Public Health Discussions on Social Media: An Evaluation of Automated Sentiment Analysis Methods (Preprint)

Background Sentiment analysis is one of the most widely used methods for mining and examining text. Social media researchers need guidance on choosing between manual and automated sentiment analysis methods. Objective Popular sentiment analysis tools based on natural language processing (NLP; VADER Valence Aware Dictionary for Sentiment Reasoning, TEXT2DATA T2D, and Linguistic Inquiry and Word Count LIWC-22), and a large language model (ChatGPT 4.0) were compared with manually coded sentiment scores, as applied to the analysis of YouTube comments on videos discussing the opioid epidemic. Sentiment analysis methods were also examined regarding ease of programming, monetary cost, and other practical considerations. Methods Evaluation methods included descriptive statistics, receiver operating characteristic (ROC) curve analysis, confusion matrices, Cohen κ, accuracy, specificity, precision, sensitivity (recall), F1-score harmonic mean, and the Matthews correlation coefficient. An inductive, iterative approach to content analysis of the data was used to obtain manual sentiment codes. Results A subset of comments were analyzed by a second coder, producing good agreement between the 2 coders’ judgments (κ=0.734). YouTube social media about the opioid crisis had many more negative comments (4286/4871, 88%) than positive comments (79/662, 12%), making it possible to evaluate the performance of sentiment analysis models in an unbalanced dataset. The tone summary measure from LIWC-22 performed better than other tools for estimating the prevalence of negative versus positive sentiment. According to the ROC curve analysis, VADER was best at classifying manually coded negative comments. A comparison of Cohen κ values indicated that NLP tools (VADER, followed by LIWC’s tone and T2D) showed only fair agreement with manual coding. In contrast, ChatGPT 4.0 had poor agreement and failed to generate binary sentiment scores in 2 out of 3 attempts. Variations in accuracy, specificity, precision, sensitivity, F1-score, and MCC did not reveal a single superior model. F1-score harmonic means were 0.34-0.38 (SD 0.02) for NLP tools and very low (0.13) for ChatGPT 4.0. None of the MCCs reached a strong correlation level. Conclusions Researchers studying negative emotions, public worries, or dissatisfaction with social media face unique challenges in selecting models suitable for unbalanced datasets. We recommend VADER, the only cost-free tool we evaluated, due to its excellent discrimination, which can be further improved when the comments are at least 100 characters long. If estimating the prevalence of negative comments in an unbalanced dataset is important, we recommend the tone summary measure from LIWC-22. Researchers using T2D must know that it may only score some data and, compared with other methods, be more time-consuming and cost-prohibitive. A general-purpose large language model, ChatGPT 4.0, has yet to surpass the performance of NLP models, at least for unbalanced datasets with highly prevalent (7:1) negative comments.