Causal Effects of Language on the Exchange of Social Support in an Online Community

Abstract

The provision of social support is a common function of many online communities, but a full understanding of the causal effect of emotion language on the provision of support requires experimental study. The frequency of positive- and negative-emotion words in simulated posts requesting emotional support was manipulated and presented to a sample of college students (N = 442) who were randomly assigned to read one of four simulated posts. Participants completed measures of the original poster's (OP's) distress, and they provided a response to the simulated post. These responses were subjected to a computerized text analysis, and their overall effectiveness was rated by two independent judges. Fewer positive-emotion and more negative-emotion words in the simulated post led to perceptions that the OP was distressed and unable to cope. Participant-generated responses to the post were highest in positive-emotion words when the simulated post was high in positive-emotion words, but low in negative-emotion words. Finally, simulated posts that were low in positive-emotion words received responses that were judged to be more effective than did simulated posts that were high in positive-emotion words. These findings have implications for understanding the role of emotion language on the exchange of online social support.

Introduction

Online communities function as support systems for many individuals,^1,2 and people often share distress through social media.^3,4 Although seeking support online is not always as helpful as in-person support is,⁵ for many people seeking social support from online communities is beneficial.^6–8 Despite the promise of online support, individuals may not receive a helpful response from community members. Invalidating responses from others, such as when an individual denies a person's emotions⁹ or offers unhelpful advice,¹⁰ may have a harmful effect on the original poster (OP).

One factor that might dictate whether the OP receives helpful responses is the specific language used in the post. Emotional support (e.g., expressing sympathy) is a common form of social support provided online,¹¹ and affective feedback occurs more often than cognitive feedback when the OP has shared an emotion.¹² This suggests that members of online communities might respond to the OP's emotions when formulating a response. The presence of few positive-emotion words (e.g., love, nice) and many negative-emotion words (e.g., hurt, ugly)¹³ cue others that a person is distressed. Numerous positive-emotion words, on the other hand, are associated with health.¹⁴ One might, therefore, surmise that posts with high levels of positive-emotion words would indicate that the support seeker is coping well and that no response is needed, whereas high levels of negative-emotion words would signal distress and, therefore, elicit a response from others.

Two naturalistic studies have addressed the relation between language use and responses received in an online setting. Arguello et al.¹⁵ found that posts with high amounts of positive-emotion words and negative-emotion words received more responses than low-emotion posts. Lewallen et al.¹⁶ found that messages with more positive-emotion words were less likely to receive a response than were messages lower in positive-emotion words. While valuable, these naturalistic studies are limited because no cause-and-effect relations can be drawn between the language used by the OP and the likelihood of receiving a response. The first purpose of our research was to manipulate experimentally the positive-emotion and negative-emotion word usage in a post requesting social support to see how potential helpers appraise the OP's need for a response.

Even if a response is provided, there is no guarantee that such a response would be helpful. Because of emotional contagion (i.e., the transmission of an emotion from one person to another¹⁷), few positive-emotion words and/or many negative-emotion words in a post might foster a negative reaction from potential helpers. In a well-known study of Facebook users, Kramer et al.¹⁸ manipulated the number of posts with positive-emotion and negative-emotion words that showed up in users' News Feeds. A reduction in negative-emotion posts led to a decrease in the negative-emotion words the Facebook user used in subsequent status updates, whereas a reduction in positive-emotion posts led to a decrease in the user's positive-emotion words. Other research supports this link between the negative emotion of one user and the negative emotions of others with whom the user has communicated through computers.^19,20

Given the possibility of emotional contagion, support-seeking posts with low levels of positive-emotion words and/or high levels of negative-emotion words might cause negative emotions in potential helpers, thus leading to responses that are themselves highly negative in tone. Research on linguistic style matching²¹ would support the possibility that posts high in negative-emotion (or positive-emotion) words would elicit responses that are also high in negative-emotion (or positive-emotion) words. Our second purpose was to examine the link between the emotional content of an original post and the emotional content of a response.

Regardless of the emotion language used in the response, would the response be helpful? High-negative and/or low-positive posts might activate empathy in potential helpers and be met with emotional understanding.²² In a nonexperimental study, Rodríguez et al.¹² found that online support providers gave more emotional support and used more empathy in response to posts with negative emotions than posts that were positive or bivalent. An alternative theoretical possibility is that emotional contagion may result in unhelpful responses to high-negative and/or low-positive posts. Given these two potential outcomes, our third purpose was to examine the role of emotion language in an original post on the perceived helpfulness of actual responses. Along with this, we explored how participants' appraisals of the OP's need for a response predicted the helpfulness of those responses.

We sought to fill a void in the literature by manipulating the emotional content of posts to understand the relationships between the language used in support-seeking posts and (a) support providers' appraisal of the importance of providing a response, (b) the language that support providers use when responding, and (c) the effectiveness of these natural responses. First, we expected that posts with high levels of negative-emotion words and low levels of positive-emotion words would be appraised as needing a response from support providers. Second, we hypothesized that participants would respond with more positive-emotion words to posts that contained more (vs. fewer) positive-emotion words and that participants would respond with more negative-emotion words to posts that contained more (vs. fewer) negative-emotion words. Third, we hypothesized that the emotion language of the original post would predict the helpfulness of the response, but, because differing hypotheses could be made (empathy vs. emotional contagion), we did not make a directional hypothesis.

Materials and Methods

Participants

College students (N = 442) at a Midwestern university in the United States participated in this study. Their average age was 22.61 years old (SD = 6.03). Nearly three-fourths (73 percent) of the sample were women (n = 319). Most (n = 371) participants were European American, 26 identified as Latino/Latina, 19 identified as African American, and 7 identified as another ethnic/cultural group.

Research design

We used a 2-by-2 between-subjects design. One independent variable was the frequency of positive-emotion words in the simulated post (high vs. low), and the other was the frequency of negative-emotion words in the simulated post (high vs. low). A high frequency was between 4.71 and 6.12 (percentage of total words), whereas a low frequency was between 0.00 and 1.45 (see Appendix for the content of the simulated posts). These posts were intended to be similar to those found on self-help message boards. Each post contained the same thematic content (specifically, a person experiencing a relationship breakup), but varied in the frequency of positive-emotion words and negative-emotion words. Dependent measures included (a) participants' perceptions of the simulated post and OP, (b) participants' use of emotion words in a generated response to the simulated post, and (c) judges' ratings of the effectiveness of the participant-generated response.

Dependent measures

Perception of the OP's distress

To address our first purpose, participants completed a questionnaire that followed the presentation of the simulated post. This scale, developed by the authors for the purpose of this study, included six items (Table 1) designed to assess the degree to which the participant believed that the OP needed support. Items used a five-point scale ranging from 1 (not at all) to 5 (extremely). The six items assessed diverse content, such as the degree to which the OP seemed distressed, the importance of the OP receiving a reply, and the participant's willingness to reply. The heterogeneity of the items was evidenced by a confirmatory factor analysis failing to support a one-factor model, χ²(9, N = 417) = 153.41, root mean squared error of approximation (RMSEA) = 0.20, comparative fit index (CFI) = 0.81, standardized root mean of the residual (SRMR) = 0.09. Thus, these items were considered as separate variables in the analyses as opposed to indicators of a single variable.

Table 1.

Means and Standard Deviations (in Parentheses) of Perceptions of the Simulated Post

	Positive emotion in simulated post
	Low		High
	Negative emotion in simulated post
Item	Low (n = 102)	High (n = 108)	Low (n = 106)	High (n = 96)
How willing would you be to reply to this online post with a supportive response?	2.35 (1.20)	2.57 (1.25)	2.30 (1.24)	2.38 (1.22)
How important do you think it is that this post receives a reply?	3.19 (1.14)	3.35 (1.12)	2.99 (1.10)	3.36 (0.91)
How distressed (suffering from anxiety, sorrow, or pain) is the person who wrote this post?	3.61 (0.69)	3.94 (0.75)	3.23 (0.76)	3.61 (0.76)
How likely could the person who wrote this post manage without a response?	3.27 (0.83)	3.06 (0.92)	3.52 (0.76)	3.24 (0.91)
To what degree do you believe this individual is able to cope (deal effectively with something difficult)?	3.09 (0.83)	2.81 (0.90)	3.45 (0.85)	3.07 (0.87)
To what degree would the person who wrote this post experience emotional harm if she did not receive a response?	2.87 (0.98)	3.11 (0.99)	2.68 (0.87)	2.91 (0.91)

Word usage in participant-generated response

To address our second purpose, we used the Linguistic Inquiry and Word Count (LIWC2007)²³ program to count the number of positive-emotion and negative-emotion words the participant used in their response to the OP. The LIWC2007 is text analysis software that counts the frequency of words in a given category, among those emotion categories. The Negative Emotion and Positive Emotion word counts are expressed as percentages of the total number of words written that are either negative or positive in connotation. Pennebaker et al.²⁴ demonstrated the external validity of LIWC word counts through high correlations with judges' ratings.

Judges' ratings of the effectiveness of the participant-generated response

To address our third purpose, two judges read all of the participant-generated responses and rated their perceived helpfulness using a six-item scale (e.g., “How effective was this response in providing emotional support to the OP?” and “To what degree was this response helpful in alleviating the distress of the OP?”). These items used a five-point scale ranging from 1 (not at all) to 5 (completely). Coefficients alpha for the two judges were 0.90 and 0.91. We averaged item responses within judge, such that high scores indicated high levels of effectiveness. The average measure intraclass correlation for the two scores was 0.81, thus suggesting adequate reliability. We, therefore, averaged the two judges' scores into a composite measure of effectiveness of the response.

Procedure

Participants were recruited by a solicitation email that was sent to all students at one university, and the sample comprised students who volunteered to participate in response to this email. As an incentive, participants had the opportunity to win a $25 gift certificate upon the completion of the study. The recruitment email contained a link to one of four surveys that varied only in the simulated post. Thus, random assignment to experimental condition took place through the receipt of one of four email recruitment messages. After clicking the link in the solicitation email, participants were presented with an informed consent page. If they gave their consent, they clicked “Next.” Upon completion of the informed consent form, participants were asked to complete a demographics questionnaire. Then participants were instructed:

On the next screen, you will be presented with an emotional post that is similar to those found in an online community. Assume this post is from a stranger. You do not know the person who wrote this post. Your task is to read the online post as if you were a member of this online community.

Then they were presented with the simulated post. After reading the post, participants were asked to “refer to the simulated post” and complete the six items that measured their perception of whether the OP deserved a supportive response. After completing these items, participants were instructed:

Now we would like you to respond to the original post as if you were a member of the online community. You are free to respond however you wish, but your response must fill the entire text box. This study is confidential and your answer will not be made public.

After writing their response, participants were debriefed. Participants' natural responses were later analyzed through the LIWC2007. In addition, two judges later read and rated the usefulness of the natural responses.

Results

Purpose 1: perceptions of the simulated posts

The effect of manipulated negative-emotion words and positive-emotion words on the six items assessing perceptions of the post was tested using a two-way multivariate analysis of variance (MANOVA). Gender did not interact with either manipulated negative-emotion words nor positive-emotion words, so results are given for the entire sample. There were multivariate main effects for positive-emotion words and negative-emotion words, but the interaction was not significant (Table 2). Means and standard deviations are displayed in Table 1. As a follow-up, we conducted a univariate analysis of variance (ANOVA) on each item with adjusted alpha levels of 0.008. Higher (vs. lower) levels of positive-emotion words in the post were associated with perceptions that the OP was experiencing less distress, F(1, 408) = 23.05, p < 0.001, and was better able to cope, F(1, 408) = 13.32, p < 0.001. Posts with more (vs. fewer) negative-emotion words led participants to view the OP as more distressed, F(1, 408) = 23.95, p < 0.001, less able to manage without a response, F(1, 408) = 8.33, p < 0.008, and less able to cope, F(1, 408) = 14.66, p < 0.001.

Table 2.

Multivariate Analyses of Variances of Positive-Emotion and Negative-Emotion Words in Simulated Post on Perceptions of the Post and Emotion Words in the Responses

Dependent variable	Effect	Wilks' Λ	F
Perceptions of the post	Negative-emotion main effect	0.93	4.85^***
	Positive-emotion main effect	0.93	4.80^***
	Interaction	0.99	0.55
Emotion words in responses	Negative-emotion main effect	0.98	4.10^*
	Positive-emotion main effect	0.99	2.95
	Interaction	0.97	6.96^**

df for perceptions of the post = (6, 403); df for emotion words in responses = (2, 437).

p < 0.05; ^**p < 0.01; ^***p < 0.001.

df, degrees of freedom.

Purpose 2: content of participant-generated responses

The association between emotion words in the simulated post and emotion words in the participants' natural responses was tested using a MANOVA. As with the above analysis, gender did not interact with manipulated negative-emotion words nor positive-emotion words, so results here are for the entire sample. There was a multivariate main effect for negative-emotion words in the original post and an interaction (Table 2). Means and standard deviations are displayed in Table 3. We followed up this significant MANOVA with two ANOVAs (with an adjusted alpha level of 0.025). For positive-emotion words in the response, there was a main effect for the OP's negative-emotion words, F(1, 438) = 7.89, p < 0.01, but this was qualified by a significant interaction, F(1, 438) = 9.16, p < 0.01. Specifically, participant responses were highest in positive-emotion words when the original post had a high level of positive-emotion words, but a low level of negative-emotion words. There were no effects of the original post on the participants' use of negative-emotion words.

Table 3.

Means and Standard Deviations (in Parentheses) of Emotion Words in Natural Responses

	Positive emotion
	Low		High
	Negative emotion
Natural responses	Low (n = 108)	High (n = 115)	Low (n = 111)	High (n = 108)
Positive emotion	3.30 (3.15)	3.37 (3.37)	5.04 (4.41)	3.07 (3.02)
Negative emotion	1.86 (2.44)	1.58 (2.02)	1.59 (1.62)	1.71 (1.88)

Purpose 3: judges' ratings of natural responses

We examined whether the emotion content of the simulated post was associated with the effectiveness of the participants' responses. Results did not differ by gender, so these results are for the entire sample. A two-way ANOVA on response effectiveness revealed a main effect for the post's positive-emotion words, F(1, 306) = 12.73, p < 0.001. Specifically, responses to simulated posts with low levels of positive-emotion words were rated by judges as being more effective (M = 3.45, SD = 0.86) than were responses to simulated posts with high levels of positive-emotion words (M = 3.13, SD = 0.72). There was no main effect for negative-emotion words, F(1, 306) = 0.03, p = 0.86, nor an interaction, F(1, 306) = 0.00, p = 0.98.

We then examined whether participants' perceptions of the OP (from Purpose 1) predicted the effectiveness of participants' own responses. We conducted a multiple regression analysis with the six perceptions of the simulated post that we measured (i.e., those listed in Table 1) predicting response effectiveness. The six variables explained a significant percentage of variance in response effectiveness, R² = 0.15, F(6, 299) = 8.79, p < 0.001. Participants' ratings of their willingness to reply to the OP, β = 0.14, p = 0.02, and the perceived importance that the OP receives a reply, β = 0.22, p < 0.01, were associated with more effective responses; none of the other four variables were predictive of response effectiveness.

Discussion

The aim of this study was to investigate the impact of language on the provision of social support in an online community. Specifically, we examined experimentally how the use of positive-emotion and negative-emotion words in a request for help affects (a) potential helpers' perceptions of the seriousness of the request, (b) the emotion language used in responses provided by helpers, and (c) the effectiveness of the helpers' responses. Our findings suggest that the use of emotion language plays a meaningful role in these outcomes.

Posts that were low in positive-emotion words and those that were high in negative-emotion words elicited perceptions that the OP was experiencing distress and an inability to cope. Because the use of emotion words is associated with people's emotional experiences,²⁵ potential helpers presumably inferred that such posts reflected the OP's level of distress. Positive (vs. negative) emotions are also indicative of an individual's coping skills and resiliency,²⁶ so using positive-emotion words may signal to others that the OP is well-equipped to cope with the problem. Thus, emotion language in an online request for help has the potential to dictate one's likelihood of receiving a response.

The emotion language in an online request for help also dictates the emotional tone of the responses the OP receives. Responses were highest in positive-emotion words when the request for help was high in positive-emotion words, but low in negative-emotion words. Finding that the use of positive-emotion words in the response mirrored the use of positive-emotion words in the original post supports both the theory of linguistic style matching²¹ and the theory of emotional contagion.¹⁷ Specifically, mimicry appears to be a naturally occurring part of human interactions,¹⁷ and this might have manifested itself in the exchange between the OP and the helper. It may also be that participants were inclined to reinforce the OP's positive outlook, as represented by a high level of positive-emotion words and a low level of negative-emotion words, by responding in a highly positive manner.

Responses to online requests for help were deemed (by our judges) to be most effective when the original post was low in positive-emotion words. Thus, just as the level of positive-emotion words in the original post affected the emotion language used in responses, it also influenced the effectiveness of those responses. Perhaps, as discussed above, posts low in positive-emotion words indicated that the OP was unable to cope²⁶; this could motivate the support provider to provide a particularly helpful response. The finding that the effectiveness of the response was partly predicted by the responder's appraisal that it is important that the OP receives a response suggests that, in general, seeking support for a problem through an anonymous, online network can lead to effective responses.

These conclusions must be considered in light of the study's limitations. Perhaps the most prominent limitation was our simulation of an online exchange. Although a simulation was necessary to provide the level of experimental control we sought, we recognize that potential helpers likely behave differently in actual online settings. It would be valuable to replicate this research in an actual online setting with participants who frequently participate in online communities. In addition, we only looked at one issue—a relationship breakup experienced by a college-aged female. Whereas we believed that this would be a representative type of problem seen on social-support online sites, we could not determine how different issues (e.g., sexual assault, unemployment) or different support seekers (e.g., men, older adults) might affect helpers' responses differently. Finally, although we did not find gender differences in online support our study, other research has.²⁷ Thus, in future research it would be useful to examine different types of support sought and provided as a function of gender.

Given the widespread use of social media as a means to meeting people's social and emotional needs, examining the processes explaining how people seek and provide emotional support is vital. Although simple counts of the emotional language used in posts requesting emotional support cannot capture the full complexity of this process, we believe that this study represents an important part of addressing this research need. We found that the emotion language used in posts requesting emotional support matters quite a bit in terms of people's perceptions that an OP needs a response, the emotion language used in a response, and the overall effectiveness of the response. We hope that future research addresses other aspects of language use on the exchange of social support in online settings so a comprehensive theory of this process may be developed.

This research also has implications for practitioners. Given the rise in online counseling over the past several years,²⁸ practitioners might wish to consider how they might respond to the language used by clients who request professional psychological help online. Although counselors will likely react to distressed OPs differently than the laypeople-participants did in this study, online counselors might still wish to consider the potential impact of their use of emotion words on the perceived effectiveness of their responses to clients' online self-disclosures. By considering emotion language used by all parties in computer-mediated communication, people's psychological needs might be most effectively met.

Footnotes

Acknowledgments

This research was based on a master's thesis conducted by S.A.B. under the direction of J.H.K. The authors thank Ryan Downing and Michelle Lutz for serving as judges in this study.

Author Disclosure Statement

No competing financial interests exist.

Appendix

References

Love

, Crook

, Thompson

, et al. Exploring psychosocial support online: a content analysis of messages in an adolescent and young adult cancer community. Cyberpsychology, Behavior, and Social Networking, 2012; 15:555–559.

White

, Dorman

. Receiving social support online: implications for health education. Health Education Research: Theory & Practice, 2011; 16:693–707.

Moreno

, Jelenchick

, Egan

, et al. Feeling bad on Facebook: depression disclosures by college students on a social networking site. Depression and Anxiety, 2011; 28:447–455.

Frison

, Eggermont

. The impact of daily stress on adolescents' depressed mood: the role of social support seeking through Facebook. Computers in Human Behavior, 2015; 44:315–325.

Ybarra

, Mitchell

, Palmer

, et al. Online social support as a buffer against online and offline peer and sexual victimization among U.S. LGBT and non-LGBT youth. Child Abuse & Neglect, 2015; 39:123–136.

Gill

, Whisnant

. A qualitative assessment of an online support community for ovarian cancer patients. Patient Related Outcome Measures, 2012; 3:51–58.

Haker

, Lauber

, Rössler

. Internet forums: a self‐help approach for individuals with schizophrenia?. Acta Psychiatrica Scandinavica, 2005; 112:474–477.

Hwang

, Ottenbacher

, Green

, et al. Social support in an Internet weight loss community. International Journal of Medical Informatics, 2010; 79:5–13.

Kool

, van Middendorp

, Boeije

, et al. Understanding the lack of understanding: invalidation from the perspective of the patient with fibromyalgia. Arthritis Care & Research, 2009; 61:1650–1656.

10.

Smithson

, Sharkey

, Hewis

, et al. Problem presentation and responses on an online forum for young people who self-harm. Discourse Studies, 2011; 13:487–501.

11.

Bautista

, Lin

TTC

. Tweeting social support messages after a non-celebrity's death: the case of the Philippines' #Fallen44. Cyberpsychology, Behavior, and Social Networking, 2015; 18:641–646.

12.

Rodríguez Hidalgo

, Tan

ESH

, Verlegh

PWJ

. The social sharing of emotion (SSE) in online social networks: a case study in Live Journal. Computers in Human Behavior, 2015; 52:364–372.

13.

Pennebaker

, Mehl

, Niederhoffer

. Psychological aspects of natural language use: our words, our selves. Annual Review of Psychology, 2003; 54:547–577.

14.

Pennebaker

, Francis

. Cognitive, emotional, and language processes in disclosure. Cognition and Emotion, 1996; 10:601–626.

15.

Arguello

, Butler

, Joyce

, et al. (2006) Talk to me: foundations for successful individual-group interactions in online communities. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, Association for Computing Machinery, New York, pp. 959–968.

16.

Lewallen

, Owen

, Bantum

, et al. How language affects peer responsiveness in an online cancer support group: implications for treatment design and facilitation. Psycho-Oncology, 2014; 23:766–772.

17.

Hatfield

, Cacioppo

, Rapson

. Emotional contagion. Current Directions in Psychological Science, 1993; 2:96–99.

18.

Kramer

ADI

, Guillory

, Hancock

. Experimental evidence of massive-scale emotional contagion through social networks. Proceedings of the National Academy of Sciences, 2014; 111:8788–8790.

19.

Guillory

, Spiegel

, Drislane

, et al. (2011) Upset now?: emotion contagion in distributed groups. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pp. 745–748.

20.

Hancock

, Gee

, Ciaciaco

. et al. (2008) I'm sad you're sad: emotional contagion in CMC. In Proceedings of the ACM conference on Computer-Supported Cooperative Work, Association for Computing Machinery, New York.

21.

Niederhoffer

, Pennebaker

. Linguistic style matching in social interaction. Journal of Language and Social Psychology, 2002; 21:337–360.

22.

Preece

, Ghozati

. (1999) Experiencing empathy online. In Rice

, Katz

, eds. The Internet and health communication:experiences and expectations. Thousand Oaks, CA: Sage, pp. 237–260.

23.

Pennebaker

, Chung

, Ireland

, et al. (2007) The development and psychometric properties of LIWC2007. Austin, TX: LIWC.net.

24.

Pennebaker

, Francis

, Booth

. (2001) Linguistic inquiry and word count: LIWC 2001. Mahwah, NJ: Erlbaum.

25.

Kahn

, Tobin

, Massey

, et al. Measuring emotional expression with the Linguistic Inquiry and Word Count. The American Journal of Psychology, 2007; 120:263–286.

26.

Fredrickson

. What good are positive emotions?. Review of General Psychology, 1998; 2:300–319.

27.

Teoh

, Chong

, Yip

CCE

, et al. Gender as moderator of the effects of online social support from friends and strangers: a study of Singaporean college students. International Perspectives in Psychology: Research, Practice, Consultation, 2015; 4:254–266.

28.

Richards

, Viganó

. Online counseling: a narrative and critical review of the literature. Journal of Clinical Psychology, 2013; 69:994–1011.