Spousal diabetes as a diabetes risk factor: A systematic review and meta-analysis

We conducted our systematic review in accordance with Meta-analysis of Observational Studies in Epidemiology (MOOSE) standards (Additional file 1) [24]. Three citation indices, Medline, Embase and Scopus, were searched using an OVID platform. The search string was developed to identify observational studies that addressed the following: ‘Are spouses of individuals with diabetes more at risk of diabetes than spouses of individuals without diabetes?’ Subject headings and keywords included ‘diabetes’ or ‘Diabetes Mellitus’ and ‘spouse’, ‘wife’, ‘husband’, ‘couple’, ‘married’ or ‘partner’ and ‘concordance’, ‘similar’, ‘correlation’ or ‘parallel’ (Additional file 2). The search strategy was limited to articles published between 1 January 1997 and 28 February 2013, arguably an era of ‘modern’ diabetes care. The language of publication was not restricted. We manually examined reference lists of retrieved studies to identify additional potentially relevant publications (that is, citation tracking). When articles included overlapping data, only the most comprehensive was retained.

Each abstract was appraised independently (KD, investigator, and SP, research assistant) for relevance. Differences in opinion were resolved by consensus and/or discussion with a third reviewer (AL, investigator). We used the following inclusion criteria: 1) cross-sectional, case–control or cohort design; 2) study population with married couples selected from public health records, or administrative, hospital or clinic databases; 3) outcomes were diabetes and/or prediabetes, defined as impaired fasting glucose (IFG) or impaired glucose tolerance (IGT) [25]; and 4) effect measures reported as adjusted or unadjusted odds ratios (OR), risk ratios, hazard ratios or rate ratios. We excluded studies that did not specifically address spousal concordance and those that reported simple linear correlations of metabolic syndrome criteria only. We excluded studies that examined between-spouse correlations for absolute glucose levels rather than diabetes given that there is high intra-individual variability within both the abnormal and normal absolute glucose range [26, 27]. In contrast, a diabetes diagnosis generally requires a clinical assessment that includes more than one glucose measurement and/or glycated hemoglobin or glucose tolerance testing [7, 28].

Study data were independently abstracted by two investigators (AL and KD) using standardized forms (first author, year of publication, source population, country, study design, funding sources, age restriction, proportion of spouses with diabetes (exposure), prevalence and/or incidence of diabetes in the other spouse (outcome), duration of exposure, and effect measures with 95% confidence intervals and variables used for adjustment, such as age, body mass index (BMI), marriage duration, socioeconomic status (SES)). We also abstracted information related to prediabetes, when reported. Where appropriate, we contacted authors to provide us with additional data.

The abstracts and method sections of non-English articles were translated with the assistance of native speakers of the respective languages and online translation tools. Study quality was evaluated using a modified Newcastle-Ottawa Assessment Scale for Nonrandomized Studies [29] that considered the following three potential biases: 1) selection (Were the exposed group and non-exposed group drawn from the same representative samples?); 2) comparability (Were the exposed and non-exposed groups comparable?); and 3) misclassification (Was the method of ascertainment of exposure status ‘gold-standard’ and similar between cases and non-cases?). For comparability scores, we assigned one star (*) if reported estimates were adjusted for age and another star (*) if estimates were adjusted for SES measures. Age was selected as the most important variable given that diabetes risk increases with age [30]. SES was chosen as the second important variable as SES in both married partners could explain some of the shared diabetes risk [31, 32]. We adapted the scoring system for two questions on the assessment scale for purposes of this study. We awarded one additional star if blood glucose testing was used to ascertain exposure under the section ‘Selection’ for question number 3 (that is, ascertainment of exposure, Additional file 3). Similarly, we awarded an additional star if blood glucose testing was used to assess outcome, under the section ‘Outcome’ for question 1 (that is, ascertainment of outcome, Additional file 3).

All data analyses were performed using STATA (version 11 StataCorp, College Station, TX, USA). We extracted the reported effect estimates (that is, ORs, incidence rate ratios, and so on) and 95% confidence intervals from each study to generate forest plots and visually inspected for heterogeneity across studies. We were interested in comparing effect estimates in models that did not adjust for BMI (that is, considered BMI to be along the causal pathway) with those that did, in order to capture associations likely mediated directly through physical activity and dietary patterns independent of BMI. Therefore, we generated forest plots and meta-analyzed effect estimates that adjusted for possible confounders (for example, age and/or SES) but not BMI separately from models that additionally included BMI.

The single longitudinal cohort study [9] that we identified was meta-analyzed with the cross-sectional prevalence studies under the following assumptions: diabetes incidence is low (<10%) and unchanging over the time period considered, study populations are in a steady-state, and the average duration of diabetes is the same for those exposed and unexposed (that is, exposure status does not influence duration). When these conditions are met, the prevalence OR approximates incidence rate ratios [33]. In sensitivity analyses, we excluded the longitudinal study to assess for any changes in the pooled estimate.

We used DerSimonian & Laird random-effects models which account for both within-study and between-study variability to estimate the pooled effect measures with 95% confidence intervals and calculated the Higgin’s I-squared statistic that provided a percentage of variance between studies attributable to chance. I-squared estimates ≥50% were interpreted as evidence of high heterogeneity [34].

Two key study strengths were identified. The first was performing systematic glucose testing on all participants as it ensured that all spouses had an equal opportunity to be detected to have diabetes. All participants underwent oral glucose tolerance testing in the study by Khan and colleagues [38] while, in the study from Kim and colleagues [37], fasting glucose measurements were used to detect diabetes. The second important study strength was the ability to capture diabetes incidence over time. The longitudinal cohort study by Hemminki and colleagues [9] followed 157, 549 subjects for an average of 14.8 years and was thus able to assess the impact of spousal diabetes on incident diabetes (Additional file 4).

Methods of diabetes ascertainment differed across studies (Table 1). Two evaluated a combined outcome that included prediabetes in addition to diabetes [37, 38]. Khan and colleagues [38] (United Kingdom, UK) performed oral glucose tolerance testing in all couples, and diabetes was distinguished from prediabetes. Stimpson and colleagues’ study [35] (Hispanic Americans) relied exclusively on self-report for diabetes. Jurj and colleagues [20] (Shanghai, China) used self-reported diabetes for wives and wife-reported diabetes for husbands. Two studies employed health administrative database diabetes definitions. Specifically, in the UK evaluation by Hippisley-Cox and colleagues [36], diabetes status was determined through a read code for diabetes or a current prescription for anti-hyperglycemic agents or insulin from electronic medical records. In the Swedish study by Hemminki and colleagues [9], the study population was defined through the Swedish Multigenerational Register; classification as diabetes in this study required a hospital discharge diagnosis of diabetes (Hospital Discharge Register) and thus did not capture non-hospitalized cases treated only in an outpatient setting.

The sixth study, a cross-sectional analysis from Korea by Kim and colleagues [37], performed fasting glucose measurements. The outcome included a value ≥6 mmol/L or self-reported use of anti-hyperglycemic medication or self-reported diabetes; diabetes was thus combined with prediabetes. The UK study by Khan and colleagues [38] permitted not only evaluation of diabetes alone but also a combined outcome with prediabetes.

In terms of sampling strategies and source populations, the Korean National Health and Nutrition Examination Survey evaluation employed a general population-based sampling strategy [37]. The Swedish study sampled individuals in a population registry but restricted analyses to those ≥39 years of age [9]. The Shanghai study focused on women with some queries on the spouses’ health status [20]. Stimpson and colleagues examined an older Hispanic origin population (age ≥65 years) [35]. The remaining two studies examined patients registered in general practice networks [36, 38]. With the exception of the Korean study by Kim and colleagues [37], studies restricted their investigation to adults ≥30 years old as older participants would be more likely to have type 2 instead of type 1 diabetes.

Two studies documented shared health-related behaviors (that is, dietary intake) within couples [20, 37]. Half of the studies included a measure of SES in adjustments [9, 20, 35]. The Shanghai-based study [20] stratified analyses by length of co-habitation (< versus ≥23 years; median); this did not change effect estimates. In the UK study by Khan and colleagues [38], all couples were married for at least 5 years.

Effect estimates for associations of spousal diabetes history with prevalent diabetes, adjusted for age and other covariates but not BMI, varied from as low as 10% (OR 1.1 (95% CI 1.0 to 1.3)) in the Shanghai-based study by Jurj and colleagues [20] to approximately 70% (OR 1.70 (95% CI 1.06 to 2.74)) in the Trent UK general practice study by Hippisley-Cox and colleagues [36] and in the American Southwest study on a Hispanic population by Stimpson and colleagues [35] (OR 1.64 (95% CI 1.07 to 2.54) diabetes in wives as outcome; OR 1.77 (95% CI 1.14 to 2.74) diabetes in husbands as outcome). Intermediate between these values was the Swedish cohort study by Hemminki and colleagues [9] (standardized incidence ratios 1.31 (95% CI 1.26 to 1.35) for men; 1.33 (95% CI 1.29 to 1.38) for women). While effect estimates that did not adjust for BMI were generally stronger than more fully-adjusted associations, the largest effect size was approximately a doubling of diabetes risk reported in the UK study by Khan and colleagues [38], wherein estimates were in fact adjusted for both age and BMI (OR 2.11 (95% CI 1.74 to 5.1)); it is possible that the estimate would have been even higher without BMI adjustment.

Two studies evaluated spousal associations for the combined outcome prediabetes/diabetes [37, 38]. The Korean study by Kim and colleagues reported an OR for prediabetes/diabetes of 1.92 (95% CI 1.55 to 2.37) in women and 1.94 (95% CI 1.57 to 2.40) in men after adjusting for age; no BMI-adjusted associations were reported. The UK study by Khan and colleagues demonstrated that the risk of prediabetes/diabetes was also more than twofold for those with a spouse with prediabetes/diabetes (OR 2.32 (95% CI 1.87 to 3.98); adjusted for age and BMI).

We excluded Kim and colleagues’ study from the meta-analysis as separate outcomes for prediabetes and diabetes were not reported. The remaining five studies evaluated a total of 75,498 couples with mean ages of 52 to 74 years. Studies ranged in sample size from 503 to 66,130 couples. By random-effects analyses, the overall effect estimate for diabetes in those with a spousal diabetes history was 1.26 (95% CI 1.08 to 1.45; adjusted for age and/or other covariates but not BMI; Figure 2). There was some evidence of heterogeneity (Higgin’s I-squared statistic = 65.4%, P-value = 0.03). The pooled adjusted association adjusted for BMI in addition to other covariates was 1.18 (95% CI 0.97 to 1.40; Figure 3) with less suggestion of heterogeneity (I-squared statistic = 9.3%, P-value = 0.35).

We employed a broad search strategy without language restriction. Relevant citations in retrieved articles were also examined. Study selection, quality assessment and data abstraction were performed by at least two individuals. The studies were determined to be of medium to high quality and were conducted in different regions around the world involving different ethnocultural groups. Compared to the meta-analysis by Di Castelnuovo and colleagues, our diabetes-related search string (Additional file 2) was more detailed, including ‘diabetes’ and other diabetes-related search terms in addition to ‘glucose’, given our specific focus on spousal diabetes concordance. Importantly, their included studies on diabetes (n = 3) formed a subset of our meta-analysis (n = 5) and did not include the study by Khan and colleagues who performed comprehensive assessments of glucose tolerance and demonstrated the highest effect size.

We did not include unpublished studies in our analyses as these generally lack the methodological rigor of published studies [54]. Some of the heterogeneity observed in the meta-analysis could be attributed to differing ethnocultural composition of study populations, diabetes/prediabetes ascertainment methods, study design, reference groups and characteristics of participants used to adjust effect estimates. Unmeasured confounders/mediating variables such as dietary information, physical activity level, marriage duration and time of diagnosis were not uniformly obtained across all included studies. Therefore, in pooling effect estimates, we generated random-effects models that accounted for between-study and within-study variability. Given the small number of studies, we were unable to perform meta-regression or subgroup analyses to describe the effect of other study characteristics on outcome measures or statistically explore the possibility of publication bias [55]. Results from individual studies should also be interpreted with caution as differences observed may be merely chance findings [56]; for example, although studies differed in ethnocultural composition, there were not sufficient numbers of studies within individual ethnocultural groups for definitive conclusions about any ethnocultural variations in spousal concordance. Only one study [37] reported unadjusted effect measures and, therefore, meta-analyses could only be performed for confounder-adjusted estimates. Individual studies may have potential limitations that impact the accuracy of our findings. For example, determination of diabetes or prediabetes status was more rigorous in some studies than others. Only two studies performed systematic glucose testing on all participants [37, 38]. Another study likely captured only more advanced diabetes cases as its diabetes definition required a hospital discharge diagnosis [9]; while the probable under detection is expected to be similar for individuals with or without a spousal diabetes history, it potentially reduces power to detect spousal associations or bias effects towards the null, although this may not have been a major concern given the large sample size.

Conversely, spouses of diabetes patients could have greater understanding of diabetes and seek medical assistance in the event of relevant symptoms. Similarly, physicians may enforce greater surveillance for these spouses; this detection bias could inflate estimates of association. Two studies that identified diabetes cases from electronic health records did not distinguish between type 1 and type 2 diabetes [9, 36]. However given that around 95% of diabetes in adults is type 2, this unlikely made a difference to the results. The single longitudinal cohort study by Hemminki and colleagues [9] demonstrated an effect estimate similar to the overall effect estimate identified across the cross-sectional studies, suggesting that the influence of incidence-prevalence bias (Neyman bias) associated with not capturing undiagnosed, mild or fatal diabetes cases in cross-sectional studies may be minimal when making inferences in relation to diabetes risk [57].

Spousal history appears to be a robust signal for diabetes risk that may facilitate diabetes detection. Better understanding of underlying mechanisms of concordance could allow the development of tailored strategies to leverage shared risk to achieve health behavior change. Several studies have indicated spousal concordance with respect to BMI [58‐63], consumption of fat and fiber [60] and physical activity [64, 65]. Shared behaviors and risk profiles may be present already at the time of marriage, through an assortative mating process wherein individuals with similar physical (for example, body mass index), ethnocultural, social (for example, social class) and behavioral (for example, eating and physical activity behaviors) characteristics may be more likely to become partners. Additionally or alternatively, spouses may shape one another’s behaviors over time or be influenced by common external factors (for example, life events, physical environment, social network), contributing to diabetes concordance. An examination of the effects of duration of marriage on spousal diabetes concordance could provide some insight in terms of the importance of changes in health behavior that occur during marriage. However, there was little information on marriage duration in the studies examined. There is, however, evidence for spousal correlations of weight change over time [65‐67]. In an analysis of 32 years of follow-up data from the Framingham cohort, Christakis and Fowler demonstrated that development of obesity in a spouse increased one’s risk of obesity by 37%, comparable to the 40% risk increase from the development of obesity in a sibling [43].

Even more compelling are so-called ‘ripple effects’ described by Gorin and colleagues where interventions delivered to one spouse are demonstrated to affect the other [68]. For example, in the Women’s Health Trial, the husbands of women in the low-fat dietary intervention arm reduced their fat intake and body weight to a greater extent than the husbands of women in the control arm [69]. In the National Institutes of Health-funded Look AHEAD trial that examined the effects of weight loss on vascular disease events in diabetes patients, approximately 25% of the spouses of participants in the intensive intervention arm lost 5% or more of baseline weight compared to less than 10% of spouses of participants in the control arm [68]. This body of evidence suggests that not only can spousal diabetes concordance be leveraged to increase detection of diabetes and related risk factors, but also that diabetes prevention strategies could capitalize on within-couple influences.

Three possible strategies to examine spousal diabetes concordance and its underlying mechanisms include a prospective cohort study with more detailed data collection complemented by qualitative assessment, analysis of historical cohort data and analysis of diabetes prevention trial follow-up data. In a prospective cohort study (that is, examination of a group of married couples over time wherein half have type 2 diabetes in one partner at baseline), married couples could undergo systematic evaluation of health behaviors (for example, dietary intake interviews, food frequency questionnaires, pedometer or accelerometer-based assessments of physical activity), anthropometric measures (weight, height, fat mass), sociodemographic profiles (ethnocultural background, immigration status, education, occupation, income), living arrangements and glucose handling (oral glucose tolerance testing) for accurate classification of diabetes and prediabetes. Periodic reassessment would allow capture of incident prediabetes and diabetes to determine the impact of factors such as marriage duration and degree and duration of shared health-related behaviors. Such a study would be strengthened by in-depth interviews or focus group discussions to ascertain participants’ perceptions of concordance and its underpinnings. One could also examine spousal diabetes concordance and its relationship to marriage duration using a historical cohort design, similar to that employed by Christakis and Fowler to assess obesity concordance with Framingham cohort data. Third, evaluations for ripple effects could be conducted among individuals and spouses involved in diabetes prevention trials, namely the Diabetes Prevention Program, the Finnish Diabetes Prevention Study, the India Diabetes Prevention program and a Japan lifestyle intervention program, wherein dietary and physical activity interventions lead to relative reductions of 28% to 67% in diabetes incidence over an average of four years [46, 48, 70, 71]; benefits of lifestyle intervention can persist beyond ten years [72]. It is possible that spouses of those randomized to the lifestyle intervention arms in these trials experienced lower incidence rates of diabetes than spouses of control arm participants.