Familial aggregation of multimorbidity in Sweden: national explorative family study

General description of Swedish registers

The Swedish national registers used for data extraction were: Swedish Multigeneration Register, containing data on familial relationships and index persons born in 1932 or later; National Patient Register, giving all hospital discharge diagnoses from 1964 to 2015, with nationwide coverage from 1987 and hospital outpatient diagnoses from 2001 to 2015; Total Population Register, containing data on date of death, marital status, education, and migration with high degree of coverage; and the Swedish Cause of Death Register (1961-2015).13–18 The databases were linked, as previously described,19–21 through the personal identity number, which is seldom incorrect.18 Up to January 2008, an estimated 75 638 (0.56%) individuals had changed their personal identity number compared with the estimated 13 500 000 personal identity numbers in Sweden from 1969 to 31 December 2007.18

Study design, study population, and study period

This nationwide retrospective cohort family study was conducted from 1997 to 2015. This study period was chosen because complete nationwide coverage of the National Patient Register was available and we also wanted to avoid problems with the ICD (international classification of disease) codes by using ICD-10 (10th version) only. A common cause of an incorrect personal number is immigration.18 We therefore only included Swedish born individuals that could be linked to their Swedish born first, second, and third degree relatives to minimise this source of error. Only individuals that could be linked to both of their biological Swedish born parents were included to further minimise the problem of incorrect personal identity numbers.

In the national Swedish Multigeneration Register, families with two full siblings were identified. We also retrieved data on pairs of children born in Sweden between 1948 and 2005 to parents born between 1932 and 1985 in Sweden. Age selection of parents was done so that cousins could be included. We excluded families with members who died or emigrated before 1997 or emigrated before the age of 17. Both biological parents had to be known. Full siblings were then linked to related families and to the half siblings and cousins of the full siblings (we refer to full siblings as siblings in this paper). The same criteria were applied for related families: we excluded all half siblings or cousins who were not born in Sweden, had emigrated before the age of 17 years, or had non-Swedish born parents.

Four different datasets were created: twins, siblings, half siblings, and cousins. Twins were dizygotic and monozygotic twins and were defined as full siblings born on the same date. In the datasets, all relative pairs were entered twice (ie, all sibling pairs, all twin pairs, all half sibling pairs, and all cousin pairs, as described previously).19 20 22 23 This double entry approach is a common procedure in genetics.22 23 Zygosity was assessed only from sex (ie, twins of different sexes were considered to be dizygotic twins). We had no access to zygosity data. We allowed the same person to be included in more than one family relationship.

Multimorbidity score

No standard approach for the measurement of multimorbidity exists,4 and therefore the selection and definition of morbidities were partly subjective and dependent on the data available.4 Any disease recommended as a core for a multimorbidity measure and major long term disorder was therefore included, based on the cross sectional study of Barnett et al.4 The multimorbidity score was modified and adapted to the Swedish ICD-10 codes after Barnett et al (online supplemental table S1).4 Infectious diseases (ie, viral hepatitis) were not included because their main cause is not genetic. To compensate for the exclusion of viral hepatitis and to increase the number of diseases, we added five more common non-communicable diseases or conditions associated with considerable morbidity: osteoporosis, arthrosis, gout, obesity, and pancreatic diseases. All five conditions also have a potential polygenic background. The clinicians in our team (BZ, JS, AH, and KS) agreed on the added diagnoses after discussion. Also, psoriasis was counted as a non-communicable disease separate from dermatitis and eczema.

Online supplemental table S1 lists the diagnoses and ICD-10 codes used. One point was awarded for each of the 45 non-communicable disorders and theoretically, individuals could be assigned a multimorbidity score between 0 and 45 points. No patient, however, had more than 20 points (online supplemental table S2). Individuals with two or more points were considered to have multimorbidity, in agreement with recent definitions.1 2 4 Familial associations for one, two, three, four, and five or more non-communicable diseases were also assessed. Diagnoses in the National Patient Register are coded according to the Swedish ICD-10 system (adapted from the World Health Organizationn ICD classification system).15 The Swedish Hospital Discharge Register has almost 90% overall validity or positive predictive values.15

Statistical analysis

We investigated the crude and adjusted familial associations between multimorbidity scores (one, two, three, four, and five or more diseases) for twins, siblings, half siblings, and cousins and multimorbidity (yes/no) with logistic regression.24 The ambiguity in unselected samples for which trait in twins, siblings, half siblings, and cousins should be used as the dependent variable and which should be used as the independent variable is frequently resolved with double entry.22 23 Each twin, sibling, half sibling, or cousin is entered twice in the data, and each member of a twin, sibling, half sibling, or cousin pair provides a dependent variable once and an explanatory variable once. Although the consistency of the regression estimates for heritability and environmental influences is not affected by double entry, the standard errors of the coefficients are biased and need to be adjusted.22 23 In the logistic regression models in this study, we used the variance covariance cluster method with Stata, which calculated robust standard errors with families as clusters.25 The variance covariance cluster method specifies that the standard errors allow for intragroup correlation, relaxing the usual requirement that the observations are independent. The results obtained with this technique are also (in this case) consistent with the results obtained with the standard technique to double the variances obtained from the double entry approach to correct for the dependence between the two reversed order entries.22 23 These latter results are thus not reported.

The study of inheritance starts with an individual who is affected by a genetic condition or who is concerned that they are at risk of a condition. This person is referred to as the proband. In this study, we compared the familial odds of multimorbidity in people with multimorbidity compared with people without multimorbidity. Thus a proband could be affected or unaffected by multimorbidity. Results are reported as familial odds ratios (95% confidence intervals); that is, the odds ratio comparing relatives with an affected relative (ie, affected proband) with relatives with an unaffected relative (ie, unaffected proband).24 Familial odds ratios for multimorbidity were calculated for relatives of individuals who had a multimorbidity score of two or more (affected proband) compared with relatives of individuals with a multimorbidity score of none or, at most, one disease (unaffected proband). Familial odds ratios were also calculated according to sex and age. Complex traits have been reported to show stronger inheritance at younger ages.8

Models were adjusted for year of birth, sex, region of birth (county, online supplemental table S3), and level of educational achievement (as an indicator of socioeconomic status and lifestyle related factors). Level of educational achievement was categorised into four groups: unknown; ≤9 years of education (elementary school); 10-11 years of education (vocational upper secondary education); and (4) ≥12 years of education (secondary school, college, or university). Unknown education was kept as one category and was not added because of the possibility of missing not at random (ie, the probability that missing data for a specific variable could be related to the values of this variable itself).26

To determine whether the observed odds ratios could be explained by unobserved confounders rather than causal effects, E values were calculated.27 An E value is defined as the minimum strength of association that an unmeasured confounder would need to have with both the predictor and the outcome to fully explain a specific predictor-outcome association.27 A large E value indicates that considerable unmeasured confounding would be needed to explain an effect estimate. A small E value indicates that little unmeasured confounding would be needed to explain an effect estimate.27 The lower 95% confidence interval limits for the E values were also calculated. The average genetic resemblance for twins, both monozygotic and dizygotic, was determined as 0.66 with Weinberg’s differential method. Each relative pair was assigned their average genetic resemblance (ie, 0.66 for twin pairs, 0.5 for sibling pairs, 0.25 for half sibling pairs, and 0.125 for cousin pairs).

Principal component analysis is a statistical technique for exploratory studies, features extraction, dimensionality reduction, and data compression.28 29 Briefly, to identify disease clusters, an initial principal component decomposition followed by a factor analysis with a principal factor method and an oblique promax rotation was used on the correlation matrix of tetrachoric correlations between 45 diagnoses in patients.28 29 The statistical section in the online supplemental file provides a detailed description. Statistical significance was set at P<0.05 and all tests were two tailed. Data were analysed with SAS version 9.4 (SAS Institute, Cary, NC) and Stata version 16.1 (StataCorp, College Station, TX).

Patient and public involvement

Owing to regulatory and study design constraints, neither patients or members of the public were involved in the design, or conduct, or reporting, or dissemination plans of this research. The patients’ data were anonymised so we will not be able to disseminate the results to individual participants. The research team will use academic and public dissemination channels to inform the public of the results. Our findings will be shared on social media and relevant websites.

Familial risk of multimorbidity

Table 2 shows the familial crude and adjusted odds ratios for multimorbidity (score of ≥2) and corresponding E values for twins, siblings, half siblings, and cousins of affected probands with one to five or more diseases (proband) in analyses that were compared with relatives with no diseases (unaffected proband). We found a clear progressive response, with higher familial odds ratios when the relatives of the affected proband had higher scores. For example, among twins, the adjusted odds ratio for multimorbidity was 1.73 for twin relatives with a multimorbidity score of one, 2.84 for a multimorbidity score of two, 4.09 for a multimorbidity score of three, 4.63 for a multimorbidity score of four, and 6.66 for a multimorbidity score of five or more. A declining response was also seen depending on the average genetic resemblance. The adjusted odds ratio was 6.66 (95% confidence interval 4.97 to 8.93) for multimorbidity (two or more diseases) if the multimorbidity score was five or more among twins, 2.71 (2.62 to 2.81) for siblings, 1.47 (1.41 to 1.52) for half siblings, and 1.23 (1.21 to 1.26) for cousins. The corresponding adjusted E values with the lower 95% confidence interval limits were 12.80 (9.41) for twins, 4.86 (4.70) for siblings, 2.30 (2.17) for half siblings, and 1.76 (1.71) for cousins (table 2).

Table 3 shows the odds ratios and corresponding E values for multimorbidity (≥2 score) if the proband relative had multimorbidity (two or more diseases). The odds ratios in the adjusted models for multimorbidity were 2.89 (95% confidence interval 2.56 to 3.25) in twins, 1.81 (1.78 to 1.84) in siblings, 1.26 (1.24 to 1.28) in half siblings, and 1.13 (1.12 to 1.14) in cousins in those having a diagnosis of an index score of more than one (table 3). The corresponding adjusted E values were 5.23 (lower 95% confidence interval limit 4.56) for twins, 3.02 (2.96) for siblings, 1.83 (1.79) for half siblings, and 1.51 (1.49) for cousins (table 3).

Risk of multimorbidity by age and sex

The familial odds ratios were highest at younger ages among all relatives (online supplemental tables S5–S8). For example, in twins, the adjusted odds ratio was 3.22 (95% confidence interval 2.65 to 3.92) for those aged ≤20 years, 3.14 (2.50 to 3.93) for those aged 21-30 years, and 2.29 (1.87 to 2.80) for those aged >30 years at the end of follow-up.

Among twins, the strongest association was observed between male twins (adjusted odds ratio 3.44, 95% confidence interval 2.74 to 4.32); in female twins, the adjusted odds ratio was 3.60 (2.97 to 4.37; online supplemental table S5). Twins of the same sex were monozygotic or dizygotic twins. In twins of opposite sex (dizygotic), odds ratios were lower (1.92 if male proband and 1.91 if female proband). No major differences between the sexes were observed in siblings, half siblings, or cousins (online supplemental tables S6–S8).

Principal component analysis followed by factor analysis

For principal component analysis followed by factor analysis, disease clusters among individual patients and not between family members were used. The full sibling dataset was used for this analysis. The number of diseases was reduced to nine different disease clusters (online supplemental table S9 and online supplemental figure S1). We used visual inspection to find a point at which the amount of variance explained by subsequent principal component dropped off (online supplemental figure S1). After nine factors, the eigenvalues levelled off (elbow) and nine factors were therefore extracted. In table 4, the rotated factor loadings (pattern matrix) and unique variances are sorted. The loading pattern determines the factor that has the most influence on each variable. Loadings close to −1 or 1 indicate that the factor strongly influences the variable. Loadings close to 0 indicate that the factor has a weak influence on the variable. Online supplemental table S10 shows the distribution of diseases in these nine disease clusters (factor clusters 1-9). As part of this method, all conditions were allocated to one of the nine clusters. For example, in the factor 1 cluster (mainly cardiometabolic disorders), hypertension, heart failure, coronary heart disease, and diabetes had the strongest contributions in this group (table 4). In the factor 2 cluster (mental health disorders), affective disorders, anxiety, psychoactive substance misuse, and alcohol misuse disorders were the most important in this group (ie, with the largest positive loadings; table 4).

For seven of the nine disease groups identified by factor analysis (factors 1-6, and factor 8), the odds ratios were closely related to the degree of average genetic resemblance, with the highest in twins and then decreasing stepwise with the degree of relatedness; the lowest were in cousins but these values were still significant (table 5). Online supplemental tables S11–S19 show the familial crude and adjusted odds ratios for multimorbidity (two or more diseases) in twins, siblings, half siblings, and cousins of probands with one to five or more diseases in analyses that were compared with relatives with a proband with no diseases. The two most important disease cluster groups identified (ie, with the highest eigenvalue (online supplemental table S9), were factor 1 (hypertension, heart failure, coronary heart disease, diabetes, obesity, atrial fibrillation, gout, atherosclerosis, and renal disease) and factor 2 (affective disorders, anxiety, psychoactive substance misuse, alcohol misuse disorders, anorexia or bulimia, and schizophrenia disorders) (table 5). These two disease groups were strongly clustered, with odds ratios determined by average genetic resemblance (table 5 and online supplemental tables S11 and S12). The disease groups factor 3 (inflammatory bowel disease, liver disease, pancreatic disease, and ulcers), factor 4 (epilepsy, blindness and poor vision, cerebrovascular disease, cancer, and impaired or hearing loss), factor 5 (connective tissue disease, osteoporosis, thyroid disorders, and psoriasis), factor 6 (prostate disease, arthrosis, painful back condition, diverticular disease of intestine, and chronic sinusitis), and factor 8 (asthma, dermatitis and eczema, constipation, chronic obstructive pulmonary disease, and migraine) were also clustered (table 5 and online supplemental tables S13–S16 and S18). Only multimorbidity of the disease clusters factor 7 (bronchiectasis, Parkinson’s disease, glaucoma, learning disability, and irritable bowel syndrome) and factor 9 (multiple sclerosis and dementia) were not clustered (table 5 and online supplemental tables S17 and S19).

Additional information

Online supplemental tables S20-S22 show the number of observations and individuals used for the calculations in table 2, table 3, and table 5 (ie, number of individuals and number of persons at risk).

Principal findings

In this large nationwide explorative family study, based on complete and validated data sources, we have provided robust estimates correlating with average genetic resemblance, suggesting a hereditary component of multimorbidity. Hence our results support other studies indicating a genetic contribution to multimorbidity.11 12 According to a study by Khoury et al,30 even with complete correlation in exposure among first degree relatives, environmental risk factors with relative risks of <10 gave modest familial relative risks (1-2) and low recurrence risks, suggesting that the high risk of disease specific multimorbidity has a major genetic contribution. Moreover, the calculated E values in our study suggest that strong unmeasured confounders are necessary to explain our findings.27 Nevertheless, we cannot rule out familial lifestyle factors contributing to the familial aggregation of multimorbidity. This study showed the inheritance of specific multimorbidity clusters, suggesting that further genetic research, such as genome wide association studies, could be worthwhile. The study also showed that several disease clusters aggregated in families.

A plausible cause of the familial aggregation of multimorbidity is the abundance of pleiotropy for complex traits observed in genome wide association studies.9 Seven of the nine disease clusters identified with the factor analysis in individuals showed familial aggregation. The strongest aggregation of multimorbidity was observed in twins and pairs of full siblings, whereas these associations, although significant, were reduced in half siblings and cousins. The associations correlated with average genetic resemblance, which suggests that genetic factors are important in the familial aggregation of multimorbidity.19–21 Moreover, multimorbidity inheritance was graded in terms of the number of diseases in probands. The higher odds ratios in younger than in older individuals also suggests a genetic basis.8 31 Also, third degree relatives usually do not share a household and the increased familial odds ratios among cousins suggests a genetic contribution to the familial aggregation of multimorbidity. Older individuals, women, and those with a lower educational achievement had a higher multimorbidity risk across all of the family relationships. These findings confirm previous studies and the validity of this study.4 6

The first step in identifying genetic determinants in complex diseases is to study the familial aggregation of the phenotype.8 31 The Swedish Multigeneration Register offers the opportunity to explore the influence of genetic and non-genetic familial factors in a large number of families by observing first, second, and third degree relatives and the occurrence of specific phenotypes.19–21 First degree relatives (eg, full siblings) share 50% of their genes, as well as shared environmental exposures common to their family. Second degree relatives (eg, half siblings) share 25% of their genes, and third degree relatives (eg, first cousins) share 12.5% of their genes.19–21 In this study, the increasing probability of multimorbidity among relatives of affected individuals was linked to the closeness of the relationships, being strongest in twins (66% average genetic resemblance). Hence our observations support multimorbidity heredity, although the specific underlying genetic mechanisms are not known.

The agnostic method, based on principal component analysis and the oblique rotation in the factor analysis to identify disease clusters at the individual level, identified familial aggregation of multimorbidity in seven of the nine disease groups (tables 4 and 5). This finding agrees with studies showing that diseases with a higher probability of concurrency tend to share more associated genes.32 33 The cardiometabolic diseases in the factor 1 cluster were previously described to be interrelated.34 Some of the conditions included in the factor 1 cluster were in a causal pathway with coronary heart disease (eg, obesity, diabetes, hypertension, and atherosclerosis). This finding could contribute to the clustering in the factor 1 group. The mental health disorders in the factor 2 cluster are also well known to be linked.35 Our study agrees with other studies identifying mental health disorders and cardiometabolic conditions as the two most reproducible groups of multimorbidity.36 Inflammatory bowel disease and autoimmune liver disease are also known to be linked in the factor 3 cluster.37 A well known linkage exists between autoimmune disorders, such as certain connective tissue disorders, thyroid disease, and psoriasis (factor 5 cluster).38 Thus the high familial aggregation of the factor 5 cluster is not unexpected.

Eczema and asthma in the factor 8 cluster are well known to be associated.39 Chronic obstructive pulmonary disease clustered in the same group as asthma, which is a common feature for many multimorbidity studies.36 The reason for the inclusion of constipation in the factor 8 cluster is not obvious but a recent report has described an association between asthma and constipation.40 Migraine has also been shown to be associated with asthma.41 Other associations might be acquired from treatments, such as osteoporosis clustering with connective tissue diseases (ie, because of treatment with corticosteroids).42 Many studies have reported that musculoskeletal disorders cluster together (eg, arthrosis and painful back condition, factor 6 cluster in tables 4 and 5).36 These examples show the strength and proof of concept of our study design for the use of principal component analysis and factor analysis for studying disease clustering. Nevertheless, some associations are less intuitive (eg, why cerebrovascular disease clustered in the factor 4 cluster with hearing loss instead of the cardiometabolic factor 1 cluster), although a recent systematic review and meta-analysis found an association between sensorineural hearing loss and an increased risk of stroke.43 Also, hearing loss has been reported to be related to the risk of stroke but not coronary heart disease.44

Strengths and limitations of this study

In common with all epidemiological studies, the interpretation of results is constrained by time (1997-2015) and geographical location (Sweden), resulting in bias in time period and location. Our inclusion and exclusion criteria ensured that disease events in a family could be registered in the Swedish patient registers (1997-2015). This method is a limitation because families with diseases before 1997 were not included, although this bias is most likely non-differential. Asking a proband about the family history of a disease is a major source of self-report and recall bias, however, which likely exceeds the study limitations of this study design.45 Another limitation was the lack of information about lifestyle factors, such as smoking, consumption of alcohol, and physical activity. We adjusted for education that correlates with lifestyle factors, such as smoking,46 but residual bias is likely to exist. Thus we cannot rule out the contribution of family environment and lifestyle factors, such as smoking, to the familial aggregation of multimorbidity. Another limitation was that we did not distinguish between dizygotic and monozygotic twins.

We did not consider the time courses of the diseases in individuals or which of the 45 diseases in the multimorbidity scores a patient was affected by. Different diseases could cluster in different families. With the principal component analysis and factor analysis method to identify diseases that cluster in individual patients, however, we identified disease clusters with familial aggregation. Further studies to establish smaller disease clusters could be worthwhile, although our study agrees with many other reports, including a systematic review of the two major groups of diseases (mental health disorders and cardiometabolic disease).34–42

We believe that the use of principal component analysis for studying disease clusters is an advantage over k means. The k means algorithm requires some initialisation of the centroid positions.47 For most algorithms, these centroids are randomly initialised with some method, such as the Forgy method, or random partitioning, which means that repeated iterations of the algorithm can converge to vastly different results.47 We also saw this effect and therefore we did not use k means because the clustering was not reproducible, whereas principal component analysis gave identical results. Nevertheless, we found some similarities with our grouping results between k means and principal component analysis. We also tried latent class analysis for cluster analysis but because of the large numbers, the results did not converge for the latent class analysis. A known limitation of latent class analysis for cluster analysis is that it is computationally expensive, which might be inconvenient with large datasets.48 Moreover, selecting the number of clusters is a challenging task involving inevitable subjective analytical choices.48 With the principal component analysis method, however, we reproduced the two most commonly reported clusters of multimorbidity (mental health disorders and cardiometabolic disease), which shows the strength of the principal component analysis method.34–42

Some of the 45 conditions can be acute as well as chronic (eg, constipation). We could not take this limitation into consideration or the order in which patients were affected by different diseases. Other methods for studying disease clustering could give divergent results but our two main groups (mental health disorders and cardiometabolic disease) were similar to many other studies.36 A systematic review showed that cardiovascular and metabolic diseases, mental health problems, and allergic diseases were major disease clusters, regardless of the method used for studying disease clustering and multimorbidity patterns.49

The large size of our study was a major advantage. A limitation but also a strength was the use of nationwide registers with almost complete data coverage and high validity.13–18 A previous validation of the National Patient Register by the National Board of Health and Welfare showed that 85-95% of all diagnoses were in the National Patient Register.15 The use of clinical hospital diagnoses allowed for the elimination of recall bias. Recall and self-report bias are common problems in many family studies.45 The objective data in the Swedish Multigeneration Register could therefore be an advantage compared with self-reported data.13 Age selection of parents was done so that cousins could be included. Median age at the end of follow-up was 32 years (range 0-68), limiting the applicability of the results to older people. Heritability in complex traits is dependent on age, however, and young age is an advantage in an explorative study of heredity,31 including multimorbidity inheritance. The observed age dependence of the familial odds ratio in this study suggests a genetic cause and that multimorbidity is a complex trait.8 31 Acquired factors of multimorbidity would be expected to be relatively more influential with increasing age. Another limitation is that in Swedish registers, information on ethnic groups is not available. To reduce the number of incorrect personal numbers because of immigration,18 only Swedish born individuals with Swedish born parents were included, limiting the generalisability of the study.

Conclusions

The results of our study suggest that the risk of multimorbidity among relatives of affected individuals depends on the closeness of the relationship, being strongest in twins and full siblings, but still significant in third degree relatives, indicating genetic components in susceptibility to multimorbidity. We cannot rule out familial lifestyle factors contributing to the familial aggregation of multimorbidity. Principal component analysis and factor analysis were used to identify smaller disease clusters. This explorative family study suggests that map clusters of diseases should be used for the genetic study of common diseases to show new genetic patterns of non-communicable diseases.