Anang Kurnia

This study examines the statistical determinants of graduate income using Maximum Likelihood Estimation implemented through the Iteratively Reweighted Least Squares algorithm. A generalized linear model with a logit link was estimated to assess how technical competence, English proficiency, and tuition-financing type influence the probability of earning above the regional minimum wage. The IRLS procedure enabled iterative refinement of coefficient estimates and standard errors, with convergence achieved as successive updates stabilized, indicating a well-behaved likelihood surface. Results show that academic competence and English proficiency significantly increase the log-odds of higher income, highlighting the importance of human-capital attributes in early-career outcomes. Study duration also exhibits a positive and statistically significant association with the likelihood of earning above the minimum wage. In contrast, tuition-financing type shows a positive but statistically non-significant effect, suggesting that its influence on graduate income requires further investigation. Although the model identifies several meaningful determinants, its moderate discriminative capacity indicates that pre-graduation characteristics alone cannot fully explain income variation. By explicitly incorporating MLE with IRLS, this study provides a transparent statistical framework for analyzing graduate employability outcomes and contributes empirical evidence on factors associated with early-career income attainment. The findings encourage future research integrating labor-market, employer, and longitudinal career information to improve predictive performance.

Stunting is a condition of growth failure in toddlers due to chronic malnutrition, becoming an issue in various regions of Indonesia. West Nusa Tenggara Province has one of Indonesia's highest stunting prevalence rates, calculated at 29.8% in 2024, which is more than twice the national target of 14%. Appropriate and efficient policy-making for stunting reduction requires reliable estimates for small areas like districts. This study aims to produce more reliable district-level estimates of stunting prevalence for districts level area, while direct estimation from the Survei Kesehatan Indonesia (SKI) is unreliable due to limited sample size. This research develops a threefold hierarchical Bayesian (HB) Small Area Estimation (SAE) model based on the Poisson-Gamma distribution to model the number of stunted toddlers as discrete count data with overdispersion. The proposed model incorporates auxiliary variables from official sources and includes three hierarchical random effects representing district, regency/municipality, and grouped regency/municipality levels based on geographical structure. The results show that the threefold HB SAE model achieves convergent parameter estimates and provides more stable and precise district-level stunting prevalence estimates compared to direct estimators. The multilevel model performs better than one random effect model as reflected by the lower LOOIC. The findings also suggest that districts in Sumbawa Island, particularly in the eastern part and areas located farther from regency/municipality capitals, tend to have higher stunting prevalence. However, this study is limited by the assumption of independence among area-level random effects and restricted availability of auxiliary variables. This study contributes methodologically by extending Poisson-based SAE literature through the application of a threefold HB framework in stunting estimation and provides district-level stunting statistics that support evidence-based policymaking and targeted interventions aligned with SDG Target 2.2.

This paper develops a Small Area Estimation Mixed Model Neural Network (SAE-MNN) for unit-level estimation of area-level proportions under clustered binary responses. The method is motivated by a limitation of conventional generalized linear mixed model (GLMM)-based small area estimation, namely the linear specification of the fixed-effects component on the latent scale, which may be inadequate when auxiliary variables exhibit nonlinear effects and higher-order interactions. The proposed model replaces the linear predictor with a neural-network-based nonlinear function while retaining area-level random effects within a likelihood-based mixed-model framework. Thus, SAE-MNN combines nonlinear function approximation with the borrowing-strength mechanism required for coherent small area inference. The method is evaluated through a controlled simulation study and a real-data application. The simulation study considers four scenarios generated by combining two predictor structures (linear and interaction-based nonlinear) with two levels of between-area heterogeneity (small and large), and compares SAE-MNN with a conventional GLMM and a standard deep neural network (DNN). SAE-MNN performs robustly across all scenarios and is especially effective when nonlinear interaction and substantial area-level heterogeneity coexist. In the most complex scenario, SAE-MNN attains the lowest median RRMSE (8.389%), compared with 11.474% for GLMM and 18.826% for DNN, while maintaining empirical coverage between 0.89 and 0.94. In the real-data application, using KSA segment-level harvest proportions to estimate subdistrict-level monthly harvest proportions, SAE-MNN shows the closest overall agreement with the direct estimator, reproduces the main temporal pattern more faithfully than GLMM, and yields moderate shrinkage without excessive smoothing. Overall, the results indicate that SAE-MNN provides a principled compromise between the rigidity of GLMM and the purely predictive character of DNN, thereby extending hybrid small area estimation methodology to binary and proportional outcomes in data-rich settings with nonlinear auxiliary information and area-level heterogeneity.

Hierarchical item response data require a multilevel approach to capture the diversity of respondent abilities and item characteristics by separating variations between examinees and schools to make item parameter and ability estimates more precise. This study aims to develop a two-parameter logistic multilevel item response theory (MIRT 2PL) model, employing the maximum marginal likelihood method with Gauss-Hermite quadrature (MML-GHQ) to estimate the rank order of examinees’ abilities. This study specifically investigates the efficiency and accuracy of the MIRT 2PL model with MML-GHQ to predict the ability rankings of examinees. The research incorporates both simulated and empirical data. The simulation study generated item response data under a two-level hierarchical structure, where examinees were nested within schools. The population consisted of 50 schools, with 20–30 students per school and five items. Each examinee’s ability was modelled as a combination of school-level and individual-level effects, under two conditions of school variability: high (τ = 1.2) and low (τ = 0.6). Random samples were drawn from the population, and the sampling process was repeated 10 times to assess consistency. The analysis included estimating item parameters, variance components, and examinees’ abilities using the EAP approach. Model performance was evaluated using RMSE, Spearman correlation, and computation time. Results indicated that MML-GHQ produced accurate and consistent rank estimates, particularly under high school variability. Using PISA data, increasing the number of schools, students, and items in the empirical data yielded results consistent with the simulation study. In conclusion, the MIRT 2PL model with MML-GHQ offers an effective and efficient alternative for estimating ability rankings in hierarchical item response data.

This study presents the development and validation of a Bayesian hierarchical model to estimate the incidence rate of dengue fever (DF) in West Java, Indonesia. Bayesian hierarchical models offer powerful tools for handling uncertainty and regional heterogeneity, yet their implementation remains challenging—especially in complex datasets with multilevel structures. The proposed model incorporates both random intercepts (for regencies/cities) and random slopes (for year), with various prior distribution scenarios tested to ensure robustness. Among the tested predictors, population density was found to significantly influence DF incidence. Model performance evaluation using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) yielded values of 31.26 and 48.77, respectively, indicating good predictive accuracy. This research highlights the effectiveness of hierarchical Bayesian modeling for epidemiological analysis and contributes to more targeted public health strategies in dengue-endemic regions

Mixed effects model combines fixed effects and random effects, allowing for the analysis of data with both fixed and random variations. This modeling approach is widely utilized across various fields. In R, the lme4 package is commonly employed to estimate mixed effects models using Restricted Maximum Likelihood (REML). There are several methods for estimating model parameters, including Bayesian inference, which has gained prominence with ongoing research advancements. Bayesian inference using Markov Chain Monte Carlo (MCMC) is among the most widely used Bayesian methods. Bayesian inference leverages probabilistic distributions to estimate parameters.to understand the general overview of life expectancy, serving as an indicator of survival time across different continents in the Gapminder dataset, it's essential to identify relevant variables after computing mixed effects predictions using Maximum Likelihood and REML estimation. This involves predicting life expectancy by integrating both random and fixed effects, determining relevant variables after estimating the Mixed Effects Model using REML Bootstrap estimation, and identifying influential variables after estimating the Mixed Effects Model using Bayesian MCMC inference. The methods employed include REML, Bootstrapped-REML, and Bayesian MCMC. The results indicate that all inference methods can be utilized to estimate parameters, with all predictor variables influencing life expectancy, except for the population variable. Further research is recommended to utilize data with more complex predictor variables.

Economic inequality remains one of the most persistent challenges faced by Indonesia as a developing country. Previous studies have predominantly employed conventional models such as Ordinary Least Squares (OLS) or Panel Least Squares. However, these models are often inappropriate, as they fail to account for the bounded nature of inequality indices such as the Gini ratio, which ranges between 0 and 1. Beta regression offers a more appropriate alternative. In the context of panel data, Generalized Linear Mixed Models (GLMM) and Generalized Estimating Equations (GEE) are commonly used to handle correlated data; however, their integration with nonlinear models for longitudinal Beta-distributed responses remains limited. This study proposes a novel GLMM-GEE-Tree modeling approach for Beta-distributed response data. The proposed model combines GLMM (to capture individual random effects), GEE (to handle temporal correlation and provide robust marginal estimates), and Regression Trees (to address nonlinear relationships and complex interactions). The aim is to simultaneously tackle the challenges of proportional responses, panel structure, random effects, correlation, and nonlinearity. Empirical validation uses Gini ratio data from 34 Indonesian provinces spanning 2018 to 2024. The findings reveal that in this empirical data, the GLMM-GEE-Tree model outperforms alternative models, achieving an R² of 0.472 and a QIC of 13.435 and yielding the lowest AIC and BIC values.

The high economic value and increasing demand for Strobilanthes crispa have created a market incentive to adulterate it with cheaper alternatives, such as Piper aduncum, posing significant risks to consumer safety and product efficacy. This study addresses the critical challenge of multiclass adulteration detection in S. crispa from P. Aduncum using Fourier Transform Infrared (FTIR) spectroscopy combined with optimized ensemble machine learning methods. The high-dimensional spectral data were processed with advanced preprocessing techniques, particularly Standard Normal Variate (SNV) and its combinations (SNV+SG and SG+SNV), which effectively reduced baseline shifts and noise. These enhancements significantly improved spectral quality and model performance. Among the evaluated models, Light Gradient Boosting Machine (LightGBM) achieved the highest classification accuracy of 94.58% ± 2.69. Extreme Gradient Boosting (XGBoost) also demonstrated robust performance, attaining 93.38% ± 2.27 accuracy with SNV+SG preprocessing. Both boosting algorithms proved effective in handling class imbalance, as indicated by high Balanced Accuracy and G-mean values. Statistical analysis using ANOVA confirmed that the choice of model and preprocessing method significantly influenced performance, whereas the Cuckoo Search algorithm's nest had a significant effect on computational time. Feature importance analysis identified an essential subset of spectral features that consistently contributed to the classification performance of XGBoost and LightGBM. Functional group analysis revealed that distinct chemical fingerprints native to S.crispa exhibited strong aromatic N-H stretching and C-H bending. In contrast, impurities such as P.aduncum showed dominant alkyl C-H stretching and C=O stretching. These chemical distinctions provide a reliable foundation for detecting adulteration.

West Java, one of the largest provinces in Indonesia with a population exceeding 50 million, reported a poverty rate of 7.62% in 2023. Data from the national socio-economic survey or survei sosial ekonomi nasional (SUSENAS) show that poverty is multidimensional, encompassing aspects of employment, education, sanitation, housing, food security, technology, and government assistance. Addressing this complexity requires identifying the most influential factors that determine household welfare. This study applies and compares three feature selection approaches—filter, wrapper, and embedded—to the SUSENAS dataset to evaluate their effectiveness in identifying key poverty determinants. By prioritizing variables with the strongest predictive power, the study provides an evidence-based framework for more efficient and targeted poverty alleviation strategies. Results indicate that the information filter method combined with random forest (RF) and the least absolute shrinkage and selection operator (LASSO) embedded method combined with logistic regression (LR) deliver the best performance, improving model accuracy while reducing more than 65% of irrelevant features. The selected indicators highlight critical sectors such as food security, housing, and access to technology, which can serve as short-term policy priorities. In the long term, broader interventions in education, employment, sanitation, and government support are recommended. These findings demonstrate how data-driven feature selection can guide effective policy design for reducing poverty in West Java.