We obtained data for this study from the Mobile Community Health Assistance for Tenants (m.chat) project, a technology- assisted health coaching intervention designed to improve health indicators among PSH residents in Fort Worth, Texas [23]. We recruited participants via convenience sampling from 6 local housing agencies in Fort Worth. The participants were adult, English speaking, Medicaid-enrolled or eligible, and reported at least one of the following conditions in the past year: prescribed medication for psychological or emotional problems, experienced hallucinations, received a pension for a psychiatric disability, or reported at least moderate levels of depression (>9 on the Patient Health Questionnaire). Participants met monthly with a health coach who helped to set goals related to diet, exercise, substance use, medication compliance, social support, and recreation or leisure. We gave the opportunity to participate in the EMA portion of the project to a subgroup of participants who scored ≥4 on the Rapid Estimate of Adult Literacy in Medicine-Short Form (indicating >6th grade English literacy level). This subset of clients completed EMAs each morning with questions about current emotions and setting as well as health behaviors from the previous day, including diet, exercise, substance use, leisure time activities, medication compliance, and social interactions. We used an EMA protocol based on that of a previous study with homeless smokers [20]. We provided participants with a smartphone and granted unlimited voice, short message service text, and 2 GB data for their personal use. While enrolled in the EMA portion of m.chat, participants received up to 15 “Chat Bucks” each month, proportional to the percentage of days they completed the assessment (1 Chat Buck=US $1 redeemable for health-related supplies; thus, participants could earn up to US $15 worth of health-related supplies each month). Provided they were compliant with at least 50% of the EMA prompts, participants could carry the phone for up to 12 months. Project resources allowed for up to 80 participants to participate in the EMA portion at any one time; when participants returned the phone (because they had reached the end of their allotted time, were failing to complete assessments, or decided they did not want to carry the phone any longer), it was reset to factory settings and offered to another participant based on the order of enrollment into the parent study.

The Institutional Review Board of the University of North Texas Health Science Center approved this project, and we assured participants of confidentiality. All participants gave informed consent.

For analysis, we included 155 participants who completed a total of 18,357 daily assessments between May 1, 2016, and April 30, 2017. On average, individuals received 139 daily assessments or prompts (range 14-334) and completed 106 assessments (range 4-322). The sample was split almost evenly between males (n=77) and females (n=78), and the average age was 52 (SD 8) years.

The mobile app alerted participants to complete an assessment 30 minutes after the participant’s self-reported waking time. We asked the participants to complete the assessment within 30 minutes of the initial alert; they had the option to “snooze” an assessment request 3 times each day before the EMA would be counted as missed. Below are the questions that were presented in the daily EMA (we have only presented the questions or response options considered in the analyses).

The circumplex model of affect [24,25] was used to categorize each emotion in a 2-dimensional circular space, containing dimensions for arousal and valence (Figure 1). Valence, in the context of emotions, is defined as the intrinsic attractiveness or averseness of an event, object, or situation. Likewise, arousal is the state of being physiologically and mentally alert, awake, and attentive. In a recent refinement of the model using regression [26], the circumplex model was quantitatively visualized as a circular space of radius of 1 unit within a 2-dimensional Cartesian coordinate system, which assigns scores for valence and arousal for each emotion. This model includes a comprehensive list of mood items commonly considered in behavioral sciences. We obtained scores for valence and arousal in the circumplex from this model for each of the 9 emotion items considered (Table 1). In Figure 2, these 9 emotions are presented within the circumplex, depicting their valence and arousal coordinates. During each daily assessment, we created composite scores of valence and arousal as weighted sums of responses from all 9 emotion questionnaire items, with the reported emotion scores serving as weights. These two composite scores were the outcomes of this study. Specifically, let a subject’s response in Likert-type scale for I feel happy, I feel frustrated, I feel sad, I feel worried, I feel restless, I feel excited, I feel calm, I feel bored, and I feel sluggish be denoted by l 1, l 2, l 3, l 4, l 5, l 6, l 7, l 8, and l 9, respectively. Then, using Table 1, the composite score for valence is 0.95× l 1−0.5× l 2−0.95× l 3−0.15× l 4−0.15× l 5+0.7× l 6+0.75× l 7−0.4× l 8−0.15× l 9. The corresponding composite score for arousal is 0.15× l 1+0.4× l 2−0.4× l 3−0.3× l 4+0.3× l 5+0.7× l 6−-0.7× l 7−0.8× l 8−0.5× l 9. In Table 2, for each emotion, we have provided the mean of the subject means, SD of the subject means (between subjects), and mean of the subject SDs (within subjects). All questions in the various domains (eg, diet and physical activity) described in the Instruments and Measures section, except the emotions items, were considered as potential predictors of the outcomes. To reduce the large number of predictors in the model, we combined some of the variables within the same domain to create the following new variables as predictors: number of servings of healthy diet, number of servings of sweets, and number of minutes of total physical activity. In Table 3, for each momentary predictor, we have provided the mean of the subject means, SD of the subject means (between subjects), and mean of the subject SDs (within subjects). It is important to observe that even though the EMA emotions questions asked about present emotions, predictors were recalled values from the previous day.

In that sense, the predictors are not strictly momentary, but will be referred to as momentary variables for statistical modeling and analysis. Individual demographic characteristics (ie, age at the onset of the EMA study, sex, and race) were considered time invariant for the duration of the study. Race was dichotomized as white and nonwhite, as 94.5% (147/155) of participants were either white or African American individuals.

In the general statistical model for the analysis, for each outcome, we denote the response on the t^th assessment from the i^th subject by Y_it , the value of the j^th demographic predictor (out of k total number of predictors) from the i^th subject by X_ij and the value of the momentary predictor on the t^th assessment from the i^th subject by Z_it . Then the hierarchical model can be presented as follows:

Analyses of the associations between momentary variables and valence and arousal were performed, controlling for the 3 demographic predictors of age, sex, and race. The results for the valance and arousal outcomes are presented in Tables 4 and 5, respectively (all the P values correspond to 2-tailed tests). For the predictors with an unadjusted P value<.05, we also presented Bonferroni-corrected P values (in parentheses). None of the demographic variables predicted either outcome at a statistically significant level after the extreme Bonferroni adjustment. However, the effect of race on valence barely missed significance after adjustment (unadjusted P value=.007; adjusted P value=.08), with white individuals reporting much lower valence scores on average.

Minutes spent doing physical activity the previous day was a statistically significant predictor of both valence and arousal, with expected higher scores for increased physical activity. Time spent in meaningful group interaction the previous day was not a statistically significant predictor of either valence or arousal. Time spent in meaningful one-on-one social interaction the previous day was a statistically significant predictor of both valence and arousal, with expected higher scores for more interaction time. Hours spent sitting the previous day was a statistically significant predictor of both valence and arousal, with an expected lower score for an increase in time spent sitting. Number of total servings of fruits and vegetables consumed the previous day was a statistically significant predictor of both valence and arousal, with expected higher scores for greater servings. Number of total servings of sugar-sweetened beverages and desserts the previous day was not a statistically significant predictor of either valence or arousal. Adherence to medication the previous day was a statistically significant predictor of both valence and arousal, with higher scores for adherence. Any tobacco usage the previous day was a statistically significant predictor of only valence; on average, smoking a higher number of cigarettes resulted in lower valence scores.

Even though the model controls for demographic characteristics in analyzing the effect of momentary variables on valence and arousal, it is worthwhile to explore how much influence the demographic predictors have on the momentary predictors. A strong influence of demographic predictors on the momentary predictors can make the regression coefficients unstable and hard to interpret. Unlike in a standard multiple regression framework, in our hierarchical model, the influence cannot be measured directly by studying the multicollinearity properties and other standard regression diagnostics. Instead, the amount of influence can be indirectly measured by analyzing two additional models: one with only demographic predictors and one with only momentary predictors. The change in values of the estimated regression coefficients in the full models compared with the two isolated models described above can be used to assess the influence and the robustness of the coefficients.

For the sake of brevity, we did not present the actual results from the two isolated models here, but the results are remarkably consistent with our findings from the combined model in the previous section. Not only do the statistical significances of the momentary predictors match closely but also the individual estimates of the regression coefficients are surprisingly close. The individual estimates of the regression coefficients are very close for the demographic predictors as well. The observed consistency provides fairly strong evidence on the orthogonality of the demographic predictors from the momentary predictors.

These findings provide an important glimpse into factors that affect valence and arousal in a population of individuals residing in supportive housing. To our knowledge, this is the first study to examine the connection between emotions and other factors among people with mental health disorders and a history of chronic homelessness. This underserved population is often excluded from research studies due to co-occurring mental and physical disorders, resulting in substantial gaps in our understanding of their health and health behaviors.

Our analyses provide a number of observations about the relationships between health behaviors and subsequent emotions. First, we found that physical activity was significantly associated with positive emotions the following day. This finding is consistent with the literature showing the association of moderate physical activity with improved and maintained mood [27] as well as decreased symptoms of depression and anxiety [28]. We also found a positive relationship between fruit and vegetable intake and emotions the following day. Similarly, a study using data from the Canadian Community Health Survey found a significant association of fruit and vegetable intake with lower odds of depression and psychological distress [29]. Taken together, these results argue that a coordinated program to improve physical activity and diet should be a fundamental part of health interventions for people with mental health disorders and a history of chronic homelessness.

We also found that smoking cigarettes had a negative effect on valence the following day. Although nicotine may have a calming effect due to the inhibition of negative emotions such as anger [30], our results suggest that this effect may not carry forward to the following day. Research has suggested that nicotine dependency exacerbates stress [31], and a meta-analysis of changes in mental health after smoking cessation revealed that smoking cessation is associated with reduced depression, anxiety, and stress, with effect sizes equal to or larger than those of antidepressant treatment for mood disorders [32]. Thus, individuals receiving treatment for mood disorders may benefit from concurrent smoking cessation therapy.

We also found a strong relationship between the amount of time spent in individual social interactions and emotions. Interestingly, “time spent on meaningful one-on-one social interaction the previous day” was strongly associated with arousal and valence, while the “amount of time spent interacting in a group setting” was not significantly associated with emotions. This finding was unexpected given the substantial evidence that social support predicts the quality of life in many areas [33]. However, supportive housing residents are encouraged to attend support groups that address lifestyle skills, chronic disease management, and substance use. For this population, it may be that group interactions do not contribute to emotions unless the individual feels personally connected to at least one other person in the group. Thus, group interactions by themselves may not predict emotions, while individual interactions outside of the group setting may be one indicator of healthy, rewarding relationships.

In the analysis of the effect of demographics on the association between momentary predictors with valence or arousal, demographic variables had minimal effects on the regression coefficients of the momentary predictors, even when statistically significant. Hence, it is reasonable to conclude that the demographic predictors operated almost independently of the momentary variables in terms of influencing emotions.

Finally, it would be possible to study the association of emotional affect with subsequent same-day behaviors, for instance, examining the effects of emotions in the morning on smoking or drinking later in the day. Such analyses are beyond the scope of this paper, but we plan to investigate these associations in a future manuscript.

Our study had a number of limitations. Notably, our protocol included only daily morning assessments. Thus, we were not able to examine within-day variability. However, unlike other EMA studies, which typically run for a few days to, at most, a few weeks, our study ran up to 334 days with an average of 156 days of monitoring among all participants. This allowed us to examine associations for a much longer period than most other EMA studies. In addition, our results are generalizable only to a population of individuals residing in PSH with a history of homelessness and mental health issues. It is unclear whether the findings are generalizable to other groups of people with mental health problems, let alone the population in general. Relatedly, all participants self-reported depression or a mental health condition at baseline, which may have affected emotions and mood independently of other behavioral measures. For instance, the average client reported a score of 12.62 on the Patient Health Questionnaire, indicating that most clients felt at least moderate levels of depression upon admission to the program. Finally, we cannot rule out the possibility that participating in the coaching intervention affected the relationship between behaviors and emotion. Our results must be interpreted in the context of the larger services that people were receiving in this program. Further study with a more diverse population is necessary to make any broader assertions.

Despite the limitations, our study offers an important glimpse into health behaviors that affect daily emotional arousal and valence of persons with a history of chronic homelessness and mental health problems. One of the goals of the m.chat program was to provide individual support and assistance in meeting health goals. Because mood was an important target of the program, identifying factors that predicted positive affect can help improve future iterations of programs like this. To that end, identifying modifiable behaviors associated with negative and positive moods is a first step toward improving stability and preventing future homelessness. Understanding factors associated with mood and behaviors, particularly in vulnerable populations such as formerly homeless individuals, can also help providers design more targeted treatment plans and provide more appropriate referrals to ancillary care services [34].

Notably, many of our findings are consistent with “common wisdom” drawn from other populations. Behaviors generally considered to be positive (eg, physical activities, consumption of fruits and vegetables, adherence to prescribed medication, and one-on-one social interaction) tended to enhance positive affect and restrain negative affect. The opposite was true for behaviors considered to be negative, such as smoking. In fact, the positive and negative impacts on physical health for most of these behaviors are well established, and it is noteworthy that their effects on positive and negative affect appear to be consistent with previous literature with other populations. In a separate analysis of m.chat program data, Holmes et al [35] found that participants who consumed the least Western-style foods (eg, fast food, sugar-sweetened beverages, and processed meat) had significantly lower depressive symptoms over 1 year than those who consumed the most Western-style foods. However, Holmes et al also reported no significant association between depressive symptoms and nutritious food intake and physical activity over 18 months. Because we found a relationship between fruits and vegetables, exercise, and next day mood (but not sugar and mood), the relationships between these variables may be somewhat different in the short versus long term. Such behaviors can be targeted with a goal of enhancing positive affect and restraining negative affect in order to improve the overall mental health of individuals. Such an intervention has the potential to improve treatments for individuals with mood and other psychological disorders.