For the purpose of this study, anonymized chat data from all users receiving counseling between May 2020 and July 2021 were extracted from the krisenchat database. Data extraction and preparation were performed by authors affiliated with krisenchat (ME, SS, JT, and RW) so that chat content remained within the krisenchat database. The anonymized chat data included metadata on the chat (total number of messages and words sent by users during the whole counseling process, and number of sessions) and information about the user that counselors identified and noted during the counseling process (sociodemographic information, such as gender and age, and topics of users’ concerns). krisenchat counselors were volunteers and had a background in psychosocial studies. In addition, they underwent a structured 2-month training in chat-based counseling. Regarding gender, counselors had 3 options (male, female, and diverse) to mark in their documentation. They were encouraged to record the identified gender and not the biological sex of the users. The gender “diverse” included individuals identifying as nonbinary or diverse, or indicating to be unsure about their gender identity. For more information about the study design and the nature of krisenchat, we referred to the initial evaluation study of krisenchat [20]. Linguistic variables were determined using Linguistic Inquiry and Word Count (LIWC; see below for details).

The sample examined in this study was based on the previous evaluation of krisenchat, in which the sample consisted of those who completed a subsequent feedback survey after the counseling session [20]. Thus, out of a total of 11,031 users in the above-mentioned time period, 6962 (63.1%) completed a feedback survey. The chat messages of these 6962 users were analyzed. In total, 661,131 messages (mean 94.96, SD 259.46) from 26,614 chat sessions (mean 3.82, SD 6.24) with a total word count of 8,872,154 (mean 1274.37, SD 2954.57) were analyzed.

The Medical Faculty of the University of Leipzig approved this study on August 3, 2021 (372/21-ek). Users were informed about the data protection and privacy policy of krisenchat when they first contacted the counseling service. The chat counseling only began after confirming the policy with “Yes.” Participants in the study confirmed informed consent via an opt-in function before taking part in the feedback survey.

Statistical analyses were performed using IBM SPSS Statistics version 27.0 (IBM Corp). A 2-tailed α value of .05 was applied to statistical testing. First, descriptive statistics were performed for sociodemographic variables and linguistic variables of the total sample. Additionally, Kruskal-Wallis H tests (because of nonnormality of the linguistic variables) were used to identify gender differences in use, that is, metadata for the number of sessions, messages, and words of each user. Then, a 1-way multivariate analysis of variance (1-way MANOVA) was conducted to test for gender differences in language usage controlling for word count. Gender was considered as an independent variable, and all 7 linguistic variables (ie, first-person singular pronouns, first-person plural pronouns, negations, positive emotion words, negative emotion words, insight words, and causation words) were considered as dependent variables. Post-hoc univariate ANOVAs were conducted separately for every linguistic variable. Bonferroni correction was applied to account for multiple testing. Then, binary logistic regression analysis was conducted to examine the predictive effect of linguistic variables (first-person singular and plural pronouns, negations, positive and negative emotion words, insight words, and causation words), age, and gender (recoded into a set of dummy variables with “male” as the reference variable) on the presence of psychiatric symptoms. The amount of explained variance as shown by Nagelkerke R² was interpreted as follows: R² >0.20, “acceptable” or small effect size; R² >0.40, “good” or average effect size; and R² >0.50, “very good” or large effect size [55]. Additionally, Spearman correlation coefficients (ρ) were reported between linguistic variables. Finally, with the aim to examine the deeper relationship between linguistic variables and psychiatric symptoms, explorative Spearman correlations (ρ) between all 7 linguistic variables and all categories of psychiatric symptoms (suicidality, self-harm, depression, anxiety, eating disorder symptoms, flashbacks, obsessive-compulsive symptoms, and addictive behavior) were computed and interpreted as follows: ρ=0.10, small effect size; ρ=0.30, moderate effect size; and ρ=0.50, large effect size [56].

The average user was 17 years old (mean 16.55, SD 3.45 years; range 8-25 years), and most users were female (female: 4988/5978, 83.4%; male: 881/5978, 14.7%; diverse: 109/5978, 1.8%). A large number of all users (4841/6962, 69.5%) contacted the counseling service due to psychiatric symptoms. Further concerns identified were psychosocial distress (eg, school-related problems, family-related problems, bullying, etc; 2370/6962, 34.0%) or emotional distress (eg, grief, lovesickness, anger, and loneliness; 2101/6962, 30.2%) [20].

The users participated in an average of 3.82 (SD 6.24) counseling sessions and sent an average of 94.96 (SD 259.46; range 2-11,512) messages with an average of 1274.37 (SD 2954.57) words throughout the counseling process. Additional testing indicated that there were gender differences in the numbers of sessions (χ²₂=22.849; P<.001), messages (χ²₂=14.863; P<.001), and words (χ²₂=33.036; P<.001). The results are presented in Table 1. Subsequent post-hoc tests indicated that female users attended a significantly higher number of sessions (z=−4.211; P<.001; r=0.05) and sent significantly more messages (z=−3.247; P<.001; r=0.04) and words (z=−5.349; P<.001; r=0.07) than male users, whereas there were no significant differences between female users and users identifying as diverse (number of sessions: P=.13; number of messages: P=.18; number of words: P=.22). Users identifying as diverse also attended a significantly higher number of sessions (z=−3.441; P<.001; r=0.04) and sent significantly more messages (z=−2.972; P<.001; r=0.04) and words (z=−3.639; P<.001; r=0.05) than male users.

In the total sample, the mean percentage of first-person singular pronouns among all words of a user during the whole counseling process was 11.59% (SD 2.46%), indicating that on average, more than one-tenth of all words written throughout all chat messages was a first-person singular pronoun (“I,” “me,” “my,” or “mine”). The next most used linguistic categories were positive emotion words (mean 4.85%, SD 1.70%) and insight words (mean 4.05%, SD 1.32%). The mean percentage of negations among all words of a user during the whole counseling process was 3.76% (SD 1.54%). Furthermore, negative emotion words were used with a mean percentage among all words of 3.23% (SD 1.27%). Causation words were used with a mean percentage among all words of 2.93% (SD 1.04%). Finally, first-person plural pronouns were least frequently used with a mean percentage among all words of 0.43% (SD 0.63%).

Gender-specific descriptive statistics are displayed in Table 2. A 1-way MANOVA showed statistically significant differences in linguistic variables between genders (F_{14, 11,932}=8.945; P<.001; partial η²=0.01; Wilk Λ=0.979). Post-hoc univariate ANOVAs were conducted separately for every linguistic variable. Separate ANOVAs and respective Bonferroni-corrected post-hoc tests showed that when controlling for word count, there were statistically significant differences in the use of first-person singular pronouns between genders (F_{2, 5972}=49.780; P<.001; partial η²=0.02), with female users (mean difference=0.90, 95% CI 0.68-1.12; P<.001) and users identifying as diverse (mean difference=0.98, 95% CI 0.37-1.56; P<.001) using more first-person singular pronouns than male users, whereas there was no significant difference between female users and users identifying as diverse (mean difference=0.08, 95% CI −0.66 to 0.50; P>.99). Furthermore, the use of negations differed significantly between genders (F_{2, 5972}=4.915; P=.007; partial η²=0.002), with female users using significantly more negations than male users (mean difference=−0.16, 95% CI 0.03-0.30; P=.01), while there were no significant differences in the use of negations between male users and users identifying as diverse (mean difference=0.329, 95% CI −0.05 to 0.71; P=.12) and between female users and users identifying as diverse (mean difference=0.17, 95% CI −0.20 to 0.53; P=.84). Another significant difference was found in the use of negative emotion words between genders (F_{2, 5972}=4.505; P=.01; partial η²=0.00), with female users using significantly more negative emotion words than male users (mean difference=0.12, 95% CI 0.01-0.23; P=.04; not significant after Bonferroni correction), while no significant differences in the use of negative emotion words were found between female users and users identifying as diverse (mean difference=0.22, 95% CI −0.07 to 0.52; P=.22) and between male users and users identifying as diverse (mean difference=0.11, 95% CI −0.42 to 0.21; P>.99). Finally, the results showed a significant difference in the use of insight words between genders (F_{2, 5972}=15.215; P<.001; partial η²=0.01), with female users (mean difference=0.26, 95% CI 0.14-0.37; P<.001) and users identifying as diverse (mean difference=0.41, 95% CI −0.15 to 0.47; P=.007) using significantly more insight words than male users, while no significant differences were found between female users and users identifying as diverse (mean difference=0.16, 95% CI −0.42 to 0.21; P=.67). No overall significant differences were found between genders in the use of first-person plural pronouns (F_{2, 5972}=3.006; P=.05; partial η²=0.00), positive emotion words (F_{2, 5972}=0.489; P=.61; partial η²=0.00), and causation words (F_{2, 5972}=2.434; P=.09; partial η²=0.00).

The binomial logistic regression model was statistically significant (χ²₈=25.0; P=.002), resulting in a small amount of explained variance, as shown by Nagelkerke R²=0.124 (Table 3). Of the 10 variables entered into the regression model, all but 3 contributed significantly to the presence of psychiatric symptoms: first-person singular and plural pronouns, negations, negative emotion words, causation words (all P<.001), and female gender (P=.005), while positive emotion words (P=.08), insight words (P=.90), and diverse gender (P=.57) showed no significant effects. Using first-person plural pronouns was associated with a lower likelihood of reporting psychiatric symptoms (odds ratio [OR] 0.39), as did using more causation words (OR 0.90). In contrast, a higher use of first-person singular pronouns was associated with an increased likelihood of reporting psychiatric symptoms (OR 1.05), as did using more negations (OR 1.11) or negative emotion words (OR 1.15). Finally, being female (OR 1.18) or having a higher age (OR 1.04) was also associated with an increased likelihood of the presence of psychiatric symptoms.

The results of an additional correlation analysis between all linguistic variables are reported in Table 4. Among others, significant findings included a negative association between first-person singular pronouns and first-person plural pronouns (ρ=−0.24; P<.001). In line with this, first-person singular pronouns were positively correlated with negations (ρ=0.17; P<.001) and negative emotion words (ρ=0.15; P<.001).

Finally, an exploratory correlation analysis indicated evidence for the relationship between linguistic variables and various psychiatric symptoms, which are displayed in Table 5. Suicidality, self-harm, depression, and anxiety showed the most significant correlations to linguistic variables. In particular, the use of first-person singular pronouns was positively associated with suicidality (ρ=0.11; P<.001) and self-harm (ρ=0.10; P<.001). The use of first-person plural pronouns was negatively associated with suicidality (ρ=−0.10; P<.001) and depression (ρ=−0.14; P<.001). The use of negations was positively associated with suicidality (ρ=0.18; P<.001) and self-harm (ρ=0.12; P<.001). Finally, the use of negative emotion words was positively associated with depression (ρ=0.12; P<.001) and anxiety (ρ=0.15; P<.001).

The findings of this study provide first-time valuable insights into the psycholinguistic characteristics of children, adolescents, and young adults seeking psychosocial support through a messenger-based crisis counseling service (krisenchat). Previous findings that examined psycholinguistic characteristics in association with mental health, which however focused on texts (eg, from social media or online therapies), could be identified in the chat context as well. Specifically, linguistic variables were found to be associated with the presence of psychiatric symptoms [43,44,57]. The use of first-person singular pronouns, negations, and negative emotion words increased the likelihood of the presence of psychiatric symptoms, while the use of first-person plural pronouns and causation words was associated with a lower likelihood of the presence of psychiatric symptoms. Female gender was also associated with an increased likelihood of the presence of psychiatric symptoms, which is consistent with the higher prevalence of psychiatric symptoms in women [58]. Gender differences were found, with female users exhibiting more frequent use of certain linguistic features. Previous linguistic analyses, especially in the digital context, such as social media platforms, have been performed with a focus on the presence of psychiatric symptoms or associations with psychiatric symptoms, in particular, depressive symptoms [59]. Linguistic analyses of social media have proven useful in predicting depression, anxiety, loneliness, personality disorders, or other mental health issues [43,57,60-62]. As there are no other comparative studies in this field, the focus of the below comparison of the present results with previous findings relies on correlates of linguistic variables with the presence of psychiatric symptoms, especially depression or anxiety.

Multiple implications for psychosocial chat counseling have emerged from the findings. First, counselors may use psycholinguistic analyses as an additional tool for assessing the mental health state of users. By monitoring the usage of specific linguistic features, counselors can identify individuals at risk for psychiatric symptoms and tailor interventions accordingly. For example, in cases of elevated use of first-person singular pronouns, negations, and negative emotion words, particularly for female users and users who identify as diverse, a more in-depth exploration of psychiatric symptoms, especially depression, anxiety, or suicidality, is recommended. Regardless of the actual cause, it seems advisable to monitor the frequency of first-person singular pronouns throughout the chat history. For example, establishing a word-counting function during chat counseling may provide counselors with additional valuable information. In turn, increased use of first-person plural pronouns as well as causation words can be identified as protective factors, and also expanded and used as such. For example, increased use of first-person plural pronouns on the counselor’s side might be helpful to create a sense of “unity” or belongingness. Taking social belongingness into account as a protective factor for depression, suicidality, or mental disorders in general, it seems advisable to create closeness or sociality through language to initially stabilize the user emotionally, for example, meet their need for attachment or create a base to make further recommendations [81].

Furthermore, it can be helpful to mirror statements of users in the context of active listening (positively paraphrased) or to guide the users through targeted questioning techniques into positively formulated thoughts, goals, behavioral directions, etc (ie, in concrete terms) to avoid negations. This implication is also underscored by the fact that increased use of words referring to expectation, trust, and belongingness was associated with lower depression rates [80,81]. Furthermore, the establishment of logical chains, in the sense of causation, also appears to be of great importance. Thus, attempts should be made to bring concerns, feelings, etc into a logical coherent context in the sense of stimulus-reaction chains. Moreover, the use of insight words shows no influence on the presence of psychiatric symptoms, yet reflective functioning is known to be a protective factor [82]. An emphasis on self-reflection could potentially be integrated into psychoeducation by supporting an understanding of one’s own mind as an aspect of resilience and personal agency that can be fostered through social support and professional help [82]. In practice, repeated questions about what feeling is associated with particular problems or concerns, or identifying thoughts or cognition in terms of one’s own belief patterns, can help to practice such skills.

To the best of our knowledge, this study is the first to examine language usage in a messenger-based crisis counseling service among youth and young adults. In addition to the strength of the large sample size, this study acts as a reference and comparison for further studies in this area, in part because of the use of the internationally recognized method LIWC. Nevertheless, some limitations have to be taken into account. Owing to the retrospective study design, the data rely on convenience sampling, which limits the generalizability to a more mixed-gender population. It should also be noted that the counseling service is offered in German-speaking countries, which is why cultural and linguistic differences in language usage must be taken into account, and generalizability is limited. An international comparison between counseling services in different countries, cultures, and languages would provide insights into similarities and differences. Assuming that people in crisis reach out to the chat counseling service, high use of emotion words seems somewhat expected. Since the nature of the study was cross-sectional and the words were counted across all chat messages (ie, not in chronological order or within sessions), no data can be provided on the trend of the word count for either positive or negative emotion words, which is why no indications can be derived on whether the word count relates only to the beginning of the chat or also to the progression throughout the chat counseling. Longitudinal studies examining changes in language use across consecutive chat sessions may provide further insights into these associations. Consideration must also be given to the nested data structure, which cannot be clearly read owing to the format used to provide data by LIWC. Therefore, for future studies, in addition to the trend of the word count throughout the chat counseling, the consideration of levels (eg, within a message and during a session) is recommended. Therefore, a qualitative analysis of the chats may be beneficial for providing more in-depth insights into individual language usage as well as concerns. A qualitative approach would also ensure the quality of the classification of the presence of psychiatric symptoms. Likewise, no standardized measurement instruments were used, which in turn opens up further opportunities for future research, for example, implementing symptom-specific questionnaires to examine associations between (changes in) symptom severity and language usage.

This study underlines the options, possibilities, and chances of examining psycholinguistic characteristics in psychosocial online chat counseling services for children and adolescents. The identified associations between specific linguistic features and the presence of psychiatric symptoms provide valuable insights for the development of targeted interventions. By considering psycholinguistic findings in the counseling practice, counselors may enhance their understanding of the psychological processes of users and their interventions to offer a more targeted service for children and adolescents seeking help. Nevertheless, further research is needed to investigate the mechanisms underlying linguistic patterns and explore the effectiveness of linguistic-based interventions. At the same time, this would allow further research on the link between specific indicators and changes in specific psychopathology.