The explosion of big, granular social data has enabled us to observe society from a microscopic perspective. With major events driving and reshaping our social systems, it is essential to exploit these data so that we can help drive these systems to a state of social well-being and sustainability. This thesis develops data-driven methods to study social systems from a bottom-up perspective. We divide it into two parts: one where granular, individual-level social data are available to analyze and one where they are not, so that we have to infer them. In the first part, we analyze massive social media data containing Twitter discussions around Covid-19 and climate change. We study each of these discussions from distinct but complementary perspectives. For Covid-19, we quantify the public risk perception and emotion during the pandemic by exploiting the natural language used in the tweets. We find evidence of psychophysical numbing: Twitter users increasingly fixate on mortality, but in a decreasingly emotional and increasingly analytic tone. For climate change, we quantify the polarization dynamics based on the interaction structure between Twitter users. We find two stable, highly polarized groups: climate believers and climate skeptics, whose polarization drops significantly during the "FridaysForFuture" strikes of September 2019. In the second part, we develop methods for inferring individual-level data of complex system models when the data available are noisy, aggregated, and incomplete. Assuming our model is a dynamical system, we investigate under which conditions we can infer accurate initial conditions using incomplete data. The way the data are aggregated, their levels of noise, and the model's complexity highly influence the quality of the inference. We thus propose methods to estimate individual-level data on the fly as observations become available. We validate these methods for several chaotic systems and an agent-based model of social opinion dynamics. We hope this work helps bridge the gap in designing models that better predict possible societal pathways.