After Aylan Kurdi: How Tweeting About Death,Threat,and Harm Predict Increased Expressions of Solidarity With Refugees Over Time

Abstract

Viral social media content has been heralded for its power to transform policy, but online responses are often derided as “slacktivism.” This raises the questions of what drives viral communications and what is their effect on support for social change. We addressed these issues in relation to Twitter discussions about Aylan Kurdi, a child refugee who died en route to the European Union. We developed a longitudinal paradigm to analyze 41,253 tweets posted 1 week before the images of Aylan Kurdi emerged, the week they emerged, and 10 weeks afterward—at the time of the Paris terror attacks. Tweeting about death before the images emerged predicted tweeting about Aylan Kurdi, and this, sustained by discussion of harm and threat, predicted the expression of solidarity with refugees 10 weeks later. Results suggest that processes of normative conflict and communication can be intertwined in promoting support for social change.

Keywords

social interaction threat death and dying morality intergroup dynamics

On September 2, 2015, the lifeless body of Syrian 3-year-old Aylan Kurdi¹ was discovered on a Turkish beach after his family attempted to cross the Mediterranean Sea to flee the Syrian conflict. Within 12 hours, pictures of Aylan Kurdi were retweeted over 30,000 times, reaching at least 20 million people (Vis & Goriunova, 2015). The dissemination of these images has been credited with changing the debate on immigration and galvanizing a global social movement to address the refugee crisis (Vis & Goriunova, 2015). Within weeks, there were widespread announcements of new public policy responses, with pledges from nations around the world to resettle 162,151 refugees (Amnesty International, 2016).

Through a longitudinal analysis of language on Twitter, we explored how communicating about threat and harm in the context of the death of Aylan Kurdi was related to later expressions of solidarity with refugees. We suggest that online communications are important for understanding enduring responses to events such as the refugee crisis, because they act as a springboard for mobilizing change on a global scale. Contrary to the “slacktivism” critique that suggests that low-cost, low-risk online actions foreclose engagement in higher-risk actions by satisfying the motivation to act (Lee & Hsieh, 2013), we show here that communicating online can leave a psychological legacy (Carlisle & Patton, 2013; Schumann & Klein, 2015).

Internet activism is a relatively new phenomenon, and much remains unknown about social media’s potential for mobilization (Harlow & Guo, 2014; Wojcieszak, 2009). Therefore, psychological science needs new methods to study the psychology of online engagement over time. To this end, we designed a paradigm to enable analyses of the longitudinal relationships between psychological responses to the images of Aylan Kurdi in a large volume of tweets. This approach contrasts with most approaches to understanding variations in activity on Twitter that analyze short-term peaks in sentiment or changes in tweet volume across networks or in reaction to events (see Williams & Burnap, 2016). While previous approaches attest to the volume and reach of activity, our approach allows us to make inferences about the form and nature of that activity. We document the psychological processes showing how and why online communications are related to later mobilization.

The Role of Communicating Opinions Online in the Development of Solidarity

Social psychological research has identified mechanisms by which online communications mobilize social movements (Harlow, 2012; Wojcieszak, 2009). McGarty, Thomas, Lala, Smith, and Bliuc (2014) suggested that the growth of mass opposition to government regimes in Tunisia and Egypt in 2010 and 2011 can be explained through the sharing of images of local antiregime protests from camera phones to video-sharing sites and then through satellite television. McGarty and colleagues suggest that this was an example of the process of opinion-based social identity formation, where protesters used media to communicate the opinion that the government was an illegitimate out-group that should be ousted. Similarly, Smith, Gavin, and Sharp (2015) showed that the development of the Occupy Wall Street movement was preceded by Facebook discussions in which supporters of the movement reached agreement on their opinions about how the world should address economic and social grievances (see Castells, 2012, for a detailed account).

These movements reflect what we have previously described as the formation of an identity-norm nexus (INN; Smith, Thomas, & McGarty, 2015), whereby people come to understand their commitment to end some unacceptable state of affairs as an aspect of their identity. We argue that when people perceive a normative conflict—a discrepancy between the way the world is (the descriptive norm) and the way they perceive the world ought to be (the injunctive norm)—they are motivated to communicate the discrepancy to change that state of affairs. Communicating about a normative conflict is a central, catalytic process in efforts to promote social change. Aylan Kurdi’s image encapsulated the normative conflict that people—children—are harmed while attempting to flee a violent conflict.

The issue of unacceptable harm is of central importance in motivating prochange responses. According to the theory of moral foundations (Haidt, 2007), harm is one of five broad bases of what people consider to be right or wrong and focuses attention on concern for others (Graham, Haidt, & Nosek, 2009). Accordingly, we expected that tweeting about Aylan Kurdi, a situation in which a “universal” societal injunctive norm was violated (children should not be harmed; see Haidt, 2012), would mobilize people to communicate about illegitimate harm and, therefore, express solidarity with the victim group. However, during the refugee crisis, there were also concerns raised in some quarters about threats from refugees to host communities. Many nations were concurrently on high alert for terrorist attacks. In November 2015, when the Paris terror attacks occurred, it was speculated that the attacks diminished the compassion for refugees (Lyman & Smale, 2015; Wright, 2015). Therefore, we expected that whereas communicating about illegitimate harm would be positively associated with expressions of solidarity with the victim group, communicating about external threats to oneself or others would be negatively associated with solidarity (Huddy & Feldman, 2011; Stephan & Stephan, 2000).

Thus, we predicted that tweeting about Aylan Kurdi would be positively related to expressing solidarity with refugees 10 weeks later to the extent that tweets focused on harm (Hypothesis 1) rather than threat (Hypothesis 2). To isolate the variance in solidarity with refugees explained by tweeting about harm and threat (over and above that of preexisting sympathy with refugees), we controlled for expression of prorefugee opinions prior to the emergence of the images. It was also the case that the photos of Aylan Kurdi depicted his death, and death-thought accessibility can have substantial effects on behavior (see Burke, Martens, & Faucher, 2010; McGregor et al., 1998; Vail, Arndt, Motyl, & Pyszczynski, 2012). We therefore controlled for prior talk about death in this study. Doing so allowed us to examine the specific relationship between tweeting about the image of Aylan Kurdi’s death and solidarity while controlling for prior prorefugee opinions and death-thought accessibility more generally.

Pilot Study

We used Chorus Tweetcatcher 1.3 software (Brooker, Barnett, & Cribbin, 2016) to retrieve publicly available Twitter data via the Twitter Search application program interface (API). We then used Linguistic Inquiry and Word Count (LIWC2015) software (Pennebaker, Boyd, Jordan, & Blackburn, 2015) to establish the proportion of tweets by each user at each time point that contained words in the dictionary categories. To establish the proportion of words that contained prorefugee content, we created a custom LIWC dictionary. First, in a pilot study, we established that the custom dictionary captured prorefugee content rather than refugee content in general.

Method

We conducted an a priori power analysis that established that the sample size required for an effect size (d) of 0.50 was 54 participants (to detect significance at p < .05 in a one-sample t test; Faul, Erdfelder, Lang, & Buchner, 2007). We recruited 59 participants (85% female) via a post advertising the survey on Twitter and from the Department of Psychology Research Participation Scheme for first-year undergraduates at a UK university. All participants were Twitter users. Eighty-eight percent were UK residents. Via an online survey, they rated each of the prorefugee hashtags and words/phrases in Table 1 on a scale from 1 to 7 (1 = not at all prorefugee, 4 = neutral, and 7 = completely prorefugee). These hashtags and words/phrases were selected using Chorus’ TweetVis visual analytics suite (Brooker et al., 2016), which identified the most frequently used relevant hashtags and words associated with the broad hashtag #refugees.

Table 1.

Validation of Key Terms in Prorefugee Custom Dictionary (N = 59)

Term	M	SD	Loading onto prorefugee talk	t	Mean difference from neutral midpoint of scale (4)	d	95% confidence interval for the difference
Refugeeswelcome	6.40	1.08	.54	16.45	2.40	2.22	[2.11, 2.69]
Welcome	5.04	1.67	.78	4.61	1.04	0.62	[0.59, 1.49]
Supportrefugees	6.21	1.04	.58	15.94	2.21	2.13	[1.94, 2.49]
Supportmigrants	5.73	1.52	.71	8.42	1.73	1.14	[1.32, 2.14]
Pray	5.16	1.56	.57	5.57	1.16	0.74	[0.74, 1.58]
Prayforsyria	5.73	1.21	.59	10.68	1.73	1.43	[1.41, 2.06]
Shameoneurope	5.11	1.69	.61	4.90	1.11	0.66	[0.65, 1.56]
Wehaveroom	6.21	1.07	.53	15.43	2.21	2.07	[1.93, 2.50]
Withsyria	5.84	1.01	.35	13.43	1.84	1.82	[1.56, 2.11]
Wecandobetter	5.48	1.34	.58	8.31	1.48	1.10	[1.12, 1.84]
Soukforsyria	4.67	1.36	.44	3.60	0.67	0.49	[0.30, 1.04]
Solidarity	5.14	1.51	.47	5.68	1.14	0.75	[0.74, 1.55]

Results

We conducted a series of one-sample t tests to compare the mean score for each key term against the neutral point of the scale (4; see Table 1). All of the terms were significantly associated with being prorefugee with the exception of “refugee,” t(53) = 0.87, p = .39, d = 0.19; “refugeescrisis,” t(55) = 1.67, p = .10, d = 0.22; “roeszke” (a refugee camp and border crossing into Hungary from Serbia that was widely discussed in 2015), t(53) = −0.79, p = .44, d = 0.08; “refugeesgr,” t(52) = 1.79, p = .08, d = 0.25; “syrianrefugees,” t(54) = 1.15, p = .26, d = 0.16; “refugeeconvoy,” t(55) = 0.67, p = .51, d = 0.09; “soseurope,” t(55) = 0.16, p = .87, d = 0.02; “notinmyname,” t(54) = 0.96, p = .34, d = 0.13; and “islamophobia,” t(54) = −1.23, p = .22, d = −0.17. An initial principal axis factor analysis with direct oblimin rotation indicated that the top two factors differentially represented refugee talk and prorefugee talk (the eigenvalues were λ = 6.32 and λ = 2.87, respectively, together explaining 36% of the variance). The above terms loaded onto the former rather than the latter factor and were therefore omitted from the prorefugee custom dictionary. The terms “UKhousing,” t(54) = −3.49, p = .001, d = −0.47, and “fortresseurope,” t(54) = −2.43, p = .02, d = −0.33, were significantly negatively associated with prorefugee talk. Therefore, these terms were also omitted from the dictionary. Scores for the remaining 12 key terms were entered into a second factor analysis. The eigenvalue (λ) for the prorefugee talk factor was 4.28 (explaining 33% of the variance). The terms formed a reliable measure, α = .83. Therefore, for the purposes of our main study, we created a custom dictionary to capture prorefugee talk that contained these 12 terms plus variants thereof (see Appendix A for a full list of items).

Main Study

Method

Data collection

Using a series of keyword searches in Chorus, we harvested tweets between the dates of September 6, 2015 (the week following the first tweet containing the image of Aylan Kurdi), and November 30, 2015. Using Twitter’s Search API, we could harvest a user’s past 3,200 tweets along their user timeline and 1 week of tweets for keyword searches. Initial search terms were #refugeescrisis, #migrantcrisis, #AylanKurdi, #Aylan_Kurdi, migrant crisis, Syria, refugees, #Syrianrefugees, #refugees, and variants thereof. To evolve the search with the changing narrative around the situation, we included the additional search terms #savethechildren, #helpiscoming, and #Hungary. These searches yielded 35,537 tweets from 22,212 distinct user IDs.

We next identified a subset of users for whom tweets existed before and after the publication of the images of Aylan Kurdi. These users met the following criteria: (a) Their profile indicated that they were located in the United Kingdom, (b) they had tweeted at all three time points, and (c) they had tweeted about refugees using the hashtags described above. To select this sample, we used TweetVis to compile a list of all distinct user IDs within the full data set that were located in the United Kingdom. Restricting the sample yielded a European, almost exclusively English-language tweet corpus without the need to remove tweets in languages other than English (many European Twitter users tweeted both in their national language and in English on this topic). Within the overall data set, 11,308 tweets originated from the United Kingdom. To harvest the tweets of users in the United Kingdom that were sent prior to the emergence of the images of Aylan Kurdi, we conducted an additional set of searches using the UK-based usernames in this original data set. These searches harvested each user’s previous tweets up to 1 week before the images of Aylan Kurdi were first publicized. Overall, the longitudinal data set contained 41,253 tweets divided into 3 week-long time periods.

Time 1

In the 7-day period between August 26, 2015, and September 1, 2015, we harvested the data from 459 unique usernames with UK locations. This data set contained 23,633 tweets. To control for authors’ linguistic style (Sukhoparov & Lebedev, 2014) and to ensure we analyzed only the sampled users’ own opinions, we used Chorus’ automated text analysis to classify and remove retweets (messages originating from another user that a user chose to retransmit). Overall, 9,987 (42%) of the tweets were retweets. Removing these retweets resulted in a final Time 1 data set of 13,646 tweets and 431 users (the mean original tweet frequency per username was 31.59).

Time 2

The images of Aylan Kurdi first emerged in the data set on September 2, 2015. During the 7-day period between September 2 to 8, 2015, all of the Time 1 users tweeted. This data set contained 20,703 tweets. Overall, 7,215 (34.85%) of the tweets were retweets. Removing these retweets resulted in a final Time 2 data set of 13,488 tweets and 422 users (i.e., 9 of the 431 Time 1 users only retweeted at Time 2).

Time 3

Time 3 was 10 weeks after Time 2 and immediately followed the Paris terror attacks (November 12) in the 7-day period between November 13 and 19, 2015. We chose this time point because it was popularly assumed that the terror attacks—which were erroneously rumored to be perpetrated by refugees—had reduced compassion for refugees (Lyman & Smale, 2015; Wright, 2015). If expressions of solidarity at this time were positively related to earlier tweets about Aylan Kurdi despite the salience of terrorism, this would be a pertinent response to the slacktivism critique (i.e., because tweeting about Aylan Kurdi’s images was associated with later public expressions of solidarity with refugees despite the threat from terrorism). This data set contained 20,501 tweets. Overall, 6,382 (31.12%) of the tweets were retweets. Removing these retweets resulted in a final Time 3 data set of 14,119 tweets and 373 users (all of whom had tweeted at Time 1 and Time 2).

Analytic strategy

The purposes of our analyses were (a) to explore the relationship between communicating about Aylan Kurdi at Time 2 and expressing solidarity with refugees at Time 3, controlling for the expression of prorefugee opinions and communications about death (in general; death-thought accessibility) at Time 1, and (b) to explore the processes by which communicating about Aylan Kurdi at Time 2 was related to expressing solidarity with refugees at Time 3. Thus, our analytic strategy aimed to identify content that could explain why communicating about Aylan Kurdi at Time 2 was connected to the expression of solidarity with refugees 10 weeks later, at Time 3, in the context of the Paris terror attacks, while accounting for earlier expressions of prorefugee sentiment and discussions about death in general at Time 1. To do this, we combined each user’s tweets at each time point into a single unit of observation for analyses (controlling for the number of tweets). Using LIWC2015, we then established the proportion of words by each participant at each time point that appeared in the standard and custom dictionary word categories (see the Supplemental Material available online)—including the death category (e.g., “bury,” “coffin,” “kill,” to control for the salience of death before the images emerged) and the affiliation category (e.g., “ally,” “friend,” “social,” because it was the closest LIWC category to the concept of social identification with a group; see Pennebaker et al., 2015). To index the proportion of words by each participant relating to harm and loyalty, we used the moral foundations dictionary created by Graham et al. (2009).

We created additional custom LIWC dictionaries to capture collective action content and threat content (see Appendix A for the full list of terms). One of the current authors and an independent person generated lists of words for each category on the basis of common English terms and common terms used in social psychological treatments of the two constructs. The independent lists were merged after discussion and then assessed in a sample of 14,395 English-language tweets (with spam and nontext tweets removed) drawn from tweets that included the hashtag #refugees in mid-2016 (collected independently of this study). The mean proportion of threat words in the #refugees corpus was 0.17 (SD = 1.00). The custom threat category was correlated .36 with risk in the standard LIWC2015 dictionary, which is noteworthy because perceptions of risk are central to the perception of threat (Huddy & Feldman, 2011). We concluded that this measure was related to, but not redundant with, risk. Collective action was more prevalent in the corpus (M = 0.84, SD = 2.26) and was correlated .13 with affiliation (the closest existing, but clearly distinct, construct in the standard LIWC2015 dictionary).

We further validated the threat and collective action dictionaries using Smith et al.’s (2015) data set of posts and comments on the #OccupyWallStreet Facebook event page. The mean proportion of collective action words in the #OccupyWallStreet corpus was 2.11 (SD = 5.04), and the mean proportion of threat words was 0.06 (SD = 0.54). Collective action and threat were not correlated, r(4552) = −.01, p = .50.

In Appendix B, we include Time 1 correlations from the present study. Again, threat and collective action were not significantly correlated with each other, suggesting that they are independent constructs. Furthermore, threat was positively correlated with Time 1 risk and Time 1 harm. Collective action was positively correlated with Time 1 affiliation and Time 1 anger (both LIWC2015 standard dictionary categories). The latter result is important because group-based anger, reflecting grievances about unjust circumstances, is an antecedent of collective action (van Zomeren, Postmes, & Spears, 2008). Taken together, this suggested that content pertaining to collective action was more relevant to mobilization than content related to threat.

Next, we constructed a latent Time 3 variable as our dependent variable to capture solidarity with refugees. This latent variable included the dictionary categories for prorefugee talk, collective action, loyalty, and affiliation (Fig. 1). Then, we conducted path analyses to examine the fit of the hypothesized model (Fig. 2) using the maximum likelihood procedure in AMOS (Version 18.0; Byrne, 2009). An a priori power analysis established that the sample size required for an effect size (d) of 0.50 was 86 observations (to detect significance at p < .05 in a regression model with 15 predictors; Faul et al., 2007). Missing values were estimated and replaced using the expectation-maximization algorithm in SPSS. This enabled us to examine the indirect relationships between discussion of Aylan Kurdi and expression of solidarity with refugees through discussion of issues relating to harm and threat, controlling for Time 1 expression of prorefugee opinions and discussion of death-related issues (as a control for death-thought accessibility).

Fig. 1.

Results of the confirmatory factor analysis of the latent factor for solidarity with refugees at Time 3. Standardized coefficients are shown. The unstandardized coefficient for loyalty was constrained to 0.70 to enable identifiability. Missing values were estimated and replaced using the expectation-maximization algorithm. Asterisks indicate significant paths (**p < .01).

Fig. 2.

Results from the test of the hypothesized model. Standardized coefficients are shown. Dotted lines indicate nonsignificant paths, and solid lines indicates significant paths (**p < .01, ***p < .001). Analyses controlled for volume of tweets at each time point. Missing values were estimated and replaced using the expectation-maximization algorithm. The model included correlations between Time 1 variables.

To test for indirect effects, we calculated the standardized bias-corrected bootstrap confidence intervals (for details of this bootstrapping procedure, see Cheung & Lau, 2008) using the unbiased estimates of mediation effects provided by the path model. The models’ goodness of fit was tested by using the chi-square ratio, the goodness-of-fit index (GFI; Tanaka & Huba, 1985), the root-mean-square error of approximation (RMSEA; Browne & Cudeck, 1993), and the standardized root-mean residual (SRMR; Hu & Bentler, 1999).

Results

Preliminary analyses

Means and standard deviations of the LIWC dimensions, custom dimensions, and additional metrics are in Table 2. Correlation coefficients to assess the zero-order relationships between the variables in the hypothesized model, including volume of tweets at each time point, are available in Table 3. At Time 1, the most frequently used hashtags were #refugees (89), #refugeecrisis (52), #Syria (46), and #lesbos (40). At Time 2, the most frequent hashtags were #refugeeswelcome (280), #refugees (261), #refugeecrisis (169), and #Syria (154). At Time 3, the top hashtags were #parisattacks (459), #paris (369), #isis (64), #Syria (61), and #refugees (43). This demonstrates that discussions about refugees occurred frequently at each time point, and although the Paris attacks were very widely discussed at Time 3, discussion about refugees was still present.

Table 2.

Means and Standard Deviations for LIWC2015 Dimensions and Metrics

	Time 1 (n = 431)		Time 2 (n = 422)		Time 3 (n = 373)
LIWC dimension	M	SD	M	SD	M	SD
Word count	524.08_a	711.12	509.86_a	649.89	635.77_b	718.31
Total tweets	31.66_a	53.16	26.93_b	35.90	32.00_a	38.84
Total retweets	1,681.86_a	11,298.86	4,608.61_b	24,041.13	—	—
Aylan Kurdi	—	—	0.12_a	0.46	0.01_b	0.04
Prorefugee	0.06_a	0.32	0.81_b	1.55	0.12_c	0.28
Threat	0.07_a	0.22	0.15_b	0.40	0.06_a	0.13
Collective action	0.64_a	0.72	0.76_b	0.84	0.64_a	0.71
Harm	0.14_a	0.44	0.19_a	0.38	0.36_b	0.52
Loyalty	0.14_a	0.27	0.16_a	0.34	0.17_a	0.31
Affiliation	1.94_a	1.64	2.04_a	1.69	1.86_a	1.28
Death	0.23_a	0.48	0.33_b	0.54	0.37_b	0.48

Note: Means on the same row with different subscripts differ at p < .05. For the content categories (i.e., excluding word count, total tweets, and total retweets), means represent the percentage of words that appeared in the standard and custom dictionary word categories. Total retweets were calculated at the tweet level; all other variables were calculated at the user level. LIWC2015 = Linguistic Inquiry and Word Count (Pennebaker, Boyd, Jordan, & Blackburn, 2015).

Table 3.

Zero-Order Correlations, Means, and Standard Deviations

Variable	M	SD	1	2	3	4	5	6	7	8	9	10	11	12	13
1. Time 1 death	0.24	0.62	—
2. Time 1 prorefugee	0.06	0.32	.31**	—
3. Time 2 Aylan Kurdi	0.13	0.47	.16**	.10*	—
4. Time 2 threat	0.15	0.42	.06	.02	.15**	—
5. Time 3 harm	0.36	0.50	.21**	.10	.23**	.21**	—
6. Time 3 threat	0.06	0.13	.10*	.10	.06	.17**	.36**	—
7. Time 3 solidarity (with refugees)	0.74	0.46	.10**	.06	.03	.15**	.16**	.08	—
8. Time 3 prorefugee	0.18	0.37	.07	.26**	−.06	.004	.13**	.11**	—	—
9. Time 3 loyalty	0.17	0.29	−.004	.07	.00	.19**	.17**	.15**	—	.21**	—
10. Time 3 affiliation	1.92	1.31	.08	−.02	−.01	.10	.04	−.02	—	.01	.28**	—
11. Time 3 collective action	0.67	0.77	.09	.06	.09	.09	.18**	.07	—	.09	.09	.17**	—
12. Time 1 volume of tweets	31.61	53.11	.01	−.02	−.05	.03	.07	.03	−.002	.01	−.02	−.03	.02	—
13. Time 2 volume of tweets	26.90	35.87	.03	.02	−.04	.03	.08	.09	.03	−.03	−.01	−.03	.07	.46**	—
14. Time 3 volume of tweets	32.00	38.84	.04	.03	.03	.05	−.02	.09	.07	−.03	.08	.05	.05	.04	.04

Note: Time 1 n = 431, Time 2 n = 422, and Time 3 n = 373.

p < .05. **p < .01.

Main analyses

Indexing solidarity with refugees

We conducted a confirmatory factor analysis to verify the factor structure of the proposed latent variable for solidarity (with refugees) that contained the following Time 3 word categories as observed variables: loyalty, prorefugee, collective action, and affiliation (Fig. 1). The unstandardized pathway between loyalty and solidarity was constrained to 0.70 to achieve identifiability. Results indicated that the latent variable fitted the data adequately, χ²(2) = 16.99, p < .001, GFI = .98, SRMR = .05, RMSEA = .13. All of the observed variables loaded significantly onto the factor at p ≤ .01.

Test of hypothesized model

Fit indices for the hypothesized model, controlling for volume of tweets sent at each time point, were, χ²(49) = 190.01, p < .001, GFI = .93, SRMR = .06, RMSEA = .08 (suggesting adequate, but not good, fit). Examination of the modification indices suggested that addition of paths between Time 1 prorefugee and Time 3 solidarity, β = 0.21, p = .002, and between Time 3 threat and Time 3 harm, β = 0.34, p < .001, would improve fit. Indeed, addition of these paths improved the fit of the model, χ²(47) = 126.28, p < .001, GFI = .96, SRMR = .06, RMSEA = .06. Taken together, these fit indices suggested that the relationships explained a significant amount of variance in the data (Fig. 2). There was a significant positive relationship between the Time 1 death category and the Time 2 Aylan Kurdi category, β = 0.15, p < .001. Therefore, people who mentioned death at Time 1 (before Aylan Kurdi’s death) were more likely to tweet about Aylan Kurdi the following week. However, the relationship between the Time 1 prorefugee category and the Time 2 Aylan Kurdi category was not significant, β = 0.05, p > .250, indicating that there was no association between prorefugee talk at Time 1 and posting about Aylan Kurdi a few days later. However, as expected, there was a significant positive relationship between talking about Aylan Kurdi at Time 2 and discussing harm at Time 3, β = 0.19, p = .001, and between discussing harm at Time 3 and solidarity at Time 3, β = 0.25, p = .001 (Hypothesis 1).

There was no direct association between discussing Aylan Kurdi at Time 2 and solidarity 10 weeks later, β = −0.10, p = .14. However, the results of bias-corrected bootstrapping with 2,000 resamples (Table 4) were consistent with discussion of harm at Time 3 and threat at Time 2 (but not at Time 3) mediating the relationship between discussion of Aylan Kurdi at Time 2 and solidarity at Time 3. In addition, the indirect effect of discussing threat at Time 2 on solidarity at Time 3 was significant, with discussion of harm at Time 3 but not threat at Time 3 mediating this relationship. However, Time 1 discussions of death did not have an indirect effect on Time 3 solidarity, suggesting that death-thought accessibility at Time 1 cannot explain the observed relationships. Similarly, expressing prorefugee statements at Time 1 did not have an indirect effect on expressing solidarity at Time 3.

Table 4.

Confidence Intervals for Indirect Effects

Indirect effect	Mean β	95% confidence interval
Overall effect: Time 2 Aylan Kurdi to Time 3 solidarity	0.05*	[0.03, 0.08]
Time 2 Aylan Kurdi to Time 3 solidarity through Time 2 threat	0.02*	[0.003, 0.04]
Time 2 Aylan Kurdi to Time 3 solidarity through Time 3 threat	0.01	[−0.003, 0.03]
Time 2 Aylan Kurdi to Time 3 solidarity through Time 3 harm	0.03*	[0.01, 0.06]
Overall effect: Time 2 threat to Time 3 solidarity	0.05*	[0.02, 0.08]
Time 2 threat to Time 3 solidarity through Time 3 threat	0.02	[−0.002, 0.04]
Time 2 threat to Time 3 solidarity through Time 3 harm	0.04*	[0.01, 0.07]
Overall effect: Time 1 death to Time 3 solidarity	0.003	[−0.01, 0.02]
Overall effect: Time 1 prorefugee to Time 3 solidarity	−0.001	[−0.02, 0.02]

Note: Missing values were estimated and replaced using the expectation-maximization algorithm. Analyses controlled for volume of tweets at each time point. Number of resamples = 2,000.

p < .05.

Unexpectedly, there was a significant positive relationship between discussing Aylan Kurdi and threat at Time 2, β = 0.15, p = .002, and there was a positive relationship between discussing threat at Time 2 and solidarity at Time 3, β = 0.19, p = .008. At Time 3, discussing threat did not predict solidarity, β = 0.13, p = .08. We had expected that expression of threat would be negatively associated with tweeting about Aylan Kurdi and with solidarity (Hypothesis 2) because of the narratives present at the time that linked refugees with threat to host communities. Instead, these results suggested that the threat expressed may not have reflected perceived threat from refugees; rather, it may have reflected threat to refugees. Therefore, we conducted a further analysis to explore the attributions of threat across waves.

We manually coded a subset of tweets containing threat words at each time point (Table 5). In those Time 1 tweets that contained threat content, the percentage of threat words within the tweets ranged from 3.45 (minimum) to 14.29 (maximum), with a median of 5.26. We selected 25 tweets randomly from each quartile. Two coders independently coded each tweet, and a third coder resolved any disagreements. For the purpose of establishing the threat attributions made about refugees—that is, whether the tweets focused on the threat to refugees posed (for example) by perilous journeys and murderous regimes versus threat perceived to be posed by refugees to host communities—we used five coding categories: (a) threat to refugees, (b) threat from refugees, (c) threat from terrorism, (d) mention of the refugee/migrant “crisis” with unattributed threat, and (e) none of the above. For 10 of the 300 tweets, two different codes were applied to a single tweet (meaning that the percentages in Table 5 for each wave sum to greater than 100%). We computed Siegel and Castellan’s (1988) kappa for each coder pair and then averaged it to provide a single index of interrater reliability. The resulting kappa for Time 1 indicated substantial agreement, κ = 0.87 (Landis & Koch, 1977).

Table 5.

Content Analysis of Threat Attributions

Code	Category	Percentage
Code	Category	Time 1	Time 2	Time 3
1	Threat to refugees	11	27	0
2	Threat from refugees	3	4	3
3	Threat from terrorism	1	4	32
4	Mention of refugee/migrant “crisis” with unattributed threat	24	40	15
5	None of the above	63	27	56

At Time 1, in 11% of the tweets, the context of the threat was a threat experienced by refugees, and 24% mentioned the migrant or refugee crisis without a direct attribution of threat. At Time 2 (κ = 0.70), the percentage of threat words within the tweets ranged from 3.33 to 20.00, with a median of 5.26. In 27% of the tweets, the context of the threat was a threat experienced by refugees. In only 8% of tweets was threat attributed to refugees or from terrorism. There was a significantly greater number of tweets about threat experienced by refugees than threat from refugees, χ²(1) = 20.12, p < .001. At Time 3 (κ = 0.84), the percentage of threat words within the tweets ranged from 3.70 to 20.00, also with a median of 5.26 (by coincidence, various 19 word tweets containing one threat word at each time point—hence containing 5.26% threat content—appeared at the median position at all three waves). The mode at Time 1 and Time 3 was 5% threat (1 threat word within a 20-word tweet) and 5.56% at Time 2 (1 threat word within an 18-word tweet). At Time 3, none of the tweets referred to a threat experienced by refugees, but 32% of tweets mentioned threat from terrorism.

Discussion

Tweeting about the death of Aylan Kurdi predicted expressing solidarity with refugees 10 weeks later to the extent that tweets focused on threat at the time the images emerged and harm at the time of the Paris terror attacks. This refutes the suggestion that engagement through social media need be a tidal wave of reaction that quickly subsides (cf. Morozov, 2009). Rather, communicating opinions online can be associated with substantive social and psychological changes (Harlow, 2012; McGarty et al., 2014; Smith, Gavin, & Sharp, 2015; Thomas et al., 2015; Wojcieszak, 2009). Thus, accounts of slacktivism are overly simplified; they do not take into account the way in which communicating opinions about social issues and events relate to the nature of later communications. Specifically, our results highlight the relationships between communicating about death, harm, and threat, on the one hand, and solidarity with an outgroup facing those threats, on the other.

We found a positive, direct relationship between the expression of prorefugee opinions at Time 1 and Time 3 solidarity. When controlling for this relationship, we found that tweeting about Aylan Kurdi at Time 2 was positively related to expressing solidarity with refugees 10 weeks later to the extent that discussions expressed a sense of harm. Therefore, in line with Hypothesis 1, we suggest that this is because awareness of harm focused attention on concern for refugees (Graham et al., 2009) and predicted increased expressions of solidarity (Smith, Thomas, & McGarty, 2015).

Intriguingly, expressing prorefugee opinions at Time 1 was not related to tweeting about Aylan Kurdi at Time 2. This suggests that the Aylan Kurdi situation not only moved people who were already prorefugee or had prosocial values, but also motivated those less involved. This supports our contention that Aylan Kurdi’s images violated a “universal” injunctive norm and caused a normative conflict (Smith et al., 2015).

The Time 3 data were collected just after the Paris terror attacks, when discussions about the ongoing Syrian refugee crisis were intermingled with conversations about terrorism. This timing is significant because tweeting about Aylan Kurdi was positively associated with tweeting about threat at Time 2, and there was a direct, positive relationship between threat at Time 2 and solidarity at Time 3, but not between tweeting about threat at Time 3 (the time of the terror attacks) and solidarity at Time 3. Therefore, rather than polarizing people against refugees (Hypothesis 2), communicating about threat was associated with positivity toward the group. At Time 2, the threat communicated focused on threat to refugees rather than threat posed by refugees. However, at Time 3 (when there was no relationship between communicating about threat and solidarity), the threat communications focused on threat from terrorism. Therefore, only when threat communications focused on threat to vulnerable out-groups were they associated with solidarity.

People who tweeted about death before the images emerged were more likely to tweet about Aylan Kurdi. This effect was apparent even when controlling for prorefugee sentiment. It may be that people who tweeted about death before the images emerged were more able to relate to the situation of refugees such as Aylan Kurdi and were therefore more likely to tweet about it. Indeed, Boyd, Morris, and Goldenberg (2017) found that in the presence of mortality salience, people who were open to experience were insulated from defending their world views. While one way to reduce the anxiety potentially associated with awareness of death is to avoid discussions of death (see Becker, 1973), people who were already tweeting about death would be more likely to have the resources to buffer against anxiety induced by graphic images of death and would therefore be more likely to tweet about Aylan Kurdi. This positive relationship between communicating about death and about images of the death of vulnerable others, and in turn the relationship with intergroup solidarity, presents an intriguing avenue for future research on improving intergroup relations.

We adopted a novel methodology to study these processes in a large data set, incorporating the strengths of quantitative analyses of observed communications with a commitment to examining the content of those communications. Much remains unknown about the extent to which social media communications can affect social change and how they do so. By deploying these novel longitudinal methods, we were able to demonstrate that online communications involve a psychological doubling down: Discussion of harm and threat in combination with the expression of a normative conflict (Smith, Thomas, & McGarty, 2015) is associated with mobilization potential.

There were major policy changes in many nations shortly after the images emerged (Amnesty International, 2016). Political actors respond to popular public opinion expressed through social media, so it is important to understand online activism in its own terms, regardless of whether that activism transfers to offline domains. Not all online engagement necessarily results in psychological change, and we did not consider all forms of online engagement (e.g., retweeting). Future research could examine whether the psychological responses observed here apply in the context of less effortful communications, such as retweeting and “liking.”

Conclusion

The photographs of Aylan Kurdi became emblematic of the appalling human suffering of Europe’s worst migrant crisis since the Second World War because they portrayed an unacceptable state of affairs that created a normative conflict between what is and what should be. We showed that expressing solidarity with refugees is closely intertwined with such norm violations—and solidarity is associated with communicating about the unjust harm experienced by vulnerable others.

Footnotes

Appendix A

Contents of Custom Dictionaries

Aylan Kurdi	Prorefugee	Threat	Collective Action
Aylankurdi	#refugeeswelcome	harm*	act*
#Aylankurdi	refugees welcome	risk	march*
Aylan	refugee welcome	threat*	rall*
Kurdi	welcome	danger*	protest*
#Alankurdi	support for refugee*	peril*	donat*
#Syrianchild	supportforrefugee*	hostil*	support*
Syrianchild	#supportforrefugee*	menac*	campaign*
Syrian child	support refugee*	hazard*	solidarity
Syrianboy	support migrants	advers*	mobilis*
#Syrianboy	support migrant	enem*	activis*
	supportmigrants	outsider	volunt*
	supportmigrant	crisis	participa*
	#supportmigrant*	endanger*	petition*
	pray		sign-up
	#pray*		join
	shameoneurope		affiliat*
	prayforsyria*		lobby
	#prayforsyria*		recruit
	pray for Syria*		retweet
	pray for #Syria		spread
	#pray #Syria*		stand
	#pray for #Syria*		unite
	#shameoneurope		persist
	shame on Europe		resist*
	#shame on Europe		fight
	#shame #Europe		struggle*
	shameoneurope		do
	wehaveroom		challenge
	#wehaveroom		dissent
	we have room		picket*
	welcome refugee*		boycott*
	welcomerefugees		write-in
	#welcomerefugees		demonstrat*
	#withsyria		cause*
	#wecandobetter		movement*
	we can do better		strike*
	#soukforsyria		blockade*
	UK for Syria*		sit-in
	solidarity		revol*
	#solidarity		moratori*
	#withsyria		rebel*
	with Syria*		crusad*
			drive
			canvass
			persua*
			talk
			oppose

Note: An asterisk at the end of a word or word stem signals Linguistic Inquiry and Word Count (LIWC2015) software (Pennebaker, Boyd, Jordan, & Blackburn, 2015) to ignore all subsequent letters. Consequently, for the term “welcomerefugee*,” for example, LIWC2015 also counted the phrase “welcomerefugees.” Twitter hashtags are case insensitive, but we have capitalized names to enhance readability.

Appendix B

Partial Correlations Between Time 1 Dictionary Categories

Variable	1	2	3	4	5	6
1. Prorefugee	—
2. Threat	.16**	—
3. Collective action	.28***	.02	—
4. Harm	−.01	.14**	.00	—
5. Anger	.07	.03	.20***	.19***	—
6. Risk	.16**	.19***	.38***	.27***	.21***	—
7. Affiliation	.003	−.05	.10*	−.07	.04	.07

Note: Correlations controlled for volume of tweets at Time 1.

p < .05. **p < .01. ***p < .001.

Action Editor

Ayse Uskul served as action editor for this article.

Author Contributions

L. G. E. Smith developed the research concept and design, collected and analyzed the data, and drafted the manuscript. C. McGarty and E. F. Thomas provided critical revisions and conducted coding. All authors approved the final version of the manuscript for submission.

Declaration of Conflicting Interests

The author(s) declared that there were no conflicts of interest with respect to the authorship or the publication of this article.

Supplemental Material

Additional supporting information can be found at

Notes

References

Amnesty International. (2016, February 3). Syria’s refugee crisis in numbers. Retrieved from https://www.amnesty.org/en/latest/news/2016/02/syrias-refugee-crisis-in-numbers/

Becker

(1973). The denial of death. New York, NY: The Free Press.

Boyd

Morris

K. L.

Goldenberg

J. L.

(2017). Open to death: A moderating role of openness to experience in terror management. Journal of Experimental Social Psychology, 71, 117–127. doi:10.1016/j.jesp.2017.03.003

Brooker

Barnett

Cribbin

(2016). Doing social media analytics. Big Data & Society, 3(2). doi:10.1177/2053951716658060

Browne

M. W.

Cudeck

(1993). Alternative ways of assessing model fit. In Bollen

Long

J. S.

(Eds.), Testing structural equation models (pp. 136–162). Newbury Park, CA: SAGE.

Burke

B. L.

Martens

Faucher

E. H.

(2010). Two decades of terror management theory: A meta-analysis of mortality salience research. Personality and Social Psychology Review, 14, 155–195. doi:10.1177/1088868309352321

Byrne

B. M.

(2009). Structural equation modeling with AMOS: Basic concepts, applications, and programming. London, England: Psychology Press.

Carlisle

J. E.

Patton

R. C.

(2013). Is social media changing how we understand political engagement? An analysis of Facebook and the 2008 presidential election. Political Research Quarterly, 66, 883–895. doi:10.1177/1065912913482758

Castells

(2012). Networks of outrage and hope: Social movements in the Internet age. Cambridge, England: Polity Press.

10.

Cheung

G. W.

Lau

R. S.

(2008). Testing mediation and suppression effects of latent variables. Organizational Research Methods, 11, 296–325. doi:10.1177/1094428107300343

11.

Faul

Erdfelder

Lang

Buchner

(2007). G*Power 3: A flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behavior Research Methods, 39, 175–191.

12.

Graham

Haidt

Nosek

B. A.

(2009). Liberals and conservatives rely on different sets of moral foundations. Journal of Personality and Social Psychology, 96, 1029–1046. doi:10.1037/a0015141

13.

Haidt

(2007). The new synthesis in moral psychology. Science, 316, 998–1002. doi:10.1126/science.1137651

14.

Haidt

(2012). The righteous mind: Why good people are divided by politics and religion. London, England: Penguin.

15.

Harlow

(2012). Social media and social movements: Facebook and an online Guatemalan justice movement that moved offline. New Media & Society, 14, 225–243.

16.

Harlow

Guo

(2014). Will the revolution be Tweeted or Facebooked? Using digital communication tools in immigrant activism. Journal of Computer-Mediated Communication, 19, 463–478. doi:10.1111/jcc4.12062

17.

L.-T.

Bentler

P. M.

(1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling, 6, 1–55. doi:10.1080/10705519909540118

18.

Huddy

Feldman

(2011). Americans respond politically to 9/11: Understanding the impact of the terrorist attacks and their aftermath. American Psychologist, 66, 455–467.

19.

Landis

J. R.

Koch

G. G.

(1977). The measurement of observer agreement for categorical data. Biometrics, 33, 159–174.

20.

Lee

Y. H.

Hsieh

(2013). Does slacktivism hurt activism?: The effects of moral balancing and consistency in online activism. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 811–820). New York, NY: ACM. doi:10.1145/2470654.2470770

21.

Lyman

Smale

(2015, November 15). Paris attacks shift Europe’s migrant focus to security. The New York Times. Retrieved from https://www.nytimes.com/2015/11/16/world/europe/paris-attacks-shift-europes-migrant-focus-to-security.html

22.

McGarty

Thomas

E. F.

Lala

Smith

L. G. E.

Bliuc

A.-M.

(2014). New technologies, new identities, and the growth of mass opposition in the Arab Spring. Political Psychology, 35, 725–740. doi:10.1111/pops.12060

23.

McGregor

H. A.

Lieberman

J. D.

Greenberg

Solomon

Arndt

Simon

Pyszczynski

(1998). Terror management and aggression: Evidence that mortality salience motivates aggression against worldview-threatening others. Journal of Personality and Social Psychology, 74, 590–605. doi:10.1037//0022-3514.74.3.590

24.

Morozov

(2009). Iran: Downside to the “Twitter revolution.” Dissent, 56(4), 10–14.

25.

Pennebaker

J. W.

Boyd

R. L.

Jordan

Blackburn

(2015). The development and psychometric properties of LIWC2015. Austin: University of Texas at Austin.

26.

Schumann

Klein

(2015). Substitute or stepping stone? Assessing the impact of low-threshold online collective actions on offline participation. European Journal of Social Psychology, 45, 308–322. doi:10.1002/ejsp.2084

27.

Siegel

Castellan

N. J.

(1988). Nonparametric statistics for the behavioral sciences. New York, NY: McGraw-Hill.

28.

Slovic

Västfjäll

Erlandsson

Gregory

(2017). Iconic photographs and the ebb and flow of empathic response to humanitarian disasters. Proceedings of the Natural Academy of Sciences, USA, 114, 640–644. doi:10.1073/pnas.1613977114

29.

Smith

L. G. E.

Gavin

Sharp

(2015). Social identity formation during the emergence of the occupy movement. European Journal of Social Psychology, 45, 818–832. doi:10.1002/ejsp.2150

30.

Smith

L. G. E.

Thomas

E. F.

McGarty

(2015). “We must be the change we want to see in the world”: Integrating norms and identities through social interaction. Political Psychology, 36, 543–557. doi:10.1111/pops.12180

31.

Stephan

W. G.

Stephan

C. W.

(2000). An integrated theory of prejudice. In Oskamp

(Ed.), Reducing prejudice and discrimination: The Claremont Symposium on applied social psychology (pp. 23–45). Mahwah, NJ: Erlbaum.

32.

Sukhoparov

Lebedev

(2014). Methodologies of Internet portals users’ short messages texts authorship identification based on the methods of mathematical linguistics. In 2014 IEEE 8th International Conference on Application of Information and Communication Technologies (pp. 258–263). New York, NY: IEEE.

33.

Tanaka

J. S.

Huba

G. J.

(1985). A fit index for covariance structure models under arbitrary GLS estimation. British Journal of Mathematical and Statistical Psychology, 38, 197–201.

34.

Thomas

E. F.

McGarty

Lala

Stuart

Hall

L. J.

Goddard

(2015). Whatever happened to Kony 2012? Understanding a global Internet phenomenon as an emergent social identity. European Journal of Social Psychology, 45, 356–367. doi:10.1002/ejsp.2094

35.

Vail

K. E. V.

Arndt

Motyl

Pyszczynski

(2012). The aftermath of destruction: Images of destroyed buildings increase support for war, dogmatism, and death thought accessibility. Journal of Experimental Social Psychology, 48, 1069–1081. doi:10.1016/j.jesp.2012.05.004

36.

van Zomeren

Postmes

Spears

. (2008). Toward an integrative social identity model of collective action: A quantitative research synthesis of three socio-psychological perspectives. Psychological Bulletin, 134, 504–535.

37.

Vis

Goriunova

(2015). The iconic image on social media: A rapid research response to the death of Aylan Kurdi. Retrieved from http://visualsocialmedialab.org/projects/the-iconic-image-on-social-media

38.

Williams

M. L.

Burnap

(2016). Cyberhate on social media in the aftermath of Woolwich: A case study in computational criminology and big data. British Journal of Criminology, 56, 211–238.

39.

Wojcieszak

(2009). Carrying online participation offline: Mobilization by radical online groups and politically dissimilar offline ties. Journal of Communication, 59, 564–586.

40.

Wright

(2015). Paris attacks: Refugees again at mercy of shifting public opinion. Huffington Post. Retrieved from http://www.huffingtonpost.com/kristin-wright/paris-attacks-refugees-at-mercy-of-public-opinion_b_8564890.html

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.01 MB

0.00 MB