Increases in loneliness during medical school are associated with increases in individuals’ likelihood of mislabeling emotions as negative

Abstract

Expressions of emotion represent an important and unique source of information about the states of others. Being able to effectively understand expressions of emotions to make inferences about others’ internal mental states and use these inferences to guide decision-making and behavior is critical to navigating social relationships. Loneliness, the feeling that one lacks social connection, has important functional consequences for how individuals attend to signals of emotions in others. However, it is less clear whether loneliness changes how individuals recognize emotions in others. In medical practitioners, being able to accurately recognize emotional cues from patients is critical to effectively diagnosing and reacting with care to those patients. The current study examines the relationship between changes in loneliness during medical school and students’ recognition of emotion in others. Measures of loneliness and emotion recognition were collected from 122 medical students during their first three years of medical school at the beginning and end of each academic year. Changes in loneliness were related to changes in emotion detection, with increases in loneliness being associated with decreases in the probability of accurately discriminating sad and angry faces from other expressions, decreases in the probability of mislabeling emotion expressions as happy, and increases in the probability of mislabeling other emotional expressions as pain and angry. This study suggests that changes in loneliness during medical school are associated with increases in students’ labeling emotional expressions as negative, possibly by shifting attention to cues of negative emotion and away from cues of positive emotion.

Introduction

A fundamental assumption of emotion theory is that emotion is an adaptive orienting system that evolved to guide behavior. Emotion is also, however, an interpersonal communication system that elicits response from others. Thus, emotions can be viewed as both intrapersonal and interpersonal states (Decety & Jackson, 2004; Decety & Skelly, 2014). In the context of interpersonal relationships, emotion expressions represent crucial cues used to make inferences about others’ mental states (Barrett, Adolphs, Marsella, Martinez, & Pollak, 2019). Such inferences then inform individuals’ own reactions to and behavior towards the other person (Hareli & Hess, 2012; van Kleef, 2009). For example, inferences about others’ emotional states have been demonstrated to influence a range of social behaviors, including trust, competition versus cooperation in the context of economic games (Adam & Brett, 2015; van Doorn, Heerdink, & van Kleef, 2012), pro-social and helping behaviors (Craig, 2018; Williams, 2003), and aggressive behaviors (Penton-Voak et al., 2013; Schwenck et al., 2014). Therefore, being able to effectively recognize emotions in others and use this information to interact with them is a key component in not only effectively navigating the social world, but also building and maintaining social relationships. However, how individuals recognize emotions in others is influenced by a range of factors (Barrett et al., 2019), including experience (Pollak, Cicchetti, Hornung, & Reed, 2000; Shackman & Pollak, 2005), cultural context (Gendron, Roberson, van der Vyver, & Barrett, 2014; Hess, Blaison, & Kafetsios, 2016), and current affective state (Schmid & Schmid Mast, 2010; Van Wingen et al., 2011).

Loneliness, which refers to the distressing feelings accompanying the perception that one’s needs are not being met by the quality of social relationships, is thought to play an important role in maintaining social relationships, in part through shifting attention towards certain emotional cues and away from others (Cacioppo, Hawkley, Norman, & Berntson, 2011; Hawkley & Capitanio, 2015). In the short term, loneliness signals a need for maintenance and reparation of social connections, and produces motivated psychological and behavioral changes aimed at achieving this goal (Cacioppo et al., 2006; Qualter, Brown, Munn, & Rotenberg, 2010). Indeed, loneliness has been associated with changes in attention to signals of emotion, especially those related to social threat (Bangee, Harris, Bridges, Rotenberg, & Qualter, 2014; Nowland, Robinson, Bradley, Summers, & Qualter, 2018; Shin & Kim, 2019). Specifically, loneliness has been linked to increased attention to negative social words and decreased attention to positive social words (Cacioppo et al., 2011). Additionally, individuals reporting higher levels of loneliness demonstrate greater activation of the visual cortex to unpleasant social images, indicative of increased visual attention to those images. In contrast, individuals reporting higher levels of loneliness demonstrate decreased activation of the ventral striatum to positive social images suggesting they may find positive social cues less rewarding (Cacioppo, Norris, Decety, Monteleone, & Nusbaum, 2009). The fact that individuals with higher levels of loneliness shift attention towards negative social cues, which may extend to cues of social threat, suggests that this attention shift would also change how they detect emotional states in others. Specifically, it indicates that individuals with higher levels of loneliness may be more sensitive to emotions representing cues of social threat, such as anger or pain (Hawkley & Capitanio, 2015; Norman, Hawkley, Cole, Berntson, & Cacioppo, 2011). However, there is still little evidence directly exploring how loneliness influences the detection of emotions in others. Some work does suggest that loneliness produces increased sensitivity towards emotions associated with social threat, with individuals reporting higher loneliness demonstrating increased sensitivity to pain in dislikable faces (Yamada & Decety, 2009), and adolescents reporting higher loneliness demonstrating increased sensitivity towards the detection of pain in sad and fear faces (Vanhalst, Gibb, & Prinstein, 2017). However, other literature has found no or few associations between loneliness and emotion detection in faces (Kanai et al., 2012; Lodder, Scholte, Goossens, Engels, & Verhagen, 2016), making it unclear whether loneliness does influence how individuals detect emotions in others, and if so what aspects of emotion recognition are altered.

The current study examined how changes in loneliness over an extended period of time (three years) in medical students, a population for whom being able to identify and understand emotions in others has critical implications for their ability to effectively diagnose, treat, and care for their patients (Cheng et al., 2007; Decety, Smith, Norman, & Halpern, 2014; Decety, Yang, & Cheng, 2010; Gleichgerrcht & Decety, 2013), are related to their ability to detect emotions in others. For physicians, the ability to accurately detect emotions in patients is key to 1) capturing their affective state, 2) empathizing, and 3) effectively treating them. This is especially true of negative emotions, which can act as a social cue of distress and motivate others to come to the aid of the distressed individual (Craig, 2018; Decety & Lamm, 2009; Williams & Craig, 2016). In the medical context, the role of emotional expression as a social cue becomes critical. Physicians are explicitly tasked with determining if someone is in distress and, if so, the amount and source of distress, as well as the most appropriate course of treatment. Mistakes in identifying whether a patient is in distress can lead to mistakes with treatment. Therefore, it is important to understand how physicians and medical students detect and understand emotions in others, and what factors may contribute to individual differences in these processes. Loneliness represents one such factor. Medical school has been demonstrated to increase feelings of loneliness and social isolation in students which in turn affects medical students’ clinical outcomes (Hojat, Callahan, & Gonnella, 2004; Maher et al., 2013; Schmitter, Liedl, Beck, & Rammelsberg, 2008). Loneliness in medical students is associated with reduced clinical competence (Hojat et al., 2004) and higher risk of attrition and increased burnout (Maher et al., 2013). The current longitudinal study examined how both loneliness and recognition of emotions in others changes over the course of the first three years of medical school, and whether within-individual differences in changes in loneliness relate to changes in how medical students’ accurately detect others’ emotions.

Methods

Participants

This study was part of a larger longitudinal project conducted in the greater Chicago area examining changes in students’ empathy over the course of medical school which has been described elsewhere (Smith, Norman, & Decety, 2017; Smith et al., 2019). Study participants were 128 students, ages 21 – 33 years (M = 23.35, SD = 1.84) at the start of medical school (n = 66 (51.6%) female). Of these students, 18 (14.1%) identified as Asian or Asian American, 8 (6.3%) identified as Black or African American, 5 (3.9%) identified as Hispanic or Latino, 80 (62.5%) identified as non-Hispanic White, 16 (12.5%) identified as multiracial, and 1 (0.8%) identified as Other (for SES information see Supplemental Table 1). Analyses were conducted on a subset of these students with complete data for at least three of the six time points (n = 122, 58 female). Attrition and sample composition of the subset of students included in these analyses have been discussed elsewhere (Smith et al., 2017). This sample size is comparable to that utilized in previous research and measurements were sampled more frequently than has been assessed previously (twice each year rather than only once) (Costa, Magalhães, & Costa, 2013; Riess, Kelley, Bailey, Dunn, & Phillips, 2012). Participants were compensated $60 for participation, and all provided written consent to participate at each time point. The study was approved by the University of Chicago Institutional Review Board (# IRB12-156).

Procedure

Data were collected from all participants at the beginning and end of each academic year for their first three years of medical school (2012 – 2015) for a total of six time points. At each time point, participants completed a set of online surveys and computerized tasks (online surveys were completed prior to task collection), assessing different components of empathy ((Smith et al., 2017). This current report focuses on study measures assessing medical students’ self-reported loneliness and emotion recognition behavior.

Loneliness

Participants completed the UCLA Loneliness Scale (Russell, 1996), a 20-item scale assessing individuals’ perceptions of loneliness. Items included statements such as “How often do you feel alone?” and “How often do you feel that your relationships with others are not meaningful?”. Participants rated items on a Likert scale of 1 – 4 with response options “Never”, “Rarely”, “Sometimes”, and “Often.” The UCLA Loneliness scale is widely used and has demonstrated high reliability and validity. Across time points, the questionnaire demonstrated an average Cronbach’s alpha of 0.91, indicative of high questionnaire internal consistency.

Emotion Recognition

Participants completed a task designed to assess emotion recognition. This task was modeled after those utilized in previous research that aim to assess how quickly and accurately individuals recognize different emotions (Decety, Echols, & Correll, 2009; Decety, Skelly, et al., 2014; Pollak et al., 2000). Participants were presented with 64 videos slowed to ten seconds in which actors (male and female) produced a facial expression that was either happy, sad, angry, or pained. Participants saw 16 videos for each emotion category (eight of which were portrayed by a male actor (four black and four white) and eight by a female (four black and four white)). These videos were selected from a set of 124 videos with 44 unique actors (21 white and 23 black; 21 male and 23 female). Two task versions were created matched for accuracy values collected from 62 subjects who were asked to identify the emotions in the videos in an initial validation (Version 1: M = 0.89, SE = 0.02; Version 2: M = 0.90, SE = 0.02; t = −0.40, p = 0.69). Task versions were randomized and counterbalanced across participants and time points such that participants never saw the same set of stimuli two time points in a row. Participants were asked to press the spacebar on the keyboard as soon as they recognized the emotion in each video. Once they pressed the spacebar, they were asked to identify what emotion they saw, using response options “Happy”, “Sad”, “Angry” or “Pain.” Reaction times for response to the video were in milliseconds and log transformed as a measure of speed of recognition. Participants’ responses identifying what the emotion they thought they saw was were coded for accuracy, and signal detection statistics were used to examine the nature of students’ emotion recognition performance. Each participant’s performance was measured using: hit rate (HR) which refers to the probability of selecting the response which matched the video; false alarm rate (FAR), which refers to the probability of selecting a response which did not match the video; correct rejection, which refers to the probability of not selecting an incorrect response; and miss, which refers to the probability of not selecting the correct response. The probability of hits and misses sum to 1 and the probabilities of false alarms and correct rejection rates also sum to 1. As such, the HR sufficiently describes students’ responses to correct items and FAR sufficiently describes children’s responses to incorrect items. Together, HR and FAR completely summarize the performance of a single subject in a single emotion condition.

HRs and FARs were combined into two statistics that describe individual’s sensitivity to differences between emotion expressions and either their response biases or willingness to define an ambiguous stimulus as a target. As has been done previously for similar emotion recognition tasks (Pollak et al., 2000), we used a threshold model which assumes that false alarms occur when the subject is uncertain and that hits both represent the proportion of correct identifications when the subject is certain and represent any “lucky” correct guesses during states of uncertainty (Snodgrass & Corwin, 1988). In threshold models, the discrimination index P_r represents the probability that an item will cross a recognition threshold. P_r was calculated as follows:

$$P_r=\text{HR}–\text{FAR}$$

The bias index, B_r, represents how much evidence, or certainty, an individual needs to select an emotional expression. False alarms occur when a student fails to match a target face with the emotion conveyed in the video which occurs with a probability of 1 – P_r. Given this, B_r was calculated as:

$$B_r=\text{FAR}/\left[1–\left(\text{HR}–\text{FAR}\right)\right]$$

Higher values of B_r indicate a more liberal or lax response criterion, while lower values indicate a conservative or strict criterion for selecting an emotion expression.

Statistical Analyses

Descriptive statistics for all study variables are reported in Supplemental Table 1. To assess changes in loneliness and emotion recognition over the course of the first three years of medical school, we utilized hierarchical linear modeling (HLM) techniques, a complex form of ordinary least square (OLS) regression which fits a linear function to the observed data while accounting for variation across individuals to estimate the population level rate of change based on the observed data set (Raudenbush & Bryk, 2002; Singer & Willett, 2003). Given that longitudinal data are inherently nested with time grouped within each subject (each individual has a series of outcomes for each time point), and HLM accounts for potential variation in observed outcomes across individuals, HLM was deemed the most appropriate method of analysis (Singer & Willett, 2003). Additionally, HLM can handle missing values without replacing them with values based on the distribution of the data set (imputation), which is well suited for the current data set where we did not have complete data for each subject at every time point (Singer & Willett, 2003) (for more discussion of the specific modeling approaches and how missing data were treated see Supplemental Materials). All models were run using the nlme package for R, developed for building hierarchical linear models, with full maximum likelihood estimation in R v3.2.3.

Inspection of individual participant level trajectories for loneliness, emotion recognition indices (P_r and B_r), and log transformed reaction times for responses to correctly identified videos indicated that a linear growth model was most appropriate for the data set. For each outcome measure we assessed two models. The first included only time, characterized as number of months since starting medical school, as a predictor. For emotion recognition and response time outcomes, this model also included emotion condition (happy, sad, angry, or pained) as a predictor in order to examine differences in speed of recognition, discrimination, and bias for different types of emotions. Emotion conditions were dummy coded and, given it was the only positive emotion presented, happiness was used as the comparison category. As age and gender have been associated with differences in emotion recognition (Lambrecht, Kreifelts, & Wildgruber, 2014; Mill, Allik, Realo, & Valk, 2009), the second model incorporated gender and age as additional predictors to assess any potential effects. To assess the relationship between changes in loneliness during medical school and changes in emotion recognition, we ran a model including loneliness as a time-varying predictor along with time and emotion condition (Singer & Willett, 2003). Again, we assessed two models. The first including time, emotion condition, and loneliness. The second also included gender and age. Additionally, as the emotion recognition task included two versions of the stimulus set, we ran all models controlling for counterbalancing order to account for any potential effects of the order of presentation of the sets. All predictors were mean centered to reduce issues with collinearity (Raudenbush & Bryk, 2002). Model fit was assessed using log-likelihood values and Akaike’s Information Criteria as recommended by Singer & Willett (2003). We also calculated both marginal and condition r² using methods specified in Nakagawa & Schielzeth, 2013). The marginal r² represents the variance explained by the fixed effect and the conditional r² represents the variance explained by the entire model, including both fixed and random effects. Any interaction effects were examined using simple slopes at one standard deviation above and below the mean as recommended by Aiken & West, (1991). Standardized betas (B) were calculated by standardizing all predictors at the subject level and running all models with standardized predictors.

Results

Changes in loneliness during medical school

There were no significant effects of time in medical school on loneliness (β = 0.01, B = 0.13, SE = 0.03, p = 0.636; Model Fit: AIC = 4461.81; Log Likelihood = −2224.90; marginal r² = 0.0002, conditional r² = 0.71), even when including age and gender. However, there was a significant main effect of age on loneliness (β = −0.89, B = −0.75, SE = 0.42, p = 0.038), such that younger students reported increased levels of loneliness at the start of medical school. There were no significant effects of gender on students’ reported loneliness during medical school (β = −2.43, B = −1.58, SE = 1.53, p = 0.116; Supplemental Table 2).

Changes in emotion recognition during medical school

Discrimination:

The main effect of time in the program was not significant (β = −0.0004, B = −0.003, SE = 0.0002, p = 0.132; Model Fit: AIC = −5745.16; Log Likelihood = 2884.58; marginal r² = 0.43, conditional r² = 0.59; Supplemental Table 3), suggesting that students did not demonstrate any overall changes in their ability to accurately detect emotions across emotion type over the course of medical school. However, there was a significant interaction effect of anger with time in medical school (β = −0.001, B = −0.01, SE = 0.0005, p = 0.007). Examining the simple slopes indicated that medical students demonstrated decreases in their ability to discriminate angry faces (β = −0.001, SE = 0.0004, p = 0.002), but no significant changes for happy, sad, and pain expressions (Supplemental Table 4, Figure 1). There were significant main effects for all emotion types (p < 0.001, Supplemental Table 4), such that medical students were worse at discriminating pained, angry, and sad as compared to happy faces (Figure 1). Including age and gender did not change any of these effects, and there were no significant effects of age or gender on emotion discrimination (p > 0.10; Model Fit: AIC = −5726.97; Log Likelihood = 2891.48; marginal r² = 0.44, conditional r² = 0.60; see Supplemental Table 3). Effects also did not change when controlling for task counterbalancing order (Supplemental Table 5; Model Fit: AIC = −5715.51; Log Likelihood = 2893.76; marginal r² = 0.44, conditional r² = 0.60).

Figure 1:

Change in emotion recognition indices over time in medical school by emotion (anger, happy, sad, pain). A) Emotion discrimination B) Emotion bias C) Reaction times for correct responses D) Reaction times for incorrect responses

Bias:

There was no significant main effect of time in medical school on bias towards mislabeling an emotion as another (β = −0.0003, B = −0.003. SE = 0.001, p = 0.611; Model Fit: AIC = 919.80; Log Likelihood = −447.90; marginal r² = 0.31, conditional r² = 0.31), indicating that overall, students do not change in their propensity towards mislabeling emotions during medical school. There were also no significant interaction effects of emotion types with time in medical school, suggesting this did not differ by emotion (Supplemental Table 3). There were significant main effects by emotion type (p < 0.001, see Supplemental Table 3), such that students were less likely to mislabel emotions sad, angry, and pained faces as compared to happy (Figure 1). Including age and gender did not change any of these effects and there were no significant effects of age or gender on bias towards mislabeling emotions (p > 0.10; Model Fit: AIC = 927.73; Log Likelihood = −435.86; marginal r² = 0.32, conditional r² = 0.32; see Supplemental Table 3). Effects also did not change when controlling for task counterbalancing order (Supplemental Table 5; Model Fit: AIC = 928.95; Log Likelihood = −428.48; marginal r² = 0.32, conditional r² = 0.32).

Speed of recognition (correct):

There was a significant main effect of time in medical school on students’ reaction times to the video for trials in which they correctly identified the emotion being expressed (β = −0.01, B = −0.07, SE = 0.001, p < 0.001; Model Fit: AIC = −3676.26; Log Likelihood = 1850.13; marginal r² = 0.39; conditional r² = 0.78), such that students got faster at accurately responding over time. Additionally, there were significant main effects and interactions with time in medical school for all the emotions (p < 0.001, see Supplemental Table 3). Students were slower at identifying sadness, anger, and pain as compared to happy faces (Figure 1). For the interactions, examining the simple slopes indicated that while there were significant slope changes for all emotions, students showed greater negative slope change over time for pain (β = −0.009, SE = 0.001, p < 0.001), sadness (β = −0.011, SE = 0.001, p < 0.001), and anger (β = −0.010, SE = 0.001, p < 0.001) than happiness (β = −0.006, SE = 0.001, p < 0.001), suggesting that there was less overall change in speed of identification of happy faces. Including age and gender did not change any of the observed effects (see Supplemental Table 3; Model Fit: AIC = −3666.30; Log Likelihood =1861.15; marginal r² = 0.40; conditional r² = 0.78). However, there was a significant interaction of age with pain (β = −0.02, B = −0.01, SE = 0.01, p = 0.008). Examining the simple slopes indicated that while all participants were faster at identifying pain compared to happy faces, older individuals (β = 0.15, SE = 0.04, p < 0.001) were faster than younger individuals (β = 0.21, SE = 0.04, p < 0.001) at correctly identifying pain in others. Effects did not change when controlling for task counterbalancing order (Supplemental Table 5; Model Fit: AIC = −3678.96; Log Likelihood = 1875.45; marginal r² = 0.40; conditional r² = 0.79).

Speed of recognition (incorrect):

To examine whether changes in reaction times for trials when students correctly identified the emotion were specific to accurate responses or instead representative of overall increases or decreases in speed of response, we also assessed changes in reaction times for responses to the video for trials in which they did not correctly label the emotion. Similarly to reaction times for correct trials, there was a significant main effect of time in medical school on overall reaction times (β = −0.01, B = −0.07, SE = 0.001, p < 0.001; Model Fit: AIC = 1246.43; Log Likelihood = −611.21; marginal r² = 0.13; conditional r² = 0.39), such that participants got faster at responding to the video over time, indicating that regardless of accuracy, students made faster responses as they progressed in the program. Additionally, there were similar main effects and interactions with time for sadness, pain, and anger as for reaction times for correct responses (see Supplemental Table 3), suggesting that these changes in reaction times are overall changes rather than specific to correct emotion identification. Including age and gender did not change any of these effects (Supplemental Table 3; Model Fit: AIC = 1261.89; Log Likelihood = −602.95; marginal r² = 0.13; conditional r² = 0.39). There were, however, significant interaction effects of age with sadness, anger, and pain, and of age with time in medical school, sadness, anger, and pain (Supplemental Table 3). This effect appeared to be driven by older individuals demonstrating stronger decreases in reaction times for incorrectly labeling emotions as happy and younger individuals demonstrating increases in reaction time for incorrectly labeling emotions as sad (Supplemental Figure 1). There were no significant effects of gender on reaction times for incorrect responses (Supplemental Table 3). Effects did not change when controlling for task counterbalancing order (Supplemental Table 5; Model Fit: AIC = 1258.60; Log Likelihood = −593.30; marginal r² = 0.15; conditional r² = 0.40).

Relationship between changes in loneliness and changes in emotion recognition during medical school

Discrimination:

Including loneliness as a time-varying predictor in the model did not change any of the previously observed effects (Supplemental Table 6; Model Fit: AIC = −5754.75; Log Likelihood = 2893.37; marginal r² = 0.43; conditional r² = 0.59). There was no significant main effect of loneliness on emotion discrimination (β = −0.0004, B = −0.003, SE = 0.0003, p = 0.140), suggesting that changes in loneliness were not related to changes in emotion discrimination across all emotions. However, there were emotion-specific effects of loneliness on emotion discrimination as indicated by a significant interaction of loneliness with anger (β = −0.001, B = −0.005, SE = 0.0005, p = 0.007) and sadness (β = −0.001, B = −0.004, SE = 0.0005, p = 0.027). Examining the simple slopes indicated that increases in loneliness during medical school were associated with decreases in the ability to discriminate sad (β = −0.001, SE = 0.0005, p = 0.012) and angry faces (β = −0.001, SE = 0.0005, p = 0.014) and no change in the ability to discriminate happy (β = −0.0001, SE = 0.0003, p = 0.723) or pain faces (β = 0.000, SE = 0.0005, p = 0.982) (Supplemental Table 4; Figure 2). Including age and gender did not change any of the observed effects, and age and gender had no significant effects on students’ emotion discrimination (Supplemental Table 6; Model Fit: AIC = −5734.83; Log Likelihood = 2901.41; marginal r² = 0.44; conditional r² = 0.60). Effects did not change when controlling for task counterbalancing order (Supplemental Table 7; Model Fit: AIC = −5715.51; Log Likelihood = 2893.76; marginal r² = 0.44; conditional r² = 0.60).

Figure 2:

Relationship between emotion recognition indices and changes in loneliness over time in medical school by emotion (anger, happy, sad, pain). A) Emotion discrimination B) Emotion bias C) Reaction times for correct responses D) Reaction times for incorrect responses

Bias:

Including loneliness along with time in medical school as a predictor of bias towards mislabeling emotions did not change any of the previously observed effects (Supplemental Table 6; Model Fit: AIC = 908.66; Log Likelihood = −438.33; marginal r² = 0.31; conditional r² = 0.31). There was a main effect of loneliness on bias, such that increases in loneliness (β = −0.002, B = −0.01, SE = 0.001, p = 0.018) were associated with overall decreases in bias towards mislabeling emotions. Additionally, there were emotion-specific effects, with loneliness demonstrating significant interactions with both anger (β = 0.01, B = 0.03, SE = 0.002, p = 0.002) and pain (β = 0.01, B = 0.03, SE = 0.002, p = 0.001; Figure 2). Examining the simple slopes indicated that increases in loneliness during medical school were associated with decreased bias towards mislabeling emotions as happiness (β = −0.002, SE = 0.001, p = 0.014), but increased bias towards mislabeling emotions as pain (β =0.004, SE = 0.002, p = 0.018) and anger (β = 0.003, SE = 0.002, p = 0.037) and no change in bias towards labeling emotions as sad (β =0.0004, SE = 0.002, p = 0.792). Including age and gender did not change any of the observed effects (Model Fit: AIC = 918.44; Log Likelihood = −425.22; marginal r² = 0.32; conditional r² = 0.32). There was a significant interaction effect between gender and pain (β =0.15, B = 0.03, SE = 0.07, p = 0.033). Examining the simple slopes indicated that while both males and females were significantly more likely to mislabel emotions as happy as compared to pain, males (β =−0.46, SE = 0.02, p < 0.001) were significantly less likely to mislabel emotions as pain as compared to happy than females (β = −0.31, SE = 0.05, p < 0.001). There were no other significant effects of gender or age on bias towards mislabeling emotions (Supplemental Table 6). Effects did not change when controlling for task counterbalancing order (Supplemental Table 7; Model Fit: AIC = 928.95; Log Likelihood = −428.48; marginal r² = 0.32; conditional r² = 0.32).

Speed of recognition (correct):

Including loneliness did not change any of the previously observed effects of emotion type or time in medical school on speed of recognition of emotions when students correctly labeled the emotions (Supplemental Table 6; Model Fit: AIC = −3675.55; Log Likelihood = 1853.78; marginal r² = 0.39; conditional r² = 0.78). There was no significant main effect of loneliness on reaction times (β = −0.0005, B = −0.004, SE = 0.001, p = .361). However, loneliness did interact with sadness to predict reaction times (β = −0.002, B = −0.01, SE = 0.001, p = 0.014; Supplemental Table 6; Figure 2). Examining the simple slopes indicated that increases in loneliness were associated with decreases in reaction times for correctly labeling emotions as sad (β = −0.002, SE = 0.001, p = 0.016) and little change when labeling happy faces (β = 0.000, SE = 0.001, p = 0.732). Including age and gender did not change any of the observed effects (Supplemental Table 6; Model Fit: AIC = −3662.57; Log Likelihood = 1865.28; marginal r² = 0.40; conditional r² = 0.79). Additionally, there was a significant interaction of pain with age as was observed previously (β = −0.02, SE = 0.01, p = 0.01). Effects did not change when controlling for task counterbalancing order (Supplemental Table 7; Model Fit: AIC = −3678.96; Log Likelihood = 1866.26; marginal r² = 0.40; conditional r² = 0.79).

Speed of recognition (incorrect):

Including loneliness did not change any of the previously observed effects on speed of recognition for incorrect trials (Model Fit: AIC = 1251.86, Log Likelihood = −609.93; marginal r² = 0.13; conditional r² = 0.39). There was no significant main effect of loneliness on reaction times or interactions of loneliness with emotion of the video (Figure 2; Supplemental Table 6). Including gender and age did not change any of the observed effects (Supplemental Table 6; Model Fit: AIC = 1271.82; Log Likelihood = −601.91; marginal r² = 0.13; conditional r² = 0.39). However, as demonstrated previously, there were significant interactions between time in medical school, age, and identifying sadness, anger, and pain (Supplemental Table 6). Effects did not change when controlling for task counterbalancing order (Supplemental Table 7; Model Fit: AIC = 1258.60; Log Likelihood = −593.30; marginal r² = 0.15; conditional r² = 0.40).

Discussion

This study examined the extent to which changes in loneliness are related to how individuals recognize and label emotional expressions in others. We addressed this question in a population in which accurate identification of emotions in others is important to their being able to effectively perform their job. Specifically, we examined the relationships between medical students’ reported loneliness and emotion recognition over the course of their first three years of medical school. We found that individual differences in changes in loneliness during medical school were related to changes in both participants’ ability to accurately discriminate emotions in others and their bias towards mislabeling emotions. In particular, increases in loneliness were related to a reduction in participants’ ability to discriminate angry and sad faces and increases in their bias towards mislabeling faces as anger or pain. This suggests that individuals who experience increases in loneliness both become worse at accurately detecting negative emotions in others and more likely to mislabel emotions as anger or pain (both pertinent cues of social threat). This finding has important implications as this could create a positive feedback loop in which individuals feeling more lonely are more likely to misperceive others as having a negative emotion state and less likely to accurately detect a positive emotional state, which in turn can increase feelings of loneliness. Additionally, in the context of medicine, this finding has several implications as being able to accurately recognize signals of pain in patients, and use this information to inform treatment, is a key skill for physicians (Cheng, Chen, & Decety, 2017). While we find no changes in loneliness in medical students at an aggregate level, in contrast to some previous research (Hojat et al., 2004; Maher et al., 2013; Schmitter et al., 2008), the current study does suggest that students who find the experience of medical school isolating become more likely to misinterpret expressions of emotion as pain, which may have implications for how they then treat patients.

This work is in line with previous research that has demonstrated loneliness is associated with increases in sensitivity and attention to social threat (Cacioppo et al., 2011; Norman et al., 2011). Negative emotions, and particularly expressions of pain and anger, can act as a signal to others of the presence of a potential threat (Bayet, Behrendt, Cataldo, Westerlund, & Nelson, 2018; Willis, Windsor, Lawson, & Ridley, 2015). Additionally, negative emotions such as sadness or pain can also act as a signal to others that a social group member is currently under attack and may be in need of aid (Karos, Williams, Meulders, & Vlaeyen, 2018; Williams & Craig, 2016). Interestingly, we find that decreases in individuals’ ability to discriminate emotions and increases in their bias towards labeling other emotions as negative emotions is combined with decreases in bias towards mislabeling negative emotions as happy. The fact that increases in loneliness are associated with reduced bias toward mislabeling emotions as happy parallels previous work that has suggested that loneliness not only increases attention to social threat, but also is associated with decreased attention towards positive social cues (Cacioppo et al., 2009). Overall among the population tested, bias towards mislabeling emotions as happy was quite high, and did not change much over the course of medical school. Previous work looking at emotion recognition in other populations finds a bias towards mislabeling emotions as happy (Freeman et al., 2018; Wagner et al., 2015). It is possible that the shift related to changes in loneliness represents an overall shift in attention towards negative faces or decreased sensitivity to cues of positive emotion in others. However, as we did not directly test attention but rather emotion recognition alone, it is unclear whether the observed effects are due to shifts in attention towards certain cues. Future research should directly test whether changes in attention toward negative and positive cues are related to the observed shifts in emotion detection processes.

While our results parallel previous findings of overall decreased emotion recognition accuracy in lonely individuals (Knowles, Lucas, Baumeister, & Gardner, 2015; Zysberg, 2012), they are in contrast to work that finds increased levels of loneliness are associated with increased emotion recognition accuracy (Gardner, Pickett, Jefferis, & Knowles, 2005; Lodder et al., 2016) or no effect (Kanai et al., 2012). Additionally, they somewhat differ from the few studies which have examined emotion recognition and sensitivity for specific emotion expressions that find either increased recognition of angry faces in individuals reporting increased loneliness (Lodder et al., 2016) or better detection of sadness and fear (Vanhalst et al., 2017). The current work suggests instead that increases in loneliness are associated with decreases in the ability to discriminate angry and sad faces. These differing results may be due to different approaches to characterizing emotion recognition performance. Previous research primarily has examined accuracy without taking into account false alarm rates. The current research utilized a signal detection approach, examining both at individuals’ ability to discriminate different emotional expressions from each other (accounting for false alarms) as well as their tendency to mislabel emotional expressions. Given we find that individuals become more likely to mislabel emotions as pain and anger, this suggests that previous findings related to accuracy for these emotions may be a result of individuals high in loneliness being more likely to label any emotional expression as such. Future work should continue to explore how loneliness affects emotion recognition by emotion type for both accurate labels as well as bias towards labeling one emotion over another to better understand how loneliness influences how individuals make inferences about others.

The shift towards perceiving expressions of emotion as more negative regardless of modality, suggests that, for individuals high in loneliness, changes in how they interpret social cues within their environment may initiate a positive feedback loop in which increased loneliness leads to an increased probability of perceiving negative affect in others in turn leading to increases in loneliness due to a lack of positive social feedback. This cycle could lead to increases in susceptibility to certain psychopathologies linked to shifts in attention to negative social information such as depression. Indeed, in individuals with depression, research has found decreased impaired recognition and discrimination of facial emotions, along with increased sensitivity to expressions of sadness (Anderson et al., 2011; Dalili, Penton-Voak, Harmer, & Munafò, 2015; Kohler, Hoffman, Eastman, Healey, & Moberg, 2011). In medical students and health professionals, this could be especially problematic as it may affect how interact with their patients. Being able to recognize and discriminate cues of anger, pain, sadness, and happiness, is important to effectively understanding the patients’ current emotional states. Being able to demonstrate an understanding of and empathy for patients’ emotional states in turn has been linked to improved outcomes for both patients and physicians (Brazeau, Schroeder, Rovi, & Boyd, 2010; Decety & Fotopoulou, 2015; Gleichgerrcht & Decety, 2012; Rakel et al., 2009). Physicians who are less able to appropriately recognize different emotion expressions and use that information to make inferences about patients emotional states may be less able to effectively treat their patients. However, the current research did not directly examine how the changes shift how medical students interact and are perceived by their patients which is critical to understanding how loneliness and the associated changed in emotion recognition affect the patient doctor relationship. Future work should further explore how the changes observed shift medical students and physicians behaviors within the patient physician interaction. Additionally, there is a need for exploration of the directionality of the relationships between loneliness and emotion recognition, it’s implications for social behavior, and what may be fruitful avenues of intervention. It is possible that interventions aimed at shifting attention towards positive cues of emotion in individuals reporting high levels of loneliness, may help reduced feelings of loneliness and isolation.

This study has several limitations that should be acknowledged. First, individual differences in changes in students’ loneliness may not be due to the experience of medical school, but many other potential confounding factors. However, regardless of their driving source, this work does suggest that shifts in loneliness in medical students have implications for their ability to recognize expressions of emotions in others. Additionally, while it is possible study attrition led to a bias in the effects, this seems unlikely as attrition was overall fairly low. It is also the case that our task is somewhat limited in that it focused on three negative emotions and only one positive emotion. Future work should examine a broader range of positive emotions to better understand if the current findings generalize to other positive emotions or are specific to happiness. In addition, while the task represents a better alternative to examining emotion recognition than static faces, it still does not fully capture the nuances of emotion expression in a naturalistic setting. As has been outlined extensively elsewhere (for recent review see Barrett et al., 2019), there are a number of limitations associated with posed emotions, including the fact that actors are not feeling the emotion themselves but only producing associated facial expressions. While we made sure to equally present videos using both white and black and male and female actors to limit any effects related to perceived similarity of the participant with the actor, it is still possible that individual characteristics related to perceived similarity or dissimilarity between the actor and participant influenced emotion recognition performance. Additionally, our stimuli focused solely on recognition of emotion in the face. In the future, research should examine if these relationships hold when individuals have more sources of emotional information, including cues in the voice and bodily expressions. It is also notable that the current research found little evidence for any effects of gender on emotion recognition indices, despite the fact that some previous work suggests women demonstrate superior performance by females in facial emotion recognition tasks (Hall & Matsumoto, 2004; Lambrecht et al., 2014; Montagne, Kessels, Frigerio, De Haan, & Perrett, 2005). However, this effect has primarily replicated in research examining accuracy irrespective of emotion type, and recent research examining accuracy for different emotion types is less consistent (Connolly, Lefevre, Young, & Lewis, 2019; Forni-Santos & Osório, 2015). The current study also represents a relatively limited age range and population, and the question of how changes in loneliness relate to emotion recognition should be examined in across different populations to assess generalizability.

Despite these limitations, this study provides important insight into how long-term longitudinal changes in loneliness may contribute to individual differences in how people recognize emotions in others, which has important implications for how they navigate their social interactions. It is important future work continues to investigate these questions, in a wider range of populations and across emotion expression modalities. Recent work examining auditory cues of emotion found poorer recognition for some negative emotions and better for positive, suggesting potentially interesting differences between face perception and auditory perception (Morningstar, Nowland, Dirks, & Qualter, 2019). Research examining across modalities is important for providing increased insight into how loneliness shifts social perception and interaction. In addition, the current research suggests that medical students who find medical school isolating may be less able to effectively diagnose and treat patients. Literature to date examining how medical school affects a range of affective processes, including empathy, burnout, and loneliness, in medical students has focused on changes at a population level and finds mixed results for whether medical school negatively effects these processes (Ferreira-Valente et al., 2016; Mohammadreza Hojat et al., 2009; Smith et al., 2017). The current research indicates that it may be more fruitful to examine these questions at an individual level, which in turn could help identify students most at risk for negative outcomes in medical school as well as what may be effective routes of intervention for them. Future work should continue to explore how loneliness changes emotion perception, along with how these changes then relate to social behavior.