Both voluntary and involuntary terminations are costly and disruptive for organizations. When employees willingly leave their positions, companies lose valuable talent and expertise, compounded by the need for expensive recruitment and onboarding processes to fill the vacant roles. Departures may also harm team morale and productivity, delaying projects and compromising overall organizational performance.
Prior research on voluntary turnover often takes a one-size-fits-all approach that models quitting irrespective of important employee characteristics such as personality, thinking styles, or emotional predispositions. As a social-personality psychologist, I am interested in how individual differences such as personality traits, gender, or thinking styles influence how people behave in response to situational factors, such as quitting a job or coping with COVID.
Whereas personality traits (such as those in the Big Five and DISC) are characteristic patterns of thoughts, feelings, and behaviors that describe how people interact with the world around them, thinking styles reflect how people process and internalize information in their environments (reflected in varying levels of analytical thinking, self-focus, focus on others, etc.). Thinking styles are trait-like, but they are not identical to personality. For example, a self-focused thinking style (reflected in higher rates of first-person singular pronouns I, me, and my) is characteristic of some aspects of neuroticism and linked with mental health conditions like depression and suicidality. (For more information on the complex role of first-person singular pronoun use in psychology, see Berry-Blunt and colleagues’ recent paper.)
A greater understanding of how people’s thinking styles impact their emotions and behaviors before and after leaving a job may equip I/O psychologists and HR professionals with predictive indicators that can help prevent or mitigate the organizational fallout of quitting. To explore individual differences in negative emotional language surrounding leaving a job, we collected a 150 million-word corpus of Reddit messages (posts and comments). The sample includes about 15,000 users (half of who talked about resigning or being fired from a job on a specific date, and the other half being comparison users who did not discuss quitting) written between 2010 and 2023. The longitudinal design allows us to examine language both in the weeks immediately surrounding job changes as well as habitual language styles established months before the target event (i.e., baseline language use). In our first published results using the corpus, we set out to answer a few basic questions:
How does planning to quit a job and then carrying out that plan change how people use emotional language in everyday social media messages–and how do those patterns differ across pre-pandemic and COVID eras, if at all?
How does the person interact with the situation in these cases? That is, do the emotions associated with planning and carrying out a career change differ over time for people who habitually use high versus low levels of language associated with mental health risks?
To begin looking into our own questions about job changes and to provide a resource for other researchers, we compiled the Reddit Job Change Corpus, a dataset we have made publicly available. The corpus includes all of the messages (posts and comments across all forums under the same username) from both people who reported quitting or being fired on a particular day and matched comparison users. The corpus and its code have a few key attributes:
Temporal granularity: Each user's sample of messages is anchored by a specific day on which they reported quitting their job. The corpus includes all lifetime Reddit messages from all users, allowing for fine-grained temporal analyses (spanning several years for some users).
The job change algorithm: The algorithm used to find people talking about job changes, devised by Micah Iserman, is itself useful for future research. We first found people who talked about quitting or firing on a specific day on Reddit and then used dependency checks to filter out users who talked about resigning in non-literal (e.g., hypothetical, ironic, vicarious, or fictional) ways. The code can be reused in other language samples or adapted to identify the timing of other major life events based on social media messages.
Matched comparisons: The corpus provides a matched comparison sample that pairs every user in the target set (people who quit or were fired from their jobs on a specific day) with a user who had similar posting patterns (in terms of where, when, and how verbosely they posted) but did not talk about leaving their job on Reddit.
Language Psychology Measures
For our initial language analyses, we focused on anxious and sad language in the year before and after quitting a job, excluding the messages about quitting themselves. We measured anxiety language (e.g., worried, panic) using LIWC and sadness language (e.g., grieving, depressed) using SALLEE. Both measures are based on dictionaries, or word lists, although SALLEE has additional syntax-aware capabilities that take into account the words and punctuation surrounding emotional terms.
To model individual differences in thinking styles, we used LIWC to measure two linguistic variables that tend to reflect self-focused rumination when used in social contexts like Reddit: cognitive processing language (e.g., think, realize) and first-person singular pronouns (e.g., I, me, my). Self-references like I and me are thought to reflect vulnerability to stress and may indicate less-effective ways of coping with distress. Cognitive language on social media similarly seems to indicate a predisposition to depression, perhaps because it is associated with rumination (i.e., repetitive negative thoughts, especially about the self). Using self-references or cognitive processing language at high rates might indicate that a person’s default way of thinking about the world sets them up to deal with major life changes in a distressing way.
We found that people who quit their jobs used more sadness and anxiety language than neutral comparison users in the months before and after quitting. However, those effects showed up primarily for people whose messages were highly self-focused or used high rates of cognitive processing language at baseline (averaged across messages posted 3-12 months before quitting).
The results are consistent with self-distancing research showing that using fewer "I" pronouns helps people cope with stressful or distressing experiences. The findings also dovetail with evidence that using cognitive language at high rates in social contexts (such as social media messages) can reflect mental health vulnerabilities.
Negative emotional language patterns were, for the most part, parallel between pandemic-era and pre-COVID quitting. The main difference between the two eras was that Reddit messages for both people who quit their jobs and comparison users included more sad and anxious language during versus before the pandemic, reflecting the global trauma caused by COVID.
Conclusion and Applications
We expected people who quit their jobs to show more negative emotion (especially sadness and anxiety language) in the months surrounding quitting. The results showed that that was true–but only for people who seem to have habitually self-focused or ruminative thinking styles. As such, an early indicator of employees who may be having difficulty coping and are contemplating quitting may be language that combines increased self-focus with signs of sadness and anxiety. We also found that people who quit their jobs during COVID used more negative emotional language than earlier Reddit users but otherwise followed the same approximate emotional trajectories before and after quitting.
The most fundamental lesson of findings like this is that individual differences in writing and thinking styles matter. People often simplify their social worlds by assuming that actions are caused primarily by either the situation or the person. (For example, you cut someone off in traffic because the exit signs weren't clear; other people cut you off because they're morally bankrupt.) Yet, regardless of the situation, people rarely behave uniformly. Humans are probabilistic and not perfectly internally consistent; they also bring their personal histories and dispositions with them to every social situation. Along the same lines, although quitting a job is a major life stressor for almost everyone, our results show that not everyone copes with the stress of planning and executing a job change by talking about anxiety and sadness before or after resigning.
From the point of view of employers, the message may be that it’s the quiet ones you have to observe more carefully or creatively when trying to monitor workforce wellness or predict voluntary turnover. The people in this sample who used self-focused, ruminative language showed relatively clear emotional signs that they were thinking about quitting leading up to their resignations. Employers who observe those signs could intervene (e.g., by checking in with them, providing mental health support, or helping them arrange leave) before the employee leaves for good. For people who don’t talk about themselves or their thoughts much in social contexts, it may be necessary to look for other non-emotional indicators of an impending resignation (such as decreased linguistic synchrony with coworkers in workplace communications).
In other words, accurately predicting and mitigating voluntary turnover requires understanding the psychology of employees. By using text analysis, it’s possible to unobtrusively evaluate reactions to situational factors (e.g., how working from home is affecting job satisfaction) as well as gain insights into who employees are as people, in and outside of work (e.g., whether they’re coping with major life stressors or are on the brink of a mental health crisis). Employers can use both kinds of information to proactively assist employees in navigating work and life challenges.
The findings summarized above are based on analyses of Reddit messages, but this approach can be applied to any texts written by employees, including responses to open-ended survey questions or reviews on Glassdoor. Linguistic markers of thinking styles are largely independent of contexts or topics and can therefore provide practical, actionable information across diverse sources of language data and situations.
Using the Corpus
If you’d like to learn more about our paper from the WASSA workshop or the corpus in general, the data and code are available on the Open Science Framework. The results that we describe in the paper only hint at the possibilities offered by this dataset. We would be happy to hear from you if you have any questions. If you would like to use the dataset in a published work, please cite the following paper:
Ireland, M. E., Iserman, M., & Adams, K. R. (2023, July). Sadness and Anxiety Language in Reddit Messages Before and After Quitting a Job. In Proceedings of the 13th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis (pp. 467-478). Available from https://aclanthology.org/2023.wassa-1.41.pdf