Perbandingan Metode Generalized Linear Mixed Models (GLMM) Dan Panel Data Regression (PDR) Untuk Total Kerugian Akibat Bencana Alam

Sidabutar, Renjiro Jusiyael (2026) Perbandingan Metode Generalized Linear Mixed Models (GLMM) Dan Panel Data Regression (PDR) Untuk Total Kerugian Akibat Bencana Alam. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5006221087-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (44MB) | Request a copy

Abstract

Bencana alam di kawasan Asia telah memberikan dampak sangat serius terhadap stabilitas ekonomi terutama dalam bentuk kerugian aset setiap kerusakan. Hal ini menjadi ancaman serius sehingga menjadi krusial untuk mitigasi risiko finansial dan penentuan cadangan klaim. Namun, data kerugian bencana alam sering kali memiliki karakteristik yang sulit ditangkap secara optimal oleh pemodelan sederhana. Penelitian ini bertujuan untuk membandingkan performa dua metode statistika, yaitu Generalized Linear Mixed Models (GLMM) dengan distribusi lognormal dan Panel Data Regression (PDR) dalam memodelkan total kerugian ekonomi akibat fenomena iklim di berbagai negara di Asia. Metodologi yang digunakan dalam penelitian ini mengintegrasikan data panel yang mencakup dimensi lintas wilayah (cross-section) dan dimensi waktu (time-series). Model pertama, GLMM data panel dengan distribusi lognormal mengakomodasi sifat data kerugian yang bersifat kontinu, bernilai positif, dan memiliki kurva dengan ekor memanjang kekanan. Selain itu, dianalisis dengan tambahan unsur cross section dan time series dengan mangambil Fixed Effect dan Random Effect. Model kedua, regresi data panel digunakan untuk menangkap efek spesifik antarnegara melalui pemilihan model terbaik antara Common Effect Model (CEM), Fixed Effect Model (FEM) dan Random Effect Model (REM). Data yang digunakan merupakan data sekunder yang diperoleh dari EM-DAT, World Bank, dan NOAA dengan struktur unbalanced panel. Variabel yang dianalisis dalam penelitian seperti Total Damage, Total Day, Disaster Type, Total Death, Total Affected, CPI, GDP per capita, Population Millions, Urbanization rate pct, ENSO Phase, dan Median Age. Perbandingan performa kedua model dievaluasi berdasarkan kriteria statistik Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), serta tingkat akurasi prediksi yang diukur melalui Root Mean Square Error (RMSE). Evaluasi performa model menunjukkan bahwa Regresi Data Panel dengan spesifikasi CEM memiliki kinerja yang lebih baik dibandingkan GLMM lognormal. Spesifikasi CEM mencatatkan nilai AIC dan BIC yang lebih rendah sehingga menunjukkan kebaikan model yang lebih efisien (parsimonious) dan terhindar dari over-specification. Hal ini dipicu oleh pola hubungan struktural antar variabel data yang cenderung bergerak secara linear, sehingga intersep acak pada GLMM kurang relevan dan justru menimbulkan over-specification error.
==========================================================================================================================================
Natural disasters in Asia have caused severe impacts on economic stability, particularly through substantial losses of physical assets and infrastructure. This condition poses significant challenges for financial risk mitigation and insurance claim reserve estimation. However, disaster loss data often exhibit complex characteristics that cannot be adequately captured using conventional statistical models. This study aims to compare the performance of two statistical approaches, namely the Generalized Linear Mixed Model (GLMM) with a lognormal distribution and Panel Data Regression (PDR), in modeling the total economic losses caused by natural disasters across Asian countries. The methodology integrates panel data consisting of both cross-sectional and time-series dimensions. The first approach, the lognormal GLMM, is designed to accommodate continuous, strictly positive, and right-skewed loss data with variance that increases proportionally to the mean. In addition, the model incorporates both cross-sectional and time-series effects through fixed-effect and random-effect specifications. The second approach, Panel Data Regression, is employed to capture country-specific effects by selecting the most appropriate model among the Common Effect Model (CEM), Fixed Effect Model (FEM), and Random Effect Model (REM). The study utilizes secondary data obtained from EM-DAT, the World Bank, and NOAA, organized as an unbalanced panel dataset. The variables analyzed include Total Damage, Disaster Days, Disaster Type, Total Deaths, Total Affected, Consumer Price Index (CPI), GDP per capita, Population Millions, Urbanization Rate (%), ENSO phase, and Median Age. The performance of both models is evaluated using the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and prediction accuracy measured by the Root Mean Square Error (RMSE). Model performance evaluation reveals that Panel Data Regression with a CEM specification outperforms the lognormal GLMM. CEM specification achieved lower AIC and BIC values, demonstrating a more parsimonious fit that avoids over-specification error. This finding suggests that the structural relationships among the observed variables are predominantly linear; consequently, the inclusion of random intercepts in the lognormal GLMM provides limited additional explanatory power and instead leads to model over-specification.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Asia, Bencana Alam, Distribusi Lognormal, Generalized Linear Mixed Models, Regresi Data Panel, Asia, Generalized Linear Mixed Models (GLMM), Lognormal Distribution, Natural Disaster, Panel Data Regression.
Subjects:	Q Science Q Science > Q Science (General) > Q180.55.M38 Mathematical models Q Science > QA Mathematics Q Science > QA Mathematics > QA246.8 Gaussian
Divisions:	Faculty of Mathematics, Computation, and Data Science > Actuaria > 94203-(S1) Undergraduate Thesis
Depositing User:	Renjiro Jusiyael Sidabutar
Date Deposited:	29 Jul 2026 07:46
Last Modified:	29 Jul 2026 07:46
URI:	http://repository.its.ac.id/id/eprint/139706

Actions (login required)

View Item