Differences in survival and mortality in minority ethnic groups with dementia: A systematic review and meta‐analysis

Although there are disparities in both risk of developing dementia and accessibility of dementia services for certain minority ethnic groups in the United States and United Kingdom, disparities in survival after a dementia diagnosis are less well‐studied. Our objective was to systematically review the literature to investigate racial/ethnic differences in survival and mortality in dementia.


| INTRODUCTION
Dementia is an increasingly important challenge in public health across the globe: in 2017, dementia was the leading cause of death in the United Kingdom, causing 12.7% of deaths, and the sixth-leading cause of death in the United States. 1,2 However, dementia does not affect all populations the same way. Risk of dementia incidence is thought to be higher in some minority ethnic (ME) groups in both the United States and United Kingdom. In the United States, studies have found higher risk of dementia in some African American and Hispanic/Latino groups. [3][4][5] Black and South Asian populations in the United Kingdom are also thought to have higher risk for dementia than the White British population due to higher rates of vascular risk factors. 6,7 ME groups in both countries are also thought to access care for dementia later in the disease course and at lower rates due to barriers imposed by structural and institutional racism, negative experiences with healthcare systems, discrimination, language barriers and stigma. [8][9][10] Despite evidence of disparities in risk of dementia and access to care, it is unclear how this affects the course of dementia, including survival patterns. A 2013 systematic review on predictors of mortality in dementia only identified five studies examining race/ethnicity, and only one found differences: lower mortality in African American and Latino individuals with Alzheimer's disease (AD) versus comparison groups. 11,12 Since then, additional US studies have reported reduced mortality risk in ME groups as compared to non-Hispanic White groups using health record data. 13 Total deaths attributed to dementia were also higher among the non-Hispanic White population in 2017: 70.8 per 100,000 as compared to 65.0 in the non-Hispanic Black population and 46.0 in the Hispanic population. 2 Although there are fewer studies outside the United States, some studies from the United Kingdom indicate that Asian, Mixed, and White Other ME groups may survive longer after dementia diagnosis than White British groups. 14,15 One difficulty in carrying out this research is that there is not a consensus on the language used to describe race and ethnicity; the two are often used interchangeably despite having different meanings. 16,17 While both terms are socially constructed, 'race' has historically implied that there are visible physical or biological differences among people, whereas 'ethnicity' tends to refer to social groups and include more fluid, self-defined and cultural in-group characteristics. [16][17][18] Here, we use the hybrid 'race/ethnicity' or substitute 'ethnicity' alone. Ethnicity is self-ascribed, which may not match what researchers or outside observers report based on visible characteristics. 16 However, studies do not always ask participants themselves what ethnicity they identify with. Ethnicity is also context dependent, 16 and while health disparities due to structural racism are an issue globally, 19,20 race/ethnicity groups are not comparable across countries. Thus, countries are analysed separately in this review (Further details on the interpretation and use of race and ethnicity terminology may be found in Text S1.).
Nevertheless, it is important to explore inequalities between ethnic groups in order to set priorities for policy and plan equitable health services. To date, there has been no systematic review comparing mortality rates across different racial/ethnic groups with dementia. The objective of this study is to determine whether there are racial/ethnic differences in mortality in dementia in the literature.

| Protocol and registration
The study protocol was registered in PROSPERO (CRD42018118129) (Appendix B). While this review focuses on mortality, it is part of a larger project investigating racial/ethnic disparities in both mortality and access to services in dementia. Thus, it shares a search strategy with another review exploring use of healthcare services in ME groups (PROSPERO CRD42018118132). 21 Screening for both was done concurrently.

| Inclusion and exclusion criteria
The following inclusion criteria were applied: (1) must be quantitative observational studies; (2) must compare survival or mortality across multiple race/ethnicity groups, even if not the primary study aim; (3) population followed longitudinally; (4) participants have any type of dementia or mild cognitive impairment; (5) studies including participants without dementia must provide an estimate for survival by CO ET AL.
-1641 ethnicity in the dementia-only group. Studies were excluded if: (1) the study population was defined by another unrelated disease or treatment; or (2) the outcome was specific-cause mortality rather than all-cause mortality. Cohort studies following individuals from diagnosis or first presentation to health services until death and reporting hazard ratios were further eligible for meta-analysis.

| Search
The search was conducted in Ovid on 7 November 2018 and included peer-reviewed articles from Embase, Ovid MEDLINE, Global Health and PsycINFO from inception until the search date. The shared search strategy followed this structure, including subject headings and synonyms of each term: (dementia OR Alzheimer's disease) AND (ethnicity OR race) AND (mortality OR survival OR service use). A full list of search terms is in Appendix C and PROSPERO.

| Study selection
Titles and abstracts of articles were double-screened by four authors (MC, EC, QG, AM) using Rayyan, an online application for managing systematic reviews. 22 One author (MC) screened all records and three screened 50% of the records. Any disagreements at this stage led to the paper being included in full-text screening. Full texts of articles were then independently reviewed for eligibility by two authors (MC and EC). Disagreements were discussed between the two authors, and further discussed with a third reviewer (MP), if a decision could not be reached. If more information was required, the corresponding author of the article was contacted.

| Data extraction
MC extracted data on all studies using a template in Excel. Fields extracted included study characteristics such as: country, study population, setting, type of dementia, years study completed, sample size, ethnicity groups included (maintaining terminology used by study authors), mean age at baseline, percent of female participants and main study objective. Data on study results were also extracted: outcome definition, type of analysis, follow-up time, crude and adjusted statistics, 95% confidence intervals (CIs), and other covariates studied.

| Risk of bias
The Newcastle-Ottawa Scale (NOS) was used to assess risk of bias in individual studies. 23 Two additional questions were included: (1) whether ethnicity was the main study focus (yes/no/exploratory) and (2) whether ethnicity was discussed explicitly in the study results (yes/no). For studies using health record data, it was assumed that there was little loss-to-follow-up. Because ethnicity is best defined through self-identification, 16 studies were awarded a star for 'ascertainment of exposure' if ethnicity was self-reported.

| Data synthesis
Results from included studies were tabulated. If no statistics were presented for ethnic differences in survival, we calculated these using data available in the article.
A meta-analysis combining studies which reported hazard ratios of mortality from diagnosis, onset, or first presentation to services was performed in R version 3.6.0 using 'meta', 'metafor' and 'robumeta' packages. [24][25][26] Due to expected heterogeneity (from factors such as study locations with different racial/ethnic makeup), we used a random-effects meta-analysis with inverse variance weighting.
When studies provided estimates from multiple models, the mostadjusted estimate was used. Because definitions and experiences of ME groups vary by country, we did not combine studies from different countries. Ethnicity was analysed as a subgroup, and we preserved the original ME groupings from the primary papers. See Text S1 for further information on treatment of race/ethnicity variables.
Studies often used the same comparison group to calculate hazard ratios for multiple ME groups, but a naïve meta-analysis assumes independence between each estimate. To account for this dependence, we used robust variance estimation (RVE). 27 Correlated effects weights were used, as well as Tipton's small sample correction because there were few studies. 28 Rho was set at 0.80. Both naïve and RVE estimates are reported. Further statistical subgroup analyses were not completed due to the small number of studies. As there were fewer than 10 studies in the meta-analysis, risk of publication bias was not quantified or shown as a funnel plot. 29

| RESULTS
The combined search strategy retrieved 8977 records after autodeduplication in Ovid and Endnote. After title-and-abstract screening, 8897 records were removed. Full texts of 80 articles were assessed for inclusion. Three additional articles were identified and included in full-text screening; two from known papers in background literature 15,30 and a full journal article for a conference abstract found in the search. 31,32 Twenty-two articles from 19 studies fulfilled the eligibility criteria. There was 90% agreement between the two reviewers (Cohen's kappa = 0.78). Most articles excluded were conference abstracts. Many articles were excluded because they did not compare different ethnic groups or included participants without dementia. Four studies were considered potentially eligible because ethnicity was mentioned in their analyses but were ultimately excluded because no statistics were reported or could be calculated, and authors did not respond to requests for data. [33][34][35][36] We were also unable to obtain exact statistics for two included studies, although both reported non-significant results for difference in mortality by ethnicity group. 37,38 The PRISMA flow diagram in Figure 1 includes more details on the selection process.

| Study characteristics
All but four studies were from the United States (n = 14, from 17 articles), with one from the Netherlands and three from the United Kingdom. Four articles used data from the Washington Heights Inwood Columbia Ageing Project, but their outcomes were too different to combine. [39][40][41][42] Ethnicity groups studied in the United States included: White (all studies), African American/Black (n = 13 articles/10 studies), Hispanic/ Latino (n = 10/7), Non-white (n = 4/4), Asian (n = 2/2) and Indigenous (n = 4/4). Studies did not always report whether ethnicity groups labelled as 'White' or 'Black' also included people identifying as Hispanic/Latino. In the study from the Netherlands, the authors defined ethnicity by country of birth and included Dutch, Indonesian, Turkish, Surinamese and Antillean groups. One study from the United Kingdom followed Office of National Statistics ethnicity categories, while two others used different terminology or aggregated categories.
On average, articles included 63.9% female participants (range: 44.9%-81.8%) and 32.8% minority racial/ethnic participants (range: 3.3%-92%). The average baseline age in each study (where reported) ranged from 71.5-87 years old. Twelve articles included only participants with probable/possible AD, while the rest defined dementia more broadly or included multiple dementia subtypes (n = 10).
Further details on characteristics of studies can be found in Table 1.

| Risk of bias in individual studies
Study quality was mixed (one to nine total stars), as quality was assessed according to the analysis of racial/ethnic differences and  tended to be lower in studies where this was not a main objective (n = 13). All but two studies discussed race/ethnicity results explicitly in the article text. Across studies, cohort selection quality was moderate because most articles did not use self-report or did not describe how ethnicity was ascertained (n = 6 receiving stars). Twelve articles received three or four stars for cohort selection items. Fourteen articles were given two out of two stars for the comparability item. Seven articles had three (out of three) star ratings for the outcome items; studies with more stars tended to use routine/registry data linked to national death records. NOS results are reported in Table 2 and   Table S1.

| Results of individual studies
Results of individual studies may be found in Tables 3 and 4, and

| Synthesis of results
No studies found higher risk of mortality among ME groups in any

| Meta-analysis of United States studies
We meta-analysed six studies from the United States (total 101,502 participants) which reported hazard ratios from onset, first presentation to services, or diagnosis. There were too few studies to perform subgroup analyses examining differences between these starting points. All studies included in meta-analyses adjusted for at least age and sex/gender. Figure 2 shows a forest plot of these studies.
The pooled hazard ratio of mortality for all ME groups in the with and without small sample correction found that, without including the correction, confidence intervals did not cross one.

| DISCUSSION
As far as we are aware, this is the first systematic review examining ethnicity/race differences in mortality in people with dementia. We found evidence from the meta-analysis that mortality risk may be lower in Black/African American (by 14%) and Hispanic/Latino (by 35%) groups in the United States as compared to reference groups. However, Asian American and Indigenous groups in the United States and other ME groups in other countries were less well-studied.
These findings are similar to demographic trends which have been described in populations without dementia in the United States-the 'Black-White mortality crossover' and 'Hispanic mortality paradox'. The Black-White mortality crossover posits that Black American groups have disproportionately higher risk of mortality compared to White American groups until they reach older age, after which mortality rates 'cross over' and they begin to have lower mortality risk. 56,57 It is unclear why this trend exists; theories have been proposed arguing both underlying demographic mechanisms as well as spurious results due to data biases (namely, inaccurate or incomplete recording of ME individuals' data). One demographic theory proposes that individual differences in frailty result in healthier individuals surviving past the initial higher mortality risk to older age and surpassing comparison groups' average life expectancies. [57][58][59] However, other studies have reported that systematically inaccurate reporting of age and mismatched linkages to death data, which become even more pronounced as a cohort ages and there are fewer members, may underlie these findings rather than a true crossover pattern. [60][61][62] The 'Hispanic paradox' describes the comparable-orbetter health outcomes, including longer survival, reported in Hispanic American groups despite lower-on-average socioeconomic status and education levels. [63][64][65][66] Lower mortality in Hispanic/Latino groups has been reported in general populations 64,66 and specifically in older adults. 67  Exact statistics not reported (and could not be calculated).   American subgroups. [78][79][80] Prevalence and incidence of dementia has also been found to differ across groups which are often combined under Hispanic/Latino American categories. 5 Socioeconomic and cultural differences between these groups are also masked. More detailed reporting of race/ethnicity definitions can provide additional clarity around what is being measured, although these concepts are intrinsically fluid and context dependent. 16 Future research should seek to clarify mechanisms for observed race/ethnic differences in mortality. This includes issues with reporting of variables such as ethnicity, age and death, as well as the clinical, social and economic implications of longer survival in dementia for ME older adults and their families. For example, longer survival may contribute to prolonged medical or residential care expenses, potentially exacerbating the already existing race/ethnic inequalities in out-of-pocket healthcare costs for people living with dementia and their families. 81 ME groups in the United States are also more likely to live in multigenerational homes, potentially allowing for greater familial support and reducing risk of nursing home placement, 75,[82][83][84] which has been associated with shorter survival in dementia. 45 However, prolonged survival may increase caregiver burden and stress and be detrimental to carers' mental and physical health and social and family relationships. 85 Larger caregiver burden in turn affects outcomes for the person living with dementia, including risk of nursing home placement. 84,85 Given the additional literature on disparities in risk of developing dementia and barriers to care faced by ME groups, it is important that future research and policy work to reduce inequalities and improve quality of life for ME people living with dementia and their families.

| Strengths and limitations
Agreement was high between reviewers assessing study eligibility, which may indicate that the inclusion/exclusion criteria will be replicable for future reviews/updates.
Additionally, the studies included were large, often drawing participants from registries or health record databases. This reduces concerns around low power and unrepresentative samples, although ME groups often comprise a smaller percentage of the sample, and power may still be a concern where ME group sample size is low. Certain populations may also be underrepresented if they do not access formal health services (particularly underserved minorities, who are thought to access dementia services at lower rates). 9,10,86 In research-focused AD centres, recruitment processes may differ between ME groups, potentially leading to selection biases in these cohorts as compared to in standard health services.
Because demographics and definitions of race/ethnicity vary by country, we cannot and do not generalise between countries.
Furthermore, no studies were found outside the United States and Europe. Although we did not specify language exclusions, we may have been limited by the search terms used to capture race/ethnicityother countries may use other culturally-specific language to describe race/ethnicity groups of which we are unaware.
Although we were unable to quantify it, publication bias might be less of an issue for studies where race/ethnicity is not the main effectthey may be equally likely to be published even if results comparing race/ethnicity are non-significant. However, this lack of focus on race/ ethnicity leads to other issues such as misclassification of ethnicity, lack of power, little information on how ethnicity groups differed demographically and a wider range of covariates between studies. It is also possible that other eligible studies exist which were not found in the search because ethnicity was not mentioned in the title, abstract, or subject headings.
Finally, in our meta-analysis, we chose to use the most-adjusted estimate reported in the study. This approach assumes that these estimates are more comparable, although they may not be. Studies differ greatly in their definitions and reporting of ethnicity. More consistent methodology and reporting of ethnicity in future research will be necessary to understand disparities in mortality in dementia across different race/ethnicity groups.