883	-0.27	0.7909
Female		0	.	.	.	.
PrEdClGrd		0	.	.	.	.
PrEdSmClHS		0	.	.	.	.
FRL		-0.04623	0.03708	8883	-1.25	0.2125
ProFnc	0	0	.	.	.	.
ProFnc	1	0.1951	0.05881	8883	3.32	0.0009
zCeldt		0	.	.	.	.
*FemaleProFnc**	0	0.06340	0.04752	8883	1.33	0.1822
*FemaleProFnc**	1	0.03835	0.05611	8883	0.68	0.4943
*PrEdClGrdProFnc**	0	0.1479	0.1253	8883	1.18	0.2377
*PrEdClGrdProFnc**	1	0.2608	0.1544	8883	1.69	0.0911
*PrEdSmClHSProFnc**	0	0.1508	0.05290	8883	2.85	0.0044
*PrEdSmClHSProFnc**	1	0.2672	0.06477	8883	4.13	<.0001
*zCeldtProFnc**	0	0.2515	0.02414	8883	10.42	<.0001
*zCeldtProFnc**	1	0.2173	0.02710	8883	8.02	<.0001

THE MIXED PROCEDURE
Data Set	WORK.SUB2PCT_LONG
Dependent Variable	zELA
Covariance Structures	Unstructured, Variance Components
Subject Effects	StdPseudoId, SchCode
Estimation Method	REML
Residual Variance Method	Parameter
Fixed Effects SE Method	Model-Based
Degrees of Freedom Method	Containment

Dimensions
Covariance Parameters	5
Columns in X	20
Columns in Z	4969
Subjects	1
Max Obs Per Subject	10175

Number of Observations Read	10175
Number of Observations Used	8898
Number of Observations Not Used	1277

Parameter Search
CovP1	CovP2	CovP3	CovP4	CovP5	Res Log Like	-2 Res Log Like
0.6363	-0.00940	0.01123	0.009609	0.2287	-9146.8531	18293.7063

Covariance Parameter Estimates
Cov Parm	Subject	Estimate	Standard Error	Z Value	Pr Z
UN(1,1)	StdPseudoId	0.6363	0	.	.
UN(2,1)	StdPseudoId	-0.00940	0	.	.
UN(2,2)	StdPseudoId	0.01123	0	.	.
UN(1,1)	SchCode	0.009609	0	.	.
Residual		0.2287	0	.	.

Fit Statistics
-2 Res Log Likelihood	18293.7
AIC (smaller is better)	18293.7
AICC (smaller is better)	18293.7
BIC (smaller is better)	18293.7

Solution for Fixed Effects
Effect	ProFnc	Estimate	Standard Error	DF	t Value	Pr > \|t\|
Intercept		0.03238	15327	425	0.00	1.0000
Timec		-0.02224	0.03404	2040	-0.65	0.5136
*TimecTimec**		-0.00219	0.02929	4190	-0.07	0.9404
*TimecTimecTimec*		0.002277	0.008339	4190	0.27	0.7848
*TimeTimeTimeTimec**		-0.00020	0.000738	4190	-0.27	0.7909
Female		0.03835	0.05611	4190	0.68	0.4943
PrEdClGrd		0.2608	0.1544	4190	1.69	0.0911
PrEdSmClHS		0.2672	0.06477	4190	4.13	<.0001
FRL		-0.04623	0.03708	4190	-1.25	0.2125
ProFnc	0	-0.1869	15327	4190	-0.00	1.0000
ProFnc	1	0.008199	15327	4190	0.00	1.0000
zCeldt		0.2173	0.02710	4190	8.02	<.0001
*FemaleProFnc**	0	0.02505	0.07352	4190	0.34	0.7333
*FemaleProFnc**	1	0	.	.	.	.
*PrEdClGrdProFnc**	0	-0.1129	0.1988	4190	-0.57	0.5701
*PrEdClGrdProFnc**	1	0	.	.	.	.
*PrEdSmClHSProFnc**	0	-0.1165	0.08351	4190	-1.39	0.1631
*PrEdSmClHSProFnc**	1	0	.	.	.	.
*zCeldtProFnc**	0	0.03423	0.03626	4190	0.94	0.3453
*zCeldtProFnc**	1	0	.	.	.	.

Type I Tests of Fixed Effects
Effect	Num DF	Den DF	F Value	Pr > F
Timec	1	2040	0.63	0.4270
*TimecTimec**	1	4190	10.39	0.0013
*TimecTimecTimec*	1	4190	0.01	0.9113
*TimeTimeTimeTimec**	1	4190	0.01	0.9377
Female	1	4190	5.35	0.0208
PrEdClGrd	1	4190	3.39	0.0657
PrEdSmClHS	1	4190	31.90	<.0001
FRL	1	4190	0.52	0.4690
ProFnc	1	4190	33.97	<.0001
zCeldt	1	4190	171.07	<.0001
*FemaleProFnc**	1	4190	0.10	0.7482
*PrEdClGrdProFnc**	1	4190	0.14	0.7094
*PrEdSmClHSProFnc**	1	4190	1.74	0.1877
*zCeldtProFnc**	1	4190	0.89	0.3453

Re: combining HPMIXED and MIXED for individual cross-classified growth models

SteveDenham — Mon, 04 Mar 2013 13:01:03 GMT

Without going into a lot of other interesting things, one way of putting the two procs "on the same" page would be the use of the ref= option in the class statement. Perhaps something like:

CLASS stdpseudoid schcode ProFnc(ref='1');

for both procs. This would put MIXED on the same reference as HPMIXED. Once you do this, I think you will be satisfied with all of the other things you are doing in moving the covariance parameters over to MIXED. It would be nice if the STORE statement was available for HPMIXED, as then all of the adjustments you are using MIXED for would be available directly for the HPMIXED output. Good luck with this project, and let us know if this helps at all.

Steve Denham

Re: combining HPMIXED and MIXED for individual cross-classified growth models

huanaca — Mon, 04 Mar 2013 17:10:42 GMT

Thanks very much, Steve.

I must be missing something. ProFnc is a numeric variable (0 or 1). I tried making 0 the ref group, but received a syntax error. Please see below:

NOTE: The SAS System stopped processing this step because of errors.

PROC HPMIXED DATA=sub2pct_long noclprint;

CLASS stdpseudoid schcode ProFnc(ref=0);

ERROR 22-322: Syntax error, expecting one of the following: a name, ;, -, /, :, _ALL_,

_CHARACTER_, _CHAR_, _NUMERIC_.

ERROR 76-322: Syntax error, statement will be ignored.

MODEL zELA = Timec|Timec|Timec|Timec

Female

PrEdClGrd

PrEdSmClHS

FRL

ProFnc

zCeldt /solution htype=1;

RANDOM intercept Timec /subject=stdpseudoid type=un;

RANDOM intercept /subject=schcode type=un;

PARMS /PDATA=hpmcov2pct hold=1,2,3,4,5 noiter;

RUN;

Re: combining HPMIXED and MIXED for individual cross-classified growth models

SteveDenham — Mon, 04 Mar 2013 17:35:34 GMT

The ref= syntax requires the level to be quoted, whether it is character or numeric in nature. Try

CLASS stdpseudoid schcode ProFnc(ref='0');

making sure to quote the zero.

Steve Denham

Re: combining HPMIXED and MIXED for individual cross-classified growth models

huanaca — Mon, 04 Mar 2013 17:51:03 GMT

Still no luck. I got the same error...

PROC HPMIXED DATA=sub2pct_long noclprint;

CLASS stdpseudoid schcode ProFnc(ref='0');

ERROR 22-322: Syntax error, expecting one of the following: a name, ;, -, /, :, _ALL_,

_CHARACTER_, _CHAR_, _NUMERIC_.

ERROR 76-322: Syntax error, statement will be ignored.

MODEL zELA = Timec|Timec|Timec|Timec

Female

PrEdClGrd

PrEdSmClHS

FRL

ProFnc

zCeldt /solution ;

RANDOM intercept Timec /subject=stdpseudoid type=un;

RANDOM intercept /subject=schcode type=un;

ODS OUTPUT covparms=hpmcov2pct;

RUN;

Re: combining HPMIXED and MIXED for individual cross-classified growth models

SteveDenham — Mon, 04 Mar 2013 18:31:23 GMT

Two possible workarounds:

In HPMIXED (this seems to have a real finicky taste for the ref= option):

class stdpseudoid schcode ProFnc(ref=LAST);

and leave MIXED unchanged, or

In MIXED:

class stdpseudoid schcode ProFnc(ref='0');

and leave HPMIXED unchanged, or at least without the ref= option.

I noticed something else--the fixed effects are not listed in exactly the same order in the two PROCS. I am just curious as to whether that may be affecting things, as might the htype=1 hypothesis tests. Given the likely unbalanced nature and the order dependence of Type 1 hypotheses, this may also be something to be concerned about.

Steve Denham

Re: combining HPMIXED and MIXED for individual cross-classified growth models

huanaca — Mon, 04 Mar 2013 19:21:54 GMT

Still no luck.

First, I tried the following:

PROC HPMIXED DATA=sub2pct_long noclprint;

CLASS stdpseudoid schcode ProFnc (ref=LAST);

MODEL zELA = Timec|Timec|Timec|Timec

Female

PrEdClGrd

PrEdSmClHS

FRL

ProFnc

zCeldt /solution ;

RANDOM intercept Timec /subject=stdpseudoid type=un;

RANDOM intercept /subject=schcode type=un;

ODS OUTPUT covparms=hpmcov2pct;

RUN;

But received the same error message for HPMIXED:

CLASS stdpseudoid schcode ProFnc (ref=LAST);

ERROR 22-322: Syntax error, expecting one of the following: a name, ;, -, /, :, _ALL_,

_CHARACTER_, _CHAR_, _NUMERIC_.

ERROR 76-322: Syntax error, statement will be ignored.

Then, I tried the following:

PROC HPMIXED DATA=sub2pct_long noclprint;

CLASS stdpseudoid schcode ProFnc;

MODEL zELA = Timec|Timec|Timec|Timec

Female

PrEdClGrd

PrEdSmClHS

FRL

ProFnc

zCeldt /solution ;

RANDOM intercept Timec /subject=stdpseudoid type=un;

RANDOM intercept /subject=schcode type=un;

ODS OUTPUT covparms=hpmcov2pct;

RUN;

PROC MIXED DATA=sub2pct_long noclprint covtest lognote method=reml;

CLASS stdpseudoid schcode ProFnc (ref='0');

MODEL zELA = Timec|Timec|Timec|Timec

Female

PrEdClGrd

PrEdSmClHS

FRL

ProFnc

zCeldt /solution htype=1;

RANDOM intercept Timec /subject=stdpseudoid type=un;

RANDOM intercept /subject=schcode type=un;

PARMS /PDATA=hpmcov2pct hold=1,2,3,4,5 noiter;

RUN;

But I received the same error message for MIXED:

CLASS stdpseudoid schcode ProFnc (ref='0');

ERROR 22-322: Syntax error, expecting one of the following: a name, ;, -, /, :, _ALL_,

_CHARACTER_, _CHAR_, _NUMERIC_.

ERROR 76-322: Syntax error, statement will be ignored.

Re: combining HPMIXED and MIXED for individual cross-classified growth models

SteveDenham — Tue, 05 Mar 2013 13:39:18 GMT

This is very curious and unexpected behavior. I have used the ref= option in the class statement of both procedures and have not run into this error, other than when the value was unquoted. FIRST and LAST always seem to work Since the other class variables are used only as subjects in the random statements, perhaps the global version would work:

CLASS stdpseudoid schcode ProFnc /ref=FIRST;

Steve Denham

Re: combining HPMIXED and MIXED for individual cross-classified growth models

huanaca — Tue, 05 Mar 2013 15:19:58 GMT

Thanks, Steve. I tried the global version and received the same error message. I think the reason is I have version 9.3 TS1M0, not TS1M2 (as specified below):

" NOTE : The REF= option for setting reference levels was added to the GLM, MIXED, GLIMMIX, and ORTHOREG beginning in SAS 9.3 TS1M2."

Can I trust the random effects from HP even where the fixed effects seem weird (i.e. even when it is estimating parameters for both categories of a dichotomous CLASS variable) ?

Re: combining HPMIXED and MIXED for individual cross-classified growth models

SteveDenham — Tue, 05 Mar 2013 15:28:03 GMT

Ah, a much easier question. Because your errors are gaussian, I believe that the random effect estimates are good. Proceeding from the fixed estimates, differences between levels were constant between the two methods, so variance estimates should not be affected. I still advise against htype=1, due to likely unbalance, and the potential for order entry effects as a result. (Although somewhere, Frank Harrell is rolling his eyes in disgust at this advice).

Steve Denham

Re: combining HPMIXED and MIXED for individual cross-classified growth models

huanaca — Tue, 05 Mar 2013 16:15:34 GMT

Thanks very much for your help. Much appreciated.

Re: combining HPMIXED and MIXED for individual cross-classified growth models

huanaca — Fri, 08 Mar 2013 17:37:41 GMT

Hi Steve,

I have a follow-up question I was hoping you might be able to answer...

Is it possible to get MIXED to generate hypothesis tests for covariance parameters even when it's prevented from iterating on them when given values from HPMIXED? It seems to go back to solving for the random effects, which is what takes all the time and memory, and what I'm trying to avoid by using HPMIXED and MIXED.

thank you.

Re: combining HPMIXED and MIXED for individual cross-classified growth models

SteveDenham — Fri, 08 Mar 2013 17:58:34 GMT

What sort of tests do you have in mind--being cognizant of the distributional assumptions regarding estimates of variance, etc? Once upon a time, MIXED automatically "tested" whether the individual variance estimates were different from zero as part of the default output, using an asymptotic normal assumption. Those tests were, well, wrong. Sometimes Wald tests just do not apply.

Anyway, PROC GLIMMIX enables some specific kinds of G side testing. But I would avoid tests if at all possible and consider instead the estimation approach, looking at confidence bounds on the parameters. Those should be available from HPMIXED. When you fix/hold/noiter variance-covariances, I don't think there will be any way to calculate bounds or tests on the parameters. The blup's, on the other hand, can be compared, using the ESTIMATE statement.

Steve Denham

Re: combining HPMIXED and MIXED for individual cross-classified growth models

huanaca — Fri, 08 Mar 2013 21:25:35 GMT

Well I was thinking about testing whether individual variance estimates were different from zero, using covtest. But if I were to look at the confidence bounds, I would do so with ' cl ' after the slash in the random statement? And to prevent SAS from outputting confidence intervals for every student?

RANDOM intercept Timec /subject=stdpseudoid type=un cl;

RANDOM intercept /subject=schcode type=un;

One last question if I could... Do you have any thoughts on whether it's useful/appropriate to include school-level covariates in the model even if a very small percentage (definitely less than 5%, and maybe closer to 2%) of variance is between schools? An argument in favor of including school-level covariates is that the composition of schools could still be affecting individual student performance, even if the average performance across schools doesn't vary that much.

Re: combining HPMIXED and MIXED for individual cross-classified growth models

SteveDenham — Mon, 11 Mar 2013 12:12:42 GMT

So the code with cl in it should only give three estimates (intercept, timec and their covariance) with confidence bounds. Only if you specify solution, or do the blups, would I expect to get a confidence bound for each student.

I always try to include design covariates, and the school-level covariates strike me as design based in the following sense: If you set up a designed experiment to look at these effects, you would include them. In this case, not including them would probably increase the standard errors and reduce the precision of your estimates.

Steve Denham

Re: combining HPMIXED and MIXED for individual cross-classified growth models

huanaca — Wed, 13 Mar 2013 16:40:50 GMT

Thanks again, Steve.