Social support describes the provision of resources intended to benefit a receiver’s ability to cope in times of need [16]. The most prominent functions of support are emotional (eg, comforting) or instrumental (eg, practical assistance) [17]. Recipients’ reports of support received can be distinguished from providers’ reports of support given [18]. The perspectives of receiver and provider do not necessarily closely correspond, even when support receipt and provision concern the same reported support from 2 different persons [19]. There is a growing body of research reporting positive associations between received support and health behaviors, such as dieting or increased physical activity in the general population (eg, studies by Scholz et al [20], Darlow and Xu [21], Courneya et al [22], and Molloy et al [23]). In the context of chronic illness, spouses or partners are mostly the main sources of support [24]. Support from spouses has been associated with positive outcomes for patients with chronic health conditions, such as healthier eating habits among diabetic patients [25], increased health behavior (eg, eating a healthy diet, engaging in physical activity, and avoiding highly stressful situations) in cardiovascular patients [26], and decreased risk behaviors in HIV patients [27]. In this research area, only few studies have examined social support and health behaviors on a day-to-day basis (for an exception, refer the studies by Khan et al [7], Iida et al [28], and Stephens et al [29]). Diabetes symptoms are very sensitive to lifestyle behaviors and fluctuate daily [30]. Therefore, T2DM offers an ideal context for investigating associations between partner involvement, well-being, and daily illness management. For example, focusing on daily processes revealed that relative to the previous day, spouses’ diet-related support was associated with increases in patients’ adherence to diet [29] or that daily spousal support was positively associated with patients’ daily physical activity [7]. Therefore, it seems that social support provided by a partner is especially beneficial for T2DM patients’ illness management.

Research on received social support and coping with stress has shown that being in a supportive relationship can buffer physical and psychological effects of illness-related stress in general [31] and regarding T2DM in particular [28]. In line with this finding, Stephens et al [29] found that spouses’ diet-related support was associated with decreases in diabetes-specific distress of T2DM patients. Contrary to this result, studies on received support and indicators of well-being oftentimes result in negative associations between received support and well-being (eg, studies by Bolger et al [32], Gleason et al [33], and Seidman et al [34]). As a consequence, Bolger et al [32] introduced the dyadic concept of invisible social support, which is assumed to provide the benefits of support receipt without including potential costs. According to Bolger et al [32], support is invisible to recipients when the supportive acts occur outside of their awareness (ie, one partner takes care of unexpected housework without telling the other) or the recipient may be aware of the act but may not code it as support (ie, one partner purposefully gives advice in an indirect way so as not to draw attention to the recipient’s distress or his/her inability to deal with the stressful situation). To date, the few studies that have investigated invisible support in prospective diary designs (eg, studies by Bolger et al [32], Shrout et al [35], Maisel and Gable [36], and Biehle and Mickelson [37]), observational studies (eg, studies by Howland and Simpson [38] and Girmeet al [39]), and experimental designs (eg, studies by Bolger and Amarel [40] and Kirsch and Lehman [41]) have yielded support for beneficial effects on well-being and encourage further research.

The applicability of invisible support for health outcomes has been discussed (eg, studies by Kirsch and Lehman [41], Taylor et al [42], and Westmaas et al [43]). Until now, only 2 daily diary studies focused on behavioral responses to invisible support [44,45]. In single-smoker couples, higher invisible support was associated with less distress, but with more daily cigarettes smoked after a self-set quit date [44]. In dual-smoker couples only for men invisible support was associated with less distress, but not with more smoking after a joint self-set quit date [44]. Thus, based on these first 2 studies, it seems as if invisible support can also counteract the negative effects of visible support on well-being in the context of health behavior change. However, the results with the health behavior itself emphasize the need for further understanding the interplay between invisible support and health behaviors in different health contexts. Until now, no study was conducted in the context of chronic illness, and the relevance of invisible support in T2DM with regard to diabetes-related health behaviors and well-being needs yet to be demonstrated.

Most studies so far assessed invisible support by calculating a composite score from 2 self-reports (target person and partner). On the one hand, researchers assessed invisible support dichotomously (eg, studies by Bolger et al [32] and Shrout et al [35]). In this approach, invisible support was coded when the target person reported no receipt of support, but the support provider reported provision of support. On the other hand, researchers calculated continuous invisible support by subtracting received support reported by the target person from provided support reported by partners (eg, studies by Biehle and Mickelson [37] and Lüscher et al [44,45]). Instances in which the recipient reported receiving more than the provider reported giving were collapsed to zero. However, these measures are often criticized not only because invisible support is based on self-report but also because it is merely calculated and thus potentially only a hypothetical construct. Therefore, studies in which invisible support is measured directly, observed, and coded in everyday life as has already been done in the laboratory (eg, studies by Howland and Simpson [38] and Girme [39]) are strongly needed.

The usual approach followed by social support researchers is unidirectional in that support provision of one individual (usually the healthy one), and support receipt of another individual (usually the unhealthy one) is in focus. Invisible support already introduces a dyadic conceptualization in that both perspectives of partners are usually needed for the assessment of invisible support. CDC, however, goes beyond this dyadic implementation of still individual behaviors in that CDC explicitly focuses on the joint, dyadic efforts a couple undertakes to overcome challenges and problems (eg, a study by Bodenmann [11]). There are different approaches to CDC, but all these conceptualizations share the view that CDC involves a we approach with regard to stress (eg, studies by Bodenmann [11], Acitelli and Badr [46], and Kayser et al [47]) and health behaviors (eg, studies by Johnson et al [48], Lewis et al [49], and Lipkus et al [50]). Thus, CDC explicitly refers to the couple’s perspective and comprises joint efforts to cope with a stressor at the couple’s level. In the context of diabetes management, this could mean that the couple solves all issues related to the patient’s change of behavior together. Overall, research on CDC consistently results in positive associations with well-being and relationship quality (eg, studies by Bodenmann et al [51] and Traa et al [52]). Results of the association between CDC and health behavior, however, are mixed. Some studies report positive associations between CDC and health behaviors (eg, studies by Johnson et al [48] and Rohrbaugh et al [53]; for an experimental approach, refer to the study by Lipkus et al [50]). Along similar lines, within the context of T2DM, Seidel et al [9] showed that shared expectations regarding partner's diet-related involvement were positively associated with diet adherence in male patients. Benefits of a dyadic approach of patient’s and partner’s illness representation of T2DM were also documented by Dimitraki and Karademas [54]. Yet, another experimental study in the context of diabetes management does not support the hypothesis that CDC interventions are superior to individual interventions, albeit the dyadic group proved to be better than the control group (eg, studies by Trief [55]).

There are several ways in which CDC can be measured. The usual way of assessing CDC outside the laboratory is through self-report (eg, using the respective subscale of the Dyadic Coping Inventory) [56]. At the same time, a coding system for laboratory observations exists [56]. Another objective alternative is the assessment of we-talk by counting the use of first-person plural pronouns (eg, a study by Rohrbaugh et al [57]). Indeed, in a study by Rohrbaugh et al [57] that distinguished between we-talk and self-reported CDC, authors found that it was the objectively measured we-talk, but not the self-reported communal coping that predicted heart failure symptoms and general health. Thus, this study sets out to measure CDC with 2 assessment methods combined: self-report and we-talk on a daily basis in the context of dyadic diabetes management.

With regard to an in-situ assessment of the theoretical constructs of this study (ie, visible support, invisible support, and CDC), mobile phone applications are a powerful tool for several reasons [58-64]: the widespread use of smartphones and smartwatches with various sensors and touch-based graphical user interfaces makes sophisticated assessments of theoretical constructs appealing and widely applicable. Second, the combination of sensor data of smartphones and smartwatches (eg, from the global positioning system sensor or microphone) and their proximity to their owners offers the ability to detect useful contextual information (eg, the geographic position or mood of the owner). Third, mobile phone applications are scalable, cost-effective, have low entry barriers, and are applicable to different target populations. Finally, mobile phone applications reach people in their everyday life and with an immediacy that observations using conventional research methods do not have. In recent years, researchers have increasingly begun to use smartphones and smartwatches as platforms for the assessment of health behavior. However, although there exist various mobile phone applications to monitor behaviors and outcomes related to diabetes management in general (eg, studies by Sun et al [65], Schembre et al [66], and Wang et al [67]) and individual facets of it such as physical activity (eg, studies by Joosen et al [68] and Bort-Roig et al [69]), nutrition behavior (eg, studies by Celis-Morales et al [70], Turner‐McGrievy et al [71], and Hassannejad et al [72]), well-being (eg, studies by Dubad et al [73] and Servia-Rodríguez et al [74]) or conflict in couples (eg, a study by Timmons et al [75]), mobile phone applications that use objective sensor data in combination with self-reports are so far not used for the ambulatory assessment of social support and CDC.

A prominent example of a mobile phone application for the unobtrusive assessment of natural language and communication in real life is the electronically activated recorder (EAR) [76]. The EAR collects audio snippets at random times that can be coded with regard to the content of interest for the respective studies. For example, there are applications of the EAR with regard to social support provision in couples coping with breast cancer (eg, a study by Robbins et al [77]). The focus of the EAR, however, is on auditory observation only. Novel sensor–based approaches of affect recognition (eg, studies by Betella and Verschure [78], Maass et al [79], Venkatesh et al [80], van der Heijden [81], Chapaneri and Jayaswal [82], Revathy et al [83], Koolagudi and Rao [84], Heron and Smyth [85], Spanier [86], and Diener et al [87]) can be used in combination with appropriate self-report scales such as the affective slider [78] to better understand outcome parameters of well-being in the context of diabetes management.

A number of open questions remain from the current literature on visible and invisible social support, CDC, and its ambulatory assessment by mobile phone applications that will be addressed in this study. First, no study has examined the 3 concepts of visible and invisible social support and CDC in 1 study. Thus, the unique contributions of these constructs on health behaviors and well-being have not yet been examined. In particular, with regard to health behavior change, there is insufficient knowledge on the effects of invisible social support. The second open question in the current literature on invisible support and CDC concerns the assessment in everyday life. Measures of invisible support are often criticized for being not only merely based on self-report as measures of visible support but also merely calculated from independent reports of receipt and provision of support. Thus, invisible support is potentially only a hypothetical construct, and studies in which invisible support is measured directly, observed, and coded in everyday life are strongly needed. With regard to CDC, most studies so far focused on either cross-sectional associations or longer-term associations between CDC and health behavior. Associations in everyday life using an ambulatory assessment approach have been neglected so far. Given the assumed importance of invisible support and CDC for health behaviors involved in diabetes management and well-being in T2DM patients and their partners, it is of key importance to assess these constructs in a reliable and valid objective way in everyday life to further our knowledge. Third, it is still to be investigated how to design an ambulatory assessment application for the purpose of this study, which is not only accepted by study participants in their everyday situations [79-81] but also delivers high-quality data streams that are good enough or even comparable with distinct devices (eg, high-quality microphone for affect recognition from speech in the laboratory). Fourth, available speech databases and latest research on affect recognition from speech employ usually role-taking actors and thus lack natural settings and, with it, external validity (eg, studies by Chapaneri and Jayaswal [82], Revathy et al [83], and Koolagudi and Rao [84]). Finally, multimodal approaches to affect recognition are promising, but existing research is sparse, and consistent results and approaches are still to be explored [75]. This study will address all these open questions and limitations of previous research.

The aims of this study are to address the open questions outlined above. The first aim is to examine the impact of visible and invisible support and CDC on couple’s dyadic management of T2DM (ie, health behaviors) and well-being. The second aim is to use an improved assessment approach of visible and invisible support and CDC in everyday life, which is based on observational and self-report data in situ instead of self-report only. The third aim is closely related to the first and second aims: that is, to develop an AAMC that allows to record both multimodal sensor data streams and self-reports related to the study’s core constructs in situ to better understand visible and invisible social support and CDC and the relationship between the recorded sensor data and psychological self-reports. To better understand outcome parameters of well-being in the context of diabetes management, the fourth aim is to use a multimodal affect recognition approach (ie, speech, facial expression, and heart rate variability) in combination with self-report scales.

The first aim of the study will be examined in daily life as well as in an observational setting in the laboratory, whereas the second and fourth aims refer to the experience sampling phase of this study. The results of the third aim are finally used to support the first two and the fourth aims by capturing and assessing the core study constructs by self-reports and objective measures during the experience sampling phase. The following research questions result from these study objectives: (1) What are the unique contributions of visible and invisible social support and CDC for diabetes patients’ health-related behaviors involved in diabetes management (physical activity, diet adherence, and medication adherence)? (2) What are the unique contributions of visible and invisible social support and CDC for indicators of well-being derived from multimodal data sources and self-reports captured by AAMC in diabetes patients and their partners?

This study protocol describes the design of the Dyadic Management of Diabetes (DyMand) study, funded by the Swiss National Science Foundation (CR12I1_166348/1). The study was approved by the cantonal ethic committee of the Canton of Zurich, Switzerland (Req-2017_00430). To address the aims mentioned above, this study comprises an intensive longitudinal design with 2 phases of data collection. The first phase is an experience sampling phase in romantic couple’s everyday life (7 days) following an ecological momentary assessment (EMA) approach [85]; the second phase is an observational study phase in the laboratory. Figure 1 shows the study design.

The target population of this study are 180 patients with T2DM and their romantic partners. Inclusion criteria are the medical diagnosis of T2DM of the target person with prescribed oral antidiabetic drugs and having a partner of the opposite sex without diabetes or a psychological disorder who is also willing to participate in the study. The participating couples should be in a close, committed relationship for at least 1 year and living together in 1 household for at least 6 months. Exclusion criteria include T2DM treatment with insulin, inpatient treatment, shift work of one or both partners, and insufficient knowledge of the German language.

Recruitment of patients will take place in several hospitals in Switzerland. Moreover, couples will be recruited by means of flyers in medical clinics, hospitals, private practices, pharmacies, and Schweizerische Diabetes-Gesellschaft (Swiss Diabetes Society), information provided to physicians who will inform patients actively about the study, and Web-based forums for diabetes patients and diabetes-related websites. Furthermore, we plan to invite patients through radio and television formats and health magazines.

Objective and self-report data from both partners are collected and assessed throughout the study to allow focusing on effects of both partners.

The main research aims of this study refer to between-person associations of visible support, invisible support, CDC, patient’s diabetes-related health behavior, and well-being of both partners using a dyadic approach to account for the interdependence among couple members using a 2-level statistical model for distinguishable dyads as indicated for patient-partner dyads [114]. The main analyses will be correlations and multiple regression analyses, which will be performed in SPSS and R. For the diabetes-related health behaviors, which relate to physical activity, diet, and medication adherence, we will generate a daily composite score, indicating the meeting of the recommendations for all 3 behaviors together ranging from 0 (meeting none of the recommendations) to 3 (meeting all of them). Moreover, we will also be able to analyze associations between predictors and the different behaviors assessed continuously in separate analyses. The idea of the composite score, however, takes into account that the real-life assessment method we chose to capture invisible and visible support and CDC using objective measures might result in highly ecologically valid and reliable measurements, but potentially in a rather low frequency of these predictors for the different diabetes-related health behaviors. With regard to well-being, we will consider the affective valence and arousal assessed during the experience sampling sequences or the mean scores of positive and negative affect from the end-of-day assessments. On a more exploratory level, we will also analyze day-to-day within-person and within-couple associations in further analyses using multilevel modeling. But because of the novelty of our approach and the rather short time frame of 7 days, this will not be the main focus.

To assess the relationship between the sensor data and self-reports, machine learning is applied, which is carried out in several steps. First, preprocessing of the raw sensor data involves feature extraction, feature scaling, feature selection, and dimensionality reduction. The resulting features derived from sensor data will be tested in machine learning models to predict self-reported affect and well-being. Second, the data will be split into training and test datasets to assess how derived algorithms generalize to new data [115]. The training dataset will also be split into subsets, where a k-fold cross-validation will be applied. The performance of the resulting model will then be evaluated using the test data set. This procedure will be repeated for various learning algorithms (eg, random forest, support vector machines, naive Bayes, recurrent neural networks, and feedforward neural networks). After comparing the performance across algorithms, the best overall model will be selected. We expect to produce a model that efficiently predicts affect and well-being using a multimodal compared with a unimodal approach as outlined in previous work on affective computing [99].

The sample size was calculated based on Cohen [116] and using the G*Power program [117] to secure adequate power for the primary outcomes. There are no meta-analyses available for the associations between visible support, invisible support, and CDC with health behaviors and indicators of well-being. Moreover, studies reporting results from diary studies use unstandardized effects. Thus, we base our power calculation on data from previous studies on these associations, but we are aware that the data basis is somewhat unsatisfactory to make strong conclusions about the expected effect sizes. Previous studies reported varying effect sizes for visible received support on health behavior (eg, r=0.29-0.34) [20,23]. There are only 2 studies so far that examined the association between invisible support and health behavior (eg, smoking) [44,45]. In these 2 studies, however, no standardized effect sizes are available because of the focus of within-person effects. In the case of visible and invisible social support and their relations to well-being, we draw on the experimental evidence available [40]. In a series of 3 experiments, Bolger and Amarel [40] demonstrated that visible support compared with a no-support control group was related to increases in distress (d=0.27 and d=0.66, ranging from small to medium effect sizes). Invisible support, in contrast, leads to lower distress (d=−0.63 and d=−1.09, indicating medium to large effect sizes). The difference between the effects of invisible and visible support resulted in a large effect (d=−1.09). Thus, medium effect sizes for the associations between visible and invisible support and indicators of well-being will be expected. In the case of CDC, previous studies report effect sizes with behavior of r=0.20 with exercise adherence, r=0.20 with dietary adherence in a sample of diabetes patients [48]. Moreover, in the domain of well-being, effect sizes range from r=0.20 to r=0.33 (eg, a study by Badr et al [118]).

As we will not only address bivariate associations between our predictors and criteria but aim at comparing effects between our predictors, our calculation of the power is based on a 2 dependent Pearson r’s analysis with a common index. To the best of our knowledge, no published data on the intercorrelations of visible support, invisible support, and CDC are available. Data from our own research with smoking-nonsmoking couples (excluding CDC) resulted in an association of r=−0.25 for visible and invisible support. To detect a significant association with a continuous outcome at P<.05 with a power of 1−beta=0.90, and assuming a small effect size of r=0.2, for health behavior and visible support, r=−0.2, for health behavior and invisible support and an intercorrelation of r=−0.25 between visible and invisible support, the required sample size is 164. Because effect sizes for CDC tended to be higher, the study will be powered for the smaller effect sizes of support. Previous studies applying intensive longitudinal designs resulted in very low rates of dropout even across a longer period (ie, ≤10%) [119-121]. This can be explained with a high commitment in couples willing to participate in these kinds of studies. On the basis of this experience, we will add 10% to the calculated sample size to account for potential dropout. This results in a total required number of 180 couples.

Analyses of the relationship between multimodal sensor data and self-reported data on affect and well-being are conducted with the help of machine learning. We estimate that in the best case, we will have an average of 10 completed self-reports per day for all 7 days resulting in 25,200 samples (360 individuals×10 self-reports×7 days). It is highly likely to have missing data if subjects do not complete the self-reports. Therefore, in a worst-case scenario with only 1 self-report completed per day and a corresponding 90% missing data, there will be 2520 observations left to use to train the machine learning algorithm. Owing to the fact that there exists no widely accepted sample size calculation method for machine learning approaches, the sample size of 360 individuals with up to multiple measurements on 7 consecutive days (eg, at least 2520 observations only from the end-of-day diary) lies above related work (eg, studies by Wahle et al [60] and Timmons et al [75]) and thus is assumed to be adequate for the purpose of this study.

In line with the open research data initiative of the Swiss National Science Foundation, the anonymized data of the study will be made public in a noncommercial database for replication purposes of the analyses, additional data analyses by any third parties (eg, other research groups), and quality control purposes, given that there are no ethical or legal restrictions.