Results
To investigate these questions, we analyzed a large (
n = 563,312) corpus of tweets from Twitter. We selected three politically polarizing topics: gun control (study 1), same-sex marriage (study 2), and climate change (study 3; see
SI Appendix, section 1, for more details). These topics are highly contentious in American politics and have been at the forefront of major public policy debates (
24). Because language is one direct way in which people communicate emotion, we coded the language in Twitter messages to quantify morality and emotion. Specifically, we used (and pilot tested) previously validated dictionaries (
25,
26) to count the frequency of moral and emotional words in each tweet. [Moral words are those appearing only in the moral dictionary, emotional words appear only in the emotional dictionary, and moral-emotional words (e.g., hate) are those that appear in both dictionaries (for more details, see
Methods, as well as
SI Appendix, section 1).] “Contagion” was indexed as the number of times each message was retweeted by a user for each moral/political topic (see
SI Appendix, section 1, for more details). A retweet occurs when one user shares another user’s message with his or her own social network, and represents a key form of information diffusion on Twitter (
27).
In study 1, we investigated whether moral and emotional language contained in messages predicted contagion on the topic of gun control (
n = 102,328). We measured the distinctly moral language, distinctly emotional language, and moral-emotional language for each message and fit a regression model predicting retweet rate (30% of messages were retweeted at least once). The analysis yielded no main effect of distinctly moral language [incident rate ratio (IRR) = 0.98,
P = 0.086, 95% CI = 0.95, 1.00], nor did it yield a main effect of distinctly emotional language (IRR = 1.00,
P = 0.896, 95% CI = 0.97, 1.03). Importantly, there was a significant main effect of moral-emotional language (IRR = 1.19,
P < 0.001, 95% CI = 1.14, 1.23); adding a single moral-emotional word to a given tweet increased its expected retweet rate by 19% (
Fig. 1). [The main effect of moral-emotional words remained significant after distinctly moral and distinctly emotional words were removed from the model (
SI Appendix, Tables S5–S7).]
Messages with the greatest amount of moral-emotional language had the highest expected retweet rate in the sample, even after adjusting for the effects of distinctly moral and distinctly emotional words. These results suggest that emotion is a key component for the diffusion of moral content through social networks but that the social transmission of morality is distinct from basic emotional contagion.
In study 2, we replicated these results in the domain of same-sex marriage (
n = 47,373). Again, we measured distinctly moral language, distinctly emotional language, and moral-emotional language and fit a regression model predicting retweet rate (23% of messages were retweeted). We observed no main effect of distinctly moral language (IRR = 0.99,
P = 0.540, 95% CI = 0.95, 1.03), but we did observe a main effect of distinctly emotional language (IRR = 1.15,
P < 0.001, 95% CI = 1.11, 1.20), demonstrating basic emotional contagion (
28). Adjusting for these effects, there was again a significant effect of moral-emotional language (IRR = 1.17,
P < 0.001, 95% CI = 1.09, 1.26). Adding a single moral-emotional word to a given tweet increased its expected retweet rate by 17%. Tweets with the greatest amount of moral-emotional language had the highest expected retweet rates (
Fig. 1).
In study 3, we obtained parallel results with respect to communications about climate change (
n = 413,611). We used the same methods as in the previous studies (29% of messages were retweeted). This time, we observed a main effect of distinctly moral language (IRR = 1.04,
P < 0.001, 95% CI = 1.02, 1.06), indicating a moral contagion effect in the absence of emotional language, as well as a significant main effect of distinctly emotional language (IRR = 1.08,
P < 0.001, 95% CI = 1.07, 1.09), demonstrating basic emotional contagion. As in studies 1 and 2, we also observed a significant main effect of moral-emotional language (IRR = 1.24,
P < 0.001, 95% CI = 1.22, 1.27); adding a single moral-emotional word to a given tweet increased its expected retweet rate by 24% (
Fig. 1).
Across three contentious political topics, moral-emotional language produced substantial moral contagion effects (mean IRR = 1.20, or a 20% increase in retweet rate per moral-emotional word added), even after adjusting for the effects of distinctly moral and distinctly emotional language, as well as other covariates known to affect retweet rate. [We note that there were interactions such that the presence of media lead to a relative increase in moral contagion effect (climate change), and the presence of a URL led to a relative decrease in moral contagion (climate change, same-sex marriage).] These results shed light on the types of linguistic content that can amplify messages in social networks (for a list of specific emotional and moral-emotional words that were most impactful across topics, as well as sample tweets for each word, see
SI Appendix, Tables S3 and S4).
Next, we examined whether contagion was driven by a general negativity bias or applied to positive valence as well. To measure emotional valence, we split our emotion and moral-emotion dictionaries into “positive” and “negative” emotions (
25). In the case of messages related to gun control, we observed that negative moral-emotional language (IRR = 1.19,
P < 0.001, 95% CI = 1.13, 1.25) and positive moral-emotional language (IRR = 1.09,
P = 0.053, 95% CI = 1.00, 1.18) both contributed to contagion effects. In the case of same-sex marriage, positive moral-emotional language predicted contagion (IRR = 1.92,
P < 0.001, 95% CI = 1.68, 2.20), whereas negative moral-emotional language was a negative predictor of contagion (IRR = 0.87,
P = 0.008, 95% CI = 0.79, 0.96). It is worth noting, however, that at the time of our data collection, attitudes expressed online about same-sex marriage were predominantly positive (such as those communicated with the hashtag “#lovewins”). [The phrase #lovewins led to ∼6% error in data collection for our same-sex marriage dataset due to seemingly arbitrary use of the hashtag. Removal of all tweets with #lovewins does not change results (
SI Appendix, section 1 and Table S19).] People were less likely to retweet messages about same-sex marriage that contained negative moral-emotional language (e.g., “hate”). In the case of climate change, negative moral-emotional language predicted moral contagion (IRR = 1.31,
P < 0.001, 95% CI = 1.28, 1.35), whereas positive moral-emotional language did not (IRR = 1.03,
P = 0.178, 95% CI = 0.99, 1.07). In contrast to the pattern found for same-sex marriage, when discussing climate change people were more likely to retweet negatively valenced messages, such as those referring to environmental harms caused by climate change. Thus, overall, the effects of valence on moral-emotional contagion were specific to the topic in question. See
SI Appendix, Table S11, for further model details and coefficients.
In an exploratory analysis, we also considered specific discrete emotions and their effects on social transmission. We focused on the emotions of anger and disgust because of their prominent role in communication, association with moral judgment, and distinctive relationship to moral outcomes (
29–
31). We also included sadness, a low-arousal emotion, to compare its impact to the high-arousal emotions of anger and disgust. The only consistent finding across all moral topics was that the low-arousal emotion of sadness was associated with a decrease in social transmission (mean IRR = 0.73). This pattern replicated previous work investigating the impact of discrete emotions on social transmission of online news articles (
18). The effect of anger was context-specific; it was associated with increased social transmission for the topic of climate change, which was dominated by negative emotion, but was associated with decreased social transmission for the topic of same-sex marriage, which was dominated by positive emotion. We observed no significant effects for disgust (
SI Appendix, section 3).
We also investigated the extent to which messages containing moral-emotional language transcended ideological group boundaries, as opposed to spreading largely within those boundaries. Specifically, we compared diffusion rates in retweet networks that either did or did not share the ideological orientation of the original author. We started by estimating each user’s political ideology as a continuous value using a previously validated algorithm based on follower networks (
32). For each message, we computed a retweet count based on retweeters who possessed the same ideological classification as the original author (in-group members) and a separate count based on retweeters who possessed a different ideological classification than the original author (out-group members). [One feature of this approach is that we used an ideology score of 0 as a cutoff between liberal and conservative authors and therefore as a basis for determining in-group vs. out-group rates of diffusion. This method is imperfect when it comes to analyzing tweets sent by political moderates, whose ideological estimates are close to zero. For instance, the in-group network for an author with an ideology estimate of 0.01 will be classified as conservative, whereas the in-group network for an author with an ideology estimate of −0.01 will be classified as liberal, despite the fact that these authors are extremely close to one another with respect to ideology. To address this limitation, we conducted three robustness tests by (
i) excluding all “verified” users (e.g., celebrities), to eliminate the possibility that a few well-known moderates could disproportionately sway the results; (
ii) excluding the middle 10% (in terms of ideological estimates, closest to zero) of authors in our dataset; and (
iii) excluding the middle 20% (in terms of ideological estimates, closest to zero) of authors. All three of these analyses yielded results that were highly similar to those reported in the main text, increasing our confidence that the methodological concerns discussed above did not substantially influence the findings reported here.] We then estimated a multilevel model to test whether moral-emotional contagion was stronger within the in-group retweet network than the out-group retweet network, to assess the tendency for moral-emotional messages to diffuse more widely within ideological boundaries than between them. In this model, we interacted the moral-emotional language count variable with an effects-coded classification variable indicating the in-group (as opposed to out-group). To the extent that moral-emotional language plays a larger role spreading information within in-group networks as opposed to out-group networks, we would expect to find a positive interaction coefficient (see
SI Appendix, section 2, for details).
With respect to messages about gun control, moral-emotional language did have a larger impact on retweet rates within in-group networks than out-group networks. The interaction was statistically significant (IRR = 1.20,
P = 0.049, 95% CI = 1.00, 1.45), with an estimated 20% higher diffusion rate of moral-emotional language for in-group (vs. out-group) networks (
Fig. 2). Very similar results were obtained with respect to messages about climate change (IRR = 1.34,
P = 0.001, 95% CI = 1.12, 1.60). For same-sex marriage, however, the interaction did not approach statistical significance, although the effect remained in the same direction (IRR = 1.10,
P = 0.746, 95% CI = 0.61, 1.98). These findings indicate there may be an in-group advantage (
22,
33) for moral contagion; that is, moral-emotional language may spread more widely within in-group networks than out-group networks (for a visualization of the retweet network for messages containing moral and emotional language, see
Fig. 3). The in-group advantage was also observed more consistently for moral-emotional language than nonmoral-emotional language (
SI Appendix, Table S12). To the extent that moral contagion is greater for in-group (vs. out-group) networks, it may help to explain why online discussions of moral and political topics often occur within polarized “echo chambers.”
Given past research suggesting that political conservatives may possess more homophilous online social networks than liberals (
32,
34), we also explored whether the in-group advantage for moral-emotional language would be greater in conservative (vs. liberal) social networks. Thus, we estimated a model that included a three-way interaction term involving the original author’s ideological classification, the moral-emotional language variable, and the binary in-group/out-group classification variable. Moral-emotional language increased retweets within conservative in-groups significantly more than liberal in-groups for the issue of climate change (IRR = 1.78,
P < 0.001, 95% CI = 1.35, 2.34). The three-way interaction was in the same direction for gun control and same-sex marriage, but it did not reach statistical significance for those two issues. See
SI Appendix, section 1, for more details.
Discussion
Using naturally occurring social networks on Twitter, we identified a critical role for emotion when it comes to the diffusion of moral ideas in real, online social networks. Using a large sample of tweets concerning three polarizing issues (
n = 563,312), the presence of moral-emotional words in messages increased their transmission by approximately 20% per word. The effect of moral-emotional language was observed over and above distinctly moral and distinctly emotional language as well as other factors that are known to increase online diffusion of messages. This work is consistent with accounts of moral psychology that highlight the social and emotional nature of moral discourse. It also extends current theories by identifying a social transmission process of information diffusion. In doing so, this work fosters questions pertaining to the role of social influence in the domain of morality such as how online messages can affect moral attitudes. These issues are more important than ever, given the growing use of social media for political purposes (
35).
The results of our studies also clarify how moral emotions contribute to moral contagion. One key finding was that morally framed emotional expressions explained statistical variance in diffusion of moral and political ideas over and above nonmoral-emotional expression. This finding highlights the need for an analysis of how specific emotions are functionally linked to moral outcomes, especially when it comes to the transmission of moral ideas. We have distinguished broadly between moral and nonmoral emotions, but it is likely that moral emotions can be broken down into more fine-grained subcategories [such as moral emotions that are self-conscious vs. other-condemning (
12)] that may have distinct effects on social transmission. We also observed that nonmoral emotions had a unique impact on social transmission for two out of three topics, replicating prior work demonstrating the impact of emotion on online diffusion (
18,
19). Future work should clarify how the class of moral emotions motivates individuals to share and discuss their ideas and the conditions under which moral emotions yield greater power than nonmoral emotions.
Another key finding was that the expression of moral emotion aids diffusion within political in-group networks more than out-group networks. With respect to politics, this result highlights one process that may partly explain increasing polarization between liberals and conservatives (
24). To the extent that the spread of online messages infused with moral-emotional contents is circumscribed by group boundaries, communications about morality are more likely to resemble echo chambers and may exacerbate ideological polarization. Our results also speak to recent controversies over the role of social media in creating a biased informational environment (
36). For example, the use of negative messages about rival political candidates containing strongly worded moral-emotional terms may spread more easily within (but not necessarily between) liberal or conservative social networks.
Finally, our approach illustrates the utility of bringing a social network approach to bear on questions of moral psychology. In comparison with laboratory-based studies, the social network approach offers much greater ecological validity. We were able to investigate moral discourse about contentious political topics with significant policy ramifications and to track users in complex, naturally formed social environments rather than isolated, artificial settings. Furthermore, we investigated the diffusion of ideas in digital online environments, which are becoming increasingly prominent when it comes to promoting moral and political discourse. As of 2017, Twitter is estimated to have 317 million active monthly users, and Facebook is estimated to have 1.87 billion (
37). Data collection from social media platforms can pose challenges when it comes to precisely measuring psychological constructs of interest, and the analysis of such data can be computationally intensive. For example, the effect size estimate for the issue of same-sex marriage, although robust in direction, was the most variable in size across all sensitivity analyses. This may have been due to small errors in data collection (
SI Appendix, section 1), or to the statistical clustering we described in
SI Appendix, section 2. Despite these challenges, we believe that the benefits of studying moral and political discourse in real time in naturally occurring social networks outweigh potential limitations. Future work should seek to corroborate our conclusions with more carefully controlled laboratory experiments. In particular, it is important to test the causal influence of exposure to moral-emotional language on attitudes and behavior.
Another contribution of the social network approach is that it generates a number of exciting questions for interpersonal accounts of moral judgment and behavior. Although our research program was focused on the contents of social media messages, adopting a social network approach also raises important questions about source effects and social network structure. For example, why do some communicators have more influence than others when it comes to the diffusion of moral ideas? Do more densely connected social networks spread moral ideas faster than networks with fewer ties between individuals? Our findings thus far have addressed only a sliver of important questions that can be addressed by applying a social network approach to the study of morality.
Methods
All research was conducted in accordance with the New York University (NYU) University Committee on Activities Involving Human Subjects (IRB no. 12-9058). Data collection was ruled “exempt” due to our use of public tweets only. A public tweet is a message that the user consents to be publicly available rather than only to a collection of approved followers.
To determine the presence of morality and emotion in language, we split the total corpus of words from the two dictionaries into three categories: distinctly moral (e.g., “duty”;
n = 329), distinctly emotional (e.g., “fear”;
n = 819), and moral-emotional (e.g., “hate”;
n = 159;
SI Appendix, section 1). Distinctly moral and emotional words were those that appeared only in one of the two dictionaries, whereas moral-emotional words appeared in both moral and emotion dictionaries (see
Table 1 for examples).
Pilot participants confirmed the discriminant validity of our word categories by rating each word on continuous dimensions of morality and emotion. One group of participants (n = 17) rated a random 10% subset (n = 40) of distinctly moral words from our dictionary as more “moral” than the distinctly emotional words (n = 42) [t(16) = 9.19, P < 0.001, Cohen’s d = 2.23], and they rated a random subset of moral-emotional words (n = 9) as more moral than distinctly emotional words [t(16) = 9.88, P < 0.001, d = 2.40]. A second group of pilot participants (n = 19) rated the subset of distinctly emotion words as more “emotional” than the distinctly moral words [t(18) = 3.19, P = 0.005, d = 0.73], and they rated moral-emotional words as more emotional than distinctly moral words [t(18) = 8.95, P < 0.001, d = 2.05].
As a test of robustness, we also investigated discriminant validity by asking a larger group of pilot participants (
n = 50) to make discrete categorizations of random sets of words from each category. When they were presented with unlabeled random sets of words from each (moral, emotional, moral-emotional) category and asked which word set best expressed a combination of morality and emotion, 76% of participants choose the moral-emotional set, which made that category significantly more likely to be chosen than the other category sets [χ
2(2) = 41.44,
P < 0.001]. For more details, see
SI Appendix, section 1.
To form our main predictor variables, we computed the frequency of distinctly moral, distinctly emotional, and moral-emotional words present in each tweet to determine how these factors predicted contagion as measured by retweet count (
SI Appendix, section 2). We fit a negative-binomial model with maximum likelihood estimation to account for overdispersion (
38). The majority of our data were independent (70% of users had only one message in our dataset), but there were some sources of nonindependence due to the 30% of users with more than one message in the dataset. Accounting for these sources of nonindependence revealed that the results from models that treat all data as independent are robust. For instance, dropping all sources with nonindependence produced nearly identical results to the full dataset as did within-cluster resampling via bootstrapping and the use of multilevel models. For a detailed examination of nonindependence and robustness, see
SI Appendix, section 2. Proc GENMOD in SAS 9.4 was used for all analyses and all syntax is available at
https://osf.io/59uyz/wiki/home/.
For each study, our statistical models included our three main predictor variables, as well as covariates known to affect retweet rate independent of message content (
27), including number of Twitter followers the original message author had, whether media or a URL was attached to the message, and whether the message author was a “verified” Twitter user. All predictors were grand-mean centered, and all binary variables were effects coded. For a complete list of variables entered in the model and their coefficients, see
SI Appendix, Table S5. For specific details of each model and further tests of robustness for each effect, see
SI Appendix, section 2. For all studies, the effect of moral-emotional words was almost exactly the same when covariates were included or omitted from the model (
SI Appendix, Tables S5–S10).
For our models estimating differences in the effect of moral contagion for in-group and out-group networks, we estimated a multilevel model using generalized estimating equations (
39) with an exchangeable correlation structure. See
SI Appendix, section 2, for more details.