Background
Over the last century, developed nations had made significant improvements in overall health outcomes and life expectancy. Health care equity was a primary motivation for universal health insurance systems; however, socioeconomic status (SES) has remained a determinant of health and has strong associations with morbidity and mortality [
1,
2]. Patients with low SES die younger, have higher disease burden, use less preventative health, and present later in the disease course [
3,
4]. Low SES patients also have greater health care needs compared to the general health care population [
5]. As SES is a multifaceted phenomenon, the measurement and definition of SES is complex and can range from income, education, and occupation, to harder to measure variables such as health behaviour, social participation and composite measures such as social deprivation indices [
6,
7]. In general, low SES patients have decreased access to needed health care services, even in countries with universal health coverage [
8,
9].
The end-of-life (EOL) period not only represents a period of high health care use, unmeasured differences in medical need across patient populations are also thought to be attenuated in the terminal years by virtue of the fact that all such patients die. Differences in EOL expenditures according to SES may represent variations in access to medical care, and yield insights in health care seeking behaviours, location of care, medical decision-making, and health care resource allocation [
10]. As health care expenditure increases rapidly in the time close to death, understanding SES inequalities at EOL is necessary for future health care planning to reduce such social inequalities [
11‐
13]. Accordingly, the objective of this review was to evaluate the relationship between SES and cost of health care at EOL.
Methods
Data sources and searches
The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines were followed in the conduct and reporting of this meta-analysis [
14]. We searched the following databases for eligible studies: MEDLINE, EMBASE, CINAHL, ProQuest, Web of Science, Web of Knowledge, and OpenGrey. Searches through December 2019 were conducted with no language or publication date restrictions. The following keywords were applied to the search: (income or socioeconomic or education) AND (end-of-life or last year* of life or last month* of life) AND (cost* or expenditure* or spending*). See Additional file
1 for all strategies used. Bibliographies from relevant publications were checked to identify relevant articles.
Eligibility criteria
For the qualitative analysis, we included studies that were observational studies with a sample size of over 1000 adult (18 years or older) participants to reduce sampling error, as studies often compared ≥2 socioeconomic groups. No language, publication date, or publication status restrictions were imposed. Each study had a measure of SES, including income (individual, household, or neighbourhood), education, or other proxy measures such as neighbourhood poverty rate. Notably, studies that only examined race, ethnicity, or occupations in relation to EOL cost were excluded as these measures could influence EOL costs in ways beyond what can be accounted for by income [
15]. The outcome measure was health-related costs in the last year(s) or month(s) of life. In addition, studies must investigate cost at EOL in relation to SES.
For the meta-analyses, we included articles selected in the systematic review that also satisfied the following conditions: the study provided EOL cost data for the last year of life, measured SES by income (individual, household, or neighbourhood), and measured total health care costs (total absolute or total covered by an insurance scheme) for a high- and a low-income category of participants. All studies included in the meta-analyses were peer-reviewed and not limited to the focus of a specific disease category (i.e. cancer).
Study selection
Two reviewers independently screened the titles and abstracts of all references, both were blinded to the authorship (CY, DA). Potentially relevant articles were retrieved, and full texts were screened independently by the two reviewers using an eligibility checklist for inclusion in systematic review. Disagreements were resolved by consensus. All reasons for exclusion were documented into a PRISMA flowchart [
14]. For all included studies, another eligibility checklist was used for inclusion in the meta-analyses.
Data extraction and management
For study characteristics, first author, year of publication, location, purpose, study design, and funding source were extracted from included studies. For cohort characteristics, the number of patients, inclusion/exclusion criteria, and data source used to identify patients were extracted. The type of SES measurement and data sources were extracted, as well as the EOL period (e.g. last 1 year of life) and aspects of cost examined (e.g. out-of-pocket). The main outcome measures and study conclusions were summarized. It was also noted whether regression analyses were performed, and if comorbidities or clinical complexity were adjusted for.
Comprehensiveness of reporting was assessed using the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines [
16]. Only studies that fulfilled ≥13 of 22 checklist items were included in the study.
For the meta-analysis, data was extracted to determine eligibility: currency used and year, EOL health care costs and/or regression results in highest and lowest SES categories in last year of life and spread measures (e.g. 95% confidence intervals).
Authors of two studies were contacted to provide unpublished EOL total costs for the meta-analysis, neither had the data required.
Risk of bias in individual studies
Methodological quality was assessed by both reviewers independently using a modified National Heart, Lung, and Blood Institute (NHLBI) Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies [
17]. Eight of the 14 items of the tool were used to evaluate the risk of bias in each study, which assessed study quality within the domains of study population, data collection, and data analysis. Risk of bias categories were judged by tallying the sections’ results: > 80% yes was low risk/high quality, 60–80% yes was medium risk/medium quality, and < 60% yes was high risk/low quality.
Synthesis of results
Two meta-analyses were performed using Review Manager (version 5.3): the first using direct cost data (no adjustments for confounders), and the second using regression results (adjusted for comorbidities) [
18]. The meta-analyses were restricted to high- and medium-quality studies. All costs were converted to 2015 USD using an online tool [
19].
For the meta-analysis that did not adjust for comorbidities, the principal summary measure was standardized mean difference (SMD) for the cost of last year of life between high and low SES groups. The SMD expressed the size of the intervention effect in each study relative to the variability observed in that study. An inverse-variance random-effects method was used as high levels of heterogeneity were expected due to the observational nature of studies and the variety of SES measures used [
20]. For the meta-analysis that adjusted for comorbidities, we employed the use of regression coefficients for the cost of last year of life between high and low SES groups. An inverse-variance fixed-effects method was used as the heterogeneity was not significant between the studies. For both analyses, values greater than zero indicated the degree to which high SES was associated with higher EOL cost than low SES, and values less than zero indicated the degree to which high SES was associated with lower EOL cost than low SES. Regression coefficients were divided by 10 to fit the scale of the forest plot, and the interpretation of results was scaled up tenfold in response.
Heterogeneity was tested for both meta-analyses with the Breslow-Day test using method proposed by
Higgins et al. to measure inconsistency (the percentage of total variation across studies due to heterogeneity) [
21].
Risk of bias across studies
Publication bias was assessed by evaluating a funnel plot of the effect by the inverse of its standard error. The symmetry of the plot was assessed visually. As fewer than ten studies were included in the meta-analyses, statistical tests for asymmetry were not performed [
22]. We acknowledge that other factors, such as differences in study quality or true study heterogeneity, could produce asymmetry in funnel plots [
23].
Role of the funding source
This study received no specific external funding.
Discussion
The aim of this review was to evaluate the relationship between SES and health expenditure at EOL. We found that patient SES was significantly correlated with EOL expenditures. Overall, the evidence suggested significant heterogeneity in units of cost, length of EOL period, extent of adjustment, and directionality of conclusions. One of the key factors that accounted for the variation in SES-EOL cost inequalities was adjustment for comorbidities. When unadjusted for comorbidities, low SES was associated with lower total and hospital EOL expenditures. Conversely, when adjusted for comorbidities, low SES was associated with higher total and hospital EOL expenditure. Irrespective of adjustment, low SES patients had lower specialist, out-of-pocket, and drug expenditure at EOL even within jurisdictions providing universal health coverage to its citizens.
To the best of our knowledge, our systematic review and meta-analyses were the first to examine the effects of socioeconomic variation on EOL cost of care. While SES has been shown to be an important determinant of population health across the life continuum [
43‐
47], its relationship with health-care expenditures was less consistent and more variable across jurisdictions [
48]. The focus on EOL care has also been less studied than the relationships between SES and health-care during other periods of life.
Our systematic review and meta-analyses demonstrated inconsistent results for the relationship between SES and total (and more specifically, hospital-related) EOL expenditures depending on whether or not individual studies employed comorbidity risk-adjustment methodology. Given the importance of comorbidity as a determinant of EOL expenditures [
49,
50] and the presumed higher comorbidity disease burden among socioeconomically disadvantaged populations [
24,
27,
31,
35‐
38,
40,
41], one would have hypothesized that SES inequalities in hospital-related EOL expenditures would have only become more pronounced among studies that incorporated comorbidity risk-adjustment than among studies that did not. Such was not the case; if anything, the converse was true. Several possibilities may explain such counterintuitive findings. First, “immortality bias” among lower SES patients who presented to hospital may have been present as a result of a disproportionately higher number of socioeconomically-disadvantaged patients with extensive comorbid disease burden succumbing from an out-of-hospital sudden cardiac death [
51,
52]. Such a bias may have skewed or altered the burden of comorbidity between lower and higher SES populations. Second, comorbidity adjustment, mostly measured through Charlson Comorbidity Index Score and the Johns Hopkins Aggregated Diagnostic Groups [
51,
52], may have been incomplete leading to residual unmeasured confounding. Third, unlike mortality, not all comorbidities are as predictable in their relationships with health care spending [
53]. Fourth, the propensity to hospitalize patients may have been driven by factors other than comorbidities. For example, socioeconomically-disadvantaged patients may have had fewer ambulatory and community support-systems which necessitated EOL hospitalization irrespective of comorbid disease burden. Accordingly, risk-adjustment methodology may have not appropriately accounted for the health-cost consequences associated with comorbidities in general. Finally, SES itself is a complex measure that is confounded not only by comorbidity, but also by variations in psychosocial stress, social supports, health behaviours, and health literacy. Such factors when combined with medical care access inequalities, and physician medical-decision-making biases for EOL care may have in part, or together explained the inconsistencies between those studies which did and did employ risk-adjustment methodologies. Whether SES-EOL expenditure associations should employ comorbidity risk-adjustment methodology remains unclear, as comorbidity is intertwined within the SES measurement itself.
Notwithstanding the important modulating effects of comorbidity adjustments on hospital (and total) EOL expenditures, our qualitative synthesis and meta-analyses demonstrated higher EOL ambulatory care and drug expenditures among socioeconomically advantaged than disadvantaged individuals, suggesting inequity in out-of-pocket spending. While high SES patients spent more out-of-pocket in absolute amount in the last year of life [
28], low SES patients had higher out-of-pocket expenditure proportional to their annual income [
32]. This result extended as far back as the last five years of life [
26]. In contrast, high SES patients had a higher informal caregiving cost, likely due to more availability of informal care providers and financial resources [
26]. High SES patients also spent 15% more on specialist care, and 23 to 25% more on prescription drug in the last year of life than their low SES counterparts [
24,
31]. Higher SES patients had more prescriptions filled and spent more money out-of-pocket on prescriptions [
31]. These results were not surprising, as socioeconomic differences make accessing specialist and pharmaceuticals easier for those with higher income throughout life [
54]. These differences include education, private support systems, employer-based insurance, and relative affordability of any of out-of-pocket payments [
24]. It is also important to note that most studies examining out-of-pocket and drug expenditure were based in the U.S.
Reasons for SES-EOL expenditure inequalities may be multifactorial and complex. Different models of care, durations of EOL period investigated, sociocultural preferences around death and dying, and definitions of “high” vs “low” SES may be other potential explanations for between-country differences. For example, studies on total EOL costs covered by public insurance schemes varied from weak positive SES-EOL relationships for countries that provide universal public insurance [
11,
25,
27,
30,
40], to weak inverse relationships for U.S. Medicare [
29,
33,
36‐
38]. There was some evidence that factors generating the observed inequalities in universal health care systems differed from those in market-based health care systems. For example, in the Swedish study by Hanratty and colleagues, patients with high SES had a higher median number of hospital bed days [
30], which contrasted the higher inpatient costs associated with low SES patients found in other studies [
24,
35,
41]. This suggests a higher use of hospital care in Sweden for higher SES patients that was not present in other countries.
Future studies examining the impact of SES on EOL cost should examine costs of individual service types as well as aggregated costs, and adjust for health care status using measures of clinical complexity such as comorbidities [
24]. The complexity in the associations between SES and EOL care use and cost necessitates careful examination of modulators such as location of death, cause of death, and age at death. The relationships between SES and other factors important to EOL care, such as age and race, must be explored to understand the social patterning in the terminal years of life.
Our study has important population health and health service research implications. First, the demonstration of SES-health expenditure inequalities at the EOL reinforce the notion that SES-related health and/or health-care differences persist not just at younger phases of life, but rather throughout the entire life-continuum. Second, higher EOL ambulatory and drug expenditures among socioeconomically advantaged populations irrespective of privately- or publicly-funded health care systems may underscore health service inequities that can undermine universally-funded public health care systems. SES-inequalities in health spending in universally-funded are likely smaller than those in other health care systems where affordability barriers to access medical care are more pronounced [
55,
56]. Third, our results highlight the need for further research which is designed to better understand appropriateness (or lack thereof) associated with SES differences in EOL health care spending. At this time, there is limited knowledge on the cause of socioeconomic inequalities at the EOL. Specifically, further evaluation is required to better understand the health-care seeking needs of patients and the delivery-care perspectives of their providers when managing socioeconomically diverse populations. Finally, further research must not only explore whether SES-differences in EOL health care spending impacts on outcomes (e.g. quality of life), but also address the heterogeneity in methods employed across studies. This involves establishing guidelines on factors that significantly impact results, such as health care services examined, adjustment of confounders, types of SES measures.
This study had some noteworthy limitations. First, our meta-analyses combined data across studies in order to estimate the effect of SES on EOL cost; yet, socioeconomic measures, diseases, EOL time periods, and outcomes varied across studies and were reflected in the heterogeneity (I
2). In particular, high statistical heterogeneity existed in the larger meta-analysis. In addition, the meta-analysis with adjustment for comorbidities only included two studies, which limits its statistical power. While income reflected current spending power, it may have been confounded by reverse-causality [
57]. Composite measure of SES (e.g. neighbourhood income) did not permit the study of how individual SES factors impacted health, and many studies did not justify the choice of SES measure [
58]. In addition, focusing on the last 12 months of life was an arbitrary choice and could only be considered a proxy for EOL cost. Lastly, there was a focus on high-income countries, which limit the applicability of conclusions of this review to middle- and low-income countries.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.