Indonesian Journal of Statistics and Its Applications

Comparison of Chi-Square Automatic Interaction Detector (CHAID) and Random Forest Methods in the Classification of Household Poverty Status in Central Java

Fatkhul Izzati — 2024-06-11

Central Java was in second position as the province with the highest number of poor people in Indonesia in March 2020. Poverty alleviation efforts have been carried out, but many are still not yet on target. The purpose of this study was to model the classification of household poverty status in Central Java using CHAID and random forest methods and compare the two methods. The data used in this study is data from the 2020 National Socioeconomic Survey (SUSENAS) conducted by the Central Bureau of Statistics (BPS) for Central Java. The number of poor households is much less than non-poor households. Therefore, Synthetic Minority Oversampling Technique (SMOTE) was performed to handle unbalanced data. The random forest method produced better classification performance than the CHAID method with accuracy, sensitivity, specificity, and AUC of 93,95%, 98,43%, 89,92%, and 0,9417, respectively. The important variables that build the random forest model are the floor area of the house, the age of the head of the household, cooking fuel, the place for the final disposal of feces, and ownership of the place to defecate.

Addressing multicollinearity in spatial modelling: A district level spatial analysis of pandemic COVID-19 in India.

Shalini Chandra — 2024-06-11

This study focuses on conducting spatial analysis of COVID-19 at the district level in India. Leveraging data from www.covidindia.org for confirmed cases and deaths, and integrating population characteristics from the National Family Health Survey 5 (2019-2021) and supplementary sources. The objective is to identify risk factors using spatial modelling techniques while addressing multicollinearity through principal component analysis (PCA). This study utilizes spatial analysis to identify COVID-19 hotspots and coldspots at the district level in India. It highlights highly affected districts such as Mumbai, Pune, Chennai, Kolkata, and Bengaluru, as well as low affected districts in central and north-eastern regions. The study utilized the spatial lag model (SLM), spatial error model (SEM), geographical weighted regression (GWR), and multiscale geographical weighted regression (MGWR) models to analyse the impact of demographic, socioeconomic, climatic, and comorbidity factors on COVID-19, accounting for spatial proximity. Among these models, MGWR exhibited superior performance. Key risk factors associated with the COVID-19 phenomenon identified, providing insights into the impact of household conditions, educational level of women, tobacco and alcohol consumption rates, number of health centres, and climatic factors. Moreover, the local coefficients estimated by MGWR model furnish detailed information regarding the strength and direction of the relationships between predictors and COVID-19 cases and deaths within each spatial unit. The findings emphasize the significance of addressing multicollinearity in spatial modelling. It is beneficial for accurate parameter estimation, proper interpretation of coefficients, improved spatial analysis, and providing reliable insights to support decision-making in spatial contexts.

Application of the Random forest Method to Identify Food and Beverage Industries Experiencing Raw Material Difficulties

Iman Jihad Fadillah — 2024-06-11

The food and beverage industry experienced a significant increase after the pandemic. However, challenges continue to hit this industry, especially for micro and small scale businesses. To overcome this problem, the right approach is needed. One of the first steps is to provide quality data as a basis for decision making and problem solving. However, statistical activities such as censuses and surveys often face obstacles in the form of missing values. One effective method for dealing with this is using the random forest method. This research aims to use a machine learning-based imputation method, namely the random forest method, to identify micro and small scale food and beverage industries that are experiencing raw material difficulties. The research results show that the random forest method provides accurate and consistent predictions in identifying food and beverage industries experiencing raw material difficulties. However, it is also necessary to consider the relatively long computing time for implementing this method.

Study of Small Area Estimation when Nighttime Lights as an Auxiliary Information is Measured with Error

Ardi Surya — 2024-06-11

The need for accelerated development requires rapid data collection. In today's increasingly advanced technological landscape, the utilization of big data emerges as a highly reliable solution for data collection. One exemplary form of big data is the daily capture of satellite imagery, particularly nighttime lights (NTL). NTL serves as a valuable product derived from satellite imagery and can be employed as an alternative dataset for analysis. This research utilizes Nighttime lights as an auxiliary variable to estimate the average household per capita expenditure in small areas, namely districts, employing the empirical best linear unbiased prediction Fay Herriot (EBLUP FH) method and small area estimation by incorporating measurement error effects on the covariate (SAE-ME). The study demonstrates that Nighttime lights can be employed as an alternative auxiliary variable for estimating the average per capita expenditure in districts, as evidenced by a lower RRMSE compared to direct estimation results. However, the measurement error effects on the NTL covariate should be considered by employing a model that takes into account measurement errors. The SAE-ME method provides estimated average expenditure values at the district level that closely align with BPS publications, with an average RRMSE per district of 7.5 percent.

Study of Spatial Autoregressive Regression With Heteroskedasticity Using the Generalized Method of Moments and Bayesian Approach

Abialam Koesnandy H — 2024-06-11

Spatial dependence and spatial heteroskedasticity are problems in spatial regression. Spatial autoregressive regression (SAR) concerns only to the dependence on lag. The estimation of SAR parameters containing heteroskedasticity using the maximum likelihood estimation (MLE) method provides biased and inconsistent estimators. The alternative method that can be used are generalized method of moments (GMM) and Bayesian method. GMM uses a combination of linear and quadratic moment functions simultaneously so that the computation is easier than MLE. Bayesian method solves heteroskedasticity by modeling the structure of variance-covariance matrix. The bias are used to evaluate the GMM and Bayes in estimating parameters of SAR model with heteroskedasticity disturbances in simulation data. The results show that GMM and Bayes provides the bias of parameter estimates relatively consistent and smaller with larger number of observations. GMM and Bayes methods are applied to district/city GRDP data in Indonesia. The result show GMM method with Eksponential Distance Weights (EDW) matrix produces the minimum variance and the largest pseudo-R²

Effectiveness of SMOTE-ENN to Reduce Complexity in Classification Model

Ines Riantika — 2024-06-11

A failure to produce classification models with high performance might be caused by the dataset's characteristics, such as the between-class overlapping and the class imbalance. The higher the data complexity, the more complicated it is for the algorithm to find good models. Combining the issues of class imbalance and overlapping would make the problem more challenging. To deal with this problem, this research implemented a hybrid class-balancing technique named SMOTE-ENN. This technique adds observations to the minority class to balance the class frequencies. After that, it removes some observations to reduce the degree of overlapping. The research revealed that SMOTE-ENN succeeds in doing that. We employed a random forest method to evaluate it. In 28 out of 46 cases we investigated, the new datasets generated by SMOTE-ENN could produce models with higher accuracy.