Predicting loneliness in older age using two measures of loneliness

Older people are especially vulnerable to loneliness and this has become a major public health concern for people in later life. In this paper, we propose a machine learning based approach to predict loneliness probability using two gradient boosting algorithms, XGBoost and LightGBM. The predictive models are built using data from a large nationally representative sample from, the English Longitudinal Study of Ageing (ELSA) that had seven successive waves (2002–2015). Two measures of loneliness were applied to investigate the impact of different measure strategies on the prediction of loneliness. The models achieved good performance with a high Area Under Curve (AUC) and a low Logarithmic Loss (LogLoss) on the test data, i.e. AUC (0.88) and LogLoss (0.24) using the single-item direct measure of loneliness, and AUC (0.84) and LogLoss (0.31) using the multi-item indirect measure of loneliness. A wide range of variables were investigated to identify significant risk factors associated with loneliness. Specific categories associated with important variables were also recognized by the models. Such information will further enhance our understanding and knowledge of the causes of loneliness in elderly people.