Predicting real-world emotion and health from spontaneously-assessed linguistic distancing using novel scalable technology

Abstract

Individuals who are better at regulating their emotions have been shown to have better physical and mental health outcomes. One promising emotion regulation strategy is psychological distancing, which involves appraising a stimulus with objectivity or spatial/temporal distance. Language-based psychological distancing (linguistic distancing (LD)) refers to the degree to which one implements psychological distancing naturally via language. A crucial, underexamined mechanism that may account for real-world emotion and health self-reports is spontaneous (i.e., implicit) LD. Using HealthSense, a novel, scalable, mobile health assessment application, we collected lexical transcriptions for personally-specific negative and positive events as well as emotion and health-relevant data over fourteen days (data collected in 2021) and examined how implicit LD during negative and positive events relates to well-being over time. Primary analyses revealed that higher LD during negative events was associated with lower levels of stress as well as greater emotional and physical well-being within-persons. LD during positive events on one day predicted greater reports of happiness two days later within-persons. LD during positive events was associated with fewer symptoms of depression and LD during negative events was associated with greater physical well-being between-persons. Exploratory analyses revealed that average depression, rumination, and perceived stress across the two weeks were significantly negatively associated with LD during negative events between-persons. The present results expand understanding of the relationship between LD and mental and physical health risks and motivate future research on low-burden, scalable interventions involving LD.

Introduction

Extensive research has indicated that individuals who are better at regulating their emotions have better physical and mental health outcomes (Aldao et al., 2010; Florin et al., 1985; Greer & Watson, 1985; Tianqiang Hu et al., 2014; Verzeletti et al., 2016). Indeed, positive emotion regulation is associated with fewer symptoms of depression, anxiety, physical pain, as well as enhanced social functioning, immune functioning, and general well-being (Aldao et al., 2010; Appleton et al., 2014; Cisler et al., 2010; Conley et al., 2016; Dryman & Heimberg, 2018; English et al., 2012). Emotion dysregulation has also been associated with increased risk for cardiovascular disease (Appleton & Kubzansky, 2014; Berna et al., 2014).

One particularly promising and well-studied emotion regulatory strategy is cognitive reappraisal, which involves changing one’s cognitive construal of an emotionally evocative stimulus to alter its emotional impact on behavior, experience, and/or physiology (Gross, 1998, 2015). Individuals who use cognitive reappraisal habitually tend to show positive associations with emotion and health indicators, including fewer depressive symptoms, greater self-esteem, and greater life satisfaction (Gross, 2013; Gross & John, 2003; John & Gross, 2004; Troy et al., 2010). Psychological distancing is a form of cognitive reappraisal that can be accomplished via changing one’s appraisal of an emotional stimulus by employing an objective and impartial perspective or via increasing the perceived spatial and temporal distance between oneself and the stimulus (Kross et al., 2005; Trope & Liberman, 2010). Individuals who are longitudinally trained in psychological distancing have shown reductions in self-reported negative affect and perceived stress over time (Denny, 2020; Denny & Ochsner, 2014).

Language-based evidence of psychological distancing (i.e., linguistic distancing [LD]) has been used to study the degree to which one implements psychological distancing naturally via language. LD has been studied using a standardized composite metric of one’s frequency of first-person singular pronouns, articles, discrepancy words, words greater than six letters, and present-tense verbs. These five components have shown to track “verbal immediacy” (Pennebaker & King, 1999), and as a result, the composite has been used in research related to psychological distancing (Cohn et al., 2004; Mehl et al., 2013; Nook et al., 2017, 2019; Shahane, Fagundes, et al., 2021; Shahane, Pham, et al., 2021; Shahane & Denny, 2019). First-person singular pronouns (e.g., I, my) suggest that the writing is tightly intertwined with the self, hence why this is reverse scored. Next, present-tense verbs also suggest the writing is focused on the here and now, hence why these are also reverse scored. Articles (e.g., the, an) signal objectivity—a key tactic of psychological distancing. Similarly, discrepancy words (e.g., would, should) suggest a lack of certainty or (e.g., if you compare “I should do this” to “I did this,” the latter implies more objectivity), hence why discrepancy words are reverse scored. The reason why words greater than six letters is part of the composite stems back to Information Theory (Mahowald et al., 2013; Piantadosi et al., 2011; Shannon & Weaver, 1998). Per Information Theory, longer words represent greater abstraction than shorter words (Mahowald et al., 2013; Piantadosi et al., 2011; Shannon & Weaver, 1998), and more abstract representations are associated with a greater sense of distance (Powers & LaBar, 2018). In addition, when sequences are more prolonged, entropy decreases, and English becomes more predictable as longer words, in general, carry more information (i.e., longer words are generally more specific, thereby implying greater objectivity) (Mahowald et al., 2013; Piantadosi et al., 2011).

LD has also been associated with greater emotion regulation efficacy (Nook et al., 2017) and fewer symptoms of depression, perceived stress, and difficulties in emotion regulation, and better emotional well-being, energy and vitality, social functioning, and reappraisal frequency (Shahane & Denny, 2019). Thus, LD may mitigate potentially deleterious effects of stressful life experiences, and therefore may reduce risk of validated risk factors for cardiovascular disease, such as depression and anxiety (Giltay et al., 2004; Sheps & Sheffield, 2001).

The relationship between LD and physical and mental health has been examined in the context of negative stimuli in a laboratory-based context (Nook et al., 2017; Shahane & Denny, 2019; Shahane, Pham, Lopez, & Denny, 2021; Shahane, Fagundes, Buka, & Denny, 2021). For example, we recently conducted a study in which 94 healthy young adults transcribed their thoughts while viewing negative or neutral stimuli in one of three ways: 1) by implementing objective language, 2) by implementing spatially and/or temporally far away language, 3) or by responding naturally (Shahane & Denny, 2019). Across the psychological distancing groups, we found that LD, as indexed with the standardized composite metric described by Nook and colleagues (2017), was associated with fewer symptoms of depression, perceived stress, and difficulties in emotion regulation, and better emotional well-being, energy and vitality, social functioning, and reappraisal frequency (Shahane & Denny, 2019). More recently, we found that when participants were manipulated to use objective or spatially or temporally distant language over a five-day intervention, increases in LD as assessed via novel computational algorithms from day 1 to day 5 were associated with greater increases in reappraisal success (Shahane, Pham, Lopez, & Denny, 2021), and also found a significant negative association between proinflammatory cytokines (e.g., IL-6, TNF-alpha, IFN-gamma)—markers of systemic inflammation associated with a number of diseases—and LD (Shahane et al., in press). Although these studies investigate the relationship between LD and mental and physical health over time, they differ from the present study because they were conducted within a laboratory context. Thus, a crucial, underexamined mechanism that may account for real-world emotion and health self-reports is natural language-based assessments of LD.

Thus, we were motivated to examine the relationship between LD and emotion and health indicators in a real-world, ecologically-valid context. To do so, we utilized HealthSense (Curtis et al., 2019). HealthSense is a novel, scalable, mobile health assessment application developed by engineers in the Department of Electrical and Computer Engineering at Rice University. Using HealthSense, we were able to study spontaneous real-world LD (i.e., without any instruction). Thus, the overarching goal was to leverage computational text-reading algorithms to provide insight into emotion regulatory constructs representing a promising prognostic indicator of health impacts, including mental health and cardiovascular disease risk. More specifically, we aimed to examine how spontaneous LD during negative and positive real world events relates to well-being, which was measured with ecological momentary assessments (EMAs) using multilevel models. In concert, we examined if any variance in LD predicted fluctuations in day-to-day health across negative and positive real world events over the course of two weeks. In addition, we aimed to examine if one’s affective tendencies (e.g., general tendencies to ruminate, suppress emotions, reappraise), which were measured 3 times across the two weeks, were associated with negative and positive event LD.

At present, to our knowledge no studies have examined spontaneous LD in the real world to predict mental health indicators using novel, scalable technology. In addition, the EMAs used in the present work provide ecological validity. By having participants produce the stimuli themselves with their personally-specific real-world events, we are able to examine how LD associates with authentic life situations. The participant-derived stimuli are packed with more personal meaning, and this ecological relevance may therefore reveal subtleties that are otherwise more difficult to detect with lab developed stimuli. While conditions are often tightly controlled in a lab setting, the EMAs allow us to study microprocesses that influence behavior in real-world contexts (Shiffman et al., 2008). We aimed to implement an implicit LD assessment that can parse text unobtrusively and feasibly as participants underwent EMA over a two week period. Consistent with results from Shahane & Denny (2019) that examined LD in an experimental, laboratory-based context, we predicted that spontaneous, implicitly-assessed LD would be associated with fewer symptoms of depression and perceived stress and better general well-being during negative situations experienced by participants.

Method

Participants

The total N size varied depending on the model; there were 671–714 total observations across 183–187 individuals¹ for the primary analyses (49% female, mean age = 33.95, standard deviation age = 10.52). Participants were recruited via the Prolific platform in order to have a large and diverse nationally representative United States-based sample (African American = 8%, Asian = 11%, Caucasian = 66%, Hispanic/Latinx = 6%, Native American = 0%, Biracial/Multiple Races = 7%, Other = 2%). The entire study duration was roughly 65 minutes (inclusive of all the event descriptions, EMAs, and affective tendency measures), providing participants a flat rate of $10.84. A minimum sample size of 134 yields 95% power of detecting a correlation of .30 (medium to large effect size (Cohen, 1992; Funder & Ozer, 2019)). Thus, more than 134 participants for the recruitment target were deemed sufficient to achieve 95% power based on a priori power analyses computed with G*Power (Faul et al., 2009). There were two inclusion criteria that participants had to meet in order to participate: Participants had to be at least 18 years old and also have an Android-based smartphone, which was needed for compatibility with the HealthSense app. Participants were given a full overview of the study before signing up. After signing up on Prolific, participants completed an informed consent form electronically on Qualtrics. Once participants provided their consent, participants were instructed on how to download the HealthSense app from the Google Play Store app on Android and how to register for the study using a secure unique code.

Procedure and Measures

Over two weeks, we collected health and LD data using HealthSense (Figure 1). On day one, participants first completed a series of questionnaires to assess affective tendencies—one’s general emotion tendencies: the Emotion Regulation Questionnaire (Gross & John, 2003), which assesses how frequently participants report regulating emotions via reappraisal or expressive suppression, the Rumination Response Scale (Treynor et al., 2003), which assess the degree to which participants ruminate, the Center for Epidemiologic Studies Depression Scale (CES-D; (Radloff, 1977), which assesses depressive symptomology, and the Perceived Stress Scale (PSS; (Cohen et al., 1983), which assesses perceived stress. These affective tendency questionnaires were repeated on day eight and day fifteen (Figure 1).

Figure 1.

Study schematic. Affective Tendency was measured three times on Day 1, Day 8, and Day 15. LD at negative and positive events as well as emotion and health indicators via EMAs were measured every other day starting on Day 2.

Starting on day two, participants received a ping at 4:00 PM CT every other day. First, we asked participants to describe a negative event that happened in the day as well as a positive event that happened in the day. Participants were asked to write about both positive and negative events on day two, four, six, eight, ten, twelve, and fourteen (Figure 1). The ordering of the prompts (negative description first versus positive description first) were counterbalanced each day. For example, participants could write about acing an exam for their positive event and then breaking up with their significant other for their negative event.

Next, we asked six ecological momentary assessment (EMA) questions via HealthSense to assess their health and well-being at 6:00 PM on day two, four, six, eight, ten, twelve, and fourteen (Figure 1). These questions assessed optimism/happiness, stress, depression, anxiety, general mental health, and general physical health (i.e., “How happy do you feel right now?” “How stressed do you feel right now?” “How depressed do you feel right now?” “How anxious do you feel right now?” “How do you feel emotionally right now?” “How do you feel physically right now?”) using a sliding scale of 1 (e.g., Not stressed at all) to 100 (Extremely stressed). These EMAs were modelled after mood EMA structures described in Hall and colleagues (2021) as well as previous internal studies that leveraged the HealthSense application (Hall et al., 2021). For example, as Hall and colleagues recommended, we took temporal considerations into account like including “right now” to avoid ambiguous time frames, and followed the standard 0 to 100 scale structure, which was cited as the most common response scale for EMAs. The “How [insert measure] do you feel right now” question stem was used after taking the average simplest approach among those discussed in Hall and colleagues’ systematic review, and we formatted consistently for all the EMAs to standardize. All procedures were approved by the Rice University Institutional Review Board. Data were collected in 2021.

LD Calculation

To quantify the degree of LD employed by each participant on each day, we used Pennebaker’s Linguistic Inquiry and Word Count (LIWC; Pennebaker et al., 2015), which computes percentages of words falling within particular word categories. The composite psychological distancing score was be calculated by z-scoring first-person singular pronouns (e.g., I, my), present-tense verbs, articles (e.g., the, a), discrepancy words (e.g., would, should), and words of more than six letters, which were then averaged as indicated below. The z-scored frequencies of first-person singular pronouns, present-tense verbs, and discrepancy words will be reverse-scored (multiplied by −1) and then averaged with the z-scored frequencies of articles and words of more than six letters (Cohn et al., 2004; Mehl et al., 2013; Nook et al., 2017; Pennebaker & King, 1999; Shahane & Denny, 2019). Lower LD scores indicate personal and present-focused writing, whereas higher LD scores are reflective of greater psychological distance.

Data Analysis

Primary Analyses

In the current study, experience sampling methods were implemented via multi-level mixed modeling to examine within and between person associations between LD measures, emotion and health indicators (i.e., EMAs), and affective tendencies. All analyses were conducted in Stata 15.1 (Stata Corp, 2017). Multilevel modeling allowed for modeling of nested data structures with the use of random factors and distinguishing between-person and within-person fixed effects (Bolger & Laurenceau, 2013). Specifically, a random intercept allowed for the separation of variance at two levels; level 1, which represented the variance of an EMA outcome within a participant $\left({\epsilon }_j\right)$, and level 2, which represented the variance in EMA scores between participants $\left(v_{0j}\right)$.

At level 1, the random intercept $\left({\beta }_{0j}\right)$ represented the average level of the EMA measure for person j, and residual variance $\left({\epsilon }_j\right)$, which represents the degree to which a specific score for person j deviates from their average. Fixed effects at level 1 were person mean centered $(x-\underset{\_}{x})$ and represented fluctuations from participants’ usual level of LD on the same day as the outcome variable, such that positive values indicated higher LD and negative values indicated lower LD, relative to a participant’s typical level of LD. For the current analysis, all level 1 models included a lagged dependent variable predictor $\left({\beta }_{1j}\right)$, which represents the effect of the EMA from the previous time point (two days prior) for person j. We further included both negative event $\left({\beta }_{2j}\right)$ and positive event $\left({\beta }_{3j}\right)$ LD estimates for the concurrent time point. Additionally, we included a lagged measure for LD during negative events and positive events, which provided an estimate of temporal precedence and indicated if fluctuations in LD two days prior (i.e., the previous EMA time point) to an assessment predicted within-person variance in EMA outcomes above and beyond the associations of the EMA measurement and the LD on the concurrent day. Lastly, we included an autoregressive covariance matrix to control for correlations of errors from adjacent time points. Maximum likelihood estimation was used to handle missing data, as data were considered to be missing at random (Bolger & Laurenceau, 2013).

At level two of the analysis, there was an intercept that reflects the average of individual averages of EMA outcomes $\left({\gamma }_{00}\right)$, and a variance parameter $\left(v_{0j}\right)$ that indicated the degree in which person j’s average EMA outcome varied from the sample average of average EMA outcomes. Additionally, we included the fixed effects of negative event $\left({\gamma }_{01}\right)$ and positive event $\left({\gamma }_{02}\right)$ LD, which were calculated by averaging each participants’ level of LD across time points and represent general levels of LD for each participant. Level 2 fixed effects were grand mean centered. For all models, residuals were inspected to assure that they fit the assumptions for linear modeling. These regression models at level 1 and level 2 are summarized below.

Level 1:

$$\begin{gathered} EM{A}_{ij}={\beta }_{0j}+{\beta }_{1j}LaggedEMA+{\beta }_{2j}NegativeLD+{\beta }_{3j}PositiveLD \\ +{\beta }_{4j}LaggedNegativeLD+{\beta }_{5j}LaggedPositiveLD+{\epsilon }_j \end{gathered}$$

Level 2:

$${\beta }_{0j}={\gamma }_{00}+{\gamma }_{01}NegativeLD+{\gamma }_{02}PositiveLD+v_{0j}$$

Exploratory Analyses

As an exploratory aim, we assessed whether relationships among implicit LD and emotion and health indicators (i.e., EMAs) were predicted by affective tendencies—tendencies to reappraise, suppress, perceive stress, have depressed affect, or ruminate (measured at three weekly assessments (day 1, day 8, and day 15 of the study)). Therefore, each model for the exploratory aim included a within- and between-person fixed effect for affective tendency (with each affective tendency variable being assessed in individual models), a lagged within-person predictor variable for the person centered affective tendency (which determined if within-person fluctuations in affective tendency were predictive of the EMA outcome or LD measures beyond the same week associations of the affective tendency), and a lagged person centered dependent variable to control for stability effects. Notably, the lag in these models variables were lagged one week back as the affective tendency measures were only collected on days 1, 8 and 15. Because the independent variables for these analyses were only examined at weekly intervals, only weekly level variation is able to be estimated from these data and not daily variation, such as in the primary analysis. Therefore, only weekly level EMA and LD variability is examined resulting in fewer observations. Data and corresponding scripts for data analyses are available on the Open-Science Framework at the following link: https://osf.io/pdbj9/.

Missing Data and Adjustments

Within the current study, missing data was observed among participants. We found that participants had a mean of 70% (SD = 32%) of completing linguistic distancing measures, a mean of 68% (SD = 33%) of compliance in EMA measures, and a mean of 72% (SD = 30%) in effective tendency measures. As data is required to be missing completely at random in order for the missing values to be considered ignorable, we examined whether participants with complete data differed from participants with incomplete data. We first examined whether EMA measures differed among those with missing LD measures using a multivariate analysis of variance and found that the groups did not significantly differ in EMA measures (F(6,211) = .891, p = 0.502). For the exploratory analysis, we examined whether individuals with complete affective tendency data differed in comparison to those with incomplete affective tendency using a multivariate analysis of variance; we found that the groups did not significant in LD or EMA measures (F(8, 205) = 1.208, p = .416). As there were no significant differences in the outcome variables among those with and without complete data for both the primary and exploratory analysis, data was considered missing completely at random. Therefore, maximum likelihood estimates were used in the analysis, as this method does not provide biased results when data is missing completely at random (Black et al., 2011).

The current study did not make type I error inflation adjustments. Multiple tests were conducted and therefore the experimental-wise type I error likelihood was increased. In the present work, caution should be practiced when interpreting the significant fixed effects with p-values greater than .008, as these values would be null if any type I error inflations adjustments were made and are at greater risk of being observed by chance.

Results

Descriptives

Descriptive statistics for all study variables are presented in Table 1. Figure 2A plots raw within-person variability in positive event LD and Figure 2B plots raw within-person variability in negative event LD. The average length of each response was 43 words, which was sufficient to run the linguistic analyses.

Table 1.

Descriptive Statistics for Study Variables

		Mean	SD Between	SD Within	1	2	3	4	5	6	7	8	9	10	11	12	13
1	Negative LD	−0	0.22	0.41	0.155	0.026	0.019	−0.068*	−0.036	−0.091**	0.089**	0.078*	0.083	−0.063	−0.068	−0.0512	−0.109*
2	Positive LD	0.03	0.21	0.43	0.467**	0.122	−0.020	−0.001	0.025	−0.008	0.014	−0.020	0.015	0.024	−0.017	−0.063	−0.068
3	Happy EMA	0.42	0.22	0.15	0.143*	0.173*	0.502	−0.508**	−0.507**	−0.406**	0.619**	0.286**	0.024	0.003	−0.006	−0,092	−0,007
4	Stressed EMA	0.3	0.19	0.18	−0.1	−0.1	0.06	0.465	0.412**	0.531**	−0.477**	−0.218**	−0.066	0.059	0.037	0.131**	0.103*
5	Depressed EMA	0.23	0.18	0.16	−0.168*	−0.181**	−0.1	0.789**	0.560	0.413**	−0.454**	−0.226**	−0.056	0.054	0.045	0.063	0.043
6	Anxious EMA	0.28	0.19	0.17	−0.1	0	0.05	0.853**	0.722**	0.529	−0.432**	−0.161**	−0.108*	−0.001	−0.054	0.069	0.037
7	Emotional Well-Being EMA	0.41	0.21	0.16	0.11	0.13	0.922**	0.04	−0.135*	0.04	0.466	0.349**	0.046	−0.037	0.037	−0.062	0.039
8	Physical Well-Being EMA	0.37	0.2	0.14	0.162*	0.174*	0.757**	0.09	0	0.12	0.766**	0.51	0.054	−0.069	0.029	−0.105*	−0.052
9	Reappraisal	22.5	8.14	2.65	0.06	0.08	0.486**	0.11	0.01	0.1	0.512**	0.453**	0.599	0.069	0.057	−0.063	−0.002
10	Suppression	12.1	5.89	1.74	0.08	0.09	0.153*	0.326**	0.346**	0.252**	0.134	0.179**	0.507**	0.84	0.239**	0.227**	0.105*
11	Rumination	35.8	14.9	4.11	−0.1	−0.1	0.09	0.503**	0.578**	0.494**	0.1	0.1	0.530**	0.612**	0.843	0.416**	0.302***
12	Depression	14.2	10.4	3.74	−0.160*	−0.160*	−0.241**	0.505**	0.719**	0.484**	−0.245**	−0.168*	0.165*	0.442**	0.779**	0 .841	0.342**
13	Perceived Stress	12.8	7.38	3.59	−0.142*	−0.1	−0.143*	0.613**	0.660**	0.549**	−0.138*	−0.1	0.304**	0.475**	0.818**	0.844**	0.689

Note:

^*p < .05

^**p < .01

Figure 2.

Figure 2A plots raw within-person variability in positive event LD. Figure 2B plots raw within-person variability in negative event LD.

Primary Analyses

Results from the primary analysis are presented in Table 2. Overall, across all EMA measures, there was a significant level of random variance between participants and a significant level of residual variance within persons. Intraclass correlation (ICC) values for EMA outcomes ranged from .43 – .55. Across all EMAs, lagged dependent variables had a significant negative coefficient indicating that higher levels of EMAs on one day were predictive of decreases in that EMA two days later. Within each model, participants had a maximum of 6 observations per person, as the lagging technique did not allow for the assessment of variance at the first time point, and there were an average of 3.7 observations per person. Lastly, observations across models ranged from 671–714 across 183–187 people.

Table 2.

Mixed models for positive event and negative event linguistic distancing estimating EMA measures

	Happy		Stressed		Depressed		Anxious		Emotional Well-Being		Physical Well-Being
Within Fixed Effects	b (p-value)	CI	b (p-value)	CI	b (p-value)	CI	b (p-value)	CI	b (p-value)	CI	b (p-value)	CI
Intercept	0.585		0.386		0.294		0.361		0.569		0.510
Lagged DV	−0.168** (<.001)	[−0.249, −0.087]	−0.254** (<.001)	[− 0.335, −0.175]	−0.198** (<.001)	[−0.275, −0.121]	− 0.213** (<.001)	[−0.293, −0.133]	− 0.168** (<.001)	[−0.254, −0.083]	− 0.191** (<.001)	[−0.273, −0.110]
Positive Event LD	0.006 (.710)	[−0.039, 0.027]	>−0.001 (>.999)	[−0.039, 0.039]	0.025 (.136)	[−0.058, 0.008]	0.005 (.782)	[−0.031, 0.042]	−0.004 (.811)	[−0.040, 0.031]	0.007 (.659)	[−0.023, 0.037]
Negative Event LD	0.020 (.207)	[−0.011, 0.051]	−.050** (.008)	[− 0.087, −0.013]	−0.024 (.141)	[−0.056, 0.008]	−0.025 (.172)	[−0.059, 0.011]	0.036* (.041)	[0.001, 0.069]	0.031* (.034)	[0.002, 0.060]
Positive Event LD Lagged	0.040* (.022)	[0.006, 0.074]	0.010 (.636)	[−0.030, 0.049]	−0.008 (.648)	[−0.042, 0.026]	−0.008 (.694)	[−0.045, 0.030]	0.016 (.401)	[−0.021, 0.053]	0.014 (.384)	[−0.017, 0.045]
Negative Event LD Lagged	0.014 (.358)	[−0.016, 0.045]	0.013 (.459)	[−0.022, 0.049]	0.001 (.939)	[−0.030, 0.032]	0.015 (.408)	[−0.019, 0.049]	0.009 (.581)	[−0.024, 0.043]	0.022 (.120)	[−0.006, 0.050]
Between Fixed Effects	b	CI	b	CI	b	CI	b	CI	b	CI	b	CI
Positive Event LD	0.099 (.111)	[−0.023, 0.220]	−0.101 (.126)	[− 0.229, 0.028]	−0.149* (.028)	[− 0.281, − 0.016]	−0.072 (.308)	[−0.212, 0.067]	0.029 (.629)	[−0.090, 0.149]	0.020 (.770)	[−0.111, 0.151
Negative Event LD	0.098 (.101)	[−0.019, 0.215]	−0.058 (.361)	[−0.182, 0.066]	−0.106 (.104)	[−0.233, 0.022]	−0.058 (.396)	[−0.193, 0.076]	0.114† (.053)	[0.023, 0.251]	0.150* (.035)	[0.010, 0.262]
Random Factors		CI		CI		CI		CI		CI		CI
Residual (within)	0.026	[0.022, 0.031]	0.047	[0.037, 0.059]	0.029	[0.024, 0.035	0.048	[0.037, 0.063]	0.032	[0.026, 0.040]	0.026	[0.021, 0.031]
Variance (between)	0.025	[0.018, 0.034]	0.019	[0.010, 0.034]	0.031	[0.023, 0.042]	0.023	[0.013, 0.042]	0.021	[0.014, 0.032]	0.031	[0.022, 0.042]
Auto-regression CV	0.142	[−0.017, 0.294]	0.417	[0.257, 0.555]	0.259	[0.089, 0.414]	0.503	[0.338, 0.637]	0.266	[0.066, 0.446]	0.322	[−0.139, −0.484]

Note:

^*p< .05

^**p < .01

^†p< .10

When examining associations between negative event LD and emotion and health indicators (i.e. EMAs), we found that on days in which participants engaged in high levels of LD during negative events, they reported less stress and greater emotional and physical wellbeing, relative to days in which they engaged in less LD during negative events (b = .050, p = .008, 95% CI [0.087,0.013]; b = .036, p = .041, 95% CI [0.001, 0.070]; b = .031, p = .034, 95%CI [0.002, 0.060], respectively). There were no significant associations between lagged LD during negative events and EMA outcome measures. However, we found that on days in which participants engaged in higher levels of LD during positive events, they experienced greater happiness two days later (b = .040, p = .022, 95% CI [0.006, .074]). There were no other significant associations between lagged positive event LD and other EMA outcomes. There were no significant associations between positive event LD and same day EMA outcomes.

When assessing the between-person associations between LD and EMA outcomes, we found that average LD during positive events across the two weeks was negatively associated with general levels of reported depression (b = −0.149, p = .028, 95% CI [−0.281, −0.016]). Additionally, average participant LD during negative events across the two weeks was positively associated with general levels physical wellbeing (b = −0.136, p = .035, 95% CI [0.010, 0.262]). Average LD during positive and negative events were not significantly associated with any of the other EMA outcomes.

Exploratory Analyses

Throughout the models, all EMA outcomes and LD measures had a significant amount of variance between participants and significant amount of within-person variance within the empty models without predictor variables. Additionally, for the EMA measures, ICCs across the weekly time points ranged from .46 - .56. The ICCs for LD during negative events and LD during positive events were .12 and .15, respectively. These ICCs suggest that the variability of LD was heavily influenced by within-person fluctuations of LD relative to between-person. Within each model, participants had a maximum of two observations per person, as the lagging technique did not allow for the assessment of variance at the first time point, and there was an average of 1.6 observations per person with 192–231 total observations across 122–136 people.

Results from the exploratory analysis are described below. Reappraisal and suppression within- and between-person associations in EMA measures and LD can be found in the Supplemental Materials.

Rumination estimating within- and between-person associations in EMA measures and LD

Within participants, positive fluctuations in rumination were significantly associated with within-person decreases in same week happiness (b = −0.008, 95% CI [−0.013, −0.002], p = 0.008), as well within-person increases in same week depression and stress (b = 0.008, 95% CI [0.003, 0.013], p = 0.003; b = 0.009, 95% CI [0.002, 0.015], p = 0.009, respectively). Fluctuations in rumination in the same week were not significantly related to within-person changes in physical or emotional well-being, anxiety, and LD (ps > 0.05). There were no significant associations between lagged daily fluctuations in rumination and within-person changes in EMA outcomes or LD measures (ps > 0.05).

Between participants, greater average rumination across the two weeks was significantly negatively associated with general levels of emotional and physical well-being, happiness, and negative event LD (b = −0.007, 95% CI [−0.009, −0.004], p < 0.001; b = −0.005, 95% CI [−0.008, −0.003], p < 0.001; b = −0.005, 95% CI [−0.008, −0.002], p < 0.001, b = −0.007, 95% CI [−0.012, −0.002], p = 0.007, respectively), as well as significantly positively associated with general levels of depression, anxiety, and stress (b = 0.009, 95% CI [0.006, 0.011], p < 0.001; b = 0.008, 95% CI [0.005, 0.011], p < 0.001; b = 0.006, 95% CI [0.003, 0.009], p < 0.001, respectively). Average levels of rumination were not significantly associated with general levels of positive event LD (ps > 0.05).

Depressive symptoms (CES-D) estimating within- and between-person associations in EMA measures and LD

Within participants, increases in depressive symptoms were associated with decreases in positive emotion and physical well-being and happiness on same week EMA measures (b = −0.010, 95% CI [−0.017, −0.003], p = 0.005; b = −0.008, 95% CI [−0.014, −0.002], p = 0.011; b = −0.012, 95% CI [−0.018, −0.005], p < 0.001, respectively), as well as increases in within-person changes of depression and anxiety on same week EMA measures (b = 0.009, 95% CI [0.003, 0.016], p = 0.002; b = 0.008, 95% CI [0.0003, 0.015], p = 0.039, respectively). Additionally, positive daily fluctuations in depressive symptoms were predictive of within-person decreases in emotional well-being and happiness EMA measures one week later (b = −0.007, 95% CI [−0.013, −0.001], p = 0.016; b = −0.007, 95% CI [−0.012, −0.0009], p = 0.023, respectively), beyond the associations of same day depressive symptoms fluctuations. There were no other significant associations between daily fluctuations in depressive symptoms and within-person changes in EMA outcomes and LD (ps > 0.05).

Between participants, greater average levels of depressive symptoms across the two weeks were significantly negatively associated with general levels of emotional and physical well-being, happiness, and negative event LD (b = −0.006, 95% CI [−0.009, −0.003], p < 0.001; b = −0.009, 95% CI [−0.011, −0.006], p < 0.001; b = −0.007, 95% CI [−0.010, −0.005], p < 0.001; b = −0.010, 95% CI [−0.016, −0.004], p = 0.001, respectively), as well as positively associated with general levels of stress, depression, anxiety (b = 0.007, 95% CI [0.004, 0.010], p < 0.001; b = 0.012, 95% CI [0.010, 0.015], p < 0.001; b = 0.009, 95% CI [0.005, 0.012], p < 0.001, respectively). Average levels of depressive symptoms were not significantly associated with LD during positive events.

Perceived stress estimating within- and between-person associations in EMA measures and LD

Within participants, daily positive fluctuations in perceived stress were significantly associated with decreases in within-person negative event LD for same week measurements and physical well-being (b = −0.022, 95% CI [−0.039, −0.006], p = 0.002; b = −0.008, 95% CI [−0.013, −0.0005], p = 0.034). There were no other same week within-person associations between perceived stress and EMA measures or positive event LD (ps > 0.05). However, daily positive fluctuations in perceived stress were predictive of decreases in emotional well-being and happiness two days later (b = −0.009, 95% CI [−0.016, −0.003], p = 0.005; b = −0.007, 95% CI [−0.013, −0.002], p = 0.013). There were no other significant lagged associations between perceived stress and EMA or LD measures (ps > 0.05).

Between participants, greater average levels of perceived stress during the two weeks were significantly negatively associated with general levels of emotional and physical well-being EMA measures, happiness, and negative event LD (b = −0.014, 95% CI [−0.018, −0.010], p < 0.001; b = −0.010, 95% CI [−0.015, −0.006], p < 0.001; b = −0.013, 95% CI [−0.018, −0.009], p < 0.001; b = −0.015, 95% CI [−0.024, −0.005], p = 0.002, respectively), as well as positively associated with general levels of stress, depression, and anxiety EMA measures (b = 0.016, 95% CI [0.012, 0.021], p < 0.001; b = 0..017, 95% CI [0.013, 0.021], p < 0.001; b = 0.018, 95% CI [0.014, 0.023], p < 0.001, respectively).

Discussion

Summary of Findings

We aimed to assess emotion, health, and well-being over time and their relationship to a linguistic signature of psychological distancing while participants described positive and negative events using a novel mobile health assessment application. We found that within-person effects of LD during negative events were negatively associated with stress and positively associated with emotional and physical well-being. We further found that within-person lagged effects of LD during positive events predicted happiness. In addition, between participants, we found that LD during negative events was positively associated with emotional and physical well-being, and we found that LD during positive events was negatively associated with symptoms of depression.

Next, we found that average rumination over time was significantly negatively associated with LD during negative events. In addition, average symptoms of depression over time were significantly negatively associated with LD during negative events. Within participants, greater increases in perceived stress were negatively associated with LD during negative events; between participants, greater perceived stress during the two weeks was significantly negatively associated with LD during negative events.

While correlational, these results suggest that individuals who spontaneously implement LD to a greater extent when describing negative and positive events may be buffered against perceiving situations as stressful, and instead report greater mental and physical well-being. While these results demonstrate such effects for the first time in a real-world context using spontaneously assessed LD, they are consistent with prior work on the psychological mechanisms and affective consequences of LD. For example, Nook and colleagues found that engaging in LD led to reductions in self-reported negative affect (Nook et al., 2017). Further, we previously found that LD was associated with fewer symptoms of depression, perceived stress, and difficulties in emotion regulation, and better emotional well-being, energy and vitality, social functioning, and reappraisal frequency (Shahane & Denny, 2019). In addition, a meta-analysis and review article established a modest and significant association between first-person singular pronoun use and depression (Berry-Blunt et al., 2021; Edwards & Holtzman, 2017). Furthermore, a high-powered study showed that distanced self-talk as opposed to immersed self-talk reduced emotional reactivity when individuals reflected on negative experiences (Orvell et al., 2021). Indeed, the literature taken together coalesces on linguistic shifts as a relatively effortless route to emotion regulation (Orvell et al., 2019).

Further, these results suggest that such buffering effects may go above and beyond short term in-the-moment benefits, as there may be delayed or lagged benefits of LD a few days later. In the current work we found that within-person effects of LD during positive events predicted happiness reports two days later. The lagged models used in the present study have been used in similar contexts in previous work among a variety of populations (Berking et al., 2008; Kobayashi et al., 2020; Lucas-Molina et al., 2020; Silva et al., 2018). Although previous work addresses different research questions with different populations of interest, the present study is consistent with prior work in showing prospective effects of emotion regulation on well-being (Berking et al., 2008; Kobayashi et al., 2020; Lucas-Molina et al., 2020; Silva et al., 2018). The present study is novel, however, in specifically showing lagged effects of LD and its impact on health.

Furthermore, the rumination findings in particular shed light on how and when LD may be most useful for clinicians. Specifically, rumination is a cognitive process that appears to play a role in symptom maintenance of internalizing psychopathologies, particularly among depressive disorders (Gotlib & Joormann, 2010; Visted et al., 2018). Given the negative association observed in the current study between rumination and LD during negative events, the present findings may be clinically relevant, as they support the further investigation of therapeutic techniques that specifically target rumination processes to potentially improve affect and reduce symptom maintenance of internalizing disorders (Querstret & Cropley, 2013).

Possible Pathways and Mechanisms

Indeed, there are many possible pathways and mechanisms at play. Given the results of the present work and previous literature, and although speculative, it seems that individuals may engage more LD during negative and stressful events as a way to cope. By engaging increased psychological distance between themselves and the stimulus, people may be able to regulate emotion better and therefore potentially have better health outcomes. However, for positive events, there is typically less need to implement LD since the individual may want the stimulus to be perceived as close (i.e., the positive event may be tightly intertwined with the person’s identity). The severity of the negative event is also likely to impact the ease of LD use. As a result, the many possible pathways are likely a product of the interaction between person, situation, and strategy factors (i.e., in this case, LD) (Doré et al., 2016).

More broadly, mental and physical health risks are of ever-growing importance due to the multifaceted nature of real-world stressors. The EMA measures in the present work included stress, happiness, anxiety, as well as emotional and physical health, and they all have causal relationships with more complicated diseases like cardiovascular disease, cancer, and diabetes (Bickett & Tapp, 2016; Black & Garbutt, 2002; Conley et al., 2016; Fagundes et al., 2013, 2017; Fournier et al., 2002; Giltay et al., 2004; Steptoe & Kivimäki, 2012). Understanding mechanisms to mitigate risks associated with daily stressors is of particular importance. While recent work has investigated the relationship between proinflammatory cytokines–a measure of systemic inflammation associated with a number of illnesses–and LD (Shahane et al., in press), to our knowledge the present work is the first to examine LD in relation to health in the real world with stimuli that are personally-specific. Thus, while correlational, the present work extends previous investigations of how language may be a flexible, low burden means to regulate emotion, mitigate poor health, and improve well-being.

Constraints on Generality

There are several limitations and areas for future work in the present study. Because of the “real world” nature of the study, we were unable to acquire physiological data indicative of physical health risks and thus relied on self-report. Future work should assess converging evidence using validated biomarkers that may be able to be assessed in a mobile context (e.g., heart rate variability) to provide greater insight into the biobehavioral mechanisms of LD. Further, while the current study contained a fairly diverse sample of adults in the United States, future work may assess these effects in clinical or stressed populations specifically. Psychological distancing has shown potential in clinical populations at risk for specific disorders like bipolar disorder or major depression (Kross & Ayduk, 2017). Future work may probe the value and implications of LD translationally across diverse populations, both in particular clinical populations as well as transdiagnostically. Because the present study did not use an experimental design, it is important to acknowledge that the within-person associations we observed between the study variables may be due to mechanisms that are not related to the theory proposed. Specifically, it is possible that the dependent variables we examined influence LD, as opposed to LD influencing EMA measures. It is also possible that a third time-varying factor is driving the within-person associations that we observe. For example, the current study did not assess the intensity of events that were used for positive and negative LD events. It is possible that difference between the intensities of emotions during the events influenced the affective EMA measures, as well as influenced how much distancing a person implemented during data collection. It is important to note that the current study did not make type I error inflation adjustments. Given that there were multiple tests conducted, the experimental-wise type I error likelihood was increased. In the present work, caution should be practiced when interpreting the significant fixed effects with p-values greater than .008, as these values would be null if any type I error inflations adjustments were made and are at greater risk of being observed by chance. Therefore, future experimental work that assesses if instruction in shifting language via LD leads to real-world changes in emotion, health, and well-being would allow for us to make stronger causal inferences regarding the role LD has on day-to-day changes in affect and well-being.

Conclusion

Overall, the present results illustrate novel interrelationships among natural language use, emotion regulation via psychological distancing, and real-world, ecologically-valid self-reports of emotion, health, and well-being using a scalable mobile application. LD was associated with more adaptive emotion and health reports at concurrent times as well as in lagged models predicting happiness two days later. As a whole these results suggest the promise and utility of LD as part of an adaptive emotion regulation toolkit.