Family Typology for Adults With Type 2 Diabetes: Longitudinal Stability and Validity for Diabetes Management and Well-being

Precision medicine seeks to identify and leverage individual differences in behavior, environment, and genes to maximize treatment effects (1). Application of these principles is required to tailor and evaluate treatments to improve diabetes management and reduce diabetes distress. Adults’ management of type 2 diabetes occurs in the context of close relationships (2,3) (herein family), which affect individuals’ motivation and abilities for diabetes self-management (3–7) and psychological well-being (8–11). Systematic reviews note that family interventions have mixed and often minimal effects on adults’ diabetes outcomes (3,12–14), emphasizing the disconnect between robust observational evidence and the lack of consistent intervention effects. Heterogeneous effects may be due to heterogeneity in family functioning among adults with diabetes. Matching and tailoring interventions to family functioning could maximize benefits.

Therefore, we pursued a typological approach for a concise and useful framework to organize adults across multiple dimensions of family functioning identified as relevant in observational research, including helpful involvement (including emotional support, instrumental support, and assistance) (4–6,15), harmful involvement (including nagging, arguing, undermining, and sabotaging behaviors) (5,6,15,16), autonomy supportive communication (17,18), perceived criticism (8), collaborative coping (19), and appraisals of family support (11,20). Using these dimensions of diabetes-specific family functioning, we empirically identified four types: 1) Collaborative and Helpful, 2) Satisfied with Low Involvement, 3) Want More Involvement, and 4) Critically Involved (21). Each type included individuals representing diverse family structures and sociodemographic characteristics, and type was cross-sectionally independently associated with diabetes management (diabetes self-efficacy, diabetes medication adherence, and hemoglobin A_1c [HbA_1c]) and psychosocial well-being (diabetes distress and depressive symptoms; all P < 0.01) (21).

Typological approaches have recently emerged to characterize family among adults with diabetes (22) and other chronic diseases (23), and one study found that a family typology predicted all-cause mortality over 6 years (24). However, to our knowledge, there have not been efforts to examine type stability nor validity for disease-specific outcomes over time. Longitudinal examinations of family typologies are necessary to inform applications to accelerate precision medicine approaches in behavioral and psychosocial research and care. For instance, understanding how type is associated with outcomes over time can inform intervention design and matching. Understanding if type is a dynamic state versus stable trait can inform the frequency of assessment and intervention tailoring, getting the right intervention to the right person at the right time (1). Therefore, we sought to 1) determine whether the typology was replicable and robust to sampling variations, 2) evaluate type stability over time and determine if instability was explained by measurement error, and 3) evaluate the validity of the typology over time for diabetes management and psychosocial well-being.

This was a planned secondary analysis of a two-arm parallel-group randomized controlled trial (RCT) designed to evaluate a diabetes self-management support intervention (25). We used baseline data from all RCT participants as cross-sectional data (n = 338) and follow-up data from participants assigned to the control group (n = 165) as longitudinal data. Participants were asked to complete surveys and a mail-in HbA_1c test (CoreMedica Laboratories, Lees Summit, MO; accredited by the College of American Pathologists) at baseline and 6 and 9 months postbaseline. Participants consented to review of their electronic medical record to extract HbA_1c results collected as part of regular care. Survey measures included self-reported demographic and clinical characteristics (age, gender, race, ethnicity, income, years of education, diabetes duration, and insulin status), typology dimension measures, and diabetes management and psychosocial well-being outcomes. All participants received print materials on managing diabetes and family support and text messages with instructions to access their study HbA_1c results.

The goal of our typological approach was to develop a concise and useful framework to organize family functioning specific to adults’ diabetes (26). Criteria for evaluating typologies include the following: types are replicated using different methods or in different samples, there are meaningful conceptual differences between types, and type membership predicts differences in meaningful outcomes (26). After determining the typology replicated in the RCT sample, we combined typology dimension data from the RCT baseline sample (n = 338) with the typology development sample (n = 379) (21), resulting in a combined sample of 717 participants. Combining the samples enhances the generalizability of findings and external validity of the typology. The typology development sample was predominantly non-Hispanic White (74%) and had lower HbA_1c values (mean [SD] 7.2% [1.5%]; 55 mmol/mol [16.4 mmol/mol]) and included individuals who did not participate in a self-management support intervention study, whereas the RCT sample over-represented individuals from minority racial groups (48%) and those with elevated HbA_1c values (mean [SD] 8.7% [1.7%]; 72 mmol/mol [18.6 mmol/mol]). Thus, the combined sample is more representative of adults with type 2 diabetes than either sample independently.

For both the RCT and typology development studies, adults (age ≥18 years) with type 2 diabetes receiving outpatient care at Vanderbilt University Medical Center (operationalized as having recent clinic visits and HbA_1c tests) who were community dwelling were invited to participate. Exclusion criteria were similar across studies, including needing help with daily activities and having diagnoses or ongoing treatments that may alter family dynamics around health/diabetes (e.g., pregnancy, current or recent abuse), reduce focus on diabetes management as a result of threats to longevity (e.g., hospice services, dialysis, concurrent cancer treatment, congestive heart failure), and/or interfere with validity of self-report measures (e.g., dementia, schizophrenia).

More details on eligibility, recruitment, and procedures for each study have been published (21,25). Both studies were approved by the Vanderbilt University Institutional Review Board (200398 and 210556) and included informed consent and compensation for data collection. All analyses were conducted using R (version 4.2.2). Data sets and the typology survey tool are available from the corresponding author upon request.

We assessed nine dimensions of diabetes-specific family functioning, with measures detailed in Supplementary Table 1. The Perceptions of Collaboration Questionnaire (19) assessed three dimensions: cognitive compensation (collaboration is needed to overcome perceived deficits), interpersonal enjoyment (collaboration provides encouragement), and frequency of collaboration. The Family/Friend Involvement in Adults’ Diabetes (15) measure assessed two dimensions: helpful and harmful actions (i.e., received support) relevant to the respondent’s diabetes self-care activities. Two dimensions represent helpful and harmful perceived support: autonomy support (others support their personal agency for diabetes) assessed with the Important Other Climate Questionnaire (27), and perceived criticism assessed with four items from the Family Emotional Involvement and Criticism Scale (28), adapted to be specific to diabetes. Two dimensions (a single item each) assessed appraisals, including the respondents’ appraisal of the effectiveness of their family functioning and satisfaction with their family functioning.

The Perceived Diabetes Self-Management Scale assesses diabetes self-efficacy by querying respondents’ confidence in their ability to carry out the multiple different behaviors needed to manage diabetes and predicts multiple different self-care behaviors (29). The Adherence to Refills and Medications Scale for Diabetes assesses the degree to which respondents experience different barriers to taking their diabetes medications and has been validated with other adherence measures and HbA_1c (30). We reverse coded this measure such that higher values indicate greater diabetes medication adherence. HbA_1c values most proximal to each assessment point were pulled from the electronic medical record if one was available within ±42 days; if not, we used a mail-in kit value when available. These kits have been validated against venipuncture (31), and sensitivity analyses in our prior work indicated that type of HbA_1c test does not affect findings (32). Across assessments, the mean (SD) absolute number of days between the survey assessment date and the date of the HbA_1c test was 15.6 (11.5), and the median was 12.7 days.

Diabetes distress was assessed with the Problem Areas in Diabetes five-item version, which has satisfactory sensitivity and specificity for the recognition of elevated emotional distress related to diabetes (33). Depressive symptoms were assessed with the Patient Health Questionnaire–8, which queries symptoms using diagnostic criteria and has satisfactory sensitivity and specificity for recognition of major depression (34).

Evidence of replication includes the number of types and patterns across dimensions similar to those identified in the typology development sample (21). In cluster analysis, the number of types is a determination made across multiple statistical approaches and examinations of dimension values, iteratively, using clinical or theoretical knowledge (detailed in Mayberry et al. (21)) to determine if the clusters are meaningful. We conducted k-means cluster analyses first on the RCT sample and then on the combined sample (N = 717). We assessed the recommended number of clusters with multiple approaches, including the elbow method, the gap statistic method, and a cluster tree. Across results from all cluster analyses, we normalized the dimension values using the combined sample mean and SD for like-to-like comparisons. For these analyses, we excluded participants missing more than one dimension (1.9% in the combined sample). We also used principal components analysis as a second method to examine whether the dimensions’ correlation structure was consistent across samples (Supplementary Fig. 1).

Next, we used the combined sample to evaluate how robust type assignment was to sampling variations. We performed consensus clustering by randomly selecting 90% of participants and eight of the nine dimensions for each of 1,000 replications (35). Then, k-means clustering was performed on each replication to generate a consensus matrix. We used a heat map to visualize the consensus matrix and evaluate the robustness of the original results to these sampling variations (Supplementary Fig. 2).

For longitudinal analyses, we refined our approach to type identification and used only RCT participants. The k-means clustering is useful for identifying meaningful clusters but not for assigning type to observations not used in the clustering model (e.g., follow-up assessments). Therefore, we constructed a multinomial logistic regression (MLR) model with the multinom function in R package nnet (36). The MLR included baseline type (assigned from clustering) as the outcome and baseline dimensions as predictors, returning the probability of membership for each type based on dimension values. Details on the MLR are provided in the Supplementary Material (Supplementary Figs. 3–11), including Stata and R syntax for scoring. We assigned type at follow-up assessment using the highest MLR probability. We then used these type assignments to examine the stability of type within participants over time using cross-tabulations (i.e., from baseline to each follow-up assessment and from 6 to 9 months postbaseline). This analysis sought to determine how consistent type assignment is over time for a given participant, which provides an indication as to whether the typology of diabetes-specific family functioning reflects a dynamic state or a stable trait. Only participants with more than seven of nine dimensions nonmissing at each assessment (78% of the longitudinal sample) were included in this specific analysis, because missing data or imputation methods would add noise to stability estimates.

Next, we examined if type instability was attributable to random changes in the underlying dimensions or reflected systematic changes in diabetes-specific family functioning. We used the mean and SD of observed differences between baseline and 6- or 9-month dimension values as parameters in 1,000 simulations. Simulations allowed dimensions to vary independently. We then applied the MLR to the simulated data sets to obtain type probabilities and assign type at each assessment. From this, we extracted the proportion of participants whose type remained stable in the context of independent variability in dimension values. Results across simulations provided our null distribution (Supplementary Fig. 12), built under the assumption that type is a stable characteristic and observed changes in type over time are due to independent naturally occurring changes in dimension values, consistent with measurement error. If observed stability fell within the null distribution, this would support the null hypothesis that type instability is attributable to random measurement error on included dimensions. If observed stability fell outside the null distribution, we would reject the null hypothesis in favor of the hypothesis that dimension changes over time are correlated and move systematically within participants, indicating the dynamic nature of the typology of diabetes-specific family functioning over time.

We examined longitudinal validity using two approaches to accommodate type as a dynamic state and as a stable trait. Under the dynamic state hypothesis, we ran a model in which both type and the outcome of interest were time varying. This model included baseline data from all RCT participants and longitudinal data (6- and 9-month assessments) for participants assigned to the control group only. Under the stable trait hypothesis, we examined if type at baseline was prospectively associated with outcomes of interest at 6 months. This model included data only for control group participants.

All models used generalized estimating equations with an identity link function and a working independence correlation structure to account for different numbers of assessments across participants. Models examined the effect of type on each outcome of interest, both unadjusted and then adjusted for a priori covariates: age, gender, race and ethnicity (non-Hispanic White, non-Hispanic Black, other), years of education, income, diabetes duration, and insulin use. For each outcome, omnibus P values for the effects of type were calculated with a joint Wald test with three degrees of freedom.

For outcome analyses, we imputed missing data by chained equations with 500 imputation iterations. Covariates collected at baseline had <1% missing, except for income, with 3.4% missing. Outcomes of interest assessed at baseline were missing 0.6% to 2.2% for self-reported measures and 7.8% for HbA_1c; missingness on these variables ranged from 16.3% to 18.1% at 6 months and 16.9% to 18.8% at 9 months. We conducted two separate imputation models: one for the intervention group at baseline only and one for the control group at baseline and 6 and 9 months.

Baseline demographic and clinical characteristics for the RCT baseline sample and the combined sample are listed in Table 1. Statistical approaches indicated three to five clusters would fit the data for the RCT baseline sample and, separately, for the combined sample. We compared dimension values across clusters under each of these solutions and determined both the three- and five-cluster solutions reduced meaningful differences present in the four-cluster solution (i.e., dimension values became less distinctive across types). Therefore, we determined that the four-cluster solution identified in the typology development study fit well. Under the four-cluster solution, defining characteristics (patterns on dimensions deviating greatly from the group mean) of each type were largely consistent across samples, demonstrating replication of the typology (Fig. 1). In the RCT and combined samples, type prevalence was consistent with the development sample as well: Collaborative and Helpful, ∼30%; Satisfied with Low Involvement, ∼25%; Want More Involvement, ∼30%; and Critically Involved, ∼15%. Type assignment across 1,000 replications indicated consensus for 97% of the sample, indicating type assignment was replicable and robust to variations in sampling and dimensions included.

Participants assigned to the control group (n = 165) in the RCT provided 6- and 9-month assessment data as well and had similar characteristics at baseline (Table 1). Among participants with seven or more dimensions at each follow-up assessment, ∼52% had a consistent type across assessments (baseline to 6 months, 55%; baseline to 9 months, 54%; and 6 to 9 months, 47%). Collaborative and Helpful had the highest stability (67–84%), whereas Critically Involved had the lowest (21–43%) (Table 2). There was a tendency for individuals to move into the Collaborative and Helpful type from both Critically Involved (33–43%) and Want More Involvement (24–38%) types.

The null distributions generated by simulations indicated a median 72% stability, ranging from 60% to 85%, expected due to measurement error (Supplementary Fig. 12). Our observed stability estimates (55% and 54%) fell outside the null distributions. This suggests changes in type over time reflect changes in respondents’ perceptions of family functioning (and likely actual changes in diabetes-specific family functioning), supporting the hypothesis that type is more of a dynamic state than a stable trait.

We used the Wald test with a P value of <0.05 to indicate the statistical significance of the association between type and each outcome of interest; contrasts between types and associated P values are also shown. In the longitudinal model, wherein type varied alongside outcomes over time, we found that type was associated with each outcome in both unadjusted and adjusted models (Table 3). In other words, associations between types and outcomes held even as individuals moved across types over time. Across outcomes, diabetes management and psychosocial well-being were higher among Collaborative and Helpful and Satisfied with Low Involvement and lower among Want More Involvement and Critically Involved. The only difference of note between Collaborative and Helpful and Satisfied with Low Involvement was lower diabetes self-efficacy among Satisfied with Low Involvement. Self-efficacy was lowest and HbA_1c values highest among Critically Involved and Want More Involvement, and diabetes distress was highest for Critically Involved. Supplementary Figure 13 illustrates unadjusted patterns of outcomes across types.

Longitudinal models using baseline type to predict outcomes at 6 months showed only self-efficacy was predicted by baseline type (Supplementary Table 2). These results are less informative, given the stability findings supporting the hypothesis that type is dynamic over time and the smaller sample sizes for these models.

The typology of diabetes-specific family functioning and these validation efforts advance the knowledge on family functioning and its effects on diabetes-related outcomes. First, the typology was robust and replicated across sampling and dimension variations and across samples with varying racial diversity and ranges of HbA_1c. Second, we found evidence that diabetes-specific family functioning type is a dynamic state rather than a stable trait. Only ∼50% of the sample received the same type assignment over time, and certain types (i.e., Critically Involved and Want More Involvement) were least stable and were also associated with less diabetes management and psychosocial well-being. Simulations indicated type instability was not explained by measurement error but instead likely reflected real changes. Third, in longitudinal models treating type as a dynamic state, we found independent associations between type and each outcome of interest, including measures of diabetes management (diabetes self-efficacy, diabetes medication adherence, HbA_1c) and psychosocial well-being (diabetes distress, depressive symptoms). Outcomes changed over time as type changed within participants, suggesting the typology may be a strong independent indicator across multiple outcomes. Finally, our typology can be applied as a diagnostic tool at the individual level with our MLR approach (Supplementary Material), as opposed to traditional typological approaches requiring large samples (e.g., cluster analysis, latent class analysis). Studies informing this work indicate family functioning is multidimensional, with complex interplay across these dimensions affecting outcomes (6,15,17,37,38). The typology allows us to organize this complexity to facilitate matching between individuals’ needs and treatment efforts. Across family typologies developed among adults (21,23,24) there have been four types consistently identified: a more collaborative type, a more distant/detached type, an unobligated/avoidant type, and a conflicted type, suggesting potential for generalizability across disease contexts.

Type patterns across dimensions and in outcomes reported herein are consistent with cross-sectional patterns reported in our typology development study (21). Collaborative and Helpful is indicated by high levels of collaboration and helpful involvement, with low harmful involvement and criticism, and this type had higher diabetes management and psychosocial well-being, followed closely by Satisfied with Low Involvement. Satisfied with Low Involvement is indicated by patterns of low collaboration and low involvement, combined with a lack of dissatisfaction in family functioning, and this type is also associated with higher self-management and well-being. Critically Involved and Want More Involvement are associated with low diabetes management and psychosocial well-being. Want More Involvement is indicated by lower scores across typology dimensions, like Satisfied with Low Involvement, but key differences pertain to the match between low levels of family involvement and the needs and desires of the individual with diabetes (as evidenced by higher scores on cognitive compensation, reflecting greater need for collaboration, and lower scores on effectiveness and satisfaction appraisals for Want More Involvement). This mismatch appears to be the key difference for Want More Involvement, which was associated with lower diabetes management and psychosocial well-being.

Our efforts to understand type longitudinally are unique and led to the findings that diabetes-specific family functioning is a dynamic process over time, affecting diabetes outcomes. Type changes over time are unlikely to reflect changes in whom respondents consider family. Prior evidence (21) indicates respondents are considering two or three individuals when responding to typology dimensions, with highly stable relationship types most common (partners/spouses, adult offspring, parents, siblings). Therefore, type changes are more likely attributable to changes in perceptions of and changes in family functioning, as well as in matching thereof with the respondents’ needs and desires over time. This is consistent with the concept of reciprocal determinism, within Social Cognitive Theory (39), which posits a dynamic and reciprocal interaction between an individual, their behavior, and their social environment over time (as opposed to the social environment affecting the behavior unilaterally). We see evidence for this in type changes over time and in our longitudinal validity analyses, wherein family functioning may be influenced by the individual’s appraisal of it and that interplay can affect diabetes outcomes dynamically over time. Rather than using the typology once to identify whom may benefit, it may be useful to think about repeated assessments to identify when type indicates individuals may benefit from intervention or which intervention may most meet their needs.

Aspects of our approach that weaken the typology’s generalizability include that the typology was developed and replicated using samples drawn from a large academic medical center in the mid-South, and our longitudinal analyses were conducted among individuals enrolled in an RCT to evaluate a diabetes self-management support intervention who are likely unique in important ways (e.g., help seeking, interested in a study with family aspects). Aspects strengthening generalizability include robustness to variations in sampling and the included dimensions in simulations, the racial diversity of our samples, and our combined sample representing a range of HbA_1c values. Future work should include efforts to determine if typology dimension structures hold within minority racial and ethnic groups, particularly those disproportionately affected by diabetes (e.g., among African American/Black adults and separately among Hispanic adults). Despite representation of these groups in our sample, we did not have sufficient a sample size to attempt replication of the typology within these subgroups. An important innovation is that our typology is now usable at the individual level, like a diagnostic tool; however, we acknowledge that the included dimensions are composed of 37 survey items, which may make application challenging. Furthermore, we identified substantial variability in type over 3-, 6-, and 9-month intervals, but future research should explore the timescale of type changes to inform the frequency of intervention tailoring. Future research should also explore other aspects of diabetes self-care, such as physical activity and dietary behaviors. Finally, the typology dimensions represent only the perspective of the individual with diabetes, rather than integrating perspectives from multiple family members. This limits the scope of perspectives included but makes application easier.

In their review of close relationships and the management of chronic illness, Martire and Helgeson (40) advised next steps for the field, which can be addressed with our typology, including 1) targeting a broad population of patients and tailoring the intervention to their needs based on relational and disease management indicators and/or 2) developing and delivering interventions specifically for at-risk families. We can apply this typology to inform the tailoring of the intervention form (family vs. individual) and content (relational vs. disease focused) by type. Our findings suggest two types may be better suited to an individual intervention (Satisfied with Low Involvement who may not want or need family intervention, and Critically Involvedwho may be harmed by increasing family engagement) and two types may be better suited to family intervention (Collaborative and Helpful who have engaged helpful family members, and Want More Involvement who want a more engaged family). The typology could be applied as an effect modifier within intervention studies to test these hypotheses. Type instability over time suggests dynamic (i.e., iterative, repeated) assessment and tailoring may be important to ensure interventions address current functioning as new challenges arise for families and individuals with diabetes. Our findings suggest we focus intervention development efforts on Critically Involved and Want More Involvement types but emphasize the importance of considering these as states rather than stable characteristics of individuals or their families. In conclusion, the typology is a useful tool to support the identification of individuals at times they are most in need of intervention.

Funding. This study was funded by a grant from the National Institutes of Health (R01-DK119282-S1).

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. L.S.M. wrote the first draft of the manuscript. L.S.M., S.Z., C.A.B., L.A.N., and R.A.G. were involved in the conception, design, and conduct of the study. All authors were involved in the analysis and interpretation of the results, and all authors edited, reviewed, and approved the final version of the manuscript. L.S.M. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. This work was presented at the 44th Society of Behavioral Medicine Annual Meeting, Phoenix, AZ, 28 April 2023.