proc spatialreg optimization

[Kyungjae]

I used proc spatialreg to analyze spatial econometrics model.

However, error messages kept showing up depending on model or data. "Opimization failed. Interpret the estimates with care" or 

"Optimization cannot improve the function value". In addition, although the message 'Algorithm converged' comes out, standard error, t value, and p value don`t appear. 

I specified quasi-Newton method and I specified maximum number of iterations as 1000.

When I use SAR model, this output comes out.

Model Fit Summary
Dependent Variable	t_crime
Number of Observations	116
Data Set	LKJ.D05_F
Spatial Weights	WORK.SWM
Model	SAR
Log Likelihood	-1058
Maximum Absolute Gradient	1.85E-06
Number of Iterations	167
Optimization Method	Quasi-Newton
AIC	2144
SBC	2182
Algorithm converged.
Parameter Estimates
Parameter	DF	Estimate	Standard	t Value	Approx
Error	Pr > |t|
Intercept	1	8601.954	6.352752	1354.05	<.0001
pop	1	29.05341	2.875343	10.1	<.0001
f_rate	1	794.7308	387.8651	2.05	0.0405
ac	1	-634.92	147.059	-4.32	<.0001
ad	1	380.3811	100.034	3.8	0.0001
old	1	-480.73	110.843	-4.34	<.0001
female	1	179.7138	108.2588	1.66	0.0969
college	1	-164.403	58.18005	-2.83	0.0047
mig	1	-127.122	44.54226	-2.85	0.0043
house_1	1	167.0526	66.63577	2.51	0.0122
atax	1	315.4854	81.58125	3.87	0.0001
train	1	1368.529	841.2449	1.63	0.1038
_rho	1	0.088868	0.124355	0.71	0.4748
_sigma2	0	4888967	.	.	.

However, when I use SAC model and other models, these output come out.

Model Fit Summary
Dependent Variable	t_crime
Number of Observations	116
Data Set	LKJ.D05_F
Spatial Weights	WORK.SWM
Model	SAC
Log Likelihood	-1474
Maximum Absolute Gradient	2824
Number of Iterations	363
Optimization Method	Quasi-Newton
AIC	2977
SBC	3018
ERROR: Optimization failed. Interpret the estimates with care.
Parameter Estimates
Parameter	DF	Estimate	Standard	t Value	Approx
Error	Pr > |t|
Intercept	0	7284.729	.	.	.
pop	1	27.55415	0.928053	29.69	<.0001
f_rate	1	315.0701	126.6845	2.49	0.0129
ac	1	-540.133	56.69547	-9.53	<.0001
ad	1	259.395	39.98755	6.49	<.0001
old	1	-450.053	36.2234	-12.42	<.0001
female	1	48.66603	79.86923	0.61	0.5423
college	1	-161.944	19.16263	-8.45	<.0001
mig	1	-92.4311	14.38729	-6.42	<.0001
house_1	1	148.8553	22.76375	6.54	<.0001
atax	1	307.1371	26.51714	11.58	<.0001
train	1	1597.459	270.7654	5.9	<.0001
_rho	1	-0.6494	0.117049	-5.55	<.0001
_lambda	0	0.999642	.	.	.
_sigma2	1	499441	0.128961	3872796	<.0001

or 

Model Fit Summary
Dependent Variable	t_crime
Number of Observations	116
Data Set	LKJ.D05_F
Spatial Weights	WORK.SWM
Model	SMA
Log Likelihood	-1053
Maximum Absolute Gradient	1223
Number of Iterations	281
Optimization Method	Quasi-Newton
AIC	2133
SBC	2169
WARNING: Optimization cannot improve the function value.
Parameter Estimates
Parameter	DF	Estimate	Standard	t Value	Approx
Error	Pr > |t|
Intercept	1	-3562.56	43.40649	-82.07	<.0001
pop	1	31.24189	0.270455	115.52	<.0001
f_rate	0	1113.872	.	.	.
ac	0	-631.177	.	.	.
ad	0	459.1431	.	.	.
old	0	-435.751	.	.	.
female	1	352.0277	0.973919	361.45	<.0001
college	0	-153.23	.	.	.
mig	0	-106.113	.	.	.
house_1	1	193.9406	1.662491	116.66	<.0001
atax	1	339.6874	13.22217	25.69	<.0001
train	1	1067.575	87.99192	12.13	<.0001
_lambda	0	1	.	.	.
_sigma2	0	2669892	.	.	.
Restrict1	-1	94906303	.	.	. *

I wonder how I could solve this optimization problem. Thank you.

[Rick_SAS · Accepted solution]

Hard to say since you have not included your code or data, but the most common reason for non-convergence is that the model do not fit the data. Also common is trying to fit  complex model with a small sample. I notice you only have 116 observations but your model contains 14-15 parameters.

I assume "get more data" would be hard, so try choosing a simpler model with fewer parameters.