Personalized predictions have shown promises in various disciplines but they are fundamentally constrained in their ability to generalize across individuals. These models are often trained on limited datasets which do not represent the fluidity of human functioning. In contrast, generalized models capture normative behaviors between individuals but lack precision in predicting individual outcomes. This paper aims to balance the tradeoff between one-for-each and one-for-all models by clustering individuals on mutable behaviors and conducting cluster-specific predictions of psychological constructs in a multimodal sensing dataset of 754 individuals. Specifically, we situate our modeling on social media that has exhibited capability in inferring psychosocial attributes. We hypothesize that complementing social media data with offline sensor data can help to personalize and improve predictions. We cluster individuals on physical behaviors captured via Bluetooth, wearables, and smartphone sensors. We build contextualized models predicting psychological constructs trained on each cluster's social media data and compare their performance against generalized models trained on all individuals' data. The comparison reveals no difference in predicting affect and a decline in predicting cognitive ability, but an improvement in predicting personality, anxiety, and sleep quality. We construe that our approach improves predicting psychological constructs sharing theoretical associations with physical behavior. We also find how social media language associates with offline behavioral contextualization. Our work bears implications in understanding the nuanced strengths and weaknesses of personalized predictions, and how the effectiveness may vary by multiple factors. This work reveals the importance of taking a critical stance on evaluating the effectiveness before investing efforts in personalization.

The toll from gun violence in American K-12 schools has escalated over the past 20 years. School administrators face pressure to prepare for possible active shootings, and often do so through drills, which can range from general lockdowns to simulations, involving masked “shooters” and simulated gunfire, and many variations in between. However, the broad and lasting impact of these drills on the well-being of school communities is poorly understood. To that end, this article applies machine learning and interrupted time series analysis to 54 million social media posts, both pre- and post-drills in 114 schools spanning 33 states. Drill dates and locations were identified via a survey, then posts were captured by geo-location, school social media following, and/or school social media group membership. Results indicate that anxiety, stress, and depression increased by 39–42% following the drills, but this was accompanied by increases in civic engagement (10–106%). This research, paired with the lack of strong evidence that drills save lives, suggests that proactive school safety strategies may be both more effective, and less detrimental to mental health, than drills.

Effective ways to measure employee job satisfaction are fraught with problems of scale, misrepresentation, and timeliness. Current methodologies are limited in capturing subjective differences in expectations, needs, and values at work, and they do not lay emphasis on demographic differences, which may impact people's perceptions of job satisfaction. This study proposes an approach to assess job satisfaction by leveraging large-scale social media data. Starting with an initial Twitter dataset of 1.5M posts, we examine two facets of job satisfaction, pay and supervision. By adopting a theory-driven approach, we first build machine learning classifiers to assess perceived job satisfaction with an average AUC of 0.84. We then study demographic differences in perceived job satisfaction by geography, sex, and race in the U.S. For geography, we find that job satisfaction on Twitter exhibits insightful relationships with macroeconomic indicators such as financial wellbeing and unemployment rates. For sex and race, we find that females express greater pay satisfaction but lower supervision satisfaction than males, whereas Whites express the least pay and supervision satisfaction. Unpacking linguistic differences, we find contrasts in different groups' underlying priorities and concerns, e.g., under-represented groups saliently express about basic livelihood, whereas the majority groups saliently express about self-actualization. We discuss the role of frame of reference and the "job satisfaction paradox", conceptualized by organizational psychologists, in explaining our observed differences. We conclude with theoretical and sociotechnical implications of our work for understanding and improving worker wellbeing.

Research has revealed the potential of social media as a source of large-scale, verbal, and naturalistic data for human behavior both in real-time and longitudinally. However, the in-practice utility of social media to assess and support wellbeing will only be realized when we account for extraneous factors. A factor that might confound our ability to make inferences is the phenomenon of the ``observer effect''---that individuals may deviate from their otherwise typical social media use because of the awareness of being monitored. This paper conducts a causal study to measure the observer effect in longitudinal social media use. We operationalized the observer effect in two dimensions of social media (Facebook) use---behavioral and linguistic changes. Participants consented to Facebook data collection over an average retrospective period of 82 months and an average prospective period of 5 months around the enrollment date to our study. We measured how they deviated from their expected social media use after enrollment. We obtained expected use by extrapolating from historical use using time-series (ARIMA) forecasting. We find that the deviation in social media use varies across individuals based on their psychological traits. Individuals with high cognitive ability and low neuroticism immediately decreased posting after enrollment, and those with high openness significantly increased posting. Linguistically, most individuals decreased the use of first-person pronouns, reflecting lowered sharing of intimate and self-attentional content. While some increased posting about public-facing events, others increased posting about family and social gatherings. We validate the observed changes with respect to psychological traits drawing from psychology and behavioral science theories, such as self-monitoring, public self-consciousness, and self-presentation. The findings provide recommendations to correct observer effects in social media data-driven assessments of human behavior.