topic Re: Large Gradient Values for PROC NLMIXED in Statistical Procedures

Large Gradient Values for PROC NLMIXED

Mohsen — Tue, 22 Dec 2015 23:56:22 GMT

Hi All,

I am running a nlimixed procedure in the code below but getting large Gradient values. Any tip to address it is really appreciated.

SAS Code:

title"NLMIXED model";
proc nlmixed data=library.projectfulldata ;
parms intercept=0 s0=1 b_age=0 b_gender=0 b_NUMGO=0 b_meandelay=0 b_meanrt=0 b_ADHD1=0 b_OCD1=0 b_ASD1=0 b_ADHDOCD1=0 b_ADHDASD1=0 b_stoperror1=0 b_goerror1=0 b_noresponse1=0 b_prepush1=0 b_stopcorrect1=0;
eta=exp (intercept+ (b_age)*age+(b_gender)*gender+ (b_NUMGO)*NUMGO+(b_meandelay)*meandelay+ (b_meanrt)*meanrt+(b_ADHD1)*ADHD1+(b_OCD1)*OCD1 +(b_ASD1)*ASD1 +(b_ADHDOCD1)*ADHDOCD1 +(b_ADHDASD1)*ADHDASD1 +(b_stoperror1)*stoperror1+(b_goerror1)*goerror1 +(b_noresponse1)*noresponse1 +(b_prepush1)*prepush1 +(b_stopcorrect1)*stopcorrect1 + u0);
p=eta/(1+eta);
model stopcorrect~binary(p);
random u0~N(0,s0) subject=id ;
run;

SAS OUTPUT:

SAS Output

Specifications

Data Set	LIBRARY.PROJECTFULLDATA
Dependent Variable	stopcorrect
Distribution for Dependent Variable	Binary
Random Effects	u0
Distribution for Random Effects	Normal
Subject Variable	id
Optimization Technique	Dual Quasi-Newton
Integration Method	Adaptive Gaussian Quadrature

Dimensions

Observations Used	327309
Observations Not Used	0
Total Observations	327309
Subjects	13696
Max Obs per Subject	24
Parameters	17
Quadrature Points	3

Parameter Estimates Parameter Estimate Standard Error DF t Value Pr > |t| Alpha Lower Upper Gradient

0.007191	0.03493	14E3	0.21	0.8369	0.05	-0.06127	0.07565	-1933.22
-111E-14	0.003041	14E3	-0.00	1.0000	0.05	-0.00596	0.005961	3464.237
-0.03399	0.001581	14E3	-21.50	<.0001	0.05	-0.03709	-0.03089	-39226.8
0.01530	0.007302	14E3	2.10	0.0361	0.05	0.000992	0.02962	-3119.54
-0.06348	0.001451	14E3	-43.76	<.0001	0.05	-0.06632	-0.06064	-21301.3
0.001929	0.000024	14E3	80.07	<.0001	0.05	0.001882	0.001976	611828.1
0.000160	0.000042	14E3	3.84	0.0001	0.05	0.000078	0.000242	-168051
-0.00397	0.01599	14E3	-0.25	0.8037	0.05	-0.03532	0.02737	-77.9574
-0.00032	0.05434	14E3	-0.01	0.9953	0.05	-0.1068	0.1062	-29.3978
-0.00151	0.03943	14E3	-0.04	0.9695	0.05	-0.07880	0.07578	16.8127
-0.00066	0.06224	14E3	-0.01	0.9915	0.05	-0.1227	0.1213	-8.02057
0.001201	0.06859	14E3	0.02	0.9860	0.05	-0.1333	0.1357	-49.1149
0.5016	0.01378	14E3	36.39	<.0001	0.05	0.4746	0.5286	906.2667
-0.2435	0.01857	14E3	-13.11	<.0001	0.05	-0.2799	-0.2071	327.0716
-0.07538	0.03074	14E3	-2.45	0.0142	0.05	-0.1356	-0.01513	314.5398
-0.05688	0.04590	14E3	-1.24	0.2153	0.05	-0.1469	0.03310	251.5092
0.1292	0.01291	14E3	10.01	<.0001	0.05	0.1039	0.1545	-522.615

Re: Large Gradient Values for PROC NLMIXED

mohamed_zaki — Wed, 23 Dec 2015 00:32:17 GMT

A large gradient suggests the parameter estimate is not located at the minimum and, thus, could be changed to produce a better model fit. If any gradient is much above 0.001, one should adjust the initial parameter estimates and rerun the program.

Re: Large Gradient Values for PROC NLMIXED

Mohsen — Wed, 23 Dec 2015 03:24:17 GMT

Thanks Mohamed !

To change initial values for the parameters in the SAS code which values are preferable ?

For example, if we use the current SAS output coefficient values as the initial parameter values for the modified SAS code , will it be OK ? or generally, its try and error ?

Re: Large Gradient Values for PROC NLMIXED

Rick_SAS — Wed, 23 Dec 2015 11:33:18 GMT

For your specific case, this looks like it might be a logistic regression with a random-effects term added. In that situation, I would use PROC LOGISTIC to fit a fixed-effects model, then use those estimates as the initial estimates of the random-effects model.

Re: Large Gradient Values for PROC NLMIXED

Mohsen — Wed, 23 Dec 2015 23:55:50 GMT

Thank You Rick ! Will try it.

Re: Large Gradient Values for PROC NLMIXED

SteveDenham — Wed, 30 Dec 2015 19:35:58 GMT

And I would fit it all in GLIMMIX, rather than doing the two step LOGISTIC followed by NLMIXED. I don't see anything in the NLMIXED code that couldn't be easily ported over to GLIMMIX. There is only a single subject level random effect being modeled as an additive term in the logit.

Steve Denham