Costs of multimorbidity: a systematic review and meta-analyses

The initial search was conducted on 9 October 2020, with a follow-up search on 4 January 2022. Databases searched were Web of Science, Global Health, MEDLINE, SCOPUS, The Cochrane Library, and Business Source Complete. A snowballing technique was adopted to identify further articles from within the reference lists of eligible studies. The database-specific search strings are detailed in Additional file 1 — Search strings.

The search included all quantitative studies described in full-text papers, published in English between 1 January 2010 and 4 January 2022. Papers were eligible for inclusion if they reported the costs associated with the co-existence of at least two chronic conditions. We excluded studies not specifying combinations of chronic conditions and those which had estimated incremental costs only (for full list of exclusion criteria, see Fig. 1).

Following the removal of duplicates, articles were screened based upon their titles and abstracts, using Rayyan QCRI software [24] (phase 1). Articles were subsequently transferred to Endnote X9 for full-text screening and reference management (phase 2). Both screening phases were conducted by two authors (PT and JK) using a checklist (Additional file 2 — Screening checklists), and conflicts resolved with the senior author’s input (JVO).

The quality of studies was assessed by two authors independently (PT and JK) using the Newcastle–Ottawa Quality Assessment Scale (NOS) [25]. In the case of discrepancies, consensus was reached with input from other authors.

A REDCap form, which had previously been piloted and refined, was used to facilitate the data extraction process. REDCap was chosen due to its level of data encryption, user-friendliness, and its capabilities for customization in data collection [26]. The following data were extracted: study characteristics, sample size, disease characteristics, care setting, and costs. The data cleaning process is presented in Additional file 3 — Data cleaning flow chart.

This study’s primary outcome measure is the average annual cost per patient per disease combination, comprising of direct medical costs, direct non-medical costs, and indirect costs. Direct medical cost is the cost of a defined health service or intervention and all follow-up costs for medication and medical equipment (diagnostics, hospitalization, outpatient, emergency, drugs, and equipment) [27]. Direct non-medical cost is the cost incurred in the process of seeking and after receiving health services that are not involved in the direct purchasing of medical products or services (transportation/travel costs, food, accommodation, and additional paid caregiver time) [27]. Indirect costs are those incurred as a result of losses from the disease(s) or disease management (time loss, wage loss, interest from debts/loans) [27]. Other terminologies are explained in Additional file 4 — Definition of terminologies. All costs are reported in 2021 International Dollars (denoted by I$), which is a hypothetical currency with the same purchasing power in every country, using the US as a reference [28]. First, the reported cost was inflated to 2021 local currency unit [29]; then, it was converted to International Dollar using 2021 Purchasing Power Parities (PPP) [30].

The costing perspective is important in a costing study as it determines which costs are included (direct medical/non-medical, indirect costs), the source of data, and the scope of the study. The costing perspective may reflect a patient (often out-of-pocket), an organization (provider), a health system (public or private), or all of society [31]. This review includes all costing perspectives, but mainly reports on the health system perspective as it is accounted for in the majority of studies. The health system perspective entails formal direct medical costs paid for by third-party payers or by patients [32].

To address research question 1, we tabulated all combinations of conditions and their costs as described in the studies. For one study that reported costs at baseline and follow-ups, costs were pooled to arrive at an average estimate [33].

Related conditions were grouped together. For example, Type 1 and Type 2 diabetes were classified as “Diabetes”. Mental disorder, anxiety disorder, and depression were grouped as “Mental Health conditions”. Asthma, chronic obstructive pulmonary disease (COPD), and tuberculosis (TB) were combined under the category “Respiratory diseases”. Cardiovascular disease, coronary atherosclerosis, congestive heart failure, coronary heart disease, atrial fibrillation, coronary artery disease, peripheral artery disease, myocardial infarction, heart disease/failure, cerebrovascular disease, conduction disorder or cardiac dysrhythmia, valvular disease, peripheral vascular disorders, and pulmonary circulation disorders were classified as “Heart/vascular conditions”. “Cancers” included thyroid, stomach, breast, uterus, kidney, colon and rectum, esophagus, pancreas, head and neck, other gastrointestinal, liver, ovarian, multiple myeloma, and any malignancy/tumor. Alternate groupings would have been possible, and those following a more treatment-focused perspective may have led to variation in results. However, we resorted to this approach in order to reduce the number of combinations and condense information for ease of interpretations. Grouping these conditions at an organ system level makes sense from a health system/organizational perspective considering that they show similarities related to medical specialties. For example, lung cancer is grouped together with other cancer sites and not with the respiratory diseases, considering that cancers are treated by oncologists and most other (severe) lung or respiratory conditions are treated by pulmonologists. On the other hand, we categorized hypertension separately from the heart/vascular group as it is the leading metabolic risk factor globally [34]. Moreover, most studies also report hypertension separately; therefore, following this approach allowed for cross-comparison between studies. Lastly, we did not include stroke in the heart/vascular group as it is considered a chronic disease with acute exacerbations, in which the cascade of care is important in contextualizing the costs across the patient care pathway; therefore, the cost of stroke cannot be interpreted together with other heart/vascular conditions [35].

This resulted in six main disease categories: (1) Diabetes, (2) Heart/vascular conditions, (3) Respiratory diseases, (4) Cancers, (5) Mental Health conditions, and (6) Hypertension. The first four and mental health conditions are classified by the World Health Organization as major noncommunicable diseases (NCDs) [36].

Research question 2 aimed to contextualize the variability in costs using country GDP per capita in 2020 (latest available data) [37, 38]. For this analysis, we included only the most frequently reported dyads. We used the same study eligibility criteria as for the meta-analyses (see below), with several conditions relaxed. The criteria that studies must have had the same design, and reported measures of distribution and all-cause costs were relaxed, as we could control for these factors in the model. First, we ran a linear model with annual mean direct medical costs per capita as the dependent variable and GDP per capita as the independent variable taking on fixed effects. Subsequently, we incorporated different study characteristics as random effects. Potential study characteristics that may affect costs are study, study design, data source, and country. We performed log10 transformations on costs and GDP to normalize the distribution and to stabilize the variation within groups. After testing different models and observing variance, p-value and Akaike Information Criterion (AIC), the best fit model consisted of GDP as fixed effects and study and data source as random effects. The analysis was performed in RStudio version 2021.09.2 [39].

To compare the costs of disease combinations and to identify those that resulted in high costs (research question 3), meta-analyses were conducted for the most frequently reported dyads. Studies were categorized to ensure similarities within each sub-group meta-analysis. The criteria for homogeneity were (1) same cost perspective and study design, (2) reporting annual mean direct medical cost, (3) reporting measures of distribution, (4) comparability of cost ingredients determined by recurring ingredients (hospitalization, outpatient care, emergency care, drugs), (5) studies assessed together having either all specified all-cause healthcare cost or not specified at all, and (6) studies that only assessed costs specific to the disease(s) of interest were not included. For studies that reported more than one estimate for the same dyad, these estimates were pooled before being entered into the meta-analysis — provided that the mean cost, its standard error, and the sample size corresponding to each were provided [40‐42]. Where appropriate, costs per month or per 6 months were multiplied by 12 or 2, respectively, to arrive at the estimates for 12 months [40, 42, 43].

Mean cost data were meta-analyzed assuming a normal likelihood for study-specific mean costs. Despite the non-normal nature of healthcare costs [44], the distribution of sample mean costs will approximate a normal distribution as the number of studies increases due to the Central Limit Theorem. Given the low number of studies that were available for some disease combinations, a fixed-effects model was prioritized on practical grounds, acknowledging the strength of the imposed assumption (i.e., a common underlying true cost across all studies). Random-effects models were also attempted, noting that the low number of studies may lead to convergence issues and unrealistic estimates of the between-study heterogeneity [45]. The extent of heterogeneity was estimated and presented by the means of I² [46]. All synthesis models were implemented in OpenBUGS version 3.2.3 [47] using three Markov Chain Monte Carlo chains with different starting values. Estimates were obtained from 70,000 iterations (including 20,000 burn-in). Convergence was checked using the Gelman-Rubin diagnostic, specifically with the multivariate potential scale reduction factor [32], and visually by assessing the history, chains, and autocorrelation. Vague priors were used for all parameters.

Finally, cost ingredients such as diagnostics, outpatient care, hospitalization, emergency care, and medications were tabulated and descriptively analyzed.

Charting of the study characteristics is further presented in Additional file 5. Twelve countries were represented in the 59 studies, mostly high-income countries (95%), primarily the United States (US) (n = 44) and Spain (n = 4) and one multi-country study (Table 1).

Over half of studies were cross-sectional (54%), followed by longitudinal/cohort (44%). Across the 59 studies, the primary source of data was insurance claim databases (63%). Patient survey data were used in 20 studies (34%), so were medical databases. A third of all studies linked data across multiple databases. The “other” category comprises sub-databases, for example, the cancer registry, the national inpatient sample, and the drug prescription registry [48‐50]. Most articles reported direct medical costs (98%), with few reporting direct non-medical (3%) [51, 52] or indirect costs (12%) [51‐57].

Quality assessment scores tended to be higher among cohort/longitudinal studies, and those which had used linked databases. All studies achieved the inclusion threshold of 5 points. Patient surveys had been used in a third of all studies. As acknowledged to contain a risk of bias (recall, language, time), most of the included studies which used patient survey (i.e., Medical Expenditure Panel Survey (MEPS)) [41, 58‐63] had cross-checked collected data with administrative data [64]. Several ambiguities were observed among studies. Those using data from insurance claims databases often did not specify whether reported estimates included copayments or deductibles [65‐67]. Many studies also did not specify whether the reported estimates were of adjusted or unadjusted costs, all-cause healthcare costs or disease-specific costs [60‐62]. A small number of studies did not clearly specify the year of currency, the timeframe of reported costs (per patient per month/year/two years), the sample size related to the estimate, or the measure of distribution [58, 61, 65, 68]. Finally, challenges were present in making comparisons of cost ingredients across studies, with some having broader categorizations (inpatient, outpatient, emergency care, pharmacy) and some having smaller categories (diagnostics, physician, specialist) or a mix of both [43, 65, 69].

The variation in reporting in the underlying studies presented challenges in comparing costs between different disease combinations in a homogenous way. Across the 229 sets of disease combinations, costs per year ranged from I$827 for a TB-diabetes patient in Malaysia [71] to I$147,784 for a patient with HIV and pulmonary circulation disorders in the US (Additional file 7: Table S2) [99]. Several studies reported high costs for multimorbidity in the last 6 months of life or the 6 months immediately after diagnosis. For example, cost in the last 6 months of life for a patient in the US with heart failure and diabetes is I$51,145 [72], considerably higher than that of similar patients earlier in life [33, 60, 90]. Reported costs are also high when they incorporate direct non-medical and indirect costs. In the case of patients diagnosed with stroke and diabetes in Spain, the annual cost reported per patient, which included indirect costs, was I$52,606 [51]. Of this total, direct medical costs account for merely a quarter.

Table S2: Available evidence on the costs of multimorbidity (Additional file 7)

The list of included studies in this section is in Additional file 8 — Checklist and data used for the country comparison. Studies selected for this analysis were of the same cost perspective, reported annual mean direct medical cost, and had similar cost ingredients determined by recurring ingredients (hospitalization, outpatient care, emergency care, drugs). Where appropriate, costs per month and per 24 months were multiplied by 12 and divided by 2, respectively, to arrive at the annual estimates [40, 43, 98]. Two studies reported costs from a societal perspective; hence, we removed the indirect cost from total costs for the purpose of this analysis [53, 57].

First, log10 transformed cost per capita of multimorbidity was plotted against log10 transformed country GDP per capita using a fixed-effects model (Fig. 3). In the fixed-effects model, country GDP per capita accounts for under half of the variation in cost estimates (R² = 0.47; p < 0.001). Every International Dollar increase in GDP on the log10 scale is associated with an expected increase of 1.64 International Dollars in the cost of multimorbidity on the log10 scale (p < 0.001).

Next, we used a linear mixed model with random intercept to adjust for covariates (Table 2). Country GDP per capita accounts for a high proportion of the variation in cost estimates (R² = 0.94; p < 0.001). Every International Dollar increase in GDP on the log10 scale is associated with an expected increase of 1.91 International Dollars in the cost of multimorbidity on the log10 scale. In addition, given that the observations derived from different studies, study and data source accounted for 40% and 47% of the remaining variation in costs, respectively.

The 15 studies in the meta-analyses were selected using stringent homogeneity criteria (checklist and study details in Additional file 9 — Homogeneity checklist and data used for the meta-analyses). This resulted in a small number of data points in each sub-analysis. Random-effects models did not converge for disease combinations with a low number of studies and were not pursued further. Therefore, only results of fixed-effects models are presented for consistency. All meta-analyses resulted in I² values above 95%. In the absence of other analyses utilizing a richer and more homogenous evidence base, these estimates may be the most informative to date.

Of the 11 most frequently reported dyads, patients with Cancer + Mental Health condition within the first year of cancer diagnosis had the highest average annual direct medical costs (I$85,820) and greatest intra-group variability (Table 3). This is the only care-cascade-specific sub-group among the 11 most frequently reported dyads. The cost of this dyad is two to six times higher than that of other dyads.

Patients with Hypertension + Musculoskeletal disorder incurred the lowest average annual direct medical costs (I$13,270), followed by patients with hypertension comorbid with diabetes (I$14,300), respiratory condition (I$17,180), kidney disease (I$17,740), and heart/vascular condition (I$17,880). In general, dyads involving hypertension were among the least expensive to manage.

Mental health condition when comorbid with diabetes incurred lower mean cost than when comorbid with a respiratory condition [I$20,750 (95% CI: 19,830–21,660) vs I$36,840 (95% CI: 36,440–37,250)] — the difference is statistically significant. However, when comorbid with heart/vascular condition, diabetes incurred higher mean cost than respiratory condition [I$37,090 (95% CI: 35,870–38,310) vs I$35,070 (95% CI: 34,890–35,260)] — the difference is statistically significant.

Our results further demonstrate that the costs of multimorbidity may not be additive of individual diseases. For example, the average cost of treating a patient with Diabetes + Kidney Disease (I$32,410) is observed to be larger than the summed cost of treating both {a patient with Hypertension + Diabetes (I$14,300) and a patient with Hypertension + Kidney Disease (I$17,740)}; however, this result is not statistically significant.

Studies reporting total direct medical costs most frequently included costs of hospitalization (n = 57), outpatient care (n = 56), emergency care (n = 50), and drugs (n = 40). Other cost ingredients which were less common were those for consultations, diagnostics, surgery, medical equipment, specialist services, and therapy. For studies that reported direct non-medical costs (n = 2), food and transport costs were reported frequently, while costs of social care were rarely reported [51, 52]. Indirect costs (n = 7) principally focused on productivity loss, including wage and/or time loss, with a single study additionally including the interest from debts/loans [51‐57, 89]. Detailed charting for cost ingredients is presented in Additional file 10.

The availability of accurate and detailed costing studies will be essential in the coming decades to support global efforts to address multimorbidity. This review demonstrates, however, that there is a narrow scope of research and lack of methodological standardization contributing to variability between studies and impeding meaningful comparisons.

Specifically, focusing on index disease(s) and a few pre-selected comorbidities —for example [54, 66, 67, 69, 95] — rather than exploring the costs of possible combinations from a broader population perspective limits findings to few combinations. This makes it difficult to understand the full spectrum of the economic burden of multimorbidity in a country and create a basis for comparison between countries.

Moreover, the majority of studies obtained their costs from standalone data sources. Such an approach cannot provide a holistic view of healthcare costs including aspects such as non-medical and indirect costs [104]. A “bottom-up” approach to estimating the costs of chronic diseases using a medico-administrative database would be optimal in costing for multimorbidity [105]. Patient surveys have been found to underestimate costs, and ultimately, these should be used in conjunction with other data sources to correct for recall bias and to capture any costs unknown to healthcare users [106]. The practice of linking health data across multiple interfaces [15] and steps to linking population health data with administrative claims database have been endorsed [107].

Furthermore, with costs markedly higher in the 6-month windows before death and after diagnosis, further studies taking a longitudinal design would provide a clearer understanding of how costs and drivers evolve along the disease trajectory [21]. Understanding these transition stages is essential in the design and targeting of appropriate interventions [27].

The clear lack of agreement on the scope of health services to include, or which ingredients should be costed, along with a diverse array of methodologies and definitions contribute to heterogeneity between studies. A recommended guideline to collecting and estimating costs of multimorbidity may help in standardizing the definitions, process, and components and reducing the level of variability between studies [21, 48].

Finally, the scarcity of studies on the cost of multimorbidity from LMICs raises concern over equity. LMICs have become the new hub for NCDs and more research is imperative to setting out priority health agendas and informing the re-organization of healthcare delivery to support patients with multiple needs [13].

To contextualize the possible drivers of the difference in costs among different country settings, we used GDP per capita as a determinant. While the data available for these analyses were again skewed by the predominance of American studies (14 out of 28 data points from the US), our regression showed that costs were generally higher in countries with higher GDP per capita [70]. Other studies on the costs of single chronic diseases such as diabetes have also pointed to similar findings [108]. This is not to be unexpected, and consistent with the fact that countries with higher GDP tend to have higher government spending on healthcare, driving up healthcare utilizations and costs [108]. However, the large variation in cost within a country (in this case, the US) also calls for in-depth assessment of the cost drivers of multimorbidity within a health system (system-level factors). The organization of care, e.g., standard vs integrated care, may play an important factor, though was not reported in the studies.

The high in-group variability of the costs of dyads (even when they all originate from the US — see Additional file 8) is in part due to the fact that these dyads were constructed as generic disease groups. For example, there are various cancer sites and mental health conditions with different levels of severity, treatment, duration, etc., which results in variability in costs. Additionally, the in-group variability for costs in the US might be explained by the use of different data sources and other study characteristics (study-level factors).

Although the cost of multimorbidity in part depends on the stage of the diseases alongside other factors, the knowledge of high-cost/low-cost disease combinations may inform the development of new integrated care models, where patients are classified according to their short-term/long-term need for specialist/routine care.

Using a fixed-effects model, results from the meta-analyses showed that patients with Cancer + Mental Health within the first year of cancer diagnosis had the highest average annual direct medical costs. This figure (I$85,820) is double that of another US study included in this review, which reported the cost of cancer and mental health not specific to any stage of the cascade of care (I$43,320) [103]. Indeed, treating cancers is expensive as research and development for cancer drugs and therapies are still ongoing [109] and it is even higher still during the 6–12 months following diagnosis compared to the period before or after that [110‐112]. When comorbid with depression, the overall healthcare cost of treating cancer patients has been found to increase by 113% compared to non-depressed cancer patients [113].

Dyads involving hypertension were among the least expensive to manage. For patients with kidney disease, having a comorbidity of hypertension costs less than half of having a comorbidity of diabetes. Similarly, in the case of patients with a heart/vascular condition, having hypertension costs around half of having a respiratory condition. Specifically among hypertensive patients, treatment for musculoskeletal disorder or diabetes is cheaper than kidney disease or heart/vascular condition. A study found that based on the number of episodes that occurred within a period, and the frequency of transition from primary to specialist or emergency care, conditions such as hypertension may indicate “low severity of healthcare impact” and conditions such as kidney disease, heart/vascular disease, and respiratory disease may indicate “high severity”; thus the effect on costs [13]. Many chronic conditions feature commonly in multimorbidity dyads, and whether the cost of a disease combination is more or less than the additive costs of the component diseases needs to be further explored.

Although all studies included in the meta-analyses originated from the US, our findings provide valuable data on the differences in costs for treating different disease combinations within the American healthcare system. Though healthcare systems vary considerably between countries, the difference in costs between these 11 dyads in the US provides an indication of how they may be in other countries; however, more research in this area is needed.

Our review highlights several missing pieces of the multimorbidity burden puzzle, notably that non-medical and indirect costs of multimorbidity are often not costed for. Particularly, social care is considered an important element of care for multimorbid patients given the long-term spectrum and complexity of illness, which may have lasting effects on their care needs [114]. In some settings, research has shown that social care cost may drive total care cost more than healthcare cost itself [18]. However, costing studies on multimorbidity which examine the component of social care/home care are still by and large limited [18, 115, 116]. This aspect is critical to understanding the full spectrum of multimorbidity costs to the system, especially when this has implications for vulnerable groups in society [27].

The most frequently reported cost ingredients identified were outpatient, inpatient, emergency care, and drugs. However, across studies, certain cost ingredients may overlap with one another (e.g., outpatient and GP visits, hospitalization and inpatient services, diagnostics and testing). Due to the ambiguous use of terms, it is difficult to clearly pinpoint the definition of the specific cost ingredients and what they entail.

To the best of our knowledge, our review is the first to systematically collect and quantitatively synthesize costs for disease combinations with comparability. We screened a large number of articles having conducted an extensive search. Beyond answering our research questions, we have additionally provided valuable insights into commonalities and inconsistencies in underlying methodologies between studies, and the ways in which these studies reported costs.

In the meta-analyses, the degree of heterogeneity and variation of the underlying studies limited the number of studies included, all of which were from the US. Several sub-group analyses (e.g., Respiratory + Mental health, Respiratory + Heart/vascular condition) comprised only two data points. This may have limited the plausibility of the assumed approximated normality of mean costs. Furthermore, high I² values were estimated for all disease combinations indicating a high degree of between-studies heterogeneity. Despite this, these results may still provide the best available insights though need to be interpreted in light of the highlighted methodological challenges.

The limited availability of studies from other countries led to the country comparison being dominated by studies from the US, with variation in these costs. More data from other countries are required to further examine this relationship.

Overall, we acknowledge that costs of multimorbidity may depend on many different factors including and beyond those that have been discussed in this paper. These may include individual-level factors (e.g., socio-demographic characteristics of patients, the number and severity of comorbidities, type and duration of healthcare intervention, access to care), system-level factors (e.g., the country, health system financing structure, level of care — e.g., public vs private, primary vs secondary care), and study-level factors (e.g., sampling error, costing perspective, data source, choice of cost ingredients). These details, however, were frequently unavailable, or not available to a sufficient level (e.g., severity of conditions, duration of care, type of health facility, etc.) to allow for structural reporting. Therefore, we were unable to control for some of these factors in the analysis. Despite these limitations, our review has revealed important findings.