Temperature impacts on hate speech online: evidence from 4 billion geolocated tweets from the USA

If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password

If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password

Temperature impacts on hate speech online: evidence from 4 billion geolocated tweets from the USA

A link between weather and aggression in the offline world has been established across a variety of societal settings. Simultaneously, the rapid digitalisation of nearly every aspect of everyday life has led to a high frequency of interpersonal conflicts online. Hate speech online has become a prevalent problem that has been shown to aggravate mental health conditions, especially among young people and marginalised groups. We examine the effect of temperature on the occurrence of hate speech on the social media platform Twitter and interpret the results in the context of the interlinkage between climate change, human behaviour, and mental health.
In this quantitative empirical study, we used a supervised machine learning approach to identify hate speech in a dataset containing around 4 billion geolocated tweets from 773 cities across the USA between May 1, 2014 and May 1, 2020. We statistically evaluated the changes in daily hate tweets against changes in local temperature, isolating the temperature influence from confounding factors using binned panel-regression models.
The prevalence of hate tweets was lowest at moderate temperatures (12 to 21°C) and marked increases in the number of hate tweets were observed at hotter and colder temperatures, reaching up to 12·5% (95% CI 8·0–16·5) for cold temperature extremes (–6 to –3°C) and up to 22·0% (95% CI 20·5–23·5) for hot temperature extremes (42 to 45°C). Outside of the moderate temperature range, the hate tweets also increased as a proportion of total tweeting activity. The quasi-quadratic shape of the temperature–hate tweet curve was robust across varying climate zones, income quartiles, religious and political beliefs, and both city-level and state-level aggregations. However, temperature ranges with the lowest prevalence of hate tweets were centred around the local temperature mean and the magnitude of the increases in hate tweets for hot and cold temperatures varied across the climate zones.
Our results highlight hate speech online as a potential channel through which temperature alters interpersonal conflict and societal aggression. We provide empirical evidence that hot and cold temperatures can aggravate aggressive tendencies online. The prevalence of the results across climatic and socioeconomic subgroups points to limitations in the ability of humans to adapt to temperature extremes.
Multidisciplinary studies have found an effect of temperature on aggression with empirical analyses suggesting that deviations from mild temperatures increase aggressive tendencies and conflict risk worldwide. However, conflict is nowadays not limited to the physical space alone, but is also prevalent online in the form of hate speech. People affected by hate speech have been shown to be more likely to have mental health problems or to experience an aggravation of pre-existing conditions. We analysed the effect of temperature on the occurrence of hate speech on the social media platform Twitter in the USA between 2014 and 2020.
To our knowledge, this is the first empirical study assessing the impact of temperature on online hate speech in the USA. The use of datasets from Twitter enabled the analysis of unprompted aggressions since Twitter users express their opinions online without external encouragement. Furthermore, users can tweet from any location, reducing the barrier to expressing aggression in response to temporal discomfort. In a sample of around 4 billion geolocated tweets, more than 75 million hate tweets were identified using a supervised machine learning classifier. The statistical analysis revealed a quasi-quadratic dependence of hate speech on temperature with low prevalence of hate speech observed in moderate temperatures and sharp increases in hate speech in warmer and colder temperatures. This quasi-quadratic relationship was preserved in separate analyses of temperature and hate speech in different climate zones and in the context of socioeconomic differences (income, religious adherence, and electoral outcomes). The lowest prevalence of hate speech was observed at temperatures centred around the local mean temperature and the magnitude of the increases in hate speech in hot and cold temperatures varied across climate zones.
The quasi-quadratic relationship identified in this study shows that extreme temperatures lead to more aggression online. In contrast to the majority of quantitative studies assessing physical violence, this conclusion was the same for hot temperatures and cold temperatures. Daily maximum temperatures of more than 30°C were consistently associated with substantial increases in hate speech across all climate zones and across all socioeconomic subgroups. This persistent association suggests limits to the capacity for temperature adaptation since the increases in aggression persisted even in regions where hot temperatures are common and across socioeconomic groups that have economic means to mitigate uncomfortable temperatures. Overall, the results presented in this study highlight the importance of climate change mitigation and adaptation against temperature extremes and the need to effectively combat hate speech online.
The raw data consisted of more than 4 billion tweets, spanning a timeframe from May 1, 2014 to May 1, 2020. Tweets were sampled from the 1% Twitter stream using a bounding box around the USA to extract geolocated tweets in the country. Around 1–2% of all tweets are geolocated; thus, although the stream contained all of the geolocated tweets in this period, the data only represent a small proportion of total tweet volume. For simplicity, we refer to this dataset as US1420 hereafter.
The elements of statistical learning: data mining, inference, and prediction, 2nd edn.
The US1420 dataset encompasses 6 years (May, 2014 to the end of April, 2020), two of which were leap years (2016, 2020). Considering that daily data were available for 773 cities, this yielded a total of 1 694 416 possible observations. We counted 1 694 416 observations in the US hate speech Twitter data; therefore our data were complete and we have no concerns about data sparsity in our analysis.
What do we learn from the weather? The new climate-economy literature. Does the environment still matter? Daily Temperature and income in the United States.
The funder had no role in study design, data collection, data analysis, data interpretation, or writing of the report.
A strong non-linear relationship was identified between daily maximum temperature and the percentage change in hate tweets (). Fewest hate tweets occurred between temperatures of 15°C and 18°C. The number of hate tweets remained comparably low for the directly adjacent temperature bins, but sharply increased for temperatures warmer than 27°C and colder than 6°C. On cold days with maximum temperatures between –6°C and –3°C, the number of hate tweets was approximately 12·5% (95% CI 8·0–16·5) higher than on days in the 15 to 18°C temperature range, and on hot days (42 to 45°C), the number of hate tweets was more than 22·0% (20·5–23·5) higher than days in the 15 to 18°C temperature range. On average, a city's temperature varied across 12·6 bins per year. The changes for all bins and the effects of the control variables are shown in the . For temperatures higher than 27°C and lower than 9°C, all respective regression coefficients were statistically significant at the highest level (p≤0·001). The only bin that was not significant at any significance level (p>0·1) was the 18 to 21°C bin, suggesting that the variation was below our detection level of uncertainty.
In addition to the number of hate tweets, we also used the follower-weighted number of hate tweets as the dependent variable to approximate the daily reach of hate speech. We weighed each hate tweet by the number of followers of its author. The shape of the curve was preserved, but the heat responses increased by up to 26·5% (95% CI 23·0–30·0) at high temperatures (42–45°C; ).
Results of additional analyses using varying bin widths (1°C or 5°C) were consistent with the main findings ().
To further examine the temperature–hate tweet relationship against the potential influence of weather on the overall tweet volume, we considered hate tweets as a proportion of all geolocated tweets in the sample. If the overall tweet volume is temperature-dependent such that warmer and colder temperatures result in a general increase in the number of tweets, it is possible that the proportion of hate tweets is almost constant. We tested for this by computing the proportion of geolocated hate tweets at the city level as a proportion of all US geolocated tweets in the respective city for each day. The binned panel-regression model used throughout the analysis was then applied with the daily hate tweet proportion as the dependent variable and daily maximum temperature and other weather controls as the independent variables; the resulting temperature–hate tweet response function () had the same non-linear shape as observed in . An increase in tweets of around 0·13 percentage points (95% CI 0·07–0·19) was observed for cold temperatures and an increase of around 0·18 percentage points (0·12–0·25) for warm temperatures. The mean proportion of hate tweets on all days that fell within the omitted bin (15–18°C) amounted to around 1·5%. Thus, the percentage point increases observed translated to approximately 8·6% more hate tweets in cold temperatures and around 12·1% more hate tweets in warm temperatures. This result is evidence that not only the volume of hate speech on Twitter increases in more extreme temperatures, but also that the proportion of hate in all tweets rises. Analysis of state-level aggregation confirmed these results (). The remaining analyses were conducted for both the number of hate tweets (figures 3, 4) and the proportion of hate tweets () as the dependent variable.
Our dataset comprised cities with diverse climatic conditions. These differences in mean temperature and temperature variability mean that the panel analysis was to a larger extent informed by cities with higher temperature variability. The local temperature extremes might differ from the temperature extremes in the panel analysis. To investigate the potential impact of these local differences in climate on the overall temperature–hate speech relationship, we did separate analyses for five distinct climate zones (). The general quasi-quadratic shape of the US-wide response curve () was preserved across all climate zones for which there was sufficient data coverage (A). The temperature-dependent minimum number of hate tweets and the strength of the increase in hate tweets for warm and cold temperatures differed in accordance with the individual temperature distribution of the climate zone. In the cold climate zone, which spans most of the north of the contiguous USA, a broad range of temperatures were observed annually. Accordingly, the percentage change in hate tweets was low between 6°C and 24°C. The maximum increase in hate tweets of 17·5% (95% CI 5·5 to 29·0) compared with the omitted bin (15 to 18°C) was observed between temperatures of 39°C and 42°C, which only occur rarely in this climate zone. This is likely to explain the large confidence interval. Temperatures between 24°C and 33°C were more common in this climate zone. For the 30°C to 33°C temperature bin, hate tweets increased by around 7·0% (4·5 to 9·5). At cold temperatures (–6 to –3°C), hate tweets increased by more than 12·0% (7·5 to 17·0). A similar pattern was observed in the hot-dry and mixed-humid zones across a smaller temperature range. For the mixed-humid climate zone, the maximum increase amounted to more than 11·0% (8·0 to 14·0) on cold days (0 to 3°C) and around 9·0% (6·5 to 11·0) on warm days (33 to 36°C) relative to the omitted bin (21 to 24°C). 3°C temperature bins outside this range contained less than 1·5% of days. For the hot-dry climate zone, the increase in hate tweets in hot temperatures (42°C to 45°C) was most pronounced (almost a 24% increase [22·5 to 25·5]). In cold temperatures (0 to 3°C), the number of hate tweets increased by 10·0% (–0·5 to 20·0) relative to the omitted bin (18 to 21°C); however, temperatures lower than 9°C were fairly uncommon in this zone and tweets only increased by 8·0% (7·0 to 8·5) in the 9 to 12°C bin. For the marine climate zone, the increase in tweets in response to hot temperatures was stronger than the increase in response to cold temperatures, reaching up to 11·0% (7·5 to 15·0) at temperatures higher than 33°C. For the hot-humid climate zone, the curve had a V-shape with sharp increases on both sides of the omitted bin (24 to 27°C), with a 10·5% increase (4·5 to 16·5) in hate tweets in cold temperatures (6 to 9°C) and a 15·0% (13·0 to 17·0) increase in hot temperatures (36 to 39°C).
For all climate zones, we observed that the lowest incidence of hate tweets (omitted bin) coincided with the mean temperature across the time period (A). This could suggest that the hate tweet increases are dependent on temperatures we are used to. However, independent of the local mean temperature and distributions, hot temperatures of more than 30°C led to significant increases in hate tweets of at least 7%, pointing to potential limits in adaptation to hot temperatures.
When using the proportion of hate tweets as the dependent variable, the shape of the curves were largely preserved ().
Putting faith in hate: when religion is the source or target of hate speech.
Across all socioeconomic subgroups considered, the relationship between daily maximum temperature and hate tweets had a distinct and robust non-linear shape, which is in its general form independent of economic, religious, and political differences. Independent of income quartiles, increases in the number of hate tweets at cold and warm temperature extremes fell generally between 10% and 15%. The only exception was that the medium–low income group experienced higher maximum temperatures than the other groups and and increases in hate tweets of up to 22·0% (95% CI 20·5–23·5) were observed in this temperature range (42 to 45°C; ). For all income quartiles, the mean temperature coincided approximately with the minimum bin. The non-linear shape of the temperature–hate speech relationship was also preserved when data were categorised into cities with predominantly Catholic or Evangelical religious beliefs () with similar responses observed for extreme temperatures. The observed relationship also holds independently of the 2016 election results (). The increase in hate tweets in colder temperatures was slightly more pronounced in cities belonging to counties that had a Democratic majority in the 2016 election, reaching up to 13·5% (7·0–20·5) at temperatures between –6°C and –3°C and only around 10·0% (7·0–13·0) in cities belonging to counties that had a Republican majority in the 2016 election. By contrast, the maximum increase in hate tweets in hot temperatures amounted to around 16·0% (12·5–19·5) in cities with a Democratic majority (at temperatures of 36 to 39°C) and 20·5% (18·5–22·5) in cities with a Republican majority (at temperatures of 42 to 45°C). Although the mean temperatures were within 1°C of each other, Republican majority cities had more extremely hot days, which is likely to explain the differing heat response. The analysis using the proportion of hate tweets as the dependent variable is included in the . Overall, the results were preserved for the hate-share analysis.
In addition to the analysis of socioeconomic differences, we analysed the temperature–hate tweet response by administrative unit of US census divisions, which are frequently used for data collection and analysis (). All temperature–hate tweet response curves at the census division level had the non-linear shape observed across all analyses with increases for cold days falling between 5% and 30% and increases for hot days reaching between 6% and 30%, depending on the census division.
In this study, we found that hate speech increased in absolute volume, and also as a proportion of total tweeting activity, at temperature extremes. The quasi-quadratic shape of the temperature–hate tweet curve was robust across varying climate zones, income quartiles, and religious and political beliefs.
Daily maximum temperature might not always match the temperature experienced by the Twitter user due to residential heating or cooling. This potential measurement error in the independent variables is, however, more likely to attenuate than to increase the magnitude of our estimates.
Overall, the results presented in this study highlight not only the importance of climate change mitigation and adaptation against temperature extremes, but also the need to effectively combat hate speech online and to provide resources for people who are affected. Further work is needed to understand the nature of online abuse, to analyse the most prevalent types of hate, which topics it relates to, who is targeted, and who authors it.
AS and LW designed the study. AS processed the climate and Twitter data. All authors contributed to the interpretation and presentation of the results. AS wrote the manuscript with contributions from LW. AS and LW revised the manuscript.
This study was funded by the Volkswagen Foundation. We thank Kelsey Barton-Henry for fruitful discussion.
Creative Commons Attribution – NonCommercial – NoDerivs (CC BY-NC-ND 4.0)