CAUCHY –Jurnal Matematika Murni dan Aplikasi Volume 5(1)(2017), Pages 15-19 p-ISSN: 2086-0382; e-ISSN: 2477-3344 Submitted: 10 October 2017 Reviewed: 18 October 2017 Accepted: 2 November 2017 DOI: http://dx.doi.org/10.18860/ca.v5i1.4393 Geographically Weighted Regression (GWR) Modelling with Weighted Fixed Gaussian Kernel and Queen Contiguity for Dengue Fever Case Data Grissila Yustisia1, Waego Hadi Nugroho 1, Maria Bernadetha Theresia Mitakda 1 1Department of Statistics, Brawijaya University, Malang Email: y_grissila@yahoo.com ABSTRACT Regression analysis is a method for determining the effect of the response and predictor variables, yet simple regression does not consider the different properties in each location. Methods Geographically Weighted Regression (GWR) is a technique point of approach to a simple regression model be weighted regression model. The purpose of this study is to establish a model using Geographically Weighted Regression (GWR) with a weighted Fixed Gaussian Kernel and Queen Contiguity in cases of dengue fever patients and to determine the best weighting between the weighted Euclidean distance as well as the Queen Contiguity based on the value of R2. Results from the study showed that the modeling Geographically Weighted Regression (GWR) with a weighted Fixed Gaussian Kernel showed that all predictor variables affect the number of dengue fever patients, whereas the weighted Queen Contiguity, not all predictor variables affect the dengue fever patients. Based on the value of R2 is known that a weighted Fixed Gaussian Kernel is better used. Keywords: geographically weighted regression, fixed gaussian kernel, queen contiguity INTRODUCTION Spatial data is the measurement data containing location information. Methods Geographically Weighted Regression (GWR) is a technique point of approach to a simple regression model be weighted regression model [1]. Spatial weighting matrix is used to determine the closeness of the relationship between the regions. GWR weighting role model is very important because it represents a weighted value of research data layout. Weighted grouped by distance (distance) and region (contiguity) [2]. In GWR models often use a weighting based on the distance (distance) without considering other weighting associated with the area (contiguity). In GWR models often use a weighting based on the distance (distance) without considering other weighting associated with the area (contiguity). In general, dengue fever clustered in specific locations. Things to note about the location other than the one with the distance (distance), which is the state of the location or about the area (contiguity). Therefore, in this study will consider the distance (distance) and region (contiguity) as weighting to search for the best model in dengue cases. mailto:y_grissila@yahoo.com Geographically Weighted Regression (GWR) Modelling with Weighted Fixed Gaussian Kernel and Queen Contiguity for Dengue Fever Case Data Grissila Yustisia 16 METHODS 2.1 Testing the Effect of Spatial Testing of spatial heterogeneity can be done by using the test statistic Breusch-Pagan (BP), is based on the hypothesis H0: σ1 2= σ2 2= ⋯= σj 2= σ2 H1: there are at least one j where σj 2 ≠ σ2 If 𝐻0 true test statistic: BP= ( 1 2 )𝒇(1𝑥𝑛) 𝑇 𝒁(𝒏𝒙𝒏)(𝒁(𝑛𝑥𝑛) 𝑇 𝒁(𝒏𝒙𝒏)) -1 𝒁(𝑛𝑥𝑛) 𝑇 𝒇(𝒏𝒙𝟏)+ ( 1 T )[ 𝑒(1𝑥𝑛) 𝑇 𝑾(𝒏𝒙𝒏)𝒆(𝒏𝒙𝟏) σ2 ] 2 ~ χ(p+1) 2 (1) 2.2 Fixed Gaussian Kernel Fixed Gaussian kernel is the weighting matrix based on the proximity of the location of the observation point i and point to another location. Fixed weighted Gaussian kernel as follows [1]: 𝑤𝑖𝑗 = exp⁡[−((di𝑗/ℎ)/2) 2] (2) If the location of the i located at coordinates⁡(ui,vi) it will obtain the Euclidean distance (dij) between all locations i and j are: di𝑗=√(ui- u𝑗) 2 + (vi- v𝑗) 2 (3) Bandwidth is a circle of radius from the center point location. Methods to determine the bandwidth is Cross Validation (CV) [1]): CV= ∑ (y i -ŷ ≠i (ℎ)) 2n i=1 (4) 2.3 Queen Contiguity Observations on the adjacent locations tend to be similar compared to locations far apart, because they relate to weighted location [3]. wij = { 1,⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡𝑖𝑓⁡𝑟𝑒𝑔𝑖𝑜𝑛𝑠⁡𝑖⁡𝑎𝑛𝑑⁡𝑗⁡𝑖𝑛𝑡𝑒𝑟𝑠𝑒𝑐𝑡,⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡𝑤ℎ𝑒𝑟𝑒⁡⁡⁡𝑖 ≠ 𝑗 0,⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡𝑎𝑛𝑜𝑡ℎ𝑒𝑟⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡ wi⏟ wij = [ 𝑤11 𝑤12 … 𝑤1𝑛 𝑤21 𝑤22 ⋯ 𝑤2𝑛 ⋮ 𝑤𝑛1 ⋮ 𝑤𝑛2 ⋱ ⋮ … 𝑤𝑛𝑛 ] 𝑤1 𝑤2 ⋮ 𝑤𝑛 (5) wi = ∑⁡WQ(𝑛𝑥𝑛) wij = ⁡WQ(𝑛𝑥𝑛) wi where: wi : Sum total row wij : The weighting matrix row i column j 2.4 Geographically Weighted Regression Spatial data is the measurement data containing location information. Methods Geographically Weighted Regression (GWR) is a technique point of approach to a simple regression model be weighted regression model [1]. According to Yasin [4] model of GWR is: 𝑦𝑖 = 𝛽0(𝑢𝑖,𝑣𝑖)+∑ 𝛽𝑗(𝑢𝑖,𝑣𝑖)𝑥𝑖𝑗 + 𝑖 𝑝 𝑗=1 (𝑢𝑖,𝑣𝑖)= Coordinates (longitude, latitude) point i to a geographical location. 2.5 Testing Geographically Weighted Regression Model Parameters Geographically Weighted Regression (GWR) Modelling with Weighted Fixed Gaussian Kernel and Queen Contiguity for Dengue Fever Case Data Grissila Yustisia 17 Testing Geographically Weighted Regression model parameters is done simultaneously and partially [5]: 1. Simultaneous testing to determine the effect of predictor variables together against the response variable. If 𝐻0 true test statistic: 𝐽𝐾𝑠𝑖𝑠𝑎(𝐻1) 𝛿1 2 𝛿2 ⁄ 𝐽𝐾𝑠𝑖𝑠𝑎(𝐻0) 𝑛−(𝑝+1)⁄ ~𝐹 ( 𝛿1 2 𝛿2 ,𝑛−(𝑝+1⁡) ∗ (6) 𝛿1 = 𝑡𝑟𝑎𝑐𝑒{(𝑰−𝑳) 𝑇(𝑰−𝑳)} 𝛿2 = 𝑡𝑟𝑎𝑐𝑒{(𝑰−𝑳) 𝑇(𝑰−𝑳)}2 where 𝐽𝐾𝑠𝑖𝑠𝑎(𝐻0) = 𝒀(𝟏𝒙𝒏) 𝑻 (𝑰(𝒏𝒙𝒏) −𝑳(𝒏𝒙𝒏)) 𝑻 (𝑰(𝒏𝒙𝒏) −𝑳(𝒏𝒙𝒏))𝒀(𝒏𝒙𝟏) ⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡𝐽𝐾𝑠𝑖𝑠𝑎(𝐻1) = 𝒀(𝟏𝒙𝒏) 𝑻 (𝑰(𝒏𝒙𝒏) −𝑯(𝒏𝒙𝒏))𝒀(𝒏𝒙𝟏) 𝑯 = 𝑿(𝒏𝒙𝒑)(𝑿(𝒑𝒙𝒏) 𝑻 𝑿(𝒏𝒙𝒑)) −𝟏𝑿(𝒑𝒙𝒏) 𝑻 𝑳(𝑛𝑥𝑛) = ( 𝒙1 𝑇[𝑿𝑇𝑾(u1,v1)𝑿] −1𝑿𝑻𝑾(u1,v1) 𝒙2 𝑇[𝑿𝑇𝑾(u2,v2)𝑿] −1𝑿𝑻𝑾(u2,v2) ⋮ 𝒙𝑛 𝑇[𝑿𝑇𝑾(un,vn)𝑿] −1𝑿𝑻𝑾(un,vn)) I = identity matrix of order n using criteria reject H0 if the test statistic |𝐹| > 𝐹 (𝛼, 𝛿1 2 𝛿2 ,𝑛−(𝑝+1)) ∗ 2. Partial testing to determine which predictor variables that influence the response variable for each observation location, using the t test statistic is based on the hypothesis: H0: βj(ui,vi) = 0 H1: βj(ui,vi) ≠ 0, j = 1, 2, ⋯, p t test statistic can be written as follows: β̂j(ui,vi) σ̂√cjj ~t( n-(p+1)) (7) 𝜎2 = 𝐽𝐾𝑠𝑖𝑠𝑎 𝑑𝑏𝑔𝑎𝑙𝑎𝑡 = 𝒆′𝒆 𝑛 −(𝑝 +1) where cjj a diagonal matrix element CCT, 𝑪 = (𝑿(𝑝𝑥𝑛) 𝑇 𝑾(𝑢𝑖,𝑣𝑖)(𝑛𝑥𝑛)𝑿(𝑛𝑥𝑝)) −1 (8) Reject if the test statistic |t| > t ( α 2 , n-(p+1)) 2.5 Testing Assessment Geographically Weighted Regression Model The coefficient of determination can describe the magnitude of the response variable diversity can be explained by the predictor variables. GWR R2 value obtained by the following mathematical equation [1]: 𝑅2(𝑢𝑖,𝑣𝑖) = 𝐽𝐾𝑅𝑊 𝐽𝐾𝑇𝑊 = ∑ 𝑾𝑖𝑗(𝑦𝑖−�̂�𝑖) 2𝑝 𝑗=1 ∑ 𝑾𝑖𝑗(𝑦𝑖−�̅�𝑖) 2𝑝 𝑗=1 (9) Geographically Weighted Regression (GWR) Modelling with Weighted Fixed Gaussian Kernel and Queen Contiguity for Dengue Fever Case Data Grissila Yustisia 18 RESULTS AND DISCUSSION The results of statistical calculations Breusch-Pagan test with both weighting are presented in Table 1. Table 1. Testing Result Breusch-Pagan Table 1 shows that the critical point test χ2 with error level 𝛼 = 0,05 and degrees of freedom (p+1) is 11,707 then reject H0 so the conclusion is that there is spatial heterogeneity in the data dengue cases. Further Testing GWR model parameters simultaneously, the test results are presented in Table 2. Table 2. Model Parameter Testing Results GWR Simultaneous Table 2 shows that the predictor variables with Gaussian kernel weighting Fixed effecting simultaneously the response variable for F⁡> (3,305)𝐹0.05(⁡2,31), where the weighted predictor variables Queen Contiguity no effect along the response variable for F > (2,922) 𝐹0.05(3,30). The partial test shows throughout predictor variables with Fixed Gaussian kernel weighting affect the response variable at each location and weighted Queen Contiguity affect the response variable at each location. Comparison of methods performed to determine the best weighting. Criteria for selection of the best weighting by using⁡𝑅2 are presented in Table 3. Best weighting is weighted with the largest value of 𝑅2. Table 3. Comparison of 𝑅2 at GWR Model Rated 𝑅2 for a model with Gaussian kernel weighting Fixed bigger than Queen contiguity weighted, so that it can be concluded that the weighting Fixed weighting Gaussian kernel is better used for data dengue cases in this study. Weighting Breusch- pagan Fixed Gaussian Kernel 13.341 Queen Contiguity 18.089 Weighting F Fixed Gaussian Kernel 4.397 Queen Contiguity 0.21 Weighting 𝑹𝟐 Fixed Gaussian Kernel 0.42 Queen Contiguity 0.06 Geographically Weighted Regression (GWR) Modelling with Weighted Fixed Gaussian Kernel and Queen Contiguity for Dengue Fever Case Data Grissila Yustisia 19 CONCLUSION The GWR method with the weighted Fixed Gaussian Kernel yields a R2 greater than the weighted Queen Contiguity. This result indicates that the weighted Fixed Gaussian Kernel use for dengue fever case data in this study. REFERENCES [1] H. Yasin, "Pemilihan Variabel pada Model Geographically Weighted Regression," Jurnal Media Statistika, vol. IV, no. 2, pp. 63-72, 2011. [2] M. Ilham and I. Dwi, "Pemodelan Data Kemiskinan di Provinsi Sumatera Barat dengan Metode Geographically Weighted Regression," Jurnal Media Statistika, vol. VI, no. 1, pp. 37-49, 2013. [3] A. S. Fotheringham, C. Brundson and M. Charlthon, Geographically Weighted Regression: The Analysis of Spatially Varying Relationships, 2002. [4] L. Anselin, I. Syabri and K. Youngihn, GeoDa: An Introduction to Spatial Data Analysis, Urbana: University of Illinois, 2004. [5] J. LeSage and R. K. Pace, Introduction to Spatial Econometrics, Boca Ration: CRC Press, 2009. ABSTRACT INTRODUCTION METHODS RESULTS AND DISCUSSION CONCLUSION REFERENCES