Geographically Weighted Regression for Prediction of
Underdeveloped Regions in East Java Province Based on Poverty
Indicators
Rusdi Hidayat N
1
, Bambang Widjanarko Otok
2
, Zumarsiyah Mahsyari
2
, Siti Halimah Sa’diyah
2
, and
Dimas Achmad Fadhila
2
1
Business Administration Study Program, FISIP, UPN “Veteran” East Java, Surabaya
2
Department of Statistics, FMKSD, Institut Teknologi Sepuluh Nopember, Surabaya
Keywords: GWR, Kernel Function, Adaptive Bisquare, Underdeveloped Regions, Poverty.
Abstract: Underdevelopment problem of a region can be seen from the dimensions of the economy, human resources,
financial capability, infrastructure, accessibility, and regional characteristics. One method to see a region is
underdeveloped or not is by looking the percentage of people living in poverty in a region in the publication
data of underdeveloped regional indicators issued by the Central Bureau of Statistics (BPS). The results
showed that the percentage of people in East Java Province who are living in poverty using linear regression
is not yet appropriate. The percentage of people living in poverty spread spatially because there is
heterogeneity between the observation sites which means that the observation of a location depends on the
observation in another location with adjacent distance so that the spatial regression modeling was done with
Adaptive Bisquare Kernel function. The grouping results with GWR resulted in nine groups based on
significant variables. Each group eas characterized by life expectancy, mean years of schooling, expenditure
and literacy rate.
1 INTRODUCTION
Underdeveloped regions are an area with districts
where communities and their territories are
relatively less developed than other regions on a
national scale. The backwardness of the area can be
measured based on six main criteria: economy,
human resources, infrastructure, regional financial
capacity, accessibility and regional characteristics
(Directorate General of Underdeveloped Regions
Development, 2016). To identify whether a district
is underdeveloped can be measured using
predetermined standards referring to the Minister of
Village Regulations, Development of
Underdeveloped Regions and Transmigration No. 3
of 2016 on Technical Guidelines for the
Determination of Indicators of Underdeveloped
Regions Nationally. The purpose of this paper is to
identify the problem of backwardness of a region
based on the percentage of poor people indicator.
If a study is influenced by the spatial aspect, then
it is necessary to consider spatial data on the model.
Spatial data is data that contains location
information. In spatial data, frequent observations at
a site depends on observation in another adjacent
location (neighboring) (Anselin, 1988). The law is
the basis for reviewing problems based on the
effects of location or spatial methods. In modeling,
if the classical regression model is used as an
analytical tool on spatial data, it can lead to
inaccurate conclusions because the assumption of
error is mutually free and the assumption of
homogeneity is not met.
2 LITERATURE RIVIEW
Geographically Weighted Regression (GWR) model
is the development of a regression model where each
parameter is calculated at each observation location,
so that each location of observation has different
regression parameter values. The GWR model is an
expansion of the global regression model in which
the basic idea is derived from non-parametric
regression (May, 2006). The y response variable in
the GWR model is predicted by the predictor
898
Hidayat N, R., Otok, B., Mahsyari, Z., Saâ
˘
A
´
Zdiyah, S. and Fadhila, D.
Geographically Weighted Regression for Prediction of Underdeveloped Regions in East Java Province Based on Poverty Indicators.
DOI: 10.5220/0007553708980907
In Proceedings of the 2nd International Conference Postgraduate School (ICPS 2018), pages 898-907
ISBN: 978-989-758-348-3
Copyright
c
2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
variable and each regression coefficient depends on
the location where the data is observed. The GWR
model is stated as follows (Fotheringham et al.,
2002):
() ()
0
1
,,
p
iii kiiiki
k
yuv uvx
ββ
ε
=
=+ +
(1)
Estimation of GWR model parameters is done by
Weighted Least Squares (WLS) method through
assigning different weights to each location where
data is observed. Suppose weighting for each
location
()
,
ii
uv
is
()
,
jii
wuv
, j = 1, 2,…,n then the
parameters at the observation location
()
ii
vu ,
then
the parameters at the observation location
()
,
jii
wuv
in equation (1) and then minimize the sum of
residual squares, or in the matrix form the sum of the
residual squares is:
() () () ()
() ()()
,,2,,
, , ,
TT TT
ii ii ii ii
TT
ii ii ii
uv uv uv uv
uv uv uv
=−
+
ε W ε yW y β XW y
β XW Xβ
(2)
with:
()
()
()
()
0
1
,
,
,
,
ii
ii
ii
pii
uv
uv
uv
uv
β
β
β
⎛⎞
⎜⎟
⎜⎟
=
⎜⎟
⎜⎟
⎜⎟
⎝⎠
β
M
and
() ()() ()
()
12
,diag ,, ,,, ,
ii ii ii n ii
uv wuv w uv w uv=W L
If equation (2) is lowered to
()
,
T
ii
uvβ
and the
result is equal to zero then obtained parameter
estimator GWR model:
() () ()
1
ˆ
,, ,
TT
ii ii ii
uv uv uv
⎡⎤
=
⎣⎦
β XW X XW y
(3)
For instance
()
12
1, , , ,
T
iiiip
x
xx=x L
is the i row
element of the X matrix. Then the prediction value
for y at the observation location
()
,
ii
uv
can be
obtained in the following way:
() ()
()
()
1
ˆ
ˆ
,,,
TTT T
iiii i ii ii
yuv uv uv
==x β xXW X XW y
So for all observations can be written as follows:
()
12
ˆˆˆ ˆ
,,,
T
n
yy y==yLyL
and
()()
12
ˆˆ ˆ
ˆ
,,,
T
n
εε ε
==ε I-L
y
L
with I is an identity matrix of nxn and
()
()
()
()
()
()
()
()
()
1
111 11
1
222 22
1
,,
,,
,,
TT T
TT T
TT T
nnn nn
uv uv
uv uv
uv uv
⎛⎞
⎜⎟
⎜⎟
⎜⎟
=
⎜⎟
⎜⎟
⎜⎟
⎝⎠
xXW X XW
xXW X XW
L
xXW X XW
M
(4)
The weighting role of the GWR model is
important because this weighted value represents the
location of the observed data with each other. The
weighting scheme on GWR can use several different
methods. There is some literature that can be used to
determine the weighting for each different location
on the GWR model, such as by using Kernel
Function.
The Kernel Function is used to estimate the
parameters in the GWR model if the distance
function
(
)
j
w
is a continuous and monotonous
function down (Chasco, Garcia and Vicens, 2007).
Weights that are formed by using this kernel
function are the Gaussian Distance Function,
Exponential Function, Bisquare Function, and
Tricube Kernel Function and involve the smoothing
parameter (Lesage, 2001). Cross Validation (CV)
method to select the optimum bandwidth, which is
mathematically defined as follows:
()
()
2
1
ˆ
()
n
ii
i
CV h y y h
=
=−
with
()
i
yh
)
is the value of the appraiser
i
y where
the observations are on location
()
,
ii
uv
removed
from the estimation process. To get the value
h
the
optimal is then obtained from
h
which results in a
minimum CV value.
Hypothesis testing on the GWR model consists
of testing the suitability of the GWR model and
testing the model parameters. Testing the suitability
of the GWR model (goodness of fit) is done with the
following hypothesis:
()
0
H: ,
kii k
uv
ββ
=
(there is no significant
difference between global regression model
and GWR)
1
H :
At least, there is one
()
,
kii k
uv
ββ
for each
0,1, 2, , , and 1, 2, ,kpin==LL
(there is a significant difference between the
global regression model and GWR).
Determination of test statistics based on Residual
Sum of Square (RSS) obtained respectively below
Geographically Weighted Regression for Prediction of Underdeveloped Regions in East Java Province Based on Poverty Indicators
899
H
0
dan H
1
. Under conditions H
0
, using OLS method
obtained the following RSS value:
() ()()( )
0
ˆˆ ˆ ˆ
RSS H
T
TT
== = εε yy yy y IHy
wit
h
()
TT
=
1
HXXX X which is idempotent. Under
conditions H
1
, spatially varying regression
coefficients in equation (1) is determined by the
GWR method, to obtain the following RSS values:
() ()()
()()
1
ˆˆ ˆ ˆ
RSS H
T
T
T
T
==
=−
εε yy yy
yIL ILy
(5)
so obtained the following test statistic (Leung et al.,
2000a):
()
()
()
2
1
1
2
1
0
H
H1
RSS
F
RSS n p
δ
δ
⎛⎞
⎜⎟
⎜⎟
⎝⎠
=
−−
Below H
0
,
1
F
will follow the F distribution with
degrees of freedom
2
1
1
2
df
δ
δ
⎛⎞
=
⎜⎟
⎜⎟
⎝⎠
and
()
2
1df n p=−
. If taken the level of significance
then reject H
0
if
12
11,,df df
FF
α
<
.
with:
()()
, 1, 2
T
i
i
tr i
δ
⎛⎞
⎡⎤
=− =
⎜⎟
⎢⎥
⎣⎦
⎝⎠
IL IL
. (6)
Another alternative as a test statistic is to use the
difference in the number of residual squares below
H
0
and below H
1
(Leung et al., 2000a), i.e:
() ()
()
()
()()()
()()
011
2
11
1
1
T
T
T
T
RSSH RSSH
F
RSS H
τ
δ
τ
δ
=
⎡⎤
−−
⎣⎦
=
−−
yIH ILILy
yIL ILy
Below H
0
2
F
will follow the distribution F with
degrees of freedom
2
1
1
2
df
τ
τ
=
and
2
1
2
2
df
δ
δ
⎛⎞
=
⎜⎟
⎜⎟
⎝⎠
. If
taken the level of significance
then reject H
0
if
12
2,,df df
FF
α
.
with:
( )()()
, 1, 2
T
i
i
tr i
τ
⎛⎞
⎡⎤
=− =
⎜⎟
⎢⎥
⎣⎦
⎝⎠
IH IL IL
If it is concluded that the GWR model is
significantly different from the global regression
model, then the next step is to perform a partial test
to find out whether there is a significant influence
difference from the predictor variable
k
x
between
one location and another (May, He and Fang, 2004).
This test can be done by hypothesis:
H
0
:
() ( ) ( )
11 2 2
,, ,
kk knn
uv u v u v
ββ β
===L
(there
is no significant difference of influence from the
predictor variable
k
x
between one location and
another)
H
1
: At least, there is one,
()
, , for 1, 2,...,
kii
uv i n
β
=
()
0,1, 2, ,kp= L
which
different. (there is a significant effect difference
from the predictor variable
k
x
between one location
and another location).
To perform the above test it is determined first
variance
()
iik
vu ,
ˆ
β
(i = 1, 2, ..., n) which denoted
by:
() ()
2
2
11
11
ˆˆ
,,
11
nn
kkiikii
ii
T
kk
Vuvuv
nn
nn
ββ
==
⎛⎞
=−
⎜⎟
⎜⎟
⎝⎠
⎡⎤
=−
⎢⎥
⎣⎦
∑∑
β IJβ
with
()
()
()
()
11
22
,
,
,
,
k
k
kii
knn
uv
uv
uv
uv
β
β
β
⎛⎞
⎜⎟
⎜⎟
=
⎜⎟
⎜⎟
⎜⎟
⎝⎠
β
M
.
While the test statistic used is:
2
3
11
11
RSS(H )
T
kk k
Vtr
nn
F
δ
⎛⎞
⎡⎤
⎜⎟
⎢⎥
⎣⎦
⎝⎠
=
BI JB
with:
() ()
() ()
() ()
1
11 11
1
22 22
1
,,
,,
,,
TT T
k
TT T
k
k
TT T
knn nn
uv uv
uv uv
uv uv
⎛⎞
⎡⎤
⎣⎦
⎜⎟
⎜⎟
⎡⎤
⎜⎟
⎣⎦
=
⎜⎟
⎜⎟
⎜⎟
⎡⎤
⎣⎦
⎝⎠
eXW X XW
eXW X XW
B
eXW X XW
M
,
k
e
is column vector which is size
()
1p +
which is
worth one for the k elements for the other. Matrix
L
as in (4) and RSS (H
1
) as in the equation (5).
ICPS 2018 - 2nd International Conference Postgraduate School
900
Below H
0
, test statistic
3
F
will be distributed F
with degrees of freedom
2
1
1
2
df
γ
γ
⎛⎞
=
⎜⎟
⎜⎟
⎝⎠
and
2
1
2
2
df
δ
δ
⎛⎞
=
⎜⎟
⎜⎟
⎝⎠
with
11
T
ik k
i
tr
nn
γ
⎛⎞
⎡⎤
=−
⎜⎟
⎢⎥
⎣⎦
⎝⎠
BI JB
i=1,2 dan
i
δ
as in equation (6). Reject H
0
if
12
3,,df df
FF
α
(Leung et al., 2000a).
The testing of the significance of model
parameters at each location is done by partially
testing the parameters. This test is conducted to
determine which parameters significantly affect the
response variable. The form of the hypothesis is as
follows:
()
0
H: , 0
kii
uv
β
=
()
1
:,0
kii
Huv
β
with
1, 2, ,kp= L
The parameter estimator
()
ˆ
,
ii
uvβ
will follow the
normal multivariate distribution with the average
()
β ,
ii
uv
and the covariance variant matrix
2T
ii
σ
CC
, with:
()
()
()
1
,,
TT
iii ii
uv uv
=CXW XXW ,
so it get to:
ˆ
(,) (,)
kii kii
kk
uv uv
c
ββ
σ
~N(0,1)
with
kk
c
is the k-diagonal element of the matrix
T
ii
CC
. So the test statistic used is:
ˆ
(,)
ˆ
kii
hit
kk
uv
T
c
β
σ
=
Below H
0
T will follow the t distribution with
degrees of freedom
2
1
2
δ
δ
⎛⎞
⎜⎟
⎜⎟
⎝⎠
while
σ
ˆ
is acquired by
rooting
2
1
1
RSS(H )
ˆ
σ
δ
=
. If the given level of
significance is
, then the decision to decline H
0
or in other words parameters
()
,
kii
uv
β
significant
to the if model
,
2
hit
df
Tt
α
>
, where
2
1
2
df
δ
δ
⎛⎞
=
⎜⎟
⎜⎟
⎝⎠
.
Akaike Information Criterion Correction (AICc)
method used to select the best model defined as
follows :
()
()
()
ˆ
2ln ln2
2()
c
ntr
AIC n n n
ntr
σπ
⎧⎫
+
=++
⎨⎬
−−
⎩⎭
S
S
(7)
with :
ˆ
σ
: The standard deviation estimator value of the
maximum likelihood estimated error, ie
2
ˆ
RSS
n
σ
=
S : Where projection matrix
ˆ
=ySy
The best model selection is done by determining the
model with the smallest AICc value (Fotheringham
et al., 2002).
3 METHODOLOGY
This study used secondary data from Publication of
BPS in 2014. The research variables are presented in
Table 1 below.
Table 1: Research variable.
Variable Indicator
Y Percentage of poor people
X1 Life expectancy
X2 Average school duration
X3 Per capita population expenditure
X4 Literacy
Stages performed in the study.
1.
Description of characteristics and pattern of
percentage distribution of poor people in East
Java.
2.
Modeling percentage of poor people in East
Java with Linear Regression and GWR with
AIC criteria. The steps are as follows.
i.
Detection of muticolinearity cases.
ii.
Modeling percentage of poor people in East
Java with Linear Regression:
a.
Calculates the parameter estimator value
of the Linear regression model
b.
Perform parallery testing simultaneously
and partially.
c.
Testing the residual assumption of IIDN.
iii.
Model GWR on the percentage of the poor
in East Java:
a.
Calculate euclidian distance between
observation locations based on
geographical position. Euclidean distance
between location i located at coordinate
(u
i
, v
i
) to location j at coordinate (u
j
,v
j
)
Geographically Weighted Regression for Prediction of Underdeveloped Regions in East Java Province Based on Poverty Indicators
901
b.
Determine the optimum bandwidth with
CV criteria
c.
Determine the optimum weighting with
the gauss kernel weighted function.
d.
Calculates the value of the GWR model
parameter estimator
e.
Testing GWR parameters (partial test and
partial test)
Comparing the AICc value of the Global / Linear
Regression Model with the GWR model, the
minimum AICc value is the best model.
4 RESULT AND DISCUSSION
The description of this study includes the mean and
standard deviation of each variable which is
presented in Table 2.
Table 2: Description of research variables.
Varia
ble
Minimum Maximum Mean StDev
Y 4.59 25.8 12.1 4.99
X1 62.16 73.3 69.2 3.15
X2 3.49 10.9 7.3 1.72
X3 7143 15492 10013 2062
X4 77.93 98.5 92.3 4.84
4.1 Multicolinearity Detection
One of the requirements in regression analysis with
some predictor variables is that there is no case of
multicollinearity or there is no predictor variable that
has correlation with other predictor variables. The
detection of muticollinearity is based on the
Variance Inflation Factor (VIF) value. Here is the
VIF value of each predictor variable.
Table 3: VIF of predictor variable.
Variabel X1 X2 X3 X4
VIF 2.085 9.066 4.698 4.526
Table 3 shows that all predictor variables have a
VIF value of less than 10. This detects that there are
no multicollinearity cases or no predictor variables
that have correlation with other predictor variables.
4.2 Significance Test of Linear
Regression of Hypertension
Prevalence
Here is the test of the significance of linear
regression parameters either simultaneously or
partially to determine the effect of predictor
variables used. The hypothesis for simultaneous
parameter significance test in linear regression is as
follows.
H
0
: β
1
= β
2
=…= β
4
= 0 (parameters have no
significant effect on the mode)
H
1
: At least, there is one β
k
0; k = 1, 2,, 4 (at
least, there is one parameter that significantly
affects the model)
Table 4: ANOVA linear regression table.
Variation
Sources
Sum of
Squares
df
Mean
Square
F P
Regression 598.32 9 66.48 2.14 0.030
Residual 4418.32 142 31.11
Total 5016.64 151
Table 4 yields a value F
count
in the amount of 2,14
and p-value in the amount of 0,030. Based on the
level of significance (
α) in the amount of 5% and
F
(0,05;9;142)
in the amount of 1,946, obtained decision
reject H
0
because of value F
count
> F
(0,05;9;142)
or P-
Value < 0,05. This can be interpreted that there is at
least one parameter that has a significant effect on
the prevalence of hypertension.
Furthermore, to determine which predictor
variables that give a significant influence, then the
partial significance of the parameters were tested
and presented in Table 5. The following hypothesis
test of spatial parameter significance of linear
regression model (global)
.
H
0
: β
k
= 0,
H
1
: β
k
0, k = 1,2,3,4
Table 5. Parameter test of partial regression coefficients.
Parame-
ter
Coeffi-
cient
SE
Coeffi-
cient
T Sig.
β0 59.720 16.410 3.640 0.001
β1 0.076 0.187 0.410 0.687
β2 -1.422 0.789 -1.800 0.081
β3 -0.000 0.000 -0.140 0.891
β4 -0.454 0.179 -2.530 0.016
ICPS 2018 - 2nd International Conference Postgraduate School
902
Based on the test results in Table 5, with a
significant level (
α) in the amount of 5% and
()
()
2
0.025,33
;1
2.035
np
tt
α
−−
==
, it is obtained that all
values of T smaller than t
(0.025,33)
, except parameters
β
4
. This shows that the literacy rate variable
significantly affects the percentage of poor people.
4.3 Testing Residual Assumptions
Testing of residual assumptions is identical,
independent, and normally distributed.
4.3.1 Identical Residual Assumption Test
One assumption test in OLS regression is that
residual variance should be homoscedasticity
(identical) or case of heteroscedasticity. How to
identify the case of heteroscedasticity is to create a
regression model between residual and predictor
variables. If there are predictor variables that
significantly affect the model, then it can be said that
the residual is not identical or happened case of
heteroscedasticity. Testing identical residual
assumptions provides information that no cases of
heteroscedasticity or residual have been identical to
a significant level (
α) of 0.05 and
()()
;, 1 0,05;4,33
2.659
pn p
FF
α
−−
==
. This is because of
the P-Value (0.119) is bigger than
α (0.05) and F
(1.99) is smaller than 2.659, then there is no
heteroscedasticity.
4.3.2 Independent Residual Assumption
Test
An independent residual assumption test is used to
determine whether or not the relationship exists
between residuals. The test statistic used is Durbin-
Watson. The value of DW = 1.099 earned value
07875,2=d
with
0201,1=
L
d
and
9198,1=
U
d
. So the decision that can be taken is
Reject H0 because
0802,2)4(9198,1 =<<=
UU
ddd
. It shows
that there is a residual relationship, so that the
independent residual assumption is not met.
4.3.3 Normal Distributed Assumption Test
Normal distributed assumption test is performed by
the following Kolmogorov-Smirnov test.
H
0
: Data is normally distributed
H
1
: Data is not normally distributed
5.02.50.0-2.5-5.0
99
95
90
80
70
60
50
40
30
20
10
5
1
RESIDUAL
Percent
Mean - 1.67819E-14
StDev 2.346
N38
KS 0.109
P-Valu e >0.150
Probability Plot of Residual
Nor ma l
Figure 1. Probability plot normal residual.
Based on Figure 1, it is found that the red dots
spread close to the linear (normal) line meaning that
the data has been normally distributed. In addition, it
can also be seen from the value of P-Value is greater
0,15. So the decision that can be taken is Failed
Reject H
0
at a significant level (α) in the amount of
5%, that is, the data has fulfilled normal distributed
assumptions. Based on the results of the assumption
test, it can be concluded that the residuals in the
linear regression model (global) data have normal
distribution, but the identical and independent
assumptions are not met. So that spatial regression is
done with GWR approach.
4.4 Modeling of Spatial Regression of
Percentage of Poor People
Analysis using GWR method aims to determine the
variables that affect the percentage of poor people in
each location of observation that is the district / city
in the province of East Java. The first step to get the
GWR model is to determine the point of latitude and
longitude coordinates at each location, calculate the
euclidean distance and determine the optimum
bandwidth value based on Cross Validation (CV)
criteria. The next step is to determine the weighting
matrix with kernel function: Fixed Gaussian, Fixed
Bi-Square, Adaptive Gaussian, Adaptive Bi-Square
and estimates GWR model parameters. The
weighted matrix obtained for each location is then
used to form the model, so that different models are
obtained at each observation location.
The hypothesis test of the GWR model consists
of two tests, namely the GWR model conformity test
and the parameter significance test of the GWR
model. Here are the results of hypothesis testing
GWR model.
Geographically Weighted Regression for Prediction of Underdeveloped Regions in East Java Province Based on Poverty Indicators
903
H
0
:
kiik
vu
β
β
=),( ;
(There was no significant difference
between the linear regression model
(global) and the GWR model)
H
1 :
At least, there is one
kiik
vu
β
β
),( k = 1,
2, ….,9
(There is a significant difference between
the linear regression model (global) and the
GWR model)
Table 6. Estimated GWR on kernel function weight.
Stat
istic
Function weight
Fixed
Gaussia
n
Fixed
Bi-
square
Adaptive
Gaussian
Adaptive
Bi-
square*
MS
E
5.560 5.049 5.989 1.995
R
2
0.829 0.852 0.797 0.998
AIC
c
185.676 184.839 186.017 -14185
Table 6 shows the comparison of estimated
GWR models with different weights. The GWR
model conformity test is performed by using the
difference of sum of residual squares of GWR model
and global regression model. The GWR model will
be significantly different from the global regression
model if it can significantly reduce the number of
residual squares. Table 6 shows that the smallest
AICc value is the GWR model with the Adaptive Bi-
Square kernel function weights, that is -14185. So,
by using the level of significance α at 5% it can be
concluded that the GWR model is significantly
different from the global regression model. This
means that the GWR model with the Adaptive Bi-
Square kernel function weights more feasible to
describe the percentage of poor people in East Java
Province.
Next is a test of the significance of GWR model
parameters with Adaptive Bi-Square kernel function
weights partially to know which parameters
significantly influence the percentage of poor people
in each location of observation. The grouping of
districts with the same variables that significantly
affect the percentage of poor people is presented in
Table 7.
Table 7. T-count value in variables in each regency / city
using adaptive bisquare.
Regency/City
Predictor Variable
X1 X2 X3 X4
"Pacitan
Regency"
-0.58 0.69 -0.53 -2.89*
"Ponorogo
Regency"
0.46 -0.39 0.51 -3.06*
"Trenggalek
Regency"
1.22 -0.33 -0.24 -1.69
"Tulungagung
Regency"
2.31* 0.35 -0.88 -2.48*
"Blitar
Regency"
-1.73 0.92 -2.99* -0.39
"Kediri
Regency"
1.22 -0.07 -0.83 -0.67
"Malang
Regency"
1.37 -1.03 -1.41 -0.48
"Lumajang
Regency"
2.73* 1.22 -1.53 -2.76*
"Jember
Regency"
0.97 0.11 -1.36 -0.60
"Banyuwangi
Regency"
-0.44 0.78 -0.74 -2.85*
"Bondowoso
Regency"
1.87 -1.50 1.55 -2.46*
"Situbondo
Regency"
2.92* -0.74 1.50 -1.64
"Pasuruan
Regency"
-0.01 -0.51 0.32 -1.80
"Probolinggo
Regency"
-0.05 0.95 -1.10 -2.97*
"Sidoarjo
Regency"
3.02* -0.92 1.61 -2.00*
"Mojokerto
Regency"
0.46 0.10 -0.14 -1.00
"Jombang
Regency"
1.36 0.82 -1.17 -0.89
"Nganjuk
Regency"
0.55 -0.08 1.05 -1.16
"Madiun
Regency"
2.92* 0.67 -1.13 -2.97*
"Magetan
Regency"
1.06 -1.16 0.10 -0.61
"Ngawi
Regency"
0.50 -3.55* 0.60 1.72
"Bojonegoro
Regency"
-2.41* 2.43* -2.84* 0.38
ICPS 2018 - 2nd International Conference Postgraduate School
904
5 CONCLUSIONS
Result of modeling percentage of poor population in
East Java Province based on Regency / City using
linear regression showed that only one variable that
affect the percentage of poor people, which is
literacy rate. The percentage of poor people in East
Java Province spread spatially because there is
heterogeneity between observation locations which
means that the observation of a location depends on
the observation in other location with adjacent
distance so spatial regression modeling with
Adaptive Bisquare kernel function was done, which
produced 9 groups.
REFERENCES
Anselin, L., 1988. Spatial Econometrics: Method and
Models. Netherlands, Kluwer Academic Publishers.
Chasco, C., Garcia, I., and Vicens, J., 2007. Modeling
Spatial Variations in Household Disposible Income
with Geographically Weighted Regression. Munich
Personal RePEc Arkhive (MPRA) Working Papper
No. 1682.
Draper, N., Smith, H., 1992. Analisis Regresi Terapan.
Jakarta, PT Gramedia Pustaka Utama
Fotheringham, A.S., Brunsdon, C., and Charlton, M.,
2002. Geographically Weighted Regression. Jhon
Wiley & Sons, Chichester, UK
LeSage, J.P., 2001. A Family of Geographically Weighted
Regression. Departement of Economics University of
Toledo.
Leung, Y., Mei, C.L., and Zhang, W.X., 2000a. Statistic
Tests for Spatial Non-Stationarity Based on the
Geographically Weighted Regression Model,
Environment and Planning A, 32 9-32.
Leung, Y., Mei, C.L., and Zhang, W.X., 2000b. Testing
for Spatial Autocorrelation Among the Residuals of
the Geographically Weighted Regression.
Environment and Planning A, 32, 871-890.
Miller, H.J., 2004. Tobler’s First Law and Spatial
Analysis. Annals of the Association of America
Geographers, 94(2), 284-289.
APPENDIX
Group 1: Regency / City di in East Java with
Percentage of Poor People Not Affected by Literacy
Rate
Group 2: Regency / City in East Java with
Percentage of Poor People Not Affected by Life
Expectancy Factors, Average School Duration, Per
Capita Population Expenditure, and Literacy Rate
"Tuban
Regency"
1.81 0.35 -0.93 -2.19*
"Lamongan
Regency"
-1.80 0.01 0.47 1.48
"Gresik
Regency"
3.14* -5.10* 2.16* 1.99*
"Bangkalan
Regency"
-2.10* 0.28 -0.22 -2.63*
"Sampang
Regency"
3.37* -3.91* 3.84* 2.83*
"Pamekasan
Regency"
0.01 0.83 -0.90 -3.00*
"Sumenep
Regency"
2.90* 1.59 0.13 -2.68*
"Kediri City" 1.54 0.19 -1.43 -0.95
"Blitar City" -1.15 1.24 -1.36 -1.69
"Malang City" 0.32 0.96 -1.28 -1.56
"Probolinggo
City"
2.90* 1.59 0.09 -2.68*
"Pasuruan City" -0.15 0.92 -1.08 -3.05*
"Mojokerto
City"
0.82 -0.09 1.08 -1.13
"Madiun City" 2.20* -1.33 0.72 -2.00*
"Surabaya
City"
-0.69 -4.77* 3.48* 3.41*
"Batu City" 3.22* -5.15* 2.23* 2.04*
Geographically Weighted Regression for Prediction of Underdeveloped Regions in East Java Province Based on Poverty Indicators
905
Group 3: Regency / City in East Java with
Percentage of Poor People Affected by Life
Expectancy Figures and Literacy Rates
Group 4: Regency / City in East Java with
Percentage of Poor People Affected by Capita
Population Expenditure
Group 5: Regency / City in East Java with
Percentage of Poor People Affected by Life
Expectancy Factor
Group 6: Regency / City in East Java with
Percentage of Poor People Affected by Average
School Duration
Group 7: Regency / City in East Java with
Percentage of Poor People Affected by Average
School Duration, Per Capita Population Expenditure,
and Literacy Rate
ICPS 2018 - 2nd International Conference Postgraduate School
906
Group 8: Regency / City in East Java with
Percentage of Poor People Affected by Life
Expectancy Factors, Average School Duration, Per
Capita Population Expenditure, and Literacy Rate
Group 9: Regency / City in East Java with
Percentage of Poor People Affected by Life
Expectancy Factors, Average School Duration, and
Per Capita Population Expenditure
Geographically Weighted Regression for Prediction of Underdeveloped Regions in East Java Province Based on Poverty Indicators
907