Cost parameter
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].
Analysis and presentation of results
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.