Comparison of machine learning and deep learning models for longitudinal infant mortality rate data: a study based on real world and simulated data

Date

2026-8-12

Author

Özhayat, Okan

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

84
views

0
downloads

Cite This

Infant mortality remains a key indicator of national health and development, as accurate cross-country prediction is essential for guiding public health policy. However, comparative analyses in the literature remain constrained by reliance on either controlled simulations or isolated empirical datasets. This thesis presents a comprehensive comparison of eight algorithms: LightGBM, XGBoost, Elastic Net, CatBoost, Random Forest, a feedforward neural network, and the recurrent LSTM and GRU architectures. The balanced panel of 194 countries observed annually from 2013 to 2023 via the UNICEF Global Dataflow is used. Each algorithm is evaluated under five modeling specifications. Also, two feature configurations are used: full and correlation-filtered. This empirical analysis is complemented by a Monte Carlo simulation study spanning 48 data-generating scenarios. Each of them was replicated 100 times. Performance is assessed using RMSE, MAE, MASE, and the adjusted coefficient of determination. Results reveal a marked divergence between the two regimes. CatBoost and XGBoost achieve the most accurate and reliable predictions on the empirical panel. In the controlled simulation, they overfit and are outperformed by GRU and Elastic Net in terms of stability. The thesis's main methodological contribution is to show that simulation-based rankings offer necessary but insufficient evidence of predictive validity. Also, robust evaluation requires integrating controlled diagnostic simulation with empirical benchmarking on representative data.

Subject Keywords

Infant mortality rate, Longitudinal data, Gradient boosting, Monte carlo simulation, Machine learning

URI

https://hdl.handle.net/11511/120647

Collections

Graduate School of Natural and Applied Sciences, Thesis