Mean Levels and Variability in Affect, Diabetes Self-Care Behaviors, and Continuously Monitored Glucose: A Daily Study of Latinos with Type 2 Diabetes

Abstract

Objective

This study investigated between-person and within-person associations among mean levels and variability in affect, diabetes self-care behaviors, and continuously monitored glucose in Latinos with type 2 diabetes.

Methods

Fifty participants (mean age=57.8 [SD=11.7] years, 74% women, mean A1c=8.3% [SD=1.5%]) wore a “blinded” continuous glucose monitor for 7 days, and they responded to twice-daily automated phone surveys regarding positive affect (PA), negative affect (NA) and self-care behaviors.

Results

Higher mean levels of NA were associated with higher mean glucose (r=.30), greater % hyperglycemia (r = .34) and greater % out of range glucose (r = .34). Higher NA variability was also related to higher mean glucose (r = .34), greater % hyperglycemia (r = .44) and greater % out-of-range glucose (r = .43). Higher PA variability was related to lower % hypoglycemia (r = −.33). Higher mean levels of self-care behaviors were related to lower glucose variability (r = −.35). Finally, higher self-care behavior variability was related to greater % hyperglycemia (r =.31) and greater % out-of-range glucose (r =−.28). In multilevel regression models, within-person increases from mean levels of self-care were associated with lower mean levels of glucose (b = −7.4, 95% CI: −12.8 to -1.9), lower % hyperglycemia (b = −.04, 95% CI: −.07 to −.01), and higher % hypoglycemia (b = .02, 95% CI: .01 to .03) in the subsequent 10-hour period.

Conclusions

Near-to-real time sampling documented associations of glucose with affect and diabetes self-care that are not detectable with traditional measures.

INTRODUCTION

The primary aim of diabetes treatment is achievement of glucose levels as near to normal as possible for the prevention of long-term complications and associated early mortality. Psychosocial characteristics, particularly patient affect and self-care behaviors, are important factors in achieving glucose targets. Yet, most of what is known about these factors comes from decontextualized, retrospective, once-off self-reports of affect and behavior combined with static measures of diabetes control that may be suboptimal for the research question at hand. Affect, diabetes self-care behaviors, and glucose levels are each quite dynamic, exhibiting short-term changes on the order of minutes-to-hours. Examination of the inter-relationships among these factors might be highly informative. The study reported here employed an ecologically valid approach to investigating relationships among daily affect, daily self-care behaviors and continually measured glucose among Latinos with type 2 diabetes over 7 days. Specifically, we tested whether between-person differences in mean levels and variability in affect and diabetes self-care behaviors were associated with mean levels and variability in glucose. We also tested whether within-person (within-day) deviations from mean levels of affect and diabetes self-care behaviors predicted subsequent glucose levels and glucose variability.

Historically, there have been two main methods for assessing diabetes control, i.e., glycosylated hemoglobin A1c (A1c) and self-monitoring of blood glucose (SMBG). A1c reflects average blood glucose levels during the preceding 4 weeks to 3 months. The major limitation of A1c is that it is solely a measure of central tendency, not variability, and is not sensitive to small or recent changes in glucose. SMBG provides a snapshot of current glucose levels. Its main limitations are that blood glucose levels that occur at times when SMBG is not performed are not observed and, because SMBG data tend to be highly skewed, calculation of standard deviation of glucose is problematic. The advent of continuous glucose monitoring (CGM) now allows minute-by-minute investigation of changes in glucose levels. CGM systems collect and store glucose data in an ongoing fashion for several days at a time. The main advantages of CGM are its ability to identify fluctuations and trends that are unobservable with A1c or SMBG as well as its appropriateness for calculation of glucose variability.

Similarly, in behavioral diabetes research, diabetes self-care behaviors and affect are usually measured with static, retrospective measures. For example, the Summary of Diabetes Self-Care Activities which assesses self-reported behaviors over the past 7 days, and the Center for Epidemiological Studies Depression scale (CES-D) which assesses depressive symptoms over the past 7 days, have both been shown to be associated with A1c (1,2). Yet, inherent to retrospective self-report measures are problems of recall error, bias, and the tendency for current affect to influence reporting of past experience. Also problematic is the decontextualized nature of the psychosocial questionnaires, which are often administered in a research or medical setting. Furthermore, clinicians and patients are well aware that self-care behaviors vary (3) and relatively short-term changes in diabetes self-care behaviors can affect glucose. Some behaviors have glucose effects on the order of hours, such as eating and medication skipping. Other changes may be on the order of days to weeks, such as changes in eating and physical activity due to season (4), school holidays (5) and religious fasting (6). Affect, also, is by nature a dynamic phenomenon. Changes in affect are a normal part of the human experience. Frequent or intense changes may reflect affective dysregulation, a characteristic of mood, anxiety, and personality disorders (7). Trait measures of affect variability have been developed (e.g., (8)) yet, agreement between recalled affect variability and real-time mood changes has been shown to be poor (9).

Whereas A1c, SMBG, and retrospective self-report measures of behavior and affect are each useful and perfectly appropriate for certain research questions, we suggest that research questions pertaining to short-term associations and variability are better addressed with measures that can more adeptly capture near-to-real time changes. CGM paired with frequently repeated sampling of behavior and affect allows the investigation of dynamic, within-person associations. In this study we employed a micro-longitudinal design to pair electronic self-reports of diabetes affect and self-care with CGM for 7 days.

The relationships among affect, self-care and glucose may be particularly important for Latinos, who are almost twice as likely to have diabetes compared to non-Latino Whites, with the highest prevalence among Puerto Ricans (10). Latinos with diabetes have lower levels of glycemic control than non-Hispanic Whites (11) and they are twice as likely to be hospitalized for diabetes related complications (12). Compared to their non-Latino White counterparts with diabetes, they also have higher rates of emotional distress (13,14) and although data vary by sampling strategy and the specific sub-ethnicity under investigation, Latinos as a group have been shown to have lower levels of the diabetes self-care behaviors (15,16) that are powerful determinants of glucose control (17).

To summarize, this study investigated relationships among mean levels and variability in affect, diabetes self-care behaviors, and glucose across approximately 7 days in a primarily Puerto-Rican sample of Latinos with type 2 diabetes. Three research questions were examined: (1) Across all observations, are mean levels of affect and diabetes self-care behaviors associated with mean glucose levels and glucose variability? (2) Across all observations, are variability in affect and diabetes self-care associated with mean glucose levels and glucose variability? (3) Do within-person deviations from mean levels of affect and diabetes self-care behaviors predict short-term changes in glucose mean levels and glucose variability?

METHODS

Participants and Procedure

The UConn Health Institutional Review Board approved all procedures. The study reported here was a sub-study of the Community Health Workers Assisting Latinos Manage Stress and Diabetes (CALMS-D) trial (18,19) which tested a group stress management intervention delivered by community health workers for Latinos with type 2 diabetes. Data for the substudy were collected between 2012 and 2014 and participants were enrolled in the substudy upon their completion of the trial. Our outcomes of interest, i.e., mean and SD CGM, did not differ between treatment arms, nor did treatment arms differ on A1c, p = .37.

CALMS-D participants were recruited from the ‘Brownstone Clinic,’ an outpatient clinic at Hartford Hospital, serving low-income patients with diabetes, approximately 80% of whom are Latino. CALMS-D participants were adult Hartford residents, self-identified Latino or Hispanic, Spanish-speaking, ambulatory, with type 2 diabetes ≥6 months, and most recent HbA1c in the past year >7.0%. Chart review excluded patients for medical instability or intensive medical treatment, bipolar disorder or thought disorder, or suicide attempt or psychiatric hospitalization in the past 2 years. Face-to-face screening excluded recruits for alcohol problems as indicated by an elevated score on the CAGE questionnaire (20) or enrollment in another research study.

CALMS-D participants were recruited for this sub-study over the phone by community health workers who had already worked with them as part of CALMS-D. They had rated their literacy on a 5-point scale from 1= “excellent” to 5= “I cannot read at all” and their numeracy on 5-point scale from 1= “excellent” to 5= “I cannot read or write numbers.” Because of the literacy and numeracy demands of the IVR and CGM protocols, individuals who responded 5 to either question (N=3) were not invited to participate in the sub-study. In view of the intensity of the sub-study protocol, CALMS-D participants who had been difficult to contact or schedule for CALMS-D study visits or were lost to follow up were also not contacted for participation in the sub-study (N=12). Other reasons for exclusion included hearing loss that would make IVR phone call responding difficult (N=3), multiple medical conditions or treatments that would make the protocol unduly burdensome (N=19), schedules that would interfere with IVR reporting (e.g., shiftwork, N=6), and one CALMS-D participant had died. The remainder (N=4) were eligible but declined participation.

Of the 107 CALMS-D completers, 58 were enrolled in the sub-study and N=3 withdrew or did not follow the protocol. Of the 55 sub-study completers, 5 experienced technical difficulties with the CGM, leaving N=50 with simultaneous IVR and CGM data used in analyses reported here. Among the sub-study participants, 27 had been randomly assigned to receive the CALMS-D stress management intervention, and 23 had been assigned to a control condition. Participants in the sub-study were younger than their counterparts who took part in the parent study only (57.8 vs 62.7), p<.05. Groups did not differ on sex (p=.51), high school graduation rate (p=.51), years in the mainland US (p=.16), A1c (p=.15), insulin use (p=.85), or depressive symptoms per PHQ-8 (p=.70, (21)).

IVR Procedures

During a home visit, after informed consent for the sub-study, community health workers (CHWs) trained participants how to use the IVR, the survey for which was offered in English or Spanish. CHWs reviewed each question with participants who then practiced answering both the morning and the evening surveys. Participants who did not own or want to use their own phone were provided study phones. All were provided headsets to facilitate hands-free keypad responding. Participants were provided with a “cheat sheet” of response options to facilitate efficient responding. CHWs called the participants several times over the 7 days to promote adherence to the protocol, answer questions and resolve any problems. Participants were provided the phone number of the CHW to contact in case of any difficulties.

IVR reporting windows were set for 8–10 AM and 8–10 PM. Twice daily assessments were chosen to balance density of data collection and participant burden. Specific reporting windows were chosen as times when this population would likely not be otherwise occupied (e.g., with early morning preparing grandchildren for school, late afternoon dinner preparations). The morning window allowed the capture of experiences from early morning and from the previous night when night eating syndrome may occur (22,23). The evening window allowed the capture of the day’s experiences when most diabetes self-care would be expected to occur. The IVR system called participants at random times during these 2-hour windows. If the call was unanswered, the system continued calling regularly within the time-window. If the participant could not complete the IVR self-reports at that time (e.g., if on another phone call) they could use a keypad response to indicate that they should be called back in 15 minutes. Postponed calls were permitted until the end of the reporting window after which the report was coded as missing. If a call was disconnected, the system called back to resume the survey. IVR data were monitored daily to detect protocol non-adherence. After 7 days of IVR data collection, the CHW returned to de-instrument and debrief the participant and deliver the appropriate level of compensation. A graduated incentive system was used to promote IVR protocol adherence. Each assessment was designed to take less than 10 minutes. Participants were compensated $4 for each IVR survey, with a $2 bonus for responding to both in a given day, for a possible of $10 per day across 7 days, for a total possible of $70.

IVR Self-report Measures

Participants could choose whether to hear the IVR questions in English (18%) or Spanish (82%). Questions and response options were pilot tested and found to be clear, meaningful, and acceptable to the target population.

Affect

Participants were asked about their affect since the previous IVR assessment, including nervous, mad, sad, bored, enthusiastic, happy, calm, and relaxed according to the circumplex model of emotion (24). They responded using a 3-point scale (1= “not at all”, 2 = “a little”, and 3 = “a lot”). Positive affect (enthusiastic, happy calm and relaxed) and negative affect (nervous, mad, sad and bored) composites were created by averaging together responses to the corresponding items. Internal consistency (Cronbach’s alpha) estimates across all daily observations was .76 for negative affect and .85 for positive affect.

Diabetes self-care

Participants were asked about their diabetes self-care behaviors since the previous IVR assessment (did you check your blood sugar?, did you eat healthy foods?, did you eat meals and snacks on time?, and, were you physically active?). Response options were ‘yes’ or ‘no.’ Affirmative responses were summed for a composite diabetes self-care measure from 0–4.

Continuous Glucose Monitoring

The CGM system used was MiniMed iPRO (Medtronic MiniMed, Northridge, CA, USA) which yields 288 glucose readings per day, with a lowest detectable level of 40 mg/dL and a highest detectable level of 400 md/dL. Residing inside a permeable membrane, a subcutaneous electrode sends interstitial glucose measurements to a monitor every 10 seconds. Twice per day, participants check blood glucose using a standard glucose meter (provided) and go about their normal daily activities. Participants were blinded to CGM glucose readings to minimize reactivity.

During a home visit, the CHWs inserted the CGM and trained participants in its calibration. Whenever possible, this home visit occurred at least 24 hours prior to the visit for training the participant on IVR to ensure that the CGM was properly inserted and calibrated prior to beginning IVR data collection. Participants received health promotion products (e.g., toothbrush and toothpaste) as an incentive for wearing the CGM.

Data Management and Analysis

To address questions concerning between-person associations, the following five glucose composites were calculated by aggregating across all available CGM glucose assessments: (1) the overall mean levels, (2) the proportion of readings indicating hypoglycemia (%<70 mg/dL), (3) the proportion of readings indicating hyperglycemia (%>180 mg/dL), (4) the proportion of out-of-range readings (% either <70 mg/dL or >180 mg/dL; the term ‘out-of-range’ here refers to any glucose value that falls outside of the target range of 70–180 mg/dL) and, (5) the standard deviation of CGM glucose readings. Whereas there are a large and growing number of indices for glucose variability using CGM data (25,26), many of them are redundant (and in fact our preliminary analyses confirmed redundancy of numerous indices). Therefore we chose SD as an intuitive and clinically meaningful index that is appropriate for type 2 diabetes. CGM outcomes were calculated from raw CGM data (i.e., not extracted from the CGM summary report) so that glucose cutoffs stated above could be applied to the specific timeframes under investigation. For the IVR affect and self-care behaviors, we calculated mean levels and standard deviations by aggregating across all available morning and evening IVR reports across all days. Pearson correlations were estimated between the CGM outcomes and the IVR composites.

For our questions concerning within-person associations, CGM data were restructured to coincide with the twice-daily IVR reports of affect and diabetes self-care behaviors. As shown in Figure 1, we created concurrent and prospective glucose time intervals to align with the morning and evening IVR assessment windows. Specifically, we calculated the glucose outcomes (mean levels, standard deviations, proportion of hyperglycemic readings, proportion of hypoglycemic readings, and the proportion out-of-range) during the 2-hour IVR assessment windows (i.e., 8:00 AM to 10:00 PM and 8:00 PM to 10:00 AM) and during the 10-hour windows immediately following the morning and evening IVR windows (i.e., 10:00 AM to 8:00 PM and 10:00 PM to 8:00 AM, respectively). The 10:00 AM to 8:00 PM glucose window captured glucose during waking time when affect and diabetes self-care would be expected to occur and exert any influence on glucose. The 10:00 PM to 8:00 AM window captured nocturnal hyperglycemia, which is a strong contributor to overall poor glycemic control because it represents long-term exposure to glucose, and nocturnal hypoglycemia, which is an acute and significant clinical problem.

Figure 1.

Daily assessment windows.

Note: IVR = interactive voice response. IVR assessments included negative affect, positive affect, and diabetes self-care behaviors since the previous IVR assessment.

Given the non-independent nature of the repeated measures, we used multilevel regression models (27). Specifically, we predicted glucose outcomes occurring in the 10 hours following IVR assessments, adjusting for glucose outcomes occurring during the previous 2-hour window corresponding to the IVR assessment. To evaluate within-person effects of daily affect and self-care behavior on the glucose outcomes, we person-mean centered the predictors, i.e., subtracted individuals overall mean across all observations from the IVR assessment value under analysis. Thus, the effects of the daily predictors would correspond to changes in glucose outcomes from the assessment window to the subsequent 10-hour period as a function of deviations from individuals’ mean levels of affect or self-care behaviors. Individuals’ mean levels of the predictors were also included in the models to account for between-person variation.

Covariates

For all of the study variables, there were no significant (p ≤ .05) differences across intervention conditions, thus study condition was not included as a covariate in our models. We did adjust for time of IVR assessment by including a dummy code (0 = am, 1 = pm) in the prediction models. We also created a variable that coded for medication adherence. Specifically, we created a 3-level categorical variable reflecting whether participants took medication when they were supposed to (i.e., were adherent; 78.8% of the records), did not take medication when they were supposed to (i.e., non-adherent; 14.5% of the records) or did not need medication (6.7% of the records). This 3-level variable was recoded into two dummy codes to include in the multilevel models: adherent was the reference group (coded 0) compared to “non-adherent” and compared to “did not need medication.”

RESULTS

The sample is described in Table 1. The 50 participants were mostly female (74%), had a mean age of 57.8 years (SD = 11.7; range 21 to 83), had lived in the mainland U.S. for mean = 33.0 years (SD=14.0), and 38% spoke only Spanish. A minority (26%) had obtained a high school or advanced degree, most were not working outside the home (73% retired or disabled, 11% laid off, 11% homemakers) and modal monthly household income was $501–$1000. Mean A1c=8.3% (SD=1.5) and 57% used insulin.

Table 1.

Sample characteristics

Characteristic	N (%) or mean (SD)
Sex (female)	36 (72%)
Age (in years)	58.8 (11.9)
Marital status (married)	16 (32%)
Education level <high school	36 (72%)
Employment status
Unemployed	8 (16%)
Retired	8 (16%)
Homemaker	5 (10%)
Disabled	24 (48%)
Monthly household income
0 – $1000	31 (62%)
1001 –$2000	17 (34%)
2001 – $3000	2 (4%)
Spanish speaking only (yes)	19 (38%)
Years in mainland U.S.	33.0 (14.0)
Waist-to-hip ratio	1.0 (0.1)
Body mass index	34.8 (9.8)
Insulin using (yes)	28 (56%)

Note. N = 50

Waist to hip ratio = waist circumference in cm divided by hip circumference in cm

Body mass index = weight in kilograms divided by height in meters squared

The descriptive statistics for the core study variables are shown in Table 2. These variables were derived from all available assessments (across all days) of which participants had a mean of 1569.0 (SD = 347.7) glucose assessments and a mean of 11.0 (SD = 2.7) IVR assessments (78.6% IVR adherence). During reporting days, IVR assessments were approximately evenly distributed across time of day (i.e., 51% morning and 49% evening). They were also approximately evenly distributed across days of the week; 14% per day would reflect completely even distribution across days and we found a high of 16.6% on Fridays and a low of 11.0% on Mondays.

Table 2.

Descriptive statistics for between-person associations

			Correlations
			Glucose mg/dL	Glucose mg/dL	Glucose mg/dL	Glucose mg/dL	Out of Range Glucose
	M	SD	M	SD	<70	>180
Glucose mg/dL (mean)	193.86	52.30
Glucose mg/dL (SD)	58.87	23.08	.447**
Proportion glucose <70 mg/dL	0.02	0.03	−.266	.397**
Proportion glucose >180 mg/dL	0.50	0.27	.937**	.478**	−.188
Proportion Out of Range Glucose (<70 or >180 mg/dL)	0.52	0.26	.917**	.536**	−.058	.991**
Positive Affect Mean	2.17	0.47	−.112	−.108	.095	−.113	−.103
Negative Affect Mean	1.22	0.27	.304*	.121	−.064	.338*	.335*
Diabetes Self-care Behaviors Mean	3.13	0.88	−.213	−.345*	−.071	−.234	−.247†
Positive Affect SD	0.34	0.19	−.088	−.238†	−.330*	−.097	−.142
Negative Affect SD	0.20	0.17	.341*	.044	−.190	.444**	.426**
Diabetes Self-care Behaviors SD	0.66	0.44	.218	.307*	.130	.261†	.283*

Note. N = 50. mg/dL = milligrams per deciliter,

^**p ≤ .01,

^*p ≤ .05,

^†p ≤ .

The number of glucose assessments was negatively related to the proportion of glucose assessments <70 mg/dL, (r = −.44, p = .002). This association is likely an artifact since hypoglycemia (the numerator of the proportion) was an uncommon event and the denominator of the proportion is higher for those with more glucose readings, thus individuals with a lower number of overall readings would yield a higher proportion of hypoglycemia. The number of IVR assessments was not related to other study variables.

As shown in Table 2, glucose levels were high (mean=193.9 mg/dl) and variable (SD=52.3 mg/dl) as would be expected in a sample with diabetes, with 50% of readings in the hyperglycemic range and 2% in the hypoglycemic range.

Between-Person Associations

These analyses tested associations between mean and variability indices of all available glucose values and all available reports of negative affect and self-care behaviors, across all days. As shown in Table 2, higher mean negative affect was associated with higher mean glucose (r = .30) and higher proportion hyperglycemic (r = .34) and out-of-range readings (r = .34). Higher mean self-care behaviors were associated with lower glucose SD (r = −.35). Higher positive affect SD was associated with lower proportion of hypoglycemic readings (r = −.33).

Higher negative affect SD was associated with higher mean glucose (r = .34) and greater proportions of hyperglycemic (r = .44) and out-of-range readings (r = .43). Finally, higher self-care behavior SD was associated with higher proportion of hyperglycemic (r = .31) and out-of-range readings (r = .28). Examination of scatter plots revealed that these associations were not driven by potentially influential outliers.

Within-Person, Prospective Associations

The multilevel regression analyses tested for changes in glucose outcomes from the IVR assessment window to the subsequent 10-hour period as a function of deviations from individuals’ mean levels of affect or self-care behaviors, after adjusting for time of IVR assessment, medication adherence and individuals’ mean levels of the predictors. There were 519 morning and evening observations with corresponding IVR and glucose assessments (M = 10.4 observations per person, SD = 2.7) available for analysis. Table 3 shows the results from the models in which we included positive and negative affect and self-care behaviors, along with the covariates, simultaneously. For parsimony, we show only the effects of the person-mean centered daily affect and self-care behavior predictors. As shown, only within-person changes (i.e. deviations from mean levels) in self-care behaviors predicted subsequent glucose. Specifically, a 1 point relative increase from individuals mean levels of self-care behavior was associated 7.4 point decrease in mean glucose (95% CI −12.8 to −1.9). Similarly, relative increases in self-care behaviors were associated with lower proportions of hyperglycemic readings (b = −.04, 95% CI: −.07 to −.01) and higher proportions of hypoglycemic readings (b = .02, 95% CI: .01 to .03) in the subsequent 10-hour period. Finally, we re-estimated the models without the medication adherence dummy codes; all of the significant effects remained and the coefficients were generally unchanged.¹

Table 3.

Within-person, prospective prediction of glucose from previous affect and diabetes self-care behaviors

				95% CI
Dependent Variable	Predictor	B	SE	LL	UL	p
Glucose mean	Negative affect	3.307	8.211	−12.828	19.443	.687
	Positive affect	2.297	5.280	−8.080	12.674	.664
	Self-care behaviors	−7.381	2.799	−12.883	−1.880	.009
Glucose SD	Negative affect	−3.101	3.086	−9.166	2.964	.316
	Positive affect	2.348	1.985	−1.553	6.249	.238
	Self-care behaviors	−0.158	1.052	−2.225	1.909	.881
Proportion Glucose mg/dL <70	Negative affect	0.011	0.015	−0.019	0.041	.463
	Positive affect	0.000	0.010	−0.019	0.019	.991
	Self-care behaviors	0.018	0.005	0.008	0.028	.001
Proportion Glucose mg/dL >180	Negative affect	0.008	0.046	−0.081	0.098	.854
	Positive affect	0.035	0.029	−0.023	0.093	.233
	Self-care behaviors	−0.041	0.016	−0.072	−0.011	.008
Proportion Out of Range Glucose mg/dL	Negative affect	0.020	0.043	−0.064	0.104	.642
	Positive affect	0.035	0.028	−0.019	0.089	.201
	Self-care behaviors	−0.024	0.015	−0.052	0.005	.104

Note. B = unstandardized slope; SE = standard error; CI = confidence interval; LL = lower level, UL = upper level.

DISCUSSION

This home-based study exploited the capabilities of CGM and near-to-real time sampling of affect and behavior to study mean levels and variability of affect and self-care behaviors as they relate to mean levels and variability of glucose. There are three main sets of findings from this study.

First, in between-person analyses, mean affect and affect variability were associated with mean glucose and glucose variability. Specifically, participants reporting higher mean levels of negative affect exhibited more hyperglycemia and more out-of-range glucose readings. Participants reporting more variable negative affect had a higher proportion of hyperglycemic and out-of-range glucose values. Participants reporting more variable positive affect had a greater proportion of hypoglycemic values. We interpret these findings as suggesting that hyperglycemia is associated with more negative affectivity whereas hypoglycemia is associated with less positive affectivity.

A handful of reports have shown associations between mean levels of negative affect and glucose over days and weeks (e.g., 28,29). But, to our knowledge, only one investigation studied both variability in negative affect and variability in glucose in the same study (30). Lansing et al. (2016; 30) followed n=180 adolescents with type 1 diabetes for 14 days during which they completed diaries of negative affect and used a glucometer. Their results indicated that higher negative affect variability was associated with higher glucose SD. However, calculation of glucose variability from SMBG is problematic. Also, even perfect SMBG adherence still leaves the vast majority of daily glucose values unobserved. Thus, our use of CGM allows greater confidence in the observed link between affect and glucose variability.

Second, in between-person analyses we found that mean levels of self-care and self-care variability were associated with mean levels of glucose and glucose variability. Participants with higher self-care exhibited more stable glucose. Participants with inconsistent self-care had more hyperglycemia and more out-of-range glucose values. We are not aware of any previous study that has examined patterns of self-care over short periods of time or the association between irregular self-care and glucose. It should be noted that, for our between-person associations, temporal precedence was not established and causality cannot be inferred.

Third, within-person, prospective tests showed that daily changes in self-care (i.e., deviations from one’s own mean levels) were associated with temporally proximal changes in glucose in the following 10-hours. Specifically, increases in self-care predicted less subsequent hyperglycemia and, perhaps surprisingly, more subsequent hypoglycemia. It is notable that this within-person association between higher self-care and greater risk of hypoglycemia was not observable from between-person data. This finding in particular underscores the importance of distinguishing levels of analysis and the potential problems with drawing within-person inferences from between-person associations.

In addition to research implications for behavioral sampling and design, our findings also have clinical implications. Whereas reducing hyperglycemia is the overall clinical goal of diabetes treatment, avoidance of acute episodes of hypoglycemia and reducing glycemic variability are also very important (31,32). Improvements in glycemic variability lead to improved quality of life (33) and there is early evidence that blunting daily glucose fluctuations improves some of the mechanisms involved in the development of long-term complications (34,35). Our data also underscore that attempts to achieve euglycemia carry an inherent risk of iatrogenic hypoglycemia. Patients who are encouraged to lower blood glucose should be well educated in predicting, preventing, detecting, and treating hypoglycemia. Hypoglycemia is potentially dangerous and our data suggest that it is associated with lower positive affect, potentially punishing self-care attempts. Fear of hypoglycemia is a common reason for avoiding physical activity and insulin use and for allowing glucose levels to run higher than clinically recommended (36).

In contrast to the within-person findings for self-care behaviors, within-person deviations in affect were not prospectively associated with subsequent changes in glucose. This is in contrast to Skaff et al.’s study of 206 participants with type 2 diabetes that showed men, but not women, displayed an association between negative affect and next-morning hyperglycemia across 21-days (29). Reasons for our lack of such findings might include that our study was not powered to detect moderation by gender, their measurement included 10 negative affective states (6 more than ours) and their data collection period was 2 weeks longer than ours. Furthermore, Skaff et al. relied on SMBG which is problematic for reasons outlined above. Additionally, our null effects for affect predicting changes in glucose do not rule out other possible causal links such as the effect of changes in glucose and glucose variability on subsequent affect. Indeed, Hermanns et al. (2007; 28) followed 36 type 1 patients wearing CGM for 48 hours and found that high glucose values were associated with both negative and positive affect ratings 60 minutes later. These few studies taken together start to make clear the complexity of the dynamic processes linking glucose and affect and underscore the necessity of additional studies using micro-longitudinal designs to help elucidate these associations.

Limitations

The non-experimental nature of our design does not allow causal inference. Because the sample reported here was recruited from a single urban locale, was primarily Puerto Rican and the majority were women, our findings may have limited generalizability. Having more within-day self-reports of a broader range of experiences and exposures would have allowed for additional hypotheses to be tested regarding within-day variability and temporal sequencing of exposure, response, and glucose outcomes. However, the potential benefit of additional within-day reports had to be balanced against the added participant burden.

CGM technology measures interstitial glucose which may have a time-lag compared to blood glucose levels (of 10 minutes or longer). Whereas it cannot be ruled out, given our study’s relatively large glucose windows, such a lag time is unlikely to have appreciably influenced our results. Although participants were blinded to CGM data, they could not be blinded to the twice daily CGM calibrations, so participant reactivity to twice daily blood glucose values cannot be ruled out.

The limitations to this study are generally outweighed by the its strengths including the use of CGM rather than SMBG, the measurement of both affect and self-care behaviors, adjusting for medication necessity and adherence, examination of variability in all study variables, and investigation of a hard-to-reach and understudied population with known disparities in glucose control. Indeed, these data represent a major step forward in understanding diabetes management in Latinos who currently comprise 12.5% of the U.S. population, likely rising to 25% by the year 2050 (37). Access to culturally and linguistically appropriate mental health care and diabetes self-management education is limited. Yet, our findings suggest that affect, self-management and glucose are linked in clinically important ways. These findings, together with CALMS-D findings previously reported (18,19), call for adding mental health care to behavioral diabetes treatment models that have traditionally focused mainly on SMBG, nutrition, and medication adherence (38).

Future research should examine a larger and more diverse sample, and to the degree that participant burden allows, should sample self-reports more intensively including perceived causes of affective ratings. Characteristics of patients who exhibit more variability in affect and glucose should also be investigated.

CONCLUSIONS

Assessment of near-to-real time behaviors and experiences combined with continual glucose monitoring can reveal dynamic biobehavioral relationships with clinical relevance. Researchers are encouraged to more fully exploit these capabilities, especially in underserved populations, to increase understanding of basic diabetes management processes and to inform clinical intervention.

Acronyms

SMBG

self-monitoring of blood glucose

CGM

continuous glucose monitoring

SD

standard deviation