Predicting age groups of Twitter users based on language and metadata features
Health organizations are increasingly using social media, such as Twitter, to disseminate health messages to target audiences. Determining the extent to which the target audience (e.g., age groups) was reached is critical to evaluating the impact of social media education campaigns. The main objective of this study was to examine the separate and joint predictive validity of linguistic and metadata features in predicting the age of Twitter users. We created a labeled dataset of Twitter users across different age groups (youth, young adults, adults) by collecting publicly available birthday announcement tweets using the Twitter Search application programming interface. We manually reviewed results and, for each age-labeled handle, collected the 200 most recent publicly available tweets and user handles' metadata. The labeled data were split into training and test datasets. We created separate models to examine the predictive validity of language features only, metadata features only, language and metadata features, and words/phrases from another age-validated dataset. We estimated accuracy, precision, recall, and F1 metrics for each model. An L1-regularized logistic regression model was conducted for each age group, and predicted probabilities between the training and test sets were compared for each age group. Cohen's d effect sizes were calculated to examine the relative importance of significant features. Models containing both Tweet language features and metadata features performed the best (74% precision, 74% recall, 74% F1) while the model containing only Twitter metadata features were least accurate (58% precision, 60% recall, and 57% F1 score). Top predictive features included use of terms such as "school" for youth and "college" for young adults. Overall, it was more challenging to predict older adults accurately. These results suggest that examining linguistic and Twitter metadata features to predict youth and young adult Twitter users may be helpful for informing public health surveillance and evaluation research.