riccardomonforte7-lgtm/malawi-child-depression-ml

Machine learning project using UNICEF MICS data to classify depression-related responses among children aged 5–17 in Malawi. # Malawi Child Depression ML Machine learning project using UNICEF Multiple Indicator Cluster Survey (MICS) data to investigate whether children aged 5–17 in Malawi are likely to report any form of depression related symptoms. ## Overview This project analyses a subset of UNICEF MICS data collected in Malawi during 2019–2020. The dataset includes child-level, maternal, and household-level variables, covering aspects such as child background, child labour, household characteristics, water and sanitation, maternal background, life satisfaction, victimisation, and family context. The target variable is based on survey responses to whether the child seemed very sad or depressed. For the modelling task, the original responses were converted into a binary outcome: - **No Depression**: the child never had depressive symptoms. - **Depression** : all responses other than "never". The aim of the project was to build a **predictive model** to classify whether children reported any depressive symptoms, while also identifying the **factors** that appeared most relevant in explaining this outcome. ## Methodology The analysis was carried out in Python and included: - Exploratory Data Analysis on numeric and categorical variables - Preprocessing pipelines for different variable types - Handling of missing values and categorical encodings - Feature selection using L1-penalised Logistic Regression - Final Logistic Regression model with L2 penalty - Comparison with a Random Forest classifier - Interpretation of selected variables through model coefficients and feature importance ### Feature selection approach Feature selection was performed using a repeated L1-penalised Logistic Regression procedure. The model was repeatedly fitted on bootstrap samples of the training set. At each iteration, the L1 penalty encouraged sparse solutions by shrinking some coefficients to zero. Features were then ranked according to how frequently they were selected across the repeated fitting …