This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Mental Health, is properly cited. The complete bibliographic information, a link to the original publication on http://mental.jmir.org/, as well as this copyright and license information must be included.
Many individuals who self-injure seek support and information through online communities and mobile peer-support apps. Although researchers have identified risks and benefits of participation, empirical work linking participation in these web-based spaces to self-injury behaviors and thoughts is limited.
This study aims to investigate the relationship between behavioral and linguistic traces on a mobile peer support app and self-injury outcomes.
Natural use data and web-based surveys (N=697) assessing self-injury outcomes were collected from 268 users (aged 13-38 years; median 19; 149/268, 55.6% female) of a mobile peer-support app for 4 months. Participants were identified as having posted self-injury content using an internal classifier. Natural log data was used to predict self-injury outcomes in a series of multilevel logistic and linear regressions.
Greater engagement on a mobile peer-support app was associated with a decreased likelihood of self-injury thoughts (odds ratio [OR] 0.25, 95% CI 0.09-0.73) and fewer intentions to self-injure (
To our knowledge, this is the first study to connect naturally occurring log data to survey data assessing self-injury outcomes over time. This work provides empirical support for the relationship between participation in online forums and self-injury outcomes, and it articulates mechanisms contributing to this relationship.
Self-injury—the deliberate, self-inflicted damage of body tissue [
Mobile apps are an increasingly common way for individuals to access self-injury communities and resources, and these technologies can be used to deliver interventions [
This work employs a mixed methods approach to address the gap in knowledge on the relationship between web-based self-injury support activity and self-injury outcomes. Specifically, we combine computational and survey methods to investigate the relationship between language and behaviors on a mobile peer support app, on the one hand, and self-injury behaviors, thoughts, intentions to self-injure, and ability to resist self-injury urges over time, on the other hand. This work provides empirical support for the relationship between participation in online forums and self-injury outcomes and articulates mechanisms contributing to this relationship.
Self-injurious behaviors can serve a variety of intrapersonal and interpersonal functions [
A functional understanding of self-injury is useful when interrogating the relationship between web-based activity and self-injury because it provides guidance on factors temporally associated with the behavior. Interpersonal functions are of particular interest because social and relational factors are likely salient in web-based spaces where people provide and receive peer support [
The potential benefits of online peer-to-peer support networks for individuals with mental health conditions include receipt of social support, validation, an increased sense of belonging [
Other studies substantiate a growing concern over the potential for adverse effects. Risks of exchanging online support for self-injury include reinforcement of the behavior, excessive focus on emotional suffering and rumination, and exposure to triggering content [
There is also concern that exchanges in online communities may downplay the serious consequences of the behavior [
Regarding how participation in online communities modulates self-injury behavior, the evidence is mixed. Murray and Fox [
When thinking about characteristics and contexts associated with self-injury outcomes, 2 additional lines of work can be informative: (1) diary and EMA studies and (2) computational mental health research.
A recent review of self-injury EMA studies identified emotional, cognitive, and social contexts associated with NSSI, motives that lead to NSSI, and mechanisms that influence or predict NSSI [
Cognitive states and patterns are also empirically associated with self-injury [
Finally, contextual factors, such as interpersonal conflict and feelings of rejection, are powerful predictors of same-day self-injury thoughts and behaviors. Interestingly, a study found that although the act of revealing NSSI to others is associated with greater perceived social support, perceiving support increases the likelihood of self-injury on the following day [
Although EMA studies provide insights into the complex temporal interplay between context, cognition, emotion, and behavior, by accessing data
Previous studies have shown that behavioral patterns and linguistic features of posts in web-based communities can distinguish between people with and without a number of conditions [
Suicidal ideation has also been associated with self-attentional focus and reduced social engagement and expressions of hopelessness, anxiety, impulsiveness, and loneliness [
This study aims to leverage the strengths of the methodologies mentioned above—rich, naturally occurring web-based data and self-report survey data over 12 weeks—to explore how self-injury behaviors and thoughts are related to activity and language use of self-injurious users on the TalkLife platform, a free mobile app designed for young people with a variety of mental health concerns. The platform uses a crowdsourced peer-support model to provide users with affordable and timely support. Self-injury outcomes are modeled (eg, behavior, thoughts, intentions, ability to resist) as a function of web engagement and language manifested in content. We explore 2 dominant questions: What behavioral (research question [RQ] 1) and language (RQ2) patterns are associated with self-injury behaviors, self-injury thoughts, intentions to self-injure, and ability to resist self-injury? These 2 questions are further broken down into the following subquestions:
Next, we examine language that may be predictive of self-injury outcomes. On the basis of interpersonal models of self-injury [
To understand how the mobile app’s use is related to self-injury outcomes, we employed a mixed-methods approach utilizing surveys and naturally occurring log data over 4 months; 3 types of data are included in the analyses: (1) responses to surveys, (2) behaviors on the platform, and (3) language use in posts and comments. A description of data acquisition, relevant measures, and treatment of these measures follows.
Surveys were issued on a rolling basis for 12 weeks. The first and last surveys were administered on October 25, 2018, and January 17, 2019. Survey administration was triggered internally by a classifier identifying suspected self-injury content. Once participants’ posts were flagged, they received a prompt to answer surveys once a week for the duration of the study period. Due to this method, the total duration of the study for any given participant varied. Participants could opt out of weekly surveys at any point.
The final data set was constrained to participants who had completed at least two surveys and corresponding behavioral and language data were extracted based on this criterion. Participants who did not complete basic demographics or who did not complete at least one self-injury outcome variable in a survey were removed from the sample. The number of surveys participants completed varied (mean 2, SD 1.20; range 1-10 surveys), as did the time between surveys (mean 1.74, SD 2.15 weeks; range 1-11.6 weeks). The total number of participants included in the final analyses was 268 with 697 survey observations. Our institutional review board approved all study procedures and data security measures.
Surveys included 9 items that were administered weekly with a question to address the (1) presence of self-injury behaviors, (2) frequency of self-injury behaviors, (3) presence of self-injury thoughts, (4) intensity of self-injury intentions, and (5) ability to resist. Additionally, there were items for past experience with therapy, age of first self-injury, and demographics, including age (how old are you?), race (what is your race?), and gender (what is your gender?).
Self-injury items (1-5 above) were treated as dependent variables and demographics (age, race, and gender) were included as covariates in all models. Response categories for self-injury behavior and thoughts were binary:
Deidentified behavioral data for participants meeting the above criterion (2 or more surveys) were sourced with license and consent from the TalkLife platform. This included metadata and original posts and comments. Given that weekly surveys referred to self-injury activity in the previous week, behavioral data at 1 week before each weekly survey were extracted as the primary data for prediction. In addition to controlling for demographics, we controlled for differences in time (relative to survey number) in all analyses because of the survey administration’s rolling basis.
We focus on several measures in analysis: (1) activity level (operationalized as averages of posts, gifts, reactions, comments, likes, and users followed in the previous week), (2) posting triggering content (operationalized as the number of posts a user published with trigger warnings), and (3) viewing triggering content (operationalized as the number of times a user dismissed trigger warnings when looking at others’ posts). All of these variables were averaged at the day level and log-transformed to restore normality because of their high positive skew.
Given that variance in behavior has proven to be a meaningful independent predictor of mental health in previous work [
This variance measure provides a sense of how an individual’s log data (eg, activity, publishing, and viewing triggering content) is distributed over the week. A high variance score is a proxy for instability or more change in activity over the course of the week (eg, users are very active one day, have no activity the next, and then are moderately active). Next, we computed a
Δ =
This change variable was entered into the models as continuous, including negative and positive values. Thus, the interpretation of this variable should be as a
The final variables for the behavioral data include: (1) activity, (2) trigger posts, (3) trigger dismiss, (5) variance (×3), and (6) change between proximal and distal activity (3).
All posts and comments made by participants within the study period were preprocessed and run through the Linguistic Inquiry and Word Count (LIWC) program, a psycholinguistic text analysis tool that is frequently employed in research on mental health [
Relevant dimensions were identified from the initial literature review, including: (1) affect (eg, positive emotion and specific negative emotions [sadness and anger]), (2) social or relational (eg, mentions of family or friends), (3) affiliative language (eg, affiliation and
As with the behavioral data, variability and change were computed for select language dimensions based on previous research. On the basis of the emotional cascade model [
Descriptive statistics for the survey data are provided in
Participant characteristics (n=268).
Characteristic | Value, n (%) | |
|
||
Male | 83 (30.9) | |
Female | 149 (55.6) | |
Transgender or nonbinarya | 36 (13.4) | |
|
||
White | 164 (61.2) | |
Black | 13 (4.9) | |
Asian | 43 (16.0) | |
Otherb | 47 (17.5) |
aThe response options of transgender male-to-female (n=2), transgender female-to-male (n=8), do not identify as male or female (n=12), and not sure (n=14) were combined because of small cell sizes.
bThe response options of American Indian or Alaskan Native (n=6), Native Hawaiian or other Pacific Islander (n=3), and other (n=38) were combined because of small cell sizes.
During the 4 months of this study, 48.5% (130/268) of participants reported self-injury behaviors, and 84.7% (227/268) reported having self-injury thoughts. Of those who reported injuring, the median weekly frequency was 3 times. Overall, 79.5% (213/268) of participants reported having thoughts of self-injury without engaging in self-injury behavior, whereas only 2.6% (7/268) of participants reported self-injury behaviors without also reporting self-injury thoughts.
Before analysis, diagnostic tests were run to determine appropriate modeling and the need for further data transformation. As mentioned above, highly skewed predictor variables (activity, trigger posts, and trigger views) were corrected through log-transformation. Self-injury frequency was the only response variable to be abnormally distributed and was thus also log transformed. Multicollinearity was assessed for all variables in relation to each dependent variable, using the R package mctest. The highest variance inflation factor (VIF) factor was consistently reported for rumination (8.92-9.58), followed by self-referent language (4.22-5.18). The mean VIF for each outcome was acceptable (self-injury behavior, 3.48; self-injury thoughts, 3.43; intentions to injure, 3.32; ability to resist, 3.33; and self-injury frequency, 3.37). As multicollinearity was not detected, we proceeded with analyses without excluding any variables at the outset.
The relationship between TalkLife activity and self-injury outcomes was analyzed using multilevel analysis to account for the data’s nested structure. Survey responses, log data, and language data were nested at the participant’s level; therefore, we included random effect of participant in all analyses. A total of 5 models were run to predict (1) self-injury behavior, (2) self-injury thoughts, (3) ability to resist the urge to injure, (4) intentions to injure, and (5) behavioral frequency. Logistic regressions predicting behavior and thoughts were analyzed using the R lme4 package, and linear models predicting the ability to resist, intentions, and frequency were analyzed with the nlme package. All models were adjusted for demographics (age, gender, and race) and time points.
Given this work’s exploratory nature, we began with full models including the 31 variables described above (4 control variables, 3 log variables, 10 language variables, 8 variance measures, and 8 change scores). These full models were subsequently reduced via backward variable selection. The logged coefficients were exponentiated for easier interpretation for the binary dependent variables (behaviors and thoughts). We report the significant results in the following section.
In terms of behaviors on the web, the odds of engaging in self-injury behavior increased with the number of triggering posts published in the week before the survey. For every additional unit increase in the log of triggering posts, the odds of engaging in self-injury behavior increased nearly five-fold (OR 5.37, 95% CI 1.25-23.05;
Self-injury behavior.
Self-injury behavior | B | SE | ORa (95% CI) | |
Intercept | −0.07 | 0.64 | 0.93 (0.27-3.29) | |
|
||||
Trigger posts | 1.68b | 0.74 | 5.37 (1.25-23.05) | |
|
||||
Trigger views | −0.20b | 0.10 | 0.81 (0.68-0.98) |
aOR: odds ratio.
bThe model was adjusted for demographics (age, gender, and race) and time point. Age and race were significant at
For behaviors on the web, activity level emerged as a significant predictor of self-injury thoughts. Greater active use of the platform (as indicated by posts, comments, and likes) was associated with lower odds of reporting self-injury thoughts (OR 0.64, 95% CI 0.45-0.90;
Self-injury thoughts.
Self-injury thoughts | B | SE | OR (95% CI) | |
Intercept | 1.78a | 0.67 | 5.91 (1.59, 21.97) | |
|
||||
Activity | −0.45a | 0.18 | 0.64 (0.45-0.90) | |
Trigger posts | 2.88a | 1.22 | 17.87 (1.64-194.15) | |
|
||||
Trigger posts | −1.22a | 0.55 | 0.29 (0.10-0.86) | |
|
||||
Rumination | 0.14a | 0.06 | 1.15 (1.02-1.29) |
aThe model was adjusted for demographics (age, gender, and race) and time point. Race was significant at
For behaviors on the web, the number of trigger warnings dismissed was positively related to the ability to resist (
Several language dimensions also emerged as significant. In particular, the use of self-referent language (
Ability to resist self-injury.
Ability to resist self-injury | B | SE | |
Intercept | 2.10a | 0.24 | |
|
|||
Trigger views | 1.39a | 0.66 | |
|
|||
Ib | −0.07a | 0.03 | |
Efficacy | 0.14a | 0.06 | |
|
|||
Positive emo | 0.05a | 0.02 | |
|
|||
Trigger dismiss | −0.18a | 0.08 | |
Anger | 0.19a | 0.09 |
aThe model was adjusted for demographics (age, gender, and race) and time point. Gender (
bSelf-referent language.
We noted a negative association between activity and intention to injure (
Intentions to self-injure.
Intentions to self-injure | B | SE | |
Intercept | 2.43a | 0.26 | |
|
|||
Activity | −0.37a | 0.09 | |
Trigger views | 1.50a | 0.61 | |
|
|||
Family | −0.63a | 0.22 | |
|
|||
Trigger views | −0.06a | 0.03 |
aThe model was adjusted for demographics (age, gender, and race) and time point. Age and race were significant (
There was a positive relationship between the number of posts published with triggering content in the previous week and the frequency of self-injury behaviors reported in the following report period (
Self-injury frequency.
Self-injury frequency | B | SE | |
Intercept | 0.59a | 0.15 | |
|
|||
Trigger posts | 0.45a | 0.15 | |
|
|||
Ib | 0.02a | 0.01 |
aThe model was adjusted for demographics (age, gender, and race) and time point. Age was significantly different (
bSelf-referent language.
In this study, we employed survey responses and naturally occurring log data from a mobile peer-support platform to predict self-injury outcomes. This study fills an important gap in the research literature by connecting behavioral and language patterns to self-reported self-injury outcomes and offers new insights into the relationship between participation in online communities and self-injury.
One of the primary aims of this work was to shed light on what specific behaviors may be beneficial or detrimental; much of the work on risks and benefits of participation in online communities has been qualitative and has not rigorously examined specific web-based activities [
There are several possible explanations for these findings. The most direct explanation is that active use of TalkLife reduces self-injury thoughts and intent, which is in alignment with what the platform was designed to do. This is consistent with the work showing that individuals who engage in active use of social media derive important benefits, such as cultivating feelings of support, connections with others, and companionship [
Unlike being active on the app, posting triggering content was positively associated with self-reported self-injury thoughts and behaviors. In other words, although active use appears to be indicative of fewer thoughts and intentions to injure, the types of content posted—specifically content that has been labeled as triggering—predicted an increased risk of self-injury behaviors and thoughts. Given the nature of these analyses, it is not possible to infer the causal direction of this finding as triggering posts may have been published before or after self-injury thoughts or behavior. Nevertheless, the high temporal correlation of posting triggering content and reporting self-injury thoughts and behaviors suggests that it could be leveraged to check in on users and provide support at key junctures. The directionality of this relationship should be explored in future work.
Interestingly, viewing triggering content appears to be positively related to both abilities to resist urges to self-injure and intentions to self-injure. Although seemingly contradictory, it may be that participants with a strong intention to self-injure dismiss trigger warnings to view triggering content and dissuade themselves from engaging in the behavior. In so doing, individuals may feel more capable of resisting self-injury. Indeed, this speculation is congruent with findings from other work in which seeing or reading graphic content in web-based forums appears to assuage urges to injure [
Variance in viewing triggering content was negatively associated with the ability to resist self-injury. The more varied an individual’s viewing behavior was from day-to-day, the less they reported being able to resist self-injury. In contrast, the rate of change between proximal (the week before) and more distal (remaining time between surveys) viewing shows a negative relationship with the risk of self-injury behavior and intentions to injure. As the change score increased in magnitude (ie, more viewing in a week before a survey relative to a distal period), the likelihood of self-injury behavior and intention to self-injure decreased. Together, these findings reflect a nuanced relationship in which day-to-day variability in viewing triggering content is linked to poorer ability to resist urges—yet an increase in proximal to more distal viewing activity is related to less likelihood of self-injury and self-injury intentions. Variance reflects instability, so it may be that variance in viewing triggering content is characteristic of maladaptive coping. The change score represents a quantitative shift in weekly activity from more distal activity. Thus, from a prediction perspective, high rates of change should signal potential risk. Future work should probe this complex relationship more deeply.
In response to the second set of research questions related to language, we found (1) no relationship between affiliative language and self-injury outcomes, (2) that mentions of family were negatively related to intentions to injure, (3) that variance of positive emotions and anger were related to the ability to resist self-injury, and (4) that ruminative variance was related to self-injury thoughts.
In addition to these larger themes, self-referent language in posts was negatively associated with self-reported ability to resist self-injury. This finding is consistent with work showing that self-referent language is associated with poor mental health status [
Another key finding is that as familial language increased, intentions to injure decreased. This aligns with the literature citing family as a protective factor and family disharmony as a key risk factor for self-injury [
Patterns in emotional expressions were also observed for self-injury thoughts and the ability to resist. Specifically, participants were more likely to have thoughts of self-injury when posts in the week before a survey varied in the use of ruminative language. This is in line with other studies that found rumination instability predicted daily reports of NSSI [
Finally, we found that as the rate of change of positive emotion increased, so did the ability to resist self-injury. Expressions of positive emotion have been associated with improved well-being in previous work [
The findings highlight implications for researchers working in digital mental health, clinicians working with young people who self-injure, and platform designers. In terms of research, much of the previous work investigating web-based communication about self-injury has focused on “static entities such as websites and forums rather than the fragmented, heterogeneous and dynamic current landscape of social media” [
Our findings also emphasize the importance of considering web-based activity, as it relates to self-injury recovery in the context of treatment [
Specifically, our findings suggest that clinicians should check in with clients regarding posting and viewing self-injury content. Posting content labeled as triggering was related to self-injury behavior and thoughts in our study, and viewing such content was related to both greater ability to resist self-injury and greater intentions to injure. Of note, we found that exposure to such content is not necessarily related to greater self-injury frequency. Although these findings are preliminary, they are in line with other previous work, suggesting that conventionally
Our findings also have relevance for platform designers. As confidence in being able to identify individuals engaging in, or at risk of, self-injury increases, we can begin to think about how to deliver timely and accessible interventions. Given that most of our participants were not currently in therapy, it may be beneficial to consider integrating elements of evidence-based treatment into the platform experience (eg, psychoeducation). These elements could be optional and be offered based on the individual’s use patterns. Finally, although the tendency has been to remove graphic, or potentially triggering, content from platforms, there have been discussions about the dangers of overmoderation [
This study has several limitations that should be acknowledged. First, the methodological design imposes limitations on causality. Survey questions were framed to ask about any self-injury in the week before but did not ask about specific days, or times of day, when these events (eg, thoughts, behaviors) occurred. Therefore, it was not possible to establish a detailed timeline for when the self-injury events occurred relative to the activities on TalkLife. Future work might consider daily surveys, either through diary or EMA, for a more nuanced understanding of these temporal relationships. We also chose to aggregate our data at the week level in this study, but it would be worthwhile for future work to explore associations between self-injury outcomes and web-based activity across other periods (eg, daily, monthly).
Second, while LIWC is widely used for language analysis, it does not account for context. Several key findings on language are tentative and should be more thoroughly explored in future work. One way to do this would be to triangulate with other language analysis techniques such as the tf-idf or word co-occurrence measures derived from n-grams.
Third, we operationalized web-based activity as
There are limitations to the generalizability of these findings. TalkLife is a platform specifically intended for the exchange of support related to mental health challenges, and the extent to which our findings generalize to other social platforms is unknown. In addition, there is potential for selection bias in this study. Participants were inclined to use support apps, willing to take weekly surveys on their self-injury, and engaged in active use of the app. This suggests that participants may have been similar in ways related to their use and their readiness to change self-injury behaviors. Future work may wish to consider recruiting and incentivizing a more diverse population for a broader picture of TalkLife activity.
Our findings should also be contextualized in light of the limitations of our analytical approach. Although variable selection procedures are common in psychological and social scientific research, alternative approaches, such as stochastic search variable selection or lasso regression, may enhance the reliability of models when selecting variables from a large number of predictors [
Finally, this exploratory work was meant to identify factors associated with self-injury outcomes that could be targeted in future confirmatory research with sufficiently powered samples. The methodological approach of combining naturally occurring web-based data with self-report survey data provides new insights into the relationship between the use of a web-based peer-support platform and various self-injury outcomes.
This study investigated the relationship between web-based support activities and self-injury outcomes to identify patterns that may be beneficial or harmful. To do so, we employed a novel mixed methods approach that utilized naturally occurring language and behavioral data from a mobile peer-support app and survey data collected over 4 months. Our findings point to a nuanced set of relationships. Specifically, participants who actively engaged in TalkLife were at lower odds of reporting self-injury thoughts and intentions. However, activity level was not predictive of self-injury behavior or the ability to resist self-injury urges. Posting triggering content was associated with greater odds of participants reporting self-injury thoughts and behaviors, whereas viewing triggering content was linked to both greater abilities to resist urges and greater intentions to injure. This work provides empirical support for the relationship between participation in a web-based support platform and 5 self-injury outcomes and articulates patterns that merit consideration in future work. We hope that insights from this study will inform future research on digital mental health and platform design.
ecological momentary assessment
Linguistic Inquiry And Word Count
nonsuicidal self-injury
odds ratio
research question
variance inflation factor
This work was supported by the USDA National Institute of Food and Agriculture, Hatch project 1004268 and the Cornell Center for Social Sciences. KK completed final drafts of this manuscript while also receiving support from the National Institute of Mental Health (T32 MH115882). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
None declared.