Can social media data be used to make reasonably accurate estimates of electoral outcomes? We conducted a meta-analytic review to examine the predictive performance of different features of social media posts and different methods in predicting political elections: (1) content features; and (2) structural features. Across 45 published studies, we find significant variance in the quality of predictions, which on average still lag behind those in traditional survey research. More specifically, our findings that machine learning-based approaches generally outperform lexicon-based analyses, while combining structural and content features yields most accurate predictions.