Family Networks of Obesity and Type 2 Diabetes in Rural Appalachia

Abstract

The prevalence of obesity and diabetes has been studied in adolescent and adult populations in poor, medically underserved rural Appalachia of West Virginia. A web‐based questionnaire about obesity and diabetes was obtained in 989 family members of 210 Community Based Clinical Research (CBPR) trained adolescent members of a network of 18 science clubs, incorporating 142 families. After age‐correction in < 20 years old, 50% of both adolescents and adults were obese. The frequency distribution of obesity was trimodal. In the overall population 10.4% had type 2 diabetes, while 24% of adult, obese subjects had type 2 diabetes. A new metric, the family diabetes risk potential, identified a trimodal distribution of risk potential. In the lowest most common distribution 43% of families had a diabetic family member. In the intermediate distribution, 69% had a diabetic family member, and in the distribution with highest scores all the families had a diabetic member. In conclusion, the poorest counties of rural Appalachia are at crisis level with the prevalence of obesity and diabetes. The distribution of age‐corrected obesity and family diabetes risk potential are not normally distributed. We suggest that targeting individual family units at greatest risk offers the most efficient strategy for ameliorating this epidemic.

Introduction

There is currently a nationwide epidemic of type 2 diabetes. ¹ , ² , ³ , ⁴ Closely linked to the rising prevalence of diabetes is obesity, a risk factor for diabetes in both adults and, increasingly, in children. Obesity is a significant major public health concern in its own right. ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ The highest rates of obesity and diabetes are found among the population who are poor, have lower education, and are minority groups in the United States of America (USA) and its territories. ⁵ West Virginia (WV), which is the only state in the nation that lies entirely within the Appalachian mountains, contains a population characterized by high poverty, high unemployment, low education, and limited access to health care. ¹¹ , ¹² West Virginia is an epicenter for the obesity endemic, with the second highest obesity rate in the country after Puerto Rico. ¹² , ¹³ In the last decade, WV has reported the highest age‐adjusted incidence rate of diabetes of 13 per 1,000 population ¹¹ and the highest diabetes‐related deaths in the nation, along with a high prevalence of modifiable risk factors for shortened life expectancy. The higher than expected incidence of diabetes in WV Appalachian communities has been attributed to the extent of health disparities in these underserved populations. ¹³ , ¹⁴ , ¹⁵

The problem for the next generation is even more troubling. Family history is a commonly identified risk factor for the development of type 2 diabetes. ¹ , ² Among the children associated with diabetes, 45–80% have a parent with type 2 diabetes, and 74–90% report at least one affected first‐or second‐degree relative. ³ The prevalence of obese children and adolescents has more than doubled in the last decade ⁴ to be similar to that of adults. In WV, over 43% of children have previously been classified as at‐risk of becoming overweight. ¹² These overweight and obese children are at increased risk of developing diabetes and cardiovascular diseases in the future. Unless these trends are reversed, a third of all children born today can be expected to develop type 2 diabetes during their lifetime. ¹⁶

An initiative to better understand and modify the obesity epidemic within West Virginia has been launched in a community‐based participatory research (CBPR) program. This family‐oriented community research has elected to focus on the family structure and social network relationships as a potential target for future interventions. In order to define the initial baseline situation, a CBPR partnership has been developed between the community, a network of science clubs for high school age adolescents under the umbrella of the Health Sciences and Technology Academy (HSTA), and Clinical Translational Research (CTR) investigators from the Center for Clinical Pharmacology at the University of Pittsburgh (Pitt). The community selection of adolescents to HSTA club membership is based on an educational opportunity supported by a Science Education Partnership Award (SEPA) to help educate adolescents from the poorer members and minorities of this already disadvantaged rural population about diabetes. Within this initiative, 210 HSTA club members and their science club mentors, trained in the ethics and the conduct of CBPR, ¹⁷ have enrolled and studied family members of their community to establish the magnitude and relevance of the problem of obesity and diabetes within their own family and within their own community. It was recognized that this community has been stable for several generations. Initially populated by a migration of Scottish‐Irish pioneers that populated rural Appalachia as farmers in the 18th and 19th century, and later by an African American migration in the south of the State with the inception of the coal industry, it is a remarkably homogenous stable group within US context, from a cultural, social, and genetic perspectives. These data from responders in this community are analyzed using a novel approach that emphasizes the family as the unit of central interest to apply to future intervention studies.

Methods

The protocol design was developed during a summer workshop of club members in transition between 10th and 11th grade of high school, representing 18 HSTA Science clubs, HSTA staf, and investigators from the University of Pittsburgh. This was followed by education in principles of CBPR of lay HSTA club teachers, to be mentors at these 18 clubs and by education and training of all HSTA club members who agreed to participate in the ‘Diabetes Education to Protect and Defend of Families’ (DPD) project, at each of these 18 clubs (with on average 10 members per club). A common protocol with IRB consent/assent was approved by the University of West Virginia that covered all club sites dispersed throughout the West Virginia community. (This has been described in greater detail in this journal.) After each participating student had received training in the ethical conduct of clinical research, they were trained in an appropriate process to approach friends and adults to obtain oral and written informed consent. Each student was asked to provide his own informed assent, and obtain consent from his parents for study participation and consent for a HIPAA nondisclosure information form. Next, they completed their own structured family tree and social network.

The family tree was restricted to living members within a family unit who could be expected to influence each other's behavior, including only siblings in their own generation (not cousins), parents, uncles, and aunts in the parental generation and only grandparents in the generation above. This information included position in the family, age, gender, and zip code, but did not include any confidential health care information. Consenting individuals in the family tree were then individually approached to request their assent/consent, depending on age, to respond to conf dential health care‐related questions that enquired about self‐reported height and weight, in order to compute Body Mass Index (BMI) and self‐reported history of known diabetes. The information in this questionnaire was conf dential. The questionnaire was entered into a prepared electronic data form by HSTA club members. The collected club information was collated and transferred on a thumb drive to the HSTA administration, where questionnaires were coded, and identified data retained at that site. De‐identified coded data were transferred to the Center for Clinical Pharmacology at Pitt, and aggregated into a combined data set using an electronic Management for Clinical Translational Research (eM‐CTR) that has been developed at this site. The organized and integrated (de‐identified) data set was made available to club members and Pitt investigators for analysis.

Statistical analysis and presentation

Individual entries from questionnaires were inspected for errors and possible systemic inconsistencies using relationship plots (e.g., height vs. age, BMI vs. age, percent diabetes incidence vs. number of subject questionnaires collected in a geographic area, etc.). The first goal of this inspection was conf rming the reliability of the data collected by student researchers by showing their compliance to known/obvious or expected features (such as height‐age relationship for young subjects) or results of other studies (percentage of diabetes incidence in WV). The second rationale of this primary step (see below) was quantification of the noninformative relationships in the data and preparation of final input of informative data into detailed statistical analysis.

Height analysis

As expected, the height‐age relationship can be separated into < 20 years and ≥ 20 years intervals, based on inspection of the raw data plots. The height (H) for the < 20 years group was linearly dependent on age: eq. 1, (r= 0.63, p < 0.0001). For subsequent analyses, the actual height (H) of every subject was therefore age‐corrected, H using the linear regression of this subset.

(1)

The height of subjects of age ≥ 20 years was normally distributed around the constant mean height of 1.7 m. We therefore rescaled heights of subjects in this age group in the study cohort by the mean. These two independent correction functions provided continuous rescaling of all the subject heights into interval with normal distribution around 1.0.

BMI Analysis

In the context of the above results, inspection of the raw data plots of reported BMI confirmed the age‐dependence for the subjects younger than 20 years. The cohort was, therefore, separated into subgroups < 20 years and over ≥ 20 years. The BMI for the < 20 group was age‐corrected (BMI_N0RM) using the linear regression (r= 0.55, p < 0.0001, eq. (2)) of this subset:

(2)

Reported BMI for the ≥ 20 year age group was distributed around a mean of 30 ( Figure 3 ). We therefore normalized BMI of this group using eq. (3),

(3)

Figure 3.

Comparison between the total number of subjects recruited and number of diabetics recruited in each participating HSTA Science club.

Analysis of the distribution of BMI_N0RM in the study cohort

The frequency distribution for all age‐corrected BMI's was estimated by constructing the histogram of occurrence of 989 age‐corrected BMI values. Shapiro and Wilk's W‐test ¹⁸ rejected hypothesis about normality of this distribution with p < 0.0001. Inspection of the histogram revealed that the reason for nonnormal shape of the histogram is apparent multi‐modality. We therefore used nonlinear least square fitting (Spectracalc, ThermoGalactic) of the histogram to the sum of the minimal number of three Gaussian distribution functions that were observed to be needed to describe the histogram shape. To increase the power of the regression and reduce parameter ambiguity, we restricted all the fitted Gaussian functions to the same variance (σ= 0.15) that was consistent with the data histogram. Least squares fitting of the histogram calculated from data of complete cohort provided Gaussian functions with means at μ₁= 0.8, μ₁= 1.05 and μ₁= 1.35 age‐corrected BMI values, respectively. In the subsequent analysis steps, we investigated the distributions of study results in the subgroups of the cohort (e.g., diabetics vs. nondiabetics). We further reduced the number of degrees of freedom in the fitting of these subgroup histograms by sums of Gaussians by fixing both their widths to common values σ= 0.15 as well as their means at μ_1–3. This was shown in all cases to provide significant match between the regression function and the fitted histogram. This permitted the least square quantification of the differences between the histograms in terms of optimized estimates of three amplitudes for each distribution for different subject groups. In this way, we reduced the ambiguity of the fitted results and filtered significant information from the nonnormally distributed data.

Family network analysis

A family group was defined as a group in which complete data were available that comprised at least a boy or girl, their mother and their father. All additional members of the family with complete reported data were included in the analysis of that family.

We based the selection of mathematical tools for quantitative characterization of the family group as a basic, single informational unit upon the following requirements: (1) The ability to quantitatively capture both the network of social relations in the family, the age‐corrected BMI and presence or absence of diabetes for individual subjects. (2) Provide a mathematically rigorous way of combining the age‐corrected BMI information for individual family members into a (single) numerical family‐based descriptor, without losing network information, while capturing individual information explicably and quantitatively in a full family context. (3) Define a quantitative descriptor for comparisons of families or larger social units with diverse sizes and social structures.

The quantitative mathematical tool, which satisf es the above criteria, recruits from discrete mathematics and is the weighted graph for the family (Γ_FAMILY). ¹⁹ In this graph, each vertex V _i represents one family member (i) and the edge E_ij, connecting two vertices i and j, represents the edge as the social relationship between the two subjects i and j. Each edge in the family graph Γ_FAMILY has been assigned an age‐corrected BMI‐derived number (edge weight W(E_ij)).

For the purpose of integrating family acquired information CBPR data into a single data element, we developed a new metric, the family diabetes risk potential. Derivation of this novel parameter consists of a multi‐step sequence of operations:

In the first step, the values of the edge weights W(E_ij) are determined using the following steps: We assigned each family member to one of the three groups g ₁, g ₂ or g ₃ defined by her/his age‐corrected BMI and data‐driven intervals of age‐corrected‐BMI defined, using the Gaussian components that quantified the observed tri‐component distribution of age‐corrected‐BMI values (“trichotomization”). To define the boundaries of these age‐corrected BMI, we used the crossing points of the three overlapping Gaussian components that fit the overall distribution. This choice minimizes the unavoidable impact of the Gaussian component overlaps on the g_i‐classification of the subjects. Thus, subjects classified as g ₁ had age‐corrected BMI less than 0.95, subjects classified as g ₂ age‐corrected BMI between 0.95 and 1.21 and subjects classified as g ₃ had age‐corrected BMI greater than 1.21; we have identified the latter group as massively obese. These g_i. values of subject i were used to define the “potential” p_i of the corresponding graph vertex V_i.

In the second step, the age‐corrected BMI‐based category g ₁, g ₂, or g ₃ assigned to each subject was converted into the diabetes risk potential (β_i) assuming a standard exponential risk model in which g ₁, g ₂, or g ₃ are assumed to carry an empiric relative risk of diabetes of 1, 20 and 300, respectively. The choice of an exponential function is commonly used in disease‐risk models. In our particular case, it describes the change in frequency of diabetes with increasing BMI. The parameters were adjusted to adhere to significant separation between families with and without diabetes. These empirical relative risk values β_i are obtained from the 1, 2, and 3 values of g ₁, g ₂, or g ₃ by equation 4:

(4)

In the third step, we defined the weight of each edge W(E_ij) as the average of values of β_i of the two family members that the edge connects in the graph:

(5)

From a mathematical perspective, the Γ_f is a complete, weighted graph on N vertices, where N is the family size. Capturing the family‐related information into a weighted graph Γ_f can be used as an efficient communicative way to visualize complete so cial and health‐related information in the study cohort. For visualization in Γ_f plots, the size of the vertex is proportional to β_i, color of the vertices differentiates between diabetic and nondiabetic subjects, and the thickness of the edge lines (or their numerical labels) are defined by W(E_ij).

We have also used the family graph relating BMI and diabetes in the second stage of data analysis to derive the overall family diabetes risk potential in the final two steps of analysis:

In the fourth step, we quantitatively captured all age‐corrected BMI observations in the family graph into a single metric that characterizes that family and provides a measure that can be compared across families independently of variation in family size. The topology of a complete family graph for a large family is complicated. We have therefore used the rigorous optimization algorithm of discrete mathematics, of the maximal flow in a graph Γ_FAMILY, weighted by risk for diabetes. ²⁰ For the purposes of finding the maximal flow, the Γ_FAMILY was oriented. We selected the father of the family as the “source” vertex where the family maximal flow starts and the mother of the family as the “sink” of the sub‐graph, where the maximal flow ends. All W(E_ij) values defined the set of “throughputs” of diabetes risk flow in the family graph. Mathematical routines of Network package in Maple 12 (MapleSof, Waterloo Maple, Inc, Waterloo, Ontario, Canada) were used to calculate the maximal flow in the Γ_FAMILY. The weighted, oriented sub‐graph Γ_FAMILY can be also used as an efficient way to visualize the relevant social and disease related factors between the families. In our special case, the Γ_FAMILY is always a complete graph (all N vertices are connected by N‐1 edges to all remaining N‐1 vertices representing remaining family members). The topology of the maximal flow sub‐graph is also always identical. In the resulting maximal flow graph MF_FAMILY, father and mother vertices are connected directly by one edge of the MF_FAMILY graph; all other connections between these “input” and “output” vertices were composed by edge‐vertex‐edge features, oriented all in father, mother direction, one for each of remaining family members. The MF_FAMILY allows unambiguous definition of a single‐numbered characteristic of the BMI relations within each family, defined as the value of the maximal flow Φ_FAMILY in the FM_FAMILY graph.

In the last step of defining the family diabetes risk potential, (P_FAMILY), descriptor values comparable between families of different sizes are estimated by normalizing the value of the maximal flow Φ_FAMILY by the number of edges in the family maximal flow graph FM_FAMILY:

(6)

This family measure, where ρ_FAMILY is the family diabetes risk potential, allows direct comparison of the family diabetes risk potential between families, independent of the family size. It also provides a family index that is most sensitive to small changes in BMI of the subject at greatest risk, those with a high BMI.

Results

Analysis of individual subjects

In the fall of 2007 and spring of 2008, 25 HSTA clubs were involved in the project, with a total of 210 participating students. The total number of questionnaires handed out was 1,760. Out of these, 989 were returned, a response rate of 56%. Demographic characterization of our cohort was as follows: 40% male, 60% female, 75% Caucasians, 20% African‐Americans and 4.3% other. The age distribution was: under 20 years 33.5%, 20–40 years 20.1%, 40–60 years 32%, 60–80 years 13.4%, and older than 80 years 1%.

Height increased linearly with age between 5 years and 20 years, and thereaf er stabilized at a mean of 1.7 m. Normalization for age, using equation 1 (see methods), resulted in a height histogram with a shape defined by a narrow single Gaussian distribution with small variance ( Figure 1 ).

Figure 1.

Age‐corrected height frequency distribution in rural Appalachia.

Similarly, BMI increased with age between 5 to 20 years and thereaf er maintained a mean of 30 (previously presented in this journal). Age‐correction, using equation 2 (see methods), resulted in a mean approximately constant age‐corrected BMI in which 50% of both children and adults were defined as obese (age‐corrected BMI >1). Even after age‐correction, the distribution remained asymmetrical ( Figure 2 ). This asymmetry was best fitted to a minimal number of three Gaussian distributions with identical variances of 0.15. The means of these distributions were at age‐corrected BMI = 0.8, 1.05, and 1.35, respectively ( Figure 2a ).

Figure 2.

Normalized frequency distribution of age‐corrected BMI. (A) Left Panel = nondiabetic subjects (n= 885) (B) Right panel = diabetic subjects (n= 103).

Within this cohort, 103 subjects were self‐identified as having diabetes (10.4%). All of these diabetics had type 2 diabetes. Considering adults only, the frequency of diabetes was 13.9% and for obese adults 24.2%. The frequency distribution profile of age‐corrected BMI in these diabetic subjects also exhibited a nonnormal distribution that was consistent with the same three Gaussian distributions as nondiabetics with respect to means and variance ( Figure 2b ). However, the proportion of subjects in each distribution was different between nondiabetics and diabetics. The peak frequency in g₁ was less in nondiabetics than diabetics, while the peak frequency in the distribution and in the g₂ and g₃ was higher in diabetics. Thus, there was clear expected shif to the right in the proportion of diabetics with increasing age‐corrected BMI ( Table 1 ).

Table 1.

The frequency of diabetes in the three subgroups of age corrected BMI in rural Appalachia

Individual group	BMI group mean	Total number in cohort	% of group diabetic	Odds ratio	p value
G1	25.5	572	6	‐	‐
G2	31.5	314	16	2.67	0.01
G3	41.4	134	22	5.5	0.001

Using the frequency of diabetes in age‐corrected BMI distribution in g ₁ as a point of reference, the odds ratio for occurrence of diabetes increased in g ₂ to 2.67, and further dramatically increased in g ₃ to 5.5 ( Table 1 ).

Subject recruitment at different clubs has allowed us to determine the frequency in observing diabetes by comparing of the number of recruited subjects with diabetes to nondiabetics between clubs. The significant linear regression between the reported diabetes cases and the number of subjects recruited per club had zero intercept and a slope close to 10% ( Figure 3 ).

Family diabetes risk potential

Having analyzed the cohort of participants in this study as individuals, we next proceeded to an analysis of the family as the unit of interest. Within this cohort, there were 142 families with at least an adolescent, a mother, and a father. The majority of families had one or more additional members of their families (average 6.5 members per family). Figure 4 , lef panel, illustrates a representative example of a full weighted Γ_f for a family with two members having diabetes. The complexity of this network is substantially simplified in an oriented, maximal flow sub‐graph, Γ_FAMILY ( Figure 4 , right panel). In the family diabetes risk potential analysis, we identified 53 families with at least one report of diabetes. When the family diabetes risk potential (see methods) was compared between families with and without diabetes, there was a clear difference between diabetic and nondiabetic containing families ( Figure 5 ). The frequency distribution profile fell into three clusters. In the lowest and most common family cluster (P _family < 40) the proportion of families with diabetes was less than those without diabetes (43%). In the second cluster (P _family 40–125) the proportion of diabetic families was 69% while in the highest cluster (P _family > 125), 100% of families were diabetic.

Figure 4.

Examples of (A) full family tree represented as weighted un‐oriented graph for family with two incidences of diabetes (red vertices) and (B) corresponding maximal flow oriented sub graph. The size of each vertex is proportional to age‐corrected BMI of person represented by that vertex and the number on each edge it's weighting. (See methods).

Figure 5.

The frequency distribution of family diabetes risk potential in 55 families (open bars) or 87 families (solid bars) with diabetic members. Each group has been normalized within rural Appalachia.

Distribution of age‐corrected weight in clusters of family diabetes risk potential

The distribution of age‐corrected BMI of individual family members between the BMI subgroups and the family diabetes risk potential clusters confirmed the enrichment of g ₃ age‐corrected BMI subjects, with 37% in only 8 families, and 30% of diabetic families ( Table 2 ). When the intermediate cluster and upper cluster were combined, 100% of subjects in g ₃ and 50% of diabetics came from families with a diabetes risk score >25.

Table 2.

The number of subjects (diabetics) in each of the age‐corrected BMI subgroups in the low, intermediate, and high distributions of family diabetes risk potential.

	Number of families	Age‐corrected BMI subgroup			Total number of individuals
		g1	g2	g3
Age‐corrected BMI		<0.95	0.95–1.21	>1.21
BMI<25	99	260 (33)	115 (18)	− (1)	375 (52)
BMI = 25–120	35	177 (6)	111 (9)	64 (3)	352 (18)
BMI>120	8	22 (12)	17 (8)	37 (11)	76 (31)
Total	142	459 (51)	243 (35)	101 (15)	803 (101)

Discussion

This Community‐Based Participatory Research (CBPR) provides confirmation of progression of the prevalence of obesity and diabetes in dispersed, poor rural communities in Appalachia within the state of West Virginia. ¹ , ² , ¹¹ This study has identified an intriguing frequency distribution structure for age‐corrected BMI for both nondiabetics and diabetics that can be parsimoniously represented as three sub‐distributions. Among the distributions, the relative subject numbers in each distribution change between diabetics and nondiabetics, but the location of the mean peak and the variance within each distribution remains constant. The basis for this observation is currently unknown and, to our knowledge, has not previously been reported. Further studies are under investigation to determine if it can be explained on the basis of an interaction between family, genetic, and social factors. Finally, a novel CBPR investigator tool is developed to consider the family as a unit for risk assessment for family members to develop type 2 diabetes.

This baseline prevalence study was designed to assess the extent of the problem of obesity and diabetes in the families of the HSTA club network. It was not designed as a population study, as the overall population is not the relevant comparator for the need for change within this community. This design characteristic raises the relevant and repeatedly expressed question of generalizability. We emphasize that this study has been undertaken in the context of a rural Appalachian population. It is homogeneous with respect to social (poor) and geographic (rural) context, it has been stable with respect to recent migration shif s in population, and it was selected with a bias toward the poor and minorities (as a consequence of the HSTA mission), and toward families who actively promoted their adolescent of ‐spring toward a college education. The observations, therefore, only can be generalized to other populations with equivalent traits. As a specific illustration of one unanticipated local trait, all responders had both an adult male and female role model at home (we did not inquire about family lineage). This family structure is common to truly poor, rural communities where single parent family units are exceptionally hard to maintain. This contrasts to urban communities, where single parent units and single individuals are a more common social context. Even though these considerations limit the range of generalization, they deepen the inferences that can be drawn in a social structure where these characteristics are present. We also emphasize that even if the social context were very different, the methodology of defining relevant closely integrated social units could still be applied.

Reports of the prevalence of obesity in the community usually presents information about adults and children as separate entities, pointing out that each is increasing and making the point that the prevalence of obesity in children is catching up to adults. In the West Virginia community, this concurrent survey of both groups indicates that when BMI is age‐corrected in adolescents and compared to adults, the frequency of obesity has not only caught up to, but the proportion of severe obesity is even greater than among adults. A further ominous observation is that in both adolescents and adults, 50% of the population have now reached the definition of obesity. Thus, the prevalence is steadily progressing. ⁸ , ⁹ It is almost inevitable that the obesity problem in those currently under 20 years old will result in an increase in the prevalence of diabetes in subjects in future years, ¹⁶ unless this trend can be reversed. This young population is therefore at a crisis level in urgent need of help.

Longitudinal reports from the CDC of the changing prevalence of obesity have clearly shown that all of Appalachia represents an evolving epidemic, with its epicenter in West Virginia. ⁶ It is also well recognized that the prevalence of obesity increases with poverty and among minorities. ¹² , ¹⁵ It is, therefore, not surprising that this cohort, drawn from families that are enriched to meet these def nitions, has such a high rate of obesity. What is of particular concern is that individual tracking of weight with aging over longitudinal studies is usually associated with further weight increases over time in the absence of an intervention. Thus, the young and obese are not only at high risk now, but this risk can be expected to further increase over time unless an effective weight reduction strategy is introduced. One hopeful factor is that this new evidence, together with the CBPR strategy of having young community members embedded in their families, ensures access to and the potential to implement modification strategies to better manage weight and reduce the risk of diabetes.

The structure of the frequency distribution profiles of a variable of interest in relatively large data sets is an often overlooked resource of information. We are not aware of previous investigators using de‐convolution of the frequency distribution of age‐corrected BMI in both nondiabetics and diabetes to describe the data in terms of a probabilistic relationship of an established risk factor, age‐corrected BMI, and disease prevalence (type 2 diabetes). Even if our observation that age‐corrected BMI is nonnormally distributed is not novel, however, our use of de‐convolution analysis to identify three Gaussian distributions in which the means and variance are the same for both‐diabetic and diabetic subjects is, to our knowledge, novel.

There is one relatively straightforward hypothesis or mechanistic explanation. A multi‐component, heterogeneous, probability distribution can be explained in terms of a solution of the stochastic master equation for analysis of a birth/death interpretation of a cross‐sectional study. ²⁰ In this model, each range of a subgroup of the age‐corrected BMI frequency distribution is considered as a pool of entries that were “born” at the time of the study, and then immediately “die,” so they do not contribute to the remaining pools. If this model is accepted, the three sub‐distributions represent three sub‐populations with different rates of change in BMI over time. Thus, understanding rate of change rather than of absolute weights is likely to provide a more useful approach to understanding the basis of this epidemic.

From a pragmatic perspective, this strategy of de‐convolving data distributions also provides an unbiased, data driven approach as an alternative to arbitrary conventional group divisions into terciles or quintiles of observed parameters. From a biological perspective, the introduction of the stochastic master equation introduces a mathematically rigorous basis to quantitatively characterize and identify subset subjects for further inquiry into the biologically relevant determinants of these subsets. These new observations, together with the historical knowledge that the prevalence of obesity of this population has increased in recent years, in a population with minor immigration or emigration (i.e., has the same gene pool) raises interesting questions that merit further resolution with respect to genetic and environmental interactions. With a stable gene pool and relatively minor contributions of age, gender, or ethnicity, it is reasonable to posit that the current trend for increasing BMI and an increase in prevalence of type 2 diabetes is due to changing cultural and social factors in this community. If valid, it is reasonable to postulate that interventions to alter these cultural and social factors could reverse this epidemic.

Compelling evidence has been recently been presented to indicate that there is a communicable component to the prevalence of obesity, the major risk factor for type 2 diabetes. ²¹ Application of the graphical approach that captured social relationships was used in the longitudinal Framingham Heart Study, to infer that the chance of becoming obese increased if an individual had a friend who had previously become obese. We reasoned that in the context of rural Appalachia, the strength and importance of the family as a unit could be equally or more important. Focus on social and environmental factors provides a further impetus toward designing interventions. It is reasonable to suggest that modification of this process through interactive feedback communication that takes these networks of relationships into account would slow the spread of obesity, and thereby reduce the prevalence of diabetes. We have started to explore an approach that considers the family as the target unit for intervention, rather than only the obese member within the family.

Our rationale to emphasize the family interaction is based on the self‐interest of all family members to contribute to the family's potential well‐being. In this close‐knit population, an individual's problem becomes a family problem. A lean family member has a vested interest in encouraging his obese family members to lose weight, because ultimately they will share in caring for their less fortunate family members if they develop complications of diabetes. The already obese family member, who has a clear vested self‐interest in losing weight for himself, also has a vested interest in not becoming a burden on the family and is able to share the responsibility of losing weight through family peer pressure and encouragement. We contend that providing a measure of family diabetic risk potential in context to their own community provides a new family focus for discussion and action on risk behavior in all members of the family. We reason that if the obesity epidemic is a cultural phenomenon, the most effective sustained population management strategy will require changes in lifestyle behavior in all members of families at risk in that culture, not only in a change in those most at risk. The family targeted strategy requires new tools to test this concept.

In order to develop a new family metric of risk for diabetes, we adopted using a (weighted, oriented) maximal flow subgraph. MF_FAMILY is deterministically derived from a complete family social and disease‐risk structure captured in Γ_FAMILY. The numerical value of the maximal flow represents a new metric that is independent of family size and sensitive to BMI of family members. An advantage from a CPBR perspective ²² is it provides a novel perspective of family structure to share with the family ( Figure 4 ). The innovation of the visual presentation of the family network is that each individual participant is able to see their own BMI (size of a vertex that represents them) in relation to the rest of the family, and the figure of frequency distribution of the family diabetes risk potential allows them to see their family in relation to their community ( Figure 5 ). We suggest that a shif in cultural pressure to change weight within a family as a group entity is likely to be more effective than trying to target weight reduction strategies to only the individual obese members.

The further advantage of the family diabetes risk potential variable is its nonlinear relationship to age‐corrected BMI. It has the desirable consequence of being highly sensitive with small changes in weight of the most obese, which are the targets for intervention, and is most stable in the nonobese, who do not need to change their own weight, but do need to support their less fortunate obese family members.

It is of considerable interest, and a novel observation that the family's diabetes risk potential falls into three discrete distributions, each with higher percentages of diabetes as the potential rises ( Figure 5 ). The reason for its distribution profile is not known. It is tempting to speculate differences in gene/environment interactions. This is under further study, even prior to the initiation of weight modification programs. From a pragmatic step, the identification of the three distributions in family diabetes risk potential offers a way to efficiently target families with members at greatest risk of developing diabetes, where weight management could confer benefit. If it is assumed that this cohort is representative of rural Appalachia, then an intervention for only those families in the highest cluster would mean that an intervention that only involved 6% of families would only need to involve 10% of the population, but include one third of the g ₃ population ( Table 2 ). Of this subset, almost half would be massively obese and only 30% in the g ₁ lean group. It would also include 31% of diabetics. If the target group is expanded to include the intermediate and upper clusters, there still would be selective advantage as one third of families would include all the massively obese and 67% of the g ₂ and g ₃ groups but only have 30% of family members in the lean g ₁ group ( Table 2 ). This selection would include 50% of diabetics. Thus, a rational approach to targeting subjects that could benefit from obesity management is feasible on the basis of relatively easily acquired information.

It is noteworthy that this targeted or personalized approach would only cover half of the type 2 diabetics. At the present time we do not know if the normal weight diabetics observed in this study represent obese subjects who have lost weight on learning of their diabetes, or a particular susceptible sub‐population. The former explanation would suggest obesity is manageable in this setting and the latter would suggest new strategies would be needed. These alternatives are under further study.

It is important to emphasize that the family diabetes risk potential is a fundamentally different metric when compared to the well‐known diabetes risk score. The diabetes risk score is an individual assessment of risk for diabetes that increases with BMI and is greater with a family history of diabetes or prior gestational diabetes. The diabetes risk score was derived in a Finnish population, using a prospective 10‐year follow‐up of a population, who at the outset did not have diabetes and developed it during the period of follow‐up. ²³ The objective of the diabetes risk score is to identify high‐risk subjects for screening programs to identify diabetes.

In contrast to the diabetes risk score, the family diabetes risk potential extends a very simplistic yes/no to the question of “do you have a family member with diabetes due to an elevated to BMI,” to a much more information rich context of age‐corrected BMI of all family members. Furthermore, from a perspective of health care policy implications in weight management and diabetes prevention it changes the focus from an individual to a family. It is our contention that the relative lack of success to alter the progression of the obesity epidemic despite the knowledge of how to create individual change is disproportionate over emphasis on the obese individual with the implication “you are a weak person” and the missed opportunity to address the family as a unit at risk. We also anticipate that the methodological approach to addressing the underlying driving forces for this epidemic can be applied to building new ref nements to the metrics of obesity to include cultural preferences in diet and exercise, economics and genetics into a complex family focused assessment to better define risk, and the relative contributions of different factors to risk between families.

The family diabetes risk potential uses the family tree, which is converted into double‐weighted graph, and then processed by deterministic operation (computation of the maximal flow) into family diabetes risk potential. This represents a general prototype of analytical approach that enables the topology (i.e., family structure in the community) to be converted into a quantitative entity. This strategy has potential to be broadened by defining other interconnections between the people, families, clubs, townships, counties) to explicitly influence the value of the resulting “potential.” The addition of these numerical values of parameters used to generate the graph weights (BMI in our example, socio‐economic factors in graphs for clubs or counties, etc.) have potential in refining the discrimination of the value of this network analysis.

In conclusion, the obesity epidemic in rural Appalachia of West Virginia is a major health care problem, which now af ects 50% of the population in the poor and minority community. We suggest analytical strategies to identify a def nable structure within the frequency distribution of age‐corrected BMI in this population. We also have created an easy to use and novel mathematical tool for considering the potential for diabetes in the whole family as a unit, rather than only focus on only obese individuals. The ability of our extensive network of young and enthusiastic community investigators to use principles of CBPR within their own family units provides on attractive opportunity to extend from the current baseline observations into active intervention studies to achieve effective, sustained obesity management and diabetes prevention in the future.