Determining well-being during a crisis based on Twitter data

The objective of this paper is to determine how subjective well-being (SWB) can be predicted based on Twitter (or X) data during a crisis. Typically, a questionnaire, namely the international positive and negative affect schedule short-form (I-PANAS-SF), is employed to assess affective SWB. As natural language processing (NLP) continues to advance, two kinds of techniques (sentiment analysis tools and models based on textual representation) are employed to harness social media data for estimating affective SWB. First, we estimate affective SWB during a crisis by two popular sentiment analysis tools [i.e., valence aware dictionary for sEntiment reasoner (VADER) and TextBlob]. We evaluate the reliability of sentiment analysis tools as proxies for affective SWB and compare their performance to a manually labeled scale. Second, we conduct an exploratory data analysis to identify the hidden topics and major change points related to affective SWB. Third,we investigate five different textual representation techniques over five machine learning algorithms [i.e., elastic net regression, support vector machines, random forests, neural networks, and eXtreme gradient boosting (XGBoost)]. Finally, we identify key drivers of affective SWB predictions by analyzing feature importance. Our findings reveal that for the three target variables in affective SWB (i.e., total SWB, positive affect, and negative affect), sentiment analysis tools are suboptimal for predicting affective SWB, as evidenced by their poor performance in the evaluation metrics. The models constructed using the combination of textual representation techniques and machine learning algorithms exhibit relatively strong performance. The models with the highest predictive capability include OpenAI's Ada embedding combined with neural networks for total SWB and positive affect, and OpenAI's Ada embedding in conjunction with support vector machines for negative affect, achieving RMSE reductions of 41.60%, 41.77%, and 38.00%, respectively, compared to sentiment analysis tools. Our empirical results show that models using advanced embedding methods (RoBERTa and OpenAI's ada embedding) are promising methods for predicting affective SWB. This research addresses a gap in the literature concerning the field of predicting affective SWB in the context of a crisis by utilizing advanced NLP techniques, machine learning, and social media data.