Multimorbidity, defined as the coexistence of two or more chronic conditions, is a common phenomenon among the older population worldwide: two recent population-based studies indicated that the prevalence of multimorbidity ranges between 40% and 56% in the general population aged 65 years and older [1, 2]. Multimorbidity is known to negatively affect health outcomes including mortality, hospitalisation, and readmission [3].

Health-related quality of life (HRQL) is a health outcome measure which is increasingly used to assess the medical effectiveness of interventions and to support allocation decisions in the health care sector. Generic HRQL instruments like the EQ-5D are appropriate for non-disease-specific analyses and allow comparisons between patient groups with different medical conditions [4].

Several studies examined the effect of multimorbidity on HRQL [5–13], however research has poorly represented combinations of chronic conditions [6]. In particular, to the best of our knowledge, all studies using the EQ-5D as a measure of HRQL only considered independent disease effects [6, 7, 10–12]. Therefore, it has been argued that further studies should focus on identifying interaction effects between chronic conditions and account for the impact of specific disease combinations [5, 7]. Synergistic effects on quality of life measures other than the EQ-5D have been reported mainly for the combination of diabetes and cardiovascular problems [8, 9, 14, 15], the combination of respiratory and cardiovascular problems [5, 9], and the simultaneous presence of diabetes and hypertension [14, 16].

Studies have shown that HRQL is strongly associated with body mass index (BMI) even after adjusting for comorbidities [7, 17–20]. Many contributions analysing the relationship between BMI and HRQL incorporated the effect of BMI either as a linear term, or divided its distribution in categories according to the classification of the World Health Organisation (WHO) [21]. The first approach, however, ignores the fact that association between BMI and HRQL is usually nonlinear [13, 19, 20, 22], while in the second approach, grouping different BMI values into the same category may obscure meaningful differences within categories. Both approaches can bias findings since they may conceal the true functional form of the relationship. Furthermore, concerns have been raised about whether the WHO classification is appropriate for use in the older population [23]. In particular, the question of to what extent not only being overweight but also being underweight is associated with reduced HRQL has not been sufficiently investigated in the older population.

The purpose of this study was to clarify how different chronic conditions and pairs of conditions are associated with impairments in HRQL measured by the EQ-5D in a German general population sample of individuals aged 65 years and older. Specifically, we sought to identify and explore disease combinations which are related to HRQL over and above the independent contributions of each condition. Furthermore, we wanted to appropriately account for the association of age and BMI with HRQL by using nonparametric regression methods, i.e. without imposing a priori constraints on the functional form of this relationship.

The EQ-5D index could not be calculated for 93 (2.0%) respondents due to missing values in at least one EQ-5D item. Excluding participants with missing data in the EQ-5D index or in the covariates reduced the final sample size from 4,565 to 4,412. The 153 (3.35%) individuals excluded from analyses were on average 2.9 (CI: 1.9 - 3.9) years older than the participants with complete information and were slightly more likely to be female (p = 0.016, chi-square-test).

The mean EQ-5D index in the sample was 76.3 (SD 18.8) and the observed values covered the entire range from 3.5 to 97.7. There was a ceiling effect in the data since 1,285 (29.1%) respondents stated having no problems in any of the five EQ-5D dimensions and were hence assigned the EQ-5D index value 97.7. Moderate or severe problems were most frequently reported in the EQ-5D dimension 'pain' (62.5%), followed by 'mobility' (31.3%), 'anxiety/depression' (29.2%), 'usual activities' (22.5%) and 'self-care' (10.1%).

Respondents interviewed by telephone rated their health on average 8.0 (CI: 6.2 - 9.7) points lower than the questionnaire respondents (unadjusted for covariates).

The nonlinear effects of age and BMI on the mean EQ-5D index are represented by the estimated smooth functions and in Figure 1. Since effects were almost identical in the two regression models, we only show the curves of the main effect model. The left curve in Figure 1 expresses a nonlinear age effect with stable health until 70 years and a decline in HRQL from the age of 70. Between 70 and 85 years, HRQL decreased on average by 14 units. The right curve in Figure 1 shows that the relationship between BMI and HRQL was inverse U-shaped with the maximum HRQL located around a BMI of about 24.8 kg/m². It indicates that an increase of the BMI from 24.8 to 35 was associated with an EQ-5D utility loss of about 5.0 units. On the other hand, underweight individuals with a BMI of 18 had an average impairment of 7.1 units compared to a BMI of 24.8. The estimated curves differed only slightly when the data were split into two subsets.

Applying the regression model to the imputed data set did not alter findings and the same interactions were selected. Separately fitting the regression model to the subset of younger and to the subset of older participants, we observed that the history of a stroke had a stronger negative impact on HRQL in individuals aged 65-74 years than in individuals aged 75 years and above. In contrast, the impairments in HRQL associated with the history of cancer were more pronounced in the older age group. Furthermore, the negative effect of low BMI on HRQL was more important in the older age group (results not shown).

The stratified analysis by sex revealed that the history of a stroke had a stronger effect in men (-14.3; CI: -16.9 - -11.7) than in women (-7.4; CI: -10.5 - -4.3). In contrast, the effect of chronic bronchitis was slightly more pronounced in women (-10.5; CI: -14.4 - -7.6) than in men (-6.4; CI: -9.0 - -3.7). Moreover, HRQL for men peaked at a BMI of 26.3 kg/m² while maximum HRQL in women was observed at a BMI of 23.4 kg/m².

In our study, we found that each of the chronic conditions considered was associated with impairments in HRQL, either alone or in combination with other conditions. In agreement with the results of other studies, we observed the most severe impairments for a history of stroke [7, 10, 31] and chronic bronchitis [7, 32]. With an adjusted R² of approximately 18%, our models only explained a moderate proportion of variance. Nevertheless, the adjusted R² was equal to or better than in comparable studies [7, 12, 33].

Several researchers investigated the joint effects that specific disease combinations have on quality of life. However, to the best of our knowledge, this study is the first to explicitly examine interaction effects between chronic conditions on HRQL measured by the EQ-5D. Our analyses revealed that the combination of diabetes mellitus and coronary problems, as well as the combination of coronary problems and a stroke history were synergistically associated with HRQL. There was no subtractive interaction between diseases in our data.

The joint effect of diabetes and coronary problems on HRQL in our study reflects the substantial burden of illness caused by the combination of these two conditions. Studies have shown that persons with diabetes are at greatly increased risk of cardiovascular diseases and that the prevalence of cardiovascular complications amongst persons with diabetes is especially high in older age groups [34]. Our results complement these findings by underlining the exacerbating effect that cardiovascular diseases show on HRQL in subjects with diabetes. Synergistic effects of diabetes and coronary problems on HRQL have also been reported in studies using the SF-36 [9, 15] and the HUI3 [8], as well as in studies on disability and functional status [14, 35]. In contrast, Wee et al. reported mainly additive, but even partly subtractive effects of heart disease and diabetes on the SF-36 subscales and the SF-6D [16]. For a discussion of further synergistic relationships found in literature, see, e.g., Hodek et al. [36].

The nonsignificant main effect of diabetes in our interaction model indicates that either there is no decline in HRQL caused by diabetes without coronary comorbidity, or that the decline is too low to be detected by the EQ-5D. It has been argued that the EQ-5D detects differences due to diabetes related complications, but that it lacks sensitivity in capturing differences between diabetes treatment regimes [37]. Although there is evidence that subjects with diabetes but without comorbidities still have more impairments than subjects without diabetes [14, 15], another study found that diabetes was not associated with lower EQ-5D scores after adjusting for comorbidities [7]. Rijken et al. even observed a positive main effect of diabetes on the physical scale of the SF-36 when the negative interaction term with cardiovascular disease was accounted for [9].

We found the combination of coronary problems and the history of a stroke to also have synergistic effects on HRQL. Stroke and myocardial infarction are both mainly manifestations of atherosclerosis. Studies showed higher mortality rates and increased treatment cost when stroke occurs after myocardial infarction [38, 39]. Reversely, myocardial infarction is an important cause of death in patients with cerebrovascular disease [40]. Another study found that heart disease and stroke were synergistically associated with physical disability [35]. Our results highlight the negative impact of this disease combination on HRQL.

Very few studies on HRQL in multimorbid patients accounted for the effect of weight problems, as expressed by the BMI [7, 17, 18, 41]. And to the best of our knowledge, this study is the first to explicitly examine the functional form of the relationship between BMI and HRQL by means of semiparametric regression methods, i.e., without imposing a priori constraints on its shape such as polynomial forms or piecewise constant terms. Our analyses showed that the relationship between BMI and HRQL was inverse U-shaped and that not only overweight but also lower BMI values were associated with significantly reduced HRQL. This supports findings reported by other studies [13, 19, 20, 22]. Furthermore, our study is the first to address the nonlinear association of BMI with HRQL in older adults. Ignoring the nonlinearity would overestimate HRQL for subjects with lower BMI values, which is particularly serious in the older population where being underweight can be a severe problem [42].

The additive regression models used in our study also allowed us to explore the nonlinear relationship between age and HRQL. In our sample, age was strongly associated with the mean EQ-5D index, but the age-related decline in health was only observed from the age of 70. The negative correlation between age and HRQL, even after adjustment for the effect of chronic conditions, is supported by several studies [6, 7, 10, 43]. However, there is evidence that age per se is only a weak predictor of HRQL and that rather the increasing number and severity of chronic conditions are behind the age effect [7, 11, 36, 43]. Thus, the association between age and HRQL may become less pronounced if morbidity was assessed by a greater number of comorbidities or if disease severity was accounted for.

The data used in our analyses came from a postal questionnaire for self-completion. However, about 16% of the participants were interviewed by telephone since these people had not returned the questionnaire despite being sent a reminder. Respondents interviewed by telephone were on average older, more likely to be female and suffered from more chronic conditions than the questionnaire respondents. These three aspects are all known to be negatively associated with HRQL. In fact, the unadjusted HRQL score within the telephone respondents was nearly 8 points lower than within the questionnaire respondents. However, our multivariable regression analyses showed that the difference in HRQL between the two data collection methods persisted even after adjustment for covariates. There are two possible explanations for this finding: first, it can not be ruled out that answers to quality of life questions given by the telephone respondents may be biased due to the personal interview situation [44]. Second, it is possible that the difference is caused by unobserved comorbidities. Although the telephone interviews could increase the response rate, our study (as most population surveys) was still confronted with the problem of non-response. An extensive analysis on this issue in one of the baseline surveys has shown that non-respondents included a higher percentage of people with impaired health [45], and it can be assumed that more severely ill subjects were less likely to participate in our study. As a consequence, our results may underestimate the burden of comorbidity in the older population. Nevertheless, the prevalence of the chronic conditions in our study sample was comparable to that reported in another German study with the same age range [7].

One strength of our study is the large number of patients with cell frequencies for disease combinations that allow for the valid examination of interaction terms. Also, our study is population-based so that results are more likely to be transferable to the older general population than results obtained from general practice samples. Finally, to our knowledge, our study is the first to examine the effect of disease combinations on HRQL measured by the EQ-5D, the most frequently used instrument in economic evaluation.

A limitation of our study was that we relied on a limited list of only six chronic conditions and no psychiatric condition was amongst the considered conditions [3]. This limitation is reflected by the relatively low proportions of explained deviance in the regression models, especially for the EQ-5D item 'anxiety/depression'. Also, we did not assess dementia because questions about the diagnosis of dementia are a sensitive issue and responses may be of limited validity [46]. However, most of the comparable studies evaluating interaction effects considered a similar number of chronic conditions [9, 16, 35], and our study considered most of the common widespread diseases in western countries.

Another limitation is that the presence of chronic conditions in our study was based on self-reports. We did not use a specific, validated questionnaire; however, the case-finding questions for physician-diagnosed illness used in our questionnaire are widely used in population-based studies [8, 35, 47]. Self-reports are not as valid as medical record information, however, they have been shown to provide reasonable estimates of comorbidity in the older population [48, 49]. In an earlier follow-up of the KORA S1-S3 participants, the diagnoses of myocardial infarction, stroke, and diabetes have been validated by medical record review and the agreement was very high [50].

Furthermore, our analyses did not account for time since diagnosis or disease severity. Although long-term reductions in HRQL for patients with a history of myocardial infarction or stroke were reported in literature [51, 52], disease burdens may be higher for more recent diagnoses. Differentiating by disease duration and disease severity would permit more precise quantification of the association between individual conditions and HRQL. However, this study focused on exploring the joint effects of disease combinations, and interaction effects between conditions could no longer be described comprehensively if the effect of each diagnosis was additionally differentiated by severity or disease duration. Finally, the effects that specific disease combinations have on HRQL may be more complex than described by pairwise interaction terms. However, three-way or even higher order interactions are complicated to interpret and their estimation is likely to be unstable in our data due to small cell frequencies of some three-way combinations. Nevertheless, the pairwise disease interactions in our study can be considered as a reasonable approximation of potentially more complex dependencies [35].

The effects of chronic conditions on HRQL in the older population are not always purely additive. Our study showed that the interactions between coronary problems, diabetes mellitus, and the history of stroke caused greater impairments in HRQL measured by the EQ-5D than could be expected from the separate effects of these conditions. Our findings emphasise the importance of comorbidity prevention in order to reduce the health burden caused by the exacerbating effects of specific disease combinations.