Background
Multimorbidity is a global health challenge of increasing importance. The prevalence of multimorbidity (prevMM) is a central element in assessing the burden of disease in aging populations. Epidemiological studies of multimorbidity, most commonly defined as the co-occurrence of two or more chronic medical conditions (P2+), have been published for about 20 years now [
1,
2]. Still, prevMM in most of these studies has only limited comparability due to the different study designs and definitions of multimorbidity used.
There are many aspects of the study design that can affect the comparability of prevalence estimates, such as the setting (general population, primary care, hospital, nursing home), the data source and collection (patient self-reports /interviews, medical reports, administrative data), the definition of prevMM, and the classification of diseases included. Any of these choices may influence prevalence estimates and thus affect comparisons between different populations at the same time (regional variations) or at different time points (trend estimates). In addition, demographic and socioeconomic factors are known determinants of multimorbidity [
3,
4], especially age and gender [
5,
6]. Nevertheless, estimates of prevMM vary widely across studies, and this impedes valid comparisons and the interpretation of differences between populations and subpopulations.
Reliable data on the prevalence of multimorbidity are urgently needed to inform medical and public health planning and for assessing the effects of medical and public health interventions. In recognition of these difficulties, the demand for a standardized operationalization of multimorbidity has been voiced recently [
5,
7‐
10]. Among others, Fortin and colleagues have worked towards a more uniform definition and methodology [
7,
11].
This study aims to make recommendations for a standard format in future studies for operationalizing predictors of prevMM. Our approach is empirical, as we analyzed data from our systematic review, relating P2+ simultaneously with the determinants mentioned above using meta analytic methods [
12].
Our overall aim was to investigate which study design variables affect the measured prevalence of multimorbidity (P2+). These variables should be reported and possibly standardized in future studies. We therefore pursued two research objectives: (1) to identify a set of study design variables and demographic predictors of the prevalence of two or more chronic conditions, and (2) on the basis of (1), to make design recommendations for future studies of P2+ for optimal comparability.
Discussion
Using data from a systematic review of 45 articles from 17 countries, we analyzed various variables regarding their impact on the prevalence estimates of multimorbidity, defined as the co-occurrence of two or more chronic conditions. To our knowledge this is the first study simultaneously considering population-related and design-related variables that influence P2+. As is well-known, age is an important determinant of the prevalence of multimorbidity [
3,
5,
20]. So is gender, although that is less well-established [
21]. We quantified the prevalence difference between men and women, with women showing on average a significant nearly 3% higher prevalence of P2+. This result agrees with women reporting higher prevalence of long lasting health problems (e.g., Swiss Statistical Office:
https://www.bfs.admin.ch/bfs/de/home/statistiken/gesundheit/gesundheitszustand.html [
21]). Less well known is the influence of study variables on the prevalence of multimorbidity such as the “number of age groups” (highly significant), “setting” (significant), and “number of diseases in the classification” used (significant) [
5,
19,
22].
Before discussing these results, we would like to emphasize a few limitations in order to set the context of the validity of the findings of our study.
First, as in any systematic review, we cannot exclude bias in the search strategy used or resulting methodological bias due to the heterogeneity of the studies analyzed. A certain reassurance regarding these bias problems lies in the fact that our review is based on widely varying studies with differences in study design, instruments, scope, sample selection, assessed variables, and the language of the included studies, etc. Second, our basic units of observation were age groups, which entails several problems, such as the use of aggregated data. Thus, for example on the individual level, the age effect might be rather more pronounced than our result suggests. Third, we found considerable unexplained variation between studies, necessitating the use of random effect models. The correspondence of observed and fitted prevalences leaves room for considerable improvement. However, we believe that substantial improvement in fit can only be realized by improving the quality of studies of P2+. Thus, it might be useful to report study-specific regressions of prevalence of multimorbidity on age and gender in the future. Fourth, as in any multivariable analysis, interactions between the various variables are to be expected, which, with only 108 observations, we were unable to model. Therefore, our analysis could only provide a rough but we believe nevertheless useful description of the current situation regarding the influence of study design and demographic variables on prevMM. Finally, it has to be mentioned that, due to the scarcity of adequate data, we did not investigate the impact of socioeconomic status [
4], the prevalence of disease patterns [
23,
24], or the simultaneous use of two or more data sources [
24,
25].
Despite these limitations, we believe that this paper contributes towards improving the design of multimorbidity studies.
Significant age and gender effects in multimorbidity have been reported in several studies [
3‐
5,
20,
26‐
31]. In our study, age effect was substantial, too: average prevalence at mean ages 55 to 64 was 44.9%, whereas at mean ages 65 to 74, it was 51.3%, which is a relative increase of 14.2% within 10 years of age. Compared to this, the difference between women and men amounted to only 3.0% on average [
21].
In our multivariable analyses, the “number of age groups included in the study” also had a highly significant effect on prevMM. The effect of number of age groups on P2+ was clear, but due to an outlier at two age groups, it is not easy to interpret. Here, further research may be needed. The implementation of age groups, in other words stratification by age, would allow better age adjustment and thus more precise comparisons of populations. To our knowledge, the use of age stratification was explicitly recommended by only one research group [
32], although a priori epidemiological considerations would suggest it. Moreover, in our systematic review, 58% of all studies had only one, mostly wide age group. A few articles mentioned the impact of the study setting on prevMM [
6,
19]. Differences in prevMM between the general population and primary care practices were described, the researchers stating that a health care setting can be expected to show higher prevalence than a general public setting. Our analysis showed that “setting” as such had only a marginal influence on prevMM – that is, differences in prevMM between settings such as those mentioned above were largely due to differences in age structure between the study populations and possibly some study design variables.
The item “data source” seemed to have no relevant effect in our analysis. In contrast, other studies found “data source” (e.g., self-reported data vs. administrative data) to have a significant influence on prevMM [
19,
24,
33]. However, most of those studies did not suggest an adjustment for other important variables such as age and number of age groups.
In our analyses, “disease classification” and “number of diseases in the classification” had a significant effect on prevMM. Several research groups [
5,
8,
34] have described the impact of the number of disease categories and pointed out the need for a consensus on a common classification of chronic conditions characterizing multimorbidity. Suggestions in this direction have been for a range of single diseases between 11 [
35] and 30 [
36]. Our results indicated that studies using classifications with fewer than 25 or more than 75 chronic conditions tended to yield lower prevalence estimates and thus confirmed a need to standardize disease classification to estimate prevMM. In our opinion, this choice should lead to the highest prevalence. Therefore, an upper limit is reasonable, because for more than 74 diseases, the effect of the number of diseases in the list on the fitted P2+ decreases again.
The “type of disease classification” – another study design variable not investigated in previous studies – had highly significant effects on prevMM in our analysis. This variable can be seen as a quality criterion to indicate whether the single disease entities were classified according to internationally accepted coding systems (e.g., ICD-10) or not. To quantify the burden of multimorbidity, it seemed sensible to us to suggest choices of design variables that maximize the resulting prevalence estimates. Therefore, we propose choosing a list of chronic conditions that contains from 25 to 75 single conditions.
Other authors have suggested classifications of similar size, such as the top 20 single diseases evaluated by Prados-Torres et al. [
9] in their systematic review. The most frequent diseases were hypertension, COPD, diabetes, malignancy, stroke, dementia, depression, joint disease, anxiety, congestive heart failure, coronary heart disease, asthma, cardiac arrhythmia, thyroid disease, anemia, hearing problems, dyslipidemia, obesity, prostatic hypertrophy, and osteoporosis. In another systematic review, Sinnige et al. [
37] assessed the top 20 diseases almost identically to Prados-Torres et al. In addition, Tonelli [
36] identified a panel of 30 chronic conditions to be used in administrative data for which the best identified algorithm was of high or moderate validity. Alternatively, O’Halloran’s definition of chronicity could be useful as an underlying concept [
38], as was applied by a Spanish and an Australian research group [
34,
39].
To characterize multimorbidity, we suggest using the diseases identified in the studies named above. Such a core set of diseases and conditions could then be complemented by highly prevalent or critical chronic conditions relevant to the population under study.
The majority of the studies used in our analysis included the non-communicable chronic diseases that are highly prevalent in high income countries. But when looking at multimorbidity in a more global perspective or in low/middle income countries, other or additional relevant global chronic illnesses might have to be considered [
40].
In recognition of some of the difficulties mentioned above, a need has been voiced lately for more uniform methods to enable solid comparisons between prevMM in different populations or over time [
5,
7‐
10] and subsequently to create a solid database of prevMM. Data of that kind are urgently needed to inform medical and public health planning and, in the longer term, for assessing the effects of medical, public health, or other interventions. Criteria for a meaningful operationalization of multimorbidity, especially for epidemiological research, have been proposed by various researchers [
7,
8,
19,
32]. Recently, an international research group advocated a method for validly identifying chronic conditions in administrative data [
36] as a core set for designing observational studies.
According to our second research objective, based on our empirical evaluation and the considerations above, we derive recommendations regarding standards for future studies of the prevalence of multimorbidity. Thus, to enhance comparability of prevalence estimates, as well as to facilitate the combining of information from various studies (national and international), the following aspects should be considered in the planning of new epidemiological studies:
1.
An overview of the population under study (gender, age range, setting, other socioeconomic variables) and the study design variables (total number of persons under study data source/data collection method, period of the data collection, etc.) should be standard for studies on the prevalence of multimorbidity.
2.
Prevalence estimates should be given stratified by gender and age group to permit proper adjustment.
3.
In the case of small databases or age-related limitations in the study design, stratification into at least three age groups with definite, pre-chosen upper and lower limits should be made. Alternatively, 10-year age groups could be considered, as practiced routinely by the World Health Organization.
4.
The classification of chronic conditions used should comprise between 25 and 75 items. Both the name of the disease as well as the respective related code from an internationally accepted classification system should be documented.