Source-linked AI summary
Weather impacts expressed sentiment
Patrick Baylis, Nick Obradovich, Yury Kryvasheyeu, Haohui Chen, Lorenzo Coviello, Esteban Moro, Manuel Cebrian, James H. Fowler
TL;DR
The paper investigates whether meteorological conditions are associated with human expressed sentiment, addressing conflicting prior evidence. It analyzes 3.5 billion Facebook and Twitter posts, finding that several less-ideal weather conditions are associated with worsened sentiment even after weather-related posts are excluded.
Problem
Prior empirical studies of weather and emotional state found conflicting results, partly because large-scale data on emotional states were lacking.
Method
The study uses social-media sentiment as a correlate of emotional states and analyzes Facebook and Twitter posts from tens of millions of United States residents.
Results
Cold and hot temperatures, precipitation, narrower daily temperature ranges, humidity, and cloud cover are associated with worsened expressed sentiment, including in posts without weather-related terms.
Takeaways & Limitations
The findings provide observational evidence that less-ideal weather conditions may functionally alter human emotional states, insofar as expressed sentiment validly proxies underlying emotions.
Takeaways & Limitations
Social-media sentiment is an imperfect and noisy proxy for individuals’ underlying emotional states, because the study lacks daily self-reported emotional states.
Abstract
from arXiv · showhide
We conduct the largest ever investigation into the relationship between meteorological conditions and the sentiment of human expressions. To do this, we employ over three and a half billion social media posts from tens of millions of individuals from both Facebook and Twitter between 2009 and 2016. We find that cold temperatures, hot temperatures, precipitation, narrower daily temperature ranges, humidity, and cloud cover are all associated with worsened expressions of sentiment, even when excluding weather-related posts. We compare the magnitude of our estimates with the effect sizes associated with notable historical events occurring within our data.
Introduction
Positive emotions are associated with physical, psychological, cognitive, social, and economic benefits, and emotional states can spread through social networks.
- Positive emotions are associated with improved physiological functioning, cognitive performance, mental flexibility, social connectedness, and economic success.
- Emotional states can be transmitted through social networks, potentially amplifying their broader effects.
- The passage links mood and emotional state to human physical, psychological, and economic well-being.
9 Department of Mathematics and GISC, Universidad Carlos III de Madrid
The listed affiliation identifies Data61 and CSIRO in Australia.
- Data61 is identified as part of the Commonwealth Scientific and Industrial Research Organisation in Australia.
- The affiliation is associated with Australia.
- The passage contains affiliation information rather than research findings or methods.
11 Departments of Political Science and Medicine, UC San Diego
The paper addresses conflicting evidence on weather and emotional state by analyzing expressed sentiment in large-scale social-media data. It asks whether weather associations exist, persist after excluding weather discussion, and are comparable in magnitude to effects of other events.
- Previous studies reported conflicting weather–mood associations, with limitations including small samples, negligible effects, individual variation, and aggregation dependence.
- The paper analyzes expressed sentiment among tens of millions of United States residents across 3.5 billion Facebook and Twitter posts from 2009 to 2016.
- The study examines whether meteorological conditions associate with changes in the sentiment of human expressions.
- The research tests robustness to excluding weather-related discussion and compares weather-association magnitudes with effects of other sentiment-altering events.
Social media data
The study combines Facebook and Twitter data to analyze expressed sentiment across billions of posts. Facebook offers greater representativeness and longer-form expressions, while Twitter supports additional analysis of mechanisms and precise geographic assignment.
- The dataset contains 3.5 billion posts: 2.4 billion from Facebook and 1.1 billion from Twitter.
- Facebook data are more likely to be representative and contain text expressions revealing users’ underlying emotional states.
- Facebook status updates cover January 1, 2009 through March 31, 2012, totaling 1,176 days.
- Twitter data comprise publicly viewable messages limited to 140 characters, collected from November 30, 2013 through June 30, 2016.
- Geo-located tweets were collected through the public Streaming API and assigned to metropolitan areas using a United States bounding box and metropolitan boundaries.
Meteorological data
The study combines meteorological measures from PRISM and NCEP with LIWC-based sentiment measures aggregated from Facebook and Twitter posts across 75 U.S. metropolitan areas.
- Daily maximum temperature, temperature range, and precipitation come from PRISM, while cloud cover and relative humidity come from NCEP Reanalysis II.
- LIWC classifies whether posts contain positive or negative sentiment terms, with positive and negative sentiment treated as separate constructs.The results are similar when alternative sentiment classifiers are used.
- The analysis ends on June 30, 2016 because daily NCEP data available at writing ended on that date.
City-level
The city-level analysis models positive and negative expressed sentiment as functions of meteorological conditions while controlling for geographic, temporal, and city-specific year-month factors.
- The model estimates city-level sentiment associations with maximum temperature, precipitation, temperature range, cloud cover, and relative humidity.Flexible indicator bins are used for each meteorological measure.
- Temperature and precipitation effects are estimated with flexible bins, including 5-degree temperature bins, 1-cm precipitation bins, and 20-percentage-point bins for cloud cover and humidity.
- The model includes geographic and temporal indicators plus city-specific year-month indicators to reduce bias from correlated unobserved factors.These controls address factors such as infrastructure, leisure time, daylight, and evolving city-level economic conditions.
- Regression errors are clustered by city-year-month and day, non-climatic controls are excluded, and city-days are weighted by underlying post counts.
- Estimates are reported relative to omitted baseline categories, including 20–25°C maximum temperature, 0–5°C temperature range, and 0 cm precipitation.
Exclusion of weather terms
The analysis reruns the city-level models after filtering Twitter messages containing plausible weather references, separating weather-related discussion from broader expressed sentiment.
- Twitter posts containing plausible weather references are removed before rerunning the models to test whether results extend beyond weather discussion.The initial analyses include all expressions, including terms that may directly refer to weather.
Effect sizes in context
Weather-related conditions were associated with meaningful changes in expressed sentiment, including in posts that excluded weather terms. A below-freezing day had an effect size comparable to notable local events, while interpretation is bounded by measurement and population limitations.
- All expressed sentiment: Facebook analyses associated temperature, precipitation, humidity, and cloud cover with sentiment, including 0.104 higher positive sentiment for daily temperature ranges exceeding 15 ℃.High humidity and cloud cover were associated with lower positive and higher negative expressions.
- All expressed sentiment: Twitter reproduced the Facebook pattern with statistically significant effects, but the association between below-freezing temperatures and positive sentiment was approximately 45% as large.The Twitter effects were generally attenuated relative to Facebook.
- Expressed sentiment of non-weather messages: Weather associations persisted after excluding weather terms, although temperature and precipitation effects were slightly smaller than in the all-posts model.Temperature range, high humidity, and cloud-cover associations also remained significant but were attenuated.
- Effect sizes in context: 62%: the effect of a below-freezing day on positive expressed sentiment versus the 2015 Carolina floods in Charlotte.The comparison uses aggregate expressed sentiment and does not assess the events’ overall effects on society or individual well-being.
- Discussion: Interpretation is limited because social-media sentiment is an imperfect proxy for emotional states and the sample excludes people who do not use Facebook or Twitter.The study also covers one country with a predominantly temperate climate and high air-conditioning prevalence.
Competing Financial Interests Statement
The authors report no conflicts of interest and describe the data sources, funding, and redistribution restrictions.
- The authors declare no conflicts of interest regarding authorship or publication.
- The study used public Twitter data and previously published aggregated Facebook data, which cannot be publicly redistributed.
- Funding came from the National Science Foundation and Spain’s Ministerio de Economia y Competitividad.
Supplementary Information
Supplementary analyses test within-user associations, robustness to excluding weather terms, comparisons with other events, weather-speech patterns, and alternative sentiment classifiers.
- User-level analysis: 81,388,085 user-days show temperature and precipitation effects broadly resembling city-level results, but with attenuated magnitudes; precipitation, temperature range, and cloud cover remain significant.High humidity is not significant in the user-level model.
- No weather terms: Excluding weather terms reduces effect sizes and removes significance for cold temperatures, heavy precipitation, and humidity, while high temperature ranges remain associated with improved sentiment and cloud cover with worsened sentiment.Moderate precipitation remains associated with increased negative sentiment in this model.
- Effect sizes in context: 37%: a below-freezing day’s user-level effect size relative to Carolina flooding in Charlotte, indicating a smaller but still meaningful weather association.The comparison is made for user-level expressed sentiment.
- Rate of messages containing weather terms: Less pleasant meteorological conditions increase both city-level weather-speech rates and individual-level probabilities of weather speech.Weather-related expressions comprise a larger share of messages under these conditions.
- Correlation of classifiers of expressed sentiment: The positive sentiment metrics from LIWC, SentiStrength, and Hedonometer are positively correlated, although Hedonometer correlates less with the other two than LIWC and SentiStrength correlate with each other.The supplementary analyses also replicate results using SentiStrength and Hedonometer classifiers.