100 med1: the difference in effect between categories of medic 1 and 3 in s-th
province med2: the difference in effect between categories of medic 2 and 3 in s-th
province ULS: the effect of bad nutrition hotspot in s-th province
: The poverty level k = 1, 2, or 3 of sub district in sth province, ith district, and jth observation as response variable outcome.
: The characteristic of a sub district from sth province, ith district, and jth observation.
Nested Generalized Linear Model for ordinal response is
Nested Generalized Linear Mixed Models are also implemented for the same data
as in Nested GLM with district as the random factor, . Nested GLMM in this
application is as follows
4.4 Results and Discussion
4.4.1 Standard error of parameter estimates
Appendix 22 and Appendix 25 show the parameter estimates and their standard errors for Nested GLM and Nested GLMM, respectively.
Nested Generalized Linear Model. Figure 25 shows the standard errors of model-
based Nested GLM parameters with exchangeable, unstructured, and independent
WCM. Standard errors of unstructured correlation models tend to be smaller than standard errors of independent and exchangeable models, while standard errors
independent are the highest. The standard errors among those three models are not
significantly different. The correlation structured does not give contribution to the standard errors.
101
Figure 25 Standard errors of model-based Nested GLM parameters
Furthermore, Figure 26 shows the standard errors of robust Nested GLM parameters with exchangeable, unstructured, and independent WCM. It shows that
there is no particular pattern among exchangeable, unstructured, and independent models for robust parameter estimates. For some parameters prov2, farm11,
farm21, sch12, sch22, standard errors unstructured are extremely high. According to these two figures, parameter estimates of model-based are more stable than
those of robust. The correlation structured gives contribution to the robust standard errors.
Figure 26 Standard errors of robust Nested GLM parameters
Figures 27, 28, and 29 show standard error comparisons between model- based
and robust for exchangeable, unstructured and independent WCM,
0.00 1.00
2.00 3.00
4.00 5.00
6.00
th re
s1 th
re s2
p ro
v 1
p ro
v 2
fa rm1
1 fa
rm2 1
fa rm1
2 fa
rm1 3
fa rm2
3 sch
1 1
sch 2
1 sch
1 2
sch 2
2 sch
1 3
sch 23
me d
1 1
me d
2 1
me d
2 2
me d
1 3
me d
2 3
UL S
1 UL
S 2
UL S
3
exchangeable unstructured
independent
0.00 1.00
2.00 3.00
4.00 5.00
6.00
th re
s1 th
re s2
p ro
v 1
p ro
v 2
fa rm1
1 fa
rm2 1
fa rm1
2 fa
rm1 3
fa rm2
3 sch
1 1
sch 2
1 sch
1 2
sch 2
2 sch
1 3
sch 2
3 me
d 1
1 me
d 2
1 me
d 2
2 me
d 1
3 me
d 2
3 UL
S 1
UL S
2 UL
S 3
exchangeable unstructered
independent
102 respectively. Standard error of robust tends smaller than those of model based,
except for unstructured WCM.
Figure 27 Standard error of Nested GLM parameters of exchangeable WCM
Figure 28 Standard error of Nested GLM parameters of unstructured WCM
Figure 29 Standard error of Nested GLM parameters of independent WCM
0.00 1.00
2.00 3.00
4.00 5.00
6.00
th re
s1 th
re s2
p ro
v 1
p ro
v 2
fa rm1
1 fa
rm2 1
fa rm12
fa rm1
3 fa
rm2 3
sch 1
1 sch
2 1
sch 12
sch 2
2 sch
1 3
sch 2
3 me
d 1
1 me
d 2
1 me
d 2
2 me
d 1
3 me
d 2
3 UL
S 1
UL S
2 UL
S 3
model based robust
0.00 1.00
2.00 3.00
4.00 5.00
6.00
th re
s1 th
re s2
p ro
v 1
p ro
v 2
fa rm1
1 fa
rm2 1
fa rm1
2 fa
rm1 3
fa rm2
3 sch
1 1
sch 2
1 sch
12 sch
2 2
sch 1
3 sch
2 3
me d
1 1
me d
2 1
me d
2 2
me d
1 3
me d
2 3
UL S
1 UL
S 2
UL S
3
model based robust
0.00 1.00
2.00 3.00
4.00 5.00
6.00
th re
s1 th
re s2
p ro
v 1
p ro
v 2
fa rm1
1 fa
rm2 1
fa rm1
2 fa
rm1 3
fa rm2
3 sch
1 1
sch 21
sch 12
sch 2
2 sch
1 3
sch 2
3 m
e d
1 1
m e
d 2
1 m
e d
2 2
m e
d 1
3 m
e d
2 3
UL S
1 UL
S 2
UL S
3
model based robust
103 It is not easy to specify the correlation structure of the data, even though
according to the nature of the data, unstructured WCM is the most appropriate due to some high correlations between sub districts in the same district Appendix 28.
Until this point of analysis, it is not clear which structure is the most appropriate to the data. To continue the analysis, Table 17 is considered. This table shows the
averages of standard errors for every WCM and estimation method. If exchangeable or independent WCM is used, robust estimation is better than model
based estimation. Otherwise, if unstructured WCM is used, model based estimation is better than robust estimation.
Table 17 Averages of standard errors of the Nested GLM
exchangeable Unstructured
Independent Model Based
0.89 0.82
1.05 Robust
0.53 1.24 fluctuating
0.56
Nested Generalized Linear Mixed Model. Standard errors of parameter estimates
in Nested GLMM are equal for all WCMs, except there is a small difference for the threshold parameters of independent WCM. Figure 30 shows standard errors of
robust tend smaller than those of model based. Table 18 shows the averages of
standard errors of Nested GLMM, where robust estimation is better than model based estimation.
Figure 30 Standard errors of Nested GLMM parameters
1 2
3 4
5 6
th res
1 th
res 2
p ro
v 1
far m1
1 far
m2 1
far m1
3 far
m2 3
sc h
1 1
sc h
2 1
sc h
1 2
sc h
2 2
sc h
1 3
sc h
2 3
med1 1
med2 1
med2 2
med 1
3 med2
3 ULS1
ULS2 ULS3
model based exchangeable, unstructured, independent robust exchangeable, unstructured, independent
104 Table 18 Averages of standard errors the Nested GLMM
exchangeable Unstructured Independent Model Based
1.31 1.31
1.29 Robust
0.71 0.71
0.74
It is desirable to compare the fit of different working correlation structures within a GEE Nested GLM and Nested GLMM analysis using an informal
comparison. An informal comparison is to compare the standard error of the robust SE
R
and model-based SE
M
. If SE
R
SE
M
is close to 1, this suggests the correlation structure is correctly modeled Bishop, Die Wang 2000. This study
used the SE
R
SE
M
ratio to evaluate whether the exchangeable or unstructured WCM fit the data better than the independent correlation structure. There are no
guidelines regarding the size of the ratio, but higher ratios reflect poorer model fit. This comparison is qualitative, but it is the best approach available at this time
Koper and Manseau 2009. Figure 31 shows SE
R
SE
M
ratio for Nested GLM and GLMM, where some ratios of unstructured Nested GLM are close to 1, but it is
more fluctuated, while exchangeable and independent are more stable.
Figure 31 SE
R
SE
M
ratios for Nested GLM and GLMM
Table 19 shows averages of SE
R
SE
M
ratios from all parameter estimates. According to this table, Nested GLM with independent working correlation
structure gives the smallest ratio 0.56. Through the nature of the data, it would
1 2
3 4
5 6
th res
1 th
res 2
p ro
v 1
far m1
1 far
m2 1
far m1
3 far
m2 3
sc h
1 1
sc h
2 1
sc h
1 2
sc h
2 2
sc h
1 3
sc h
2 3
med1 1
med2 1
med2 2
med 1
3 med2
3 ULS1
ULS2 ULS3
GLM-exchangeable GLM-unstructured
GLM-independent GLMM-exch. unstr.
GLMM-independent
105 seem the data is appropriate to the unstructured correlation structure, but this value
shows that the best model of Nested GLMs is the model with independent structure.
Furthermore, random clustering effects include in modeling to control the correlation arise from recording multiple locations. The result is shown at Nested
GLMM column, where ratios of exchangeable and unstructured are slightly smaller than independent. Exchangeable and unstructured WCM is not
vehemently
2
rejected, because the ratios are smaller than independent and the difference is not significant.
Table 19 Averages of SE
R
SE
M
Nested GLM Nested GLMM
Exchange -able
Unstruc -tured
Indepen- dent
Exchange -able
Unstruc -tured
Indepen- dent
Average 0.63
1.5 0.56
0.57 0.57
0.59
Through this study, it is believed that the poverty data tends to have independent correlation structure. It can be seen through the value of SE
R
SE
M
in Nested GLM modeling, independent WCM has the smallest SE
R
SE
M
. After the random effect is included to the model Nested GLMM is used for modeling,
where the correlation structure is controlled, Nested GLMM provides better fit to the data than Nested GLM with exchangeable and unstructured WCM, showed by
the values of SE
R
SE
M
= 0,57 in GLMM, which is smaller than 0.63 and 1.5 in GLM.
4.4.2 Significance p-values