Cognition and Cognitive Changes in a Low-Income Sub-Saharan African Aging Population

Abstract

Background:

Cognition and its age-related changes remain vastly understudied in low-income countries (LICs), despite evidence suggesting that cognitive decline among aging low-income populations is a rapidly increasing disease burden often occurring at younger ages as compared to high-income countries (HICs).

Objective:

We examine patterns of cognition among men and women, 45 + years old, living in rural Malawi. We analyze how key socioeconomic characteristics predict levels of cognition and its changes as individuals get older.

Methods:

Utilizing the Mature Adults Cohort of the Malawi Longitudinal Study of Families and Health (MLSFH-MAC) collected during 2012–2017, we estimate standard regression models to analyze predictors of the age- and sex-specific levels and longitudinal changes in cognition. Cognition is assessed with a screening instrument that is adapted to this low-literacy context and measures different domains such as language, attention, or executive functioning.

Results:

Women have lower levels of cognition than men, a pattern in stark contrast to findings in HICs. Schooling and socioeconomic status increase the probability of having consistently high performance during the cognitive assessment. Cognitive decline accelerates with age and is detectable already at mid-adult ages (45–55 years). Despite lower levels of cognitive function observed among women, the pace of decline with age is similar for both genders.

Conclusion:

Women are particularly affected by poor cognition in this context. The study emphasizes the importance of prioritizing cognitive health and research on cognition among older individuals in sub-Saharan Africa LICs, to which relatively little health care resources continue to be allocated.

Keywords

Age- and sex-specific patterns Alzheimer’s disease cognition gender differences longitudinal changes low-income countries Sub-Saharan Africa

INTRODUCTION

Brain health is the fundamental resource for individuals of all ages to maintain and participate in all human activities [1], and cognition is an integral part of population health [2]. With the rapid growth of aging populations globally, age-related declines in cognition and the increase in the prevalence of dementia worldwide have a significant and growing contribution to the global burden of disease [3, 4]. The number of people affected by poor cognitive health and dementia is projected to rise sharply from about ≈57.4 million people worldwide in 2019 to > 152.8 million cases in 2050 [4 –7]. The Lancet Commission on Dementia Prevention and Care concluded that “dementia is the greatest global challenge for health and social care in the 21st century” [8]. Yet, age-related cognitive changes and dementia remain vastly understudied in low-income countries (LICs) even though research suggests that declines in cognition often affect LIC individuals in early-old ages [5–7 , 10]. Figure 1 illustrates that this dearth of LIC cognition research is importantly due to the fact that very few aging studies, and hardly any with longitudinal cognition data, cover LICs. It is of utmost urgency to expand LIC aging and cognition studies to close the mismatch between the contexts where the old are most rapidly growing (e.g., LICs), and the contexts on which most aging research is focused (e.g., high-income countries (HICs) with few middle-income countries (MICs) being the exception) (Fig. 1).

Fig. 1

Population-based aging studies by Human Development Index (HDI) and annual growth of the population aged 50 + years: Most aging studies are in relatively high-income contexts where the older populations are not growing very rapidly in global comparison. Notes: Each point represents a major (and mostly publicly available) population-based global aging study. ISA, Ibadan Study of Aging (Nigeria); SAGE, WHO Study on global AGEing and adult health (Ghana, India, S. Africa); LASI, Longitudinal Ageing Study in India; HAALSI, Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa. The full list of included studies is reported in Supplementary Table 1. Annual growth rate for the 50 + population pertains to the period 2020–40, based on United Nations World Population Prospects [11]. Although the overall population age structure will remain relatively young, LICs such as Malawi have a very rapidly growing older population. The HDI is a summary measure of average achievement in key dimensions of human development, including life expectancy, education, and per capita income. HDI is for 2019, based on World Bank [12].

Each point represents a major (and mostly publicly available) population-based global aging study. ISA, Ibadan Study of Aging (Nigeria); SAGE, WHO Study on global AGEing and adult health (Ghana, India, S. Africa); LASI, Longitudinal Ageing Study in India; HAALSI, Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa. The full list of included studies is reported in Supplementary Table 1. Annual growth rate for the 50 + population pertains to the period 2020–40, based on United Nations World Population Prospects [11]. Although the overall population age structure will remain relatively young, LICs such as Malawi have a very rapidly growing older population. The HDI is a summary measure of average achievement in key dimensions of human development, including life expectancy, education, and per capita income. HDI is for 2019, based on World Bank [12].

A further indicator of this research gap in global cognition research is the fact that Volume I of the World Alzheimer Report 2020 included eight chapters on countries, all of which are HICs, and Volume II included 84 case studies, 68 of which are in HICs, and none are focused on LICs [13]. Similarly, the National Institutes on Aging (NIA)-funded Harmonized Cognitive Assessment Protocol (HCAP) [14] includes currently only one upper MIC (Mexico), one lower MIC (India), and zero LICs. Addressing this knowledge gap is urgent as low- and middle-income countries (LMICs) have 84% of the world’s population and higher current (about 58%) and projected cases of dementia than HICs [4]. An emerging literature, including our own research on cognition in Malawi [7, 15], studies of cognition among the remote Tsimane (Bolivia) [16], and related progress on expanding cognition assessments to MICs through HCAP and the Gateway on Global Aging (g2aging) [14, 17], has started to demonstrate the feasibility of cognition and dementia-related research outside of HIC settings. The next decade represents a unique window of opportunity to expand research on cognition in LICs to shape national and international responses and prepare for globalized dementia and Alzheimer’s disease and related disorders (ADRD) risks in an aging world [18]. LICs comprise the majority of individuals living in extreme poverty worldwide [19, 20], and cognition and dementia research in LICs is urgently required to close the current mismatch between the LIC contexts where dementia is most rapidly growing and the high-income contexts where most cognition research is conducted.

We document cognition patterns and longitudinal age-related changes using population-level data from Malawi, a sub-Saharan African (SSA) LIC currently ranked near the bottom (174 out of 189) of the human development index (HDI) [21]. Malawi shares many common characteristics and experiences with other global poor (i.e., very low-income) populations as defined by the criteria used by the World Bank to classify countries/regions (see the Methods section below for additional details) [12 , 23].

Investigating cognition in a context characterized by very high levels of experienced hardship throughout life is important because prior research from middle-income and high-income contexts has documented that cognition and age-related cognitive decline are strongly affected by cumulative life-course adversities [24 –31]. Using data from the Mature Adults Cohort of the Malawi Longitudinal Study of Families and Health (MLSFH-MAC) [15, 32] for the period 2012–2017, we therefore examine age- and sex-specific patterns of cognition and their longitudinal changes over time observed among mature adults 45 years and older. In addition to 1) estimating cognitive patterns by age and sex, the objective of our analyses is to 2) investigate how key socioeconomic characteristics predict levels of cognition and 3) how cognition in this LIC context changes as individuals get older. To our best knowledge, the MLSFH-MAC cohort, comprising 1,602 participants followed over 5 years from 2012–2017, provides the first longitudinal population-based data on cognition in a SSA LIC that allows the investigation of these research questions and generates much needed evidence. Our findings are of important translational relevance since they identify possible intervention points and emphasize the importance of prioritizing cognitive health among older individuals in SSA LICs and especially older women, to which relatively few health care resources are currently allocated.

MATERIALS AND METHODS

Context

Malawi represents an important case study to investigate longitudinal changes of cognition and their predictors among adults living in resource-constraint settings. The country shares common hallmarks with other SSA LICs such as subsistence-based economy, widespread poverty, and fragile health care system that is to a large extent unprepared to address the rapidly growing burden of non-communicable diseases and the needs of the fast-growing older population. The country is predominantly rural with currently only 16% of the population living in urban areas [33]. About 15% of the Malawian population is considered “ultra-poor”, i.e., with an estimated food consumption below the minimum level of dietary energy requirements [34]. In rural areas, where the study population of the MLSFH-MAC is based, the majority of individuals engage primarily in home production of crops, complemented by some market and small-scale business activities. The current life expectancy at age 45 is ≈28 years, while the estimated healthy life expectancy at the same age is around 22 years [11, 35]. These figures suggest that mature individuals in Malawi are likely to spend fewer years disease-free and in good health. Older adults aged 45 + years can expect to live a large proportion of their remaining life expectancy subject to physical and mental health limitations impacting their daily activities and overall wellbeing [36 –38]. HIV/AIDS remains widespread, with a prevalence of about 7.4% among rural adults 15–49 years old [39]. Similar to other SSA countries, the Malawian population is also experiencing an increasing double burden of infectious diseases and non-communicable diseases especially at older ages, with the latter accounting for 38% of total deaths, and 59% of deaths above age 45 years [35].

The MLSFH-MAC study population has life-course experiences that are typical for older persons in SSA and other LICs characterized by extreme to moderate poverty conditions. For example, they have lived most of their lives with per capita incomes of less than $1/day [40]. Our older study participants were born when under-5 mortality was almost one out of three [11], and all have survived sustained poverty, repeated famines, and epidemics (including HIV/AIDS) [41]. Fertility of these older individuals was high [42]. The status of women was generally low, and women often experience worse mental health at adult and older ages than men [37, 43]. The health care system continues to be generally underfunded, particularly with respect to prevention, treatment, and care for cognitive health, ADRD, and other non-communicable diseases [44, 45].

The Mature Adults Cohort of the Malawi Longitudinal Study of Families and Health (MLSFH-MAC)

The MLSFH-MAC is an ongoing population-based cohort study of mature adults aged 45 + years who live overwhelmingly in rural communities in three districts in Malawi (Mchinji in the central, Rumphi in the northern, and Balaka in the southern regions) [15, 32]. The cohort was established in 2012 by selecting respondents from the MLSFH [32], with hitherto follow-up waves in 2013, 2017-2018, and most current follow-up implemented in late 2022 and hence beyond the scope of this analysis. The original MLSFH population upon establishing the study in 1998 was based on a probabilistic population sample, with the study sample being augmented by enrolling adolescents, parents, and new spouses of respondents in the later rounds of data collections (for details, see Kohler et al. [32]). Comparisons of the 2010 MLSFH study population, from which MLSFH-MAC study participants were drawn, with the rural samples of the Malawi Demographic and Health surveys (DHS) and the Integrated Household Survey (IHS3) reveal that in key observable characteristics the study population closely matches the rural sub-sample in the 2010 nationally representative IHS3 [15, 32].

Inclusion/exclusion criteria: The two key inclusion criteria for enrolment in the MLSFH-MAC at baseline were being a MLSFH respondent aged 45 years or older in 2012 and having been interviewed in both the 2008 and 2010 MLSFH data collection rounds. Baseline enrollment included 1,266 individuals clustered in 130 + villages representing more than 90% of the 1,402 eligible MLSFH respondents who met the enrolment criteria. The primary reason for not enrolling in the cohort were migration out of the study areas and mortality. At each follow-up, MLSFH-MAC was augmented with additional MLSFH respondents who reached eligibility. To ensure an adequate representation of HIV-positive (HIV+) individuals in the cohort, age-eligible HIV+respondents were enrolled if they participated in either the 2008 or 2010 MLSFH data collection. Detailed information on sampling procedures, comparisons of the study population with nationally representative samples, study design and study instruments are provided in the MLSFH-MAC Cohort Profile [15].

Ethics approval: The data collections of the MLSFH-MAC and MLSFH have been approved by the IRB Board at the University of Pennsylvania (IRB Protocols #815016 and #826828). In Malawi, the MLSFH-MAC and MLSFH research has been approved by the Ethics Committee of the College of Medicine, Malawi (COMREC, Protocols #P01/12/1165 and #P.04/17/2160) and the National Health Sciences Research Committee (NHSRC, Protocol #19/01/2214).

MLSFH Cognitive Assessment

Research on cognition in SSA LICs continues to be limited, and as a result, at baseline in 2012 the research team was not able to implement a comprehensive cognition assessment that was previously used and/or validated in Malawi or a similar LIC context characterized by low literacy levels especially of older adults. MLSFH-MAC hence developed and extensively pretested a cognition instrument, the MLSFH Cognitive Assessment (MCA), that is designed to capture a wide range of cognitive domains and abilities, spanning from low to high cognitive performance. The MCA was also developed to be suitable for a partially illiterate study population with low schooling levels, and the MCA implementation protocol considered many factors that can influence test performance such as local culture, education, work activities, environmental context, access to health care, etc. [46, 47]. In addition, since MLSFH-MAC is a large population-based survey and study participants are interviewed in the locations where they live, it was required that the cognition assessment can be locally implemented with well-trained non-clinician interviewers.

Cognitive domains: The MCA is modeled after several well-known screening batteries such as the Mini-Mental State Exam and the Montreal Cognitive Assessment [48 –51]. The MCA measures a range of key cognitive domains affected by cognitive aging and assesses multiple cognitive domains that align with a classification of cognitive ability domains based on the involved general processes, such as language, attention, or executive functioning [52] (Table 1). The cognitive domains in MCA are also to a large extent overlapping with the domains assessed through the Harmonized Cognition Assessment Protocol (HCAP) that is part of international collaboration funded by the NIA and that aims to measure and identify cognitive impairment and dementia in population-based samples of older individuals living in different social and economic contexts (Table 1) [14, 17]. Moreover, the MCA also provides key information on a comprehensive range of cognitive functions relevant for day-to-day activities in this population and captures a widespan of cognitive dimensions ranging from a competent cognitive functioning to cognitive impairment (Table 1A). The MCA is summarized in detail in the Supplementary Material, including examples of how the MCA was adjusted to our low-literacy study context.

Table 1

MLSFH Cognition Assessment (MCA)

A: Cognitive Domains assessed in HCAP [14]	MCA	B: Mean scores for selected components in MCA (at baseline in 2012)
	2012–17		Women	Men
Orientation Place &Time	X	Recall name of president of Malawi	0.86	0.96
Immed. &Delayed Episodic Memory	X	Repeat 3-number seq. in reverse	0.29	0.45
Recognition Memory	X	Delayed recall of sentence	0.43	0.48
Visuospatial	X	Immediate word recall: Prop.	0.80	0.80
Executive Functioning	X	recalled
Attention/Speed	X	Delayed word recall: Prop.	0.55	0.54
Language/Fluency	X	recalled
Informant Reports	–

Panel A lists the cognitive domains included in the NIA-funded Harmonized Cognitive Assessment Protocol (HCAP) and the domains included in the MLSFH Cognition Assessment (MCA). X = assessed in MCA [7, 15]. Panel B shows mean scores for selected components included in the MCA. Number sequence is (4-6-1). Word recall is five words.

Implementation: The MCA was extensively pretested before being administrated to the MLSFH-MAC study population [7, 15]. All cognition questions were administered in the local languages and the translations of the questionnaire into Chichewa, Chiyao, and Chitumbuka (the three local languages spoken in the MLSFH study areas) were pretested during focus-group interviews and pilot tests. Reverse translations ensured the accuracy of the final instruments. Respondents were also interviewed by interviewers from the same region who were fluent in the respondent’s language. Interviewers and their supervisors were extensively trained in the administration of the cognition instrument, and all parts of the questionnaire were piloted extensively before the data collection. Before administering the cognitive assessment questions, interviewers screened study participants for visual and hearing impairments that might interfere with their ability to see stimuli or hear questions and hence impact their total MCA scores.

Measurement: The maximum MCA score is 30 based on the sum of the individual scores for the specific domains that are assessed with the MCA instrument, corresponding to highest (best) cognition level. We measure baseline cognition in 2012 and use the 2013 measurements in the very few instances in which the 2012 values were missing. Table 1B illustrates mean scores of selected easily-interpretable MCA components administered in 2012. For instance, while most of the respondents could recall the name of the Malawian president (86% of women and 96% of men), only 29% of women and 45% of men were able to repeat a 3-number sequence in reverse. Less than half of the study sample was able to correctly delayed-recall a sentence. The learning-effect patterns detected in our cognition assessment are similar to patterns observed in other studies such as for instance the Health and Retirement Study (HRS) in the US, that is considered as being the “gold standard” for studying aging populations outside of clinical settings, including cognitive health among older people [53]. For instance, the verbal fluency (VF) question that requires respondents to name within one minute as many animals as possible had a mean of 10.5 words in 2012, and 11.5 words a year later in 2013 (the VF total score is the total number of correct words). The general pattern and learning effect observed here is consistent with the patterns detected in HRS that find increase in average of 0.5 nouns two years later.

Analytical approach

The focus of our analysis is on cognitive changes over time observed among mature adult respondents aged 45 years and older during a span of 5 years, from 2012 to 2017. In these longitudinal analyses, we need to account for the fact that the delayed word recall component of the MCA in 2012 and 2013 was based on five words, and it was augmented to 10 words in 2017 to increase comparability with other longitudinal studies in aging populations administering this component (i.e., HRS, the Survey of Health, Aging and Retirement in Europe (SHARE), the Longitudinal Aging Study in India (LASI), and others). We used equipercentile equating to account for this modification and convert adequately one score to the other. This approach is the most common technique for direct crossover between tests based on different scores [54, 55]. More specifically, we used the cumulative probability distribution (CDF) of the delayed word recall scores for respondents 45–54 years old to create a conversion scale that allows to convert the 10-word-based scores to the ones based on five words. The advantage of this approach is that it accounts for the relative difficulty of recalling 10 versus 5 words, and it does so by aligning the CDF of the 5-word and 10-word recall score in a reference population (i.e., respondents aged 45–54 years old in 2012). Respondents aged 45–54 years were chosen as reference population as these mid-adult individuals have better cognitive outcomes than older individuals and their cognition scores are less likely to change between 2012 and 2017. Specifically, we compute the CDF of the proportion of words individuals aged 45–54 recalled in both the 2012 and the 2017 waves.

To account for the difference in sample characteristics between the two waves, we use frequency weights based on sex, schooling, and age group (45–49 and 50–54) to make the analytical samples in both years similar. This weighting ensures that the difference in the CDF of the proportions of words recalled by the MLSFH respondents in 2012 and 2017 is driven by the complexity of the task and not by differences in sample characteristics. We then apply a conversion scale based on the wave-specific CDFs to all MLSFH-MAC respondents who were interviewed in 2017, obtaining a delayed word recall score (“5-word equivalent”) in 2017 that is comparable to the one from the previous two waves. For example, based on the CDF in the reference population, delayed-recall of 8 out of 10 words in 2017 is comparable to recalling 4.6 out of 5 words in 2012, and delayed-recall of 3 out of 10 words in 2017 compares to delayed-recall of 1.5 words in 2012 (see Supplementary Figure 1). While, of course, none of the respondents recalled 4.6 words out of 5 in 2012 in the first example, this “crosswalk” based on the CDF places the difficulty of delayed-recalling 8 out of 10 words in 2017 slightly more than halfway between recalling 4 or 5 words out of 5 in 2012. As a result, this approach allows us to compute a “5-word equivalent” for each possible response on the 10-word recall component (immediate or delayed) asked in 2017 (and vice versa).

To investigate the changes in cognition by age, we use boxplots to illustrate changes in the distribution of the cognition scores as respondents get older, with analyses being conducted separately by 10-year age groups (with age measured at baseline in 2012). In addition, we estimate a standard Hierarchical Age-Period-Cohort (HAPC) model [56]. This multilevel regression model, in which cohort effects are explicitly included and period effects implicitly incorporated in the form of age-by-cohort interactions, allows to assess the age patterns of total MCA scores by cohort. Additional details on this HAPC model and the specification we used are provided in the Supplementary Material.

We estimate ordinary least square (OLS) regressions to analyze the determinants of the levels and absolute and relative changes in cognition measured by the MCA scores. All OLS regression models include age (from which the mean has been subtracted for ease of interpretation) and region fixed-effects to account for constant regional differences in which respondents live. In regressions where we pool observations from the different years together, our specifications include study-wave fixed-effects. All reported standard errors are robust to heteroscedasticity.

In our final analyses where we focus on persistent disparities in cognition and their determinants, we partition our respondents into groups corresponding to very low, “normal”, and very high cognitive health. For this purpose, we conducted k-means clustering analysis using total MCA scores obtained in both 2012 and 2017 and formed eight different clusters. The k-means clustering analysis is a widely used data-driven technique to automatically partition a dataset into groups by minimizing the within-cluster distances between observations [57, 58]. More specifically, it is an iterative method that assigns observations to the group with the closest (in terms of Euclidean distance) centroid defined by the average of all points assigned to it. This approach has been widely used in many disciplines, including psychology, medicine, and biology [59 –61]. In our analysis, the lowest cluster corresponds to individuals who consistently performed relatively poorly in the cognitive tests across the two waves (mean MCA scores were 10.3 and 8.7 at baseline and follow-up, respectively), whereas the highest cluster captures the group of individuals who performed relatively very well (mean MCA scores were 26.1 and 25.4 at baseline and follow-up, respectively). The six in-between clusters group individuals that we characterize as having “normal” cognition (i.e., mean cognition scores were 19.6 and 18.3 at baseline and follow-up, respectively). Our results are robust to using different numbers of clusters as shown in our robustness analysis in the Supplementary Table 9 as well as defining groups based on having MCA scores one standard deviation above or below the sample means as detailed in Supplementary Figure 2 and Supplementary Table 11.

RESULTS

Analytical sample characteristics

Summary statistics of our study population at baseline are presented in Table 2. Our analytical sample includes 1,062 respondents for whom we measure cognition at baseline (2012 or 2013) and at follow-up in 2017. About 60% of our respondents are women with an average age of 58 years. The average age of men is 59 years in 2012. Over 70% of respondents were 45 to 64 years old at baseline, and only 9% of the total sample was 75 + years old. Formal schooling is low in this study population, with more than one third of respondents having never attended school. There is a clear difference in educational achievement by gender: 46% of women versus 19% of men never attended school, and 12% of men but only 3% of women completed secondary or higher education. The majority of respondents were currently married (65% of women and 97% of men). This gender difference in marital status is due to a higher proportion of women being widowed. About one third of study participants (32%) lived in a house with a metal roof, a reliable indicator of wealth in this context [62].

Table 2

Summary statistics for MLSFH-MAC analytical sample (means)

	Women	Men	Total
# of observations	627	435	1,062
Respondent age (2012)	58.21	59.18	58.61
Respondent age group (2012)
45-54	0.46	0.37	0.42
55-64	0.30	0.36	0.32
65-74	0.14	0.19	0.16
75+	0.10	0.08	0.09
Schooling attainment (none, primary, secondary+)
No formal schooling	0.46	0.19	0.35
Primary schooling	0.52	0.69	0.59
Secondary or higher	0.03	0.12	0.06
Currently married (2012)	0.65	0.97	0.78
Wealth indicator: metal/tiled roof (2012)	0.31	0.33	0.32
HIV-positive status (2012)	0.04	0.05	0.05

Age patterns of cognition

In Fig. 2, we depict the change in the total MCA scores from 2012 to 2017 by 10-year age groups. The adjacent box plots on the left (Fig. 2A) show the 2012 and 2017 distribution of the total MCA scores, by baseline age in 2012, indicating substantial variation in the overall patterns of cognition in this study population. This variation persists at older ages 65 + years. The comparison of the adjacent box plots reveals a decline in cognition during the 5-year period: in each baseline age group, the right box plot has shifted to lower values (including median, 25th and 75th percentiles), indicating that respondents on average had lower total MCA scores at follow-up in 2017 as compared to the baseline in 2012. This comparison of adjacent box plots also documents that this cognitive decline during 2012– 17 is particularly pronounced at older ages and is most notable among the 75 + years old study participants.

Fig. 2

Change in Cognitive Scores by Age. Panel A on the left shows the changes in the total MCA scores between 2012 and 2017 by age groups at baseline. Shaded areas represent interquartile ranges and the horizontal lines within the boxes represent the median value. Panel B on the right represents the predicted age patterns of total MCA scores by cohort resulting from the HAPC model. The model specifies the age trends linearly and is fully interacted with sex to allow the age patterns to be sex-specific. Note that the age variable is centered around the median age of the 7-year cohort to which the respondent belonged (see the Supplementary Material for details on the model specification).

These descriptive age and cohort patterns are confirmed by the HAPC model in Fig. 2B. This HAPC model shows the predicted age patterns of the total MCA scores during 2012– 17 resulting from a multilevel regression in which cohort effects are explicitly included and period effects implicitly incorporated in the form of age-by-cohort interactions. Figure 2B highlights three characteristics of cognitive change in our study population. First, it confirms that as individuals age, they experience accelerating cognitive decline: the slope in the predicted cognition patterns is steepest (that is more negative) for older individuals, whereas the youngest cohorts experience a more modest cognitive decline.

Second, the HAPC highlights the marked gender difference in cognitive function. In this SSA LIC context characterized by extreme poverty, women have substantially lower total MCA scores, indicating lower cognition levels as compared to men. This finding is in stark contrast to the patterns observed in high-income contexts where women tend to have better cognitive function at older ages than men [63 –65]. This gender reversal could be due to the biological demands of high fertility women have in this SSA LIC context, or unequal gender roles and their implications for resource allocation. Older women in SSA, including Malawi, are generally characterized by low social position, which is also reflected in their limited access to schooling. For instance, in our MLSFH-MAC sample, only 3% of women above age 45 years have completed secondary and higher education, in contrast to 12% of older men (Table 2). This unequal resource allocation is at least partially driven by gender norms, attitudes, and stereotypes, which perpetuate the male-female gap in health outcomes and socioeconomic position [66 –68].

Third, while our model specification allows for the cognitive trajectories to differ between genders, we find similar patterns of decline for both men and women: women have lower MCA scores on average than men, but the average decline in MCA scores is similar for both genders. This finding suggests that the lower level of cognition among women does not buffer them against a declining cognitive function with age.

To further explore the gender differences identified above, we examined the results of OLS regressions of cognition (in levels) and cognition change (absolute and relative) on gender, age (centered around its mean), and their interactions (Table 3). The outcome variable of interest in the first three columns is total MCA scores, pooling data across the three MLSFH-MAC waves when the full MCA cognition assessment was collected. The outcome variable in the next three columns is the absolute change in total MCA scores between 2012 and 2017. The outcome variable in the last three columns is the relative change in total MCA scores between 2012 and 2017, as measured by the change in total MCA scores between 2012 and 2017, divided by the baseline value.

Table 3

Age and gender differences in levels of MCA and 2012– 17 changes in MCA scores

	Total MCA scores			Change in total MCA scores			Relative change in MCA scores
	All	Male	Female	All	Male	Female	All	Male	Female
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
Age	–0.162***	–0.163***	–0.238***	–0.049**	–0.051**	–0.044***	–0.003**	–0.003**	–0.003**
	(0.010)	(0.010)	(0.009)	(0.023)	(0.023)	(0.017)	(0.001)	(0.001)	(0.001)
R is female	–2.902***			0.251			0.011
	(0.143)			(0.286)			(0.017)
R is female×Age	–0.079***				0.003			–0.000
	(0.013)			(0.028)			(0.002)
Observations	3,915	1,638	2,277	1,062	435	627	1,062	435	627
_R2	0.338	0.187	0.357	0.042	0.042	0.050	0.033	0.045	0.034

Robust standard errors are reported in parentheses with *p <0.1, **p <0.05, ***p <0.01. All regressions include region fixed effects to capture any systematic differences across regions. Regressions in which we pool observations from the three waves together (columns 1-3) include dichotomous variables that take the value one if the corresponding observation was measured in 2012, 2013, or 2017, and zero otherwise. Age is centered at its mean for ease of interpretation.

Consistent with Fig. 2, the coefficient for age in these analyses shows that older individuals have lower MCA scores as well as a more rapid decline in the total MCA scores during 2012–2017. Each additional year of age is associated with 0.16–0.24 lower MCA scores on average (columns 1–3), and on average a 0.044–0.051 more rapid decline in MCA scores during 2012–2017 (column 4–6). The analyses also show that women on average have total MCA scores that are about 2.9 points lower than the men’s scores (column 1), consistent with the reversed gender pattern identified in Fig. 2B.

While the cross-sectional age-gradient in MCA scores is higher for women as compared to men (column 1), the longitudinal analyses of cognitive change during 2012–2017 in columns 4 and 7 do not show a differential decline in cognition as measured by the MCA scores for women and men. This is confirmed in columns 5 and 6 and columns 8 and 9 where we estimate a longitudinal age gradient in MCA scores that is very similar for men and women (e.g., β =–0.051 versus –0.044 columns 5 and 6). Hence, our longitudinal follow-up data do not show that cognitive function declines deferentially as women and men get older, despite the gender differences in the cognition levels and the presence of a cross-sectional gender difference in the decline of cognition with age. It is important to point out that the longitudinal follow-up in this analysis is only for 4–5 years (2012–2017), and differential changes in cognition by gender over time might emerge over longer follow-up periods. Recognizing this limitation, the regressions in Table 3 nevertheless confirm our earlier interpretation of Fig. 2B that highlighted the lower cognition scores for women as compared to men, an accelerating decline of cognitive function for both men and women as they get older, and a lack of differential cognitive decline for men and women in the longitudinally follow-up data from 2012–2017.

Gradients in cognition and cognitive decline by socioeconomic characteristics

Table 4 expands our analyses to additionally explore gradients in cognition and cognitive decline by socioeconomic characteristics, focusing on dimensions that have received considerable attention in other contexts: socioeconomic status (SES), schooling, and marriage [28, 69]. The regression specification in Table 4 follows that of Table 3 (columns 1, 4, and 7), and includes the additional covariates to estimate the gradients by SES, marital status, and schooling. We use “having a house with a metal roof”, which is a simple but reliable indicator of wealth in this SSA LIC context [62], as an indicator of SES. We show in the Supplementary Material that this measure of wealth is highly correlated to other measures of SES that we can derive from the survey, such as the wealth index based on household’s items ownership, the interviewer’s assessment of the respondent’s household’s wealth, and the number of days respondents has eaten chicken, fish, or meat in the last week (Supplementary Table 2). Our results are robust to using average protein consumption measured across several waves (2010, 2012, and 2013) to better capture long-term nutritional status of the MLSFH-MAC respondents (Supplementary Tables 2 and 4) and captures equally well the relationship between SES and level/change in cognition (Supplementary Table 3). As another measure of socioeconomic status, we focus on current marital status at each wave in the pooled analysis and at baseline in the longitudinal analysis. Given the overall low level of education in this population, schooling is measured by whether an individual has attended any primary or secondary school.

Table 4

Associations between cognitive function and selected socioeconomic characteristics

	Total MCA scores			Change in total MCA scores			Relative change in MCA scores
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
Metal roof	1.232***			–0.228			–0.005
	(0.147)			(0.305)			(0.018)
Currently married		0.062			0.991**			0.075***
		0.197			0.386			0.026
Any schooling			3.107***			0.243			–0.000
			0.181			0.359			(0.024)
Observations	3,909	3,915	3,914	1,062	1,062	1,062	1,062	1,062	1,062
_R2	0.350	0.338	0.387	0.042	0.047	0.042	0.033	0.041	0.033

Robust standard errors are reported in parentheses with *p <0.1, **p <0.05, ***p <0.01. All regressions include sex, age, age×sex, and region fixed effects to capture any systematic differences across regions. Regressions in which we pool observations from the three waves together (columns 1-3) include dichotomous variables that take the value 1 if the corresponding observation was measured in 2012, 2013, or 2017, and zero otherwise. The independent variable in the first row is a dichotomous variable that takes the value 1 if the respondent lives in a house with a metal roof. The independent variable in the second row is a dichotomous variable that takes the value 1 if the respondent is currently married. The independent variable in the third row is a dichotomous variable that takes the value 1 if the respondent has any formal schooling.

Columns 1, 2, and 3 in Table 4 show that living in a house with a metal roof and having any schooling is associated on average with a 1.2 and 3.1 points higher total MCA scores, respectively, all with p-values lower than 0.001. These gradients in MCA scores are similar to corresponding gradients in cognition measures that have been documented in other contexts [28, 69]. Being currently married, however, does not predict total MCA score (column 2). Interestingly, however, SES (wealth) and schooling do not appear to be predictors of changes in the total MCA scores between 2012 and 2017 (columns 4–9), despite their strong associations with levels of MCA scores. Marital status is associated with absolute and relative changes in the MCA scores during 2012–2017, and being married at baseline is associated with a slower cognitive decline over time. Acknowledging again the limitations of a longitudinal follow-up of only five years, the analyses in Table 4 shows strong socioeconomic gradients in the levels of cognition in this study population that mirror analogous findings in other contexts, while only marital status predicts changes in cognition over time. SES (wealth) or schooling do not predict changes in cognition during 2012–2017.

Supplementary Tables 5 and 6 show that the associations between SES (wealth) and cognition does not appear to be different for men and women. This is also the case for schooling and applies to both levels of and changes in MCA scores. The association of being currently married with the level of cognition appears to be positive only for men (Supplementary Table 6), consistent with the finding that in terms of health men often benefit more than women [70]. In separate analyses by gender (Supplementary Table 6), current marriage, however, does predict a slower cognitive decline for women, but not for men, albeit due to the reduced sample size in these gender-specific analyses, we cannot reject that the coefficient on currently married is equal for both men and women (Supplementary Table 5).

HIV+ status is a further potentially important factor shaping cognitive patterns and cognitive decline among older individuals in high-prevalence SSA countries such as Malawi [71 –73]. We explore this association in Supplementary Table 7 recognizing the limitation of a relatively low HIV prevalence in this mature adult study population: in 2012, our study sample included only 50 individuals who were HIV+ (HIV prevalence = 4.8%), 30 in the age group 45–54, 18 in the age group 55–64 and only 2 who were 65 + years old. This low prevalence is the result of high adult mortality prior the widespread availability of antiretroviral treatment (ART) that only became available in the MLSFH study areas in 2008 [74]. Subsequently, our statistical power is limited, particularly for the longitudinal analyses. Our analyses in Supplementary Table 7 do not show statistically significant (p<0.05) differences in the level or change-over-time in cognitive scores for HIV+ individuals. Moreover, the associations between gender, age, and our three other socioeconomic characteristics (schooling, wealth, and marital status), and cognitive function in level appear similar between those who are HIV+ and those who are not (Supplementary Table 7, columns 2 and 3). These findings need to be interpreted with caution in light of the fact that HIV+ individuals are few in our study population and concentrated at younger ages (45–55 years) where cognitive decline is not yet very pronounced (Fig. 2). Our estimates are hence imprecise with large confidence intervals.

Persistent disparities in cognition and their determinants

Our final analyses in Fig. 3 and Table 5 take advantage of the longitudinal MCA data to address an important concern in analyses of cognition: measurement error. Any single performance on a cognition assessment test such as the MCA could be affected by mood, emotions, environmental distractions, daily circumstances of a respondent, or other unobserved factors. While the study team established a standardized protocol for the implementation of the MCA with extensive training of interviewers and their supervisors (see above), idiosyncratic influences on performance in the cognitive assessment certainly existed. Their importance has been documented in other contexts [75].

Fig. 3

Distributions of total MCA scores at baseline in 2012/2013 (x-axis) and followup in 2017 (y-axis). The partitioning of the respondents in 8 groups is done using k-means clustering [76]. “x” symbols represent individuals in the lowest cluster and “+” symbols represent individuals in the highest cluster.

Table 5

Predictors of the probability of belonging to the lowest or highest cluster

	Highest cluster			Lowest cluster
	(1)	(2)	(3)	(4)	(5)	(6)
Metal roof	0.074***			–0.027*
	(0.025)			(0.014)
Currently married		–0.026			–0.025
		(0.028)			(0.026)
Any schooling			0.157***			–0.075***
			(0.021)			(0.022)
Age	–0.010***	–0.010***	–0.010***	0.003**	0.003**	0.003**
	(0.002)	(0.002)	(0.002)	(0.001)	(0.001)	(0.001)
R is female	–0.133***	–0.143***	–0.096***	0.081***	0.074***	0.064***
	(0.023)	(0.025)	(0.024)	(0.014)	(0.016)	(0.014)
R is female×Age	0.004**	0.004*	0.004^**	0.008***	0.008***	0.008***
	(0.002)	(0.002)	(0.002)	(0.002)	(0.002)	(0.002)
Observations	1,062	1,062	1,062	1,062	1,062	1,062
R ²	0.126	0.119	0.144	0.171	0.170	0.181

Robust standard errors are reported in parentheses with *p <0.1, **p <0.05, ***p <0.01. All regressions include region fixed effects to capture any systematic differences across regions. The outcome variable in the first three columns is a dichotomous variable that takes the value 1 if the respondent belongs to the highest cluster (green dots in Fig. 3) and 0 otherwise. The outcome variable in the last three columns is a dichotomous variable that takes the value 1 if the respondent belongs to the lowest cluster (red dots in Fig. 3) and 0 otherwise. The independent variable in the first row is a dichotomous variable that takes the value 1 if the respondent lives in a house with a metal roof at baseline, and zero otherwise. The independent variable in the second row is a dichotomous variable that takes the value 1 if the respondent is married at baseline, and zero otherwise. The independent variable in the third row is a dichotomous variable that takes the value 1 if the respondent has any formal schooling, and zero otherwise. Age is centered at its mean for ease of interpretation.

Using our longitudinal data, we can alleviate concerns about the resulting measurement error in cognition assessments by identifying respondents who had consistently good or consistently weak performance on the MCA cognition assessment. The remainder of the study population is “in between” with MCA cognition scores that place them neither in the top or bottom part of the distribution of the repeated cognitive measurements (the method used to classify respondents was previously described in detail). Figure 3 shows the results of partitioning our sample in eight different groups, where we represent respondents who were consistently measured with high total MCA scores in 2012 and 2017 with “+” symbols (the highest cluster) and those consistently measured with low score with “×” symbols (the lowest cluster). Respondents represented by “o” are individuals who scored in the middle range, that is, they did not score consistently high or low across the two MCA measurements.

The clusters identified in Fig. 3 provide an important outcome— consistently good or consistently weak performance on the MCA— to further investigate the correlates of cognitive function in our study population.

The three first columns in Table 5 show that individuals who live in a house with a metal roof (i.e., indicator of higher SES) and attended formal schooling have a higher probability of belonging to the highest cluster (“consistently good performance on the cognitive assessment”). In contrast, being married at baseline in 2012 does not predict this probability. Women and older respondents are less likely to belong to the highest cluster as well. Mirroring these results, the last three columns of the table show the corresponding predictors of belonging to the lowest cluster (“consistently weak performance on the cognitive assessment”). Having attended formal schooling reduces the likelihood of being in the lowest cluster, while marital status is not associated with this probability. Higher SES only weakly reduces the probability of being in the lowest cluster (p <0.10). Supplementary Table 8 shows that the socioeconomic predictors of cluster membership— SES (wealth), marital status and schooling— do not differ between men and women. Supplementary Table 9 documents that our results are robust to using seven or nine clusters, instead of eight, in determining the consistently good and weak performers on the MCA cognition assessment (Fig. 3).

DISCUSSION

Low-income countries (LICs) have the most rapidly growing number of older individuals [11], and the largest increase in dementia cases in the next decades will occur outside of HICs [4 –7]. Investments are required to close the mismatch between the contexts where the old are most rapidly growing (e.g., LICs), and the contexts on which most aging research is focused (e.g., high-income countries and few middle-income countries).

The urgency to tackle this knowledge gap arises from the fact that LMICs are inadequately equipped to handle the rapidly rising burden of dementia. These countries face multiple hurdles such as limited health literacy regarding non-communicable diseases, strained healthcare systems, and a lack of formal social support systems that leave older individuals and their families, who are already grappling with poverty and other difficulties of life in LICs, bear a substantial burden of caregiving [4 –6]. Yet, the extensive research on cognition and ADRD in high-income countries is potentially of limited value in informing the necessary social policy and health system changes in LICs as findings from HICs are unlikely to generalize to the vastly different institutional, epidemiological, social, and economic contexts of LICs [77, 78]. Ultimately, to understand patterns of aging across the spectrum of economic development, it is crucial that research efforts encompass studies in low-income contexts. The NIA has recognized this and has issued an approved concept for Building Neuroscience Research Infrastructure for AD/ADRD in Africa stating that: “[T]he region’s wide variety of dietary, lifestyle and environmental exposures, as well as genetic variation, can provide valuable insights on factors that contribute to AD/ADRD. Studies of AD/ADRD in Africa will present significant research opportunities that will increase knowledge about the etiology of this family of diseases and accelerate the pace of scientific discovery. These studies can also inform intervention or prevention strategies through reciprocal innovative approaches (the bi-directional and iterative exchange of a technology, methodology, or process between two countries) for AD/ADRD mitigation within US populations.”

In light of the limited number of aging studies that currently allow analyses of cognition and dementia/ADRD in LICs or African settings (Fig. 1), this paper is important because it illustrates some of the novel findings related to cognitive health that can emerge from such studies. The Malawi Longitudinal Study of Families and Health (MLSFH) is one of the few aging studies that allow longitudinal analyses of cognition in a SSA LIC, and it represents the lowest income contexts for which such longitudinal data are available. While our analyses focus on a single country, and only on patterns of cognition, the analyses are important because the MLSFH study population reflects life-course experiences that are typical for older persons in SSA and other LICs: exposures to moderate and extreme poverty conditions, high disease burdens, limited access to health care (especially for non-communicable diseases and often non-existent for dementia), and common and repeated adversities throughout the life-course. These socioeconomic contexts, which are characteristic for a large and rapidly growing number of older persons in LICs (Fig. 1), are vastly different from the life-course contexts experienced by the populations used for the overwhelming majority of studies on cognitive health among older individuals. Analyses of aging LIC populations such as provided by the MLSFH are thus required to provide new insights into the relationships among life-course adversities, cognitive aging, and dementia risks in low-income populations, and consistent with the above-mentioned NIA approved concept, help inform the comprehensive health systems and policy changes that LICs need to adapt in order to prepare for a future with increasingly growing older population and a high prevalence of age-related diseases such as dementia and ADRD [5 , 9]. Key findings that emerge from our analyses include:

First, our study demonstrates that the collection of longitudinal detailed cognition assessments is feasible in SSA LIC contexts, despite the substantial challenges of measuring cognition in populations where older adults often live in remote areas without access to clinical health care, have low levels of schooling, are often illiterate, and frequently affected by poor mental and physical health [15]. The MCA that assesses multiple cognitive domains often affected by cognitive aging and is also adapted to a low-literacy aging population, documents substantial variation in the MCA scores— and thus cognitive function— among individuals 45 + years old, with variation persisting until the oldest ages. These findings suggest that the MCA is a suitable tool for characterizing and tracking longitudinal changes in cognition among older individuals living in an LIC.

Second, the most novel aspects of findings pertain to the analysis of longitudinal changes in cognitive function during 2012–2017. Our study is the first to describe the longitudinal changes and their predictors in a SSA LIC context. We document that cognition declines with age not just cross-sectionally, but also longitudinally as individuals age. The onset of this decline is detectable at mid-adult ages (ages 45–55 years) and accelerates as individuals get older. Furthermore, despite the fact that women have lower cognitive function than men, the absolute and relative pace of decline does not differ by gender. Hence, the lower cognitive function of women does not protect them in terms of cognitive decline: women face the same pace of decline as men, despite having on average significantly lower MCA scores.

While levels of cognition are strongly predicted by gender and socioeconomic characteristics, the longitudinal change in MCA scores during 2012–2017 has fewer predictors. As already mentioned above, there does not seem to be a gender difference in cognitive decline. In addition, neither schooling nor SES (wealth) predict the longitudinal change in cognitive function as individuals age during 2012–2017, which is contrary to our expectations that higher SES individuals will experience flatter age-related decline in cognition over time. There is some evidence that being married is associated with slower cognitive decline for women, while this is not the case for men.

Third, in the SSA LIC context we study, women are found to have lower levels of cognitive function. Although this finding is similar to patterns documented in other SSA LICs [80], it is in stark contrast to the patterns observed in high-income contexts where women tend to have better cognitive function at older ages than men [63 –65]. This gender reversal could be due to the biological demands of high fertility, unequal gender roles and their implications for resource allocation, or other systemic factors that disadvantage women in this low-income context. More in-depth analyses of gender differences in cognition in SSA LIC contexts such as Malawi are required to better understand how this gender reversal is related to differential life-course experiences of men and women, and how contextual aspects— such as, for instance, the relatively low status of women, reproductive histories, or differential access to schooling and health care— contributes to this pattern.

Fourth, in addition to gender, variation in cognitive function is related to SES, schooling, and marriage. These factors are also associated with cognition or ADRD risk in many other contexts [28, 69], with higher SES (wealth), having attended school, and being currently married associated with higher cognition levels. We find similar associations in Malawi, despite the vastly different context of the MLSFH study population from that of HICs. It is worth noting that the association between being currently married and MCA scores is positive and statistically significant for men whereas this association was not found for women. This latter finding suggests that the socioeconomic environment such as marriage may mostly benefit men in this context.

Fifth, in contrast to studies that have documented a relationship between HIV-positive status and cognitive outcomes in other contexts [81 –83], our results do not show an association between HIV status and cognition in this study population. Our findings, however, need to be interpreted with caution because the prevalence of HIV in this older sample is low due to the experience of high HIV-related mortality prior the introduction of antiretroviral treatments, and the HIV-positive cases in MLSFH-MAC are mostly observed among younger study participants age 45-49 years. Malawi is a high HIV prevalence country and with higher numbers of HIV-positive individuals reaching older ages in the next decades, investigating the relationship between cognition and HIV will be of utmost importance for future research.

Sixth, the availability of longitudinal cognition measurements allowed us to investigate persistent disparities in cognitive outcomes and their associations with socioeconomic characteristics. For this purpose, we use a clustering approach that focuses on individuals who had consistently good or consistently weak performances on the cognition assessments over the period 2012–2017. The advantage of this approach is that it reduces the impacts of idiosyncratic (random) influences that can affect performance during a cognition assessment on a particular day. Focusing on individuals who consistently perform well or poorly alleviates this concern. Our analyses document that schooling and SES (wealth) increase the probability of being in the top cluster, that is having a consistently good performance on the MCA cognitive assessment, and these factors reduce the probability of being in the bottom cluster, that is having consistently weak performance on the cognitive assessments. Moreover, despite the substantial level differences in cognitive function between men and women, these predictors do not differ between men and women.

In interpreting the above results, it is important to acknowledge several limitations. First and foremost, although the MCA instrument captures a wide span of cognitive dimensions, the instrument had to be adapted to the SSA context. The development of cognitive assessments that are suitable for such LIC contexts remain an area of active research [84]. Moreover, to achieve comparison with other aging studies, the word recall in 2017 was augmented to 10 words while at baseline in 2012 and 2013 it was based on five words. Our crosswalk method allows to convert the 2017 to the 2012 word recall scores, while also accounting for the difficulty of delayed-recalling higher number of words. Yet, there might be remaining measurement errors in the longitudinal analyses due to the change in this word recall question.

We also acknowledge that our current analyses focus on measures of cognitive function, rather than a comprehensive assessment of dementia or other neuropsychiatric symptoms frequently associated with aging [85]. Albeit of high public health relevance, a research diagnosis of dementia and an assessment of neuropsychiatric symptoms are currently not feasible in the MLSFH-MAC, or any other SSA LIC aging study. These extensions will be implemented for the first time in a SSA LIC as part of forthcoming MLSFH data collections (funded via NIA R01 AG079527), along with information on basic and instrumental activities of daily living, proxy interviews, and harmonization of the MLSFH cognitive assessment (MCA) with other global aging studies. Future MLSFH research will be able to adopt a more inclusive approach to the analyses of cognitive aging and dementia.

Also, as in most longitudinal studies, selection and attrition between baseline and follow-up potentially affect our estimates. Supplementary Table 12 shows that respondents with high MCA scores at baseline were more likely to participate in the follow-up survey and less likely to have died. To the extent that cognitive function predict participation in our surveys, it is possible that our analysis actually overestimates MCA levels and underestimates cognitive decline. However, follow-up rate between baseline and follow-up was very high and selection bias is likely to be small. Our findings also pertain to an older population living in rural settings and do not capture variation in cognition among older adults living in urban or periurban areas. Finally, our analysis focuses on estimating longitudinal changes in cognition over a 5-year period, utilizing measurements collected in 2012/2013 and 2017. The results reveal a notable cognitive decline that becomes more pronounced with age during this timeframe. These findings underscore the significance of future research encompassing extended periods of observation and gathering additional data points to more accurately capture the trajectories of cognitive change in this low-income SSA context. Forthcoming MLSFH cognition data from MLSFH will provide an opportunity to extend the observation period for this study population, enabling in the future further analyses and a more comprehensive understanding of the trajectories of cognitive change in low-income SSA contexts.

Our analysis provides the lowest income context for which population-based longitudinal studies of cognition have been conducted. Studying cognition in these contexts is important because life-course experience of the aging global poor is characterized by early-life and adult-life adversities potentially reinforcing each other as many individuals experience both. As a result, the cognitive patterns of aging individuals and their longitudinal changes in SSA LICs reflect not only normal age-related changes, but also detrimental life-course influences and shocks experienced frequently and along multiple dimensions of the living environment. Cognitive patterns at older ages in our study context are also likely affected by the specific disease environment in SSA. For instance, evidence suggest that exposure to malaria has both short term and longer term effects on cognitive development among children and adults [86, 87]. HIV/AIDS, potentially affects the central and peripheral nervous systems and neurological complications are common among HIV-positive individuals [88 –91]. Moreover, changes in cognition are also related to non-communicable disease hazards such as cardiovascular disease risk factors, heart disease, stroke, and development of metabolic syndrome, all representing health conditions that are rapidly increasing among the global poor [92]. Although investigating the life-course and co-morbidity predictors of cognition is beyond the scope of the present analysis, the distinct patterns we identified in our study point to the importance that future research on cognition among the global poor needs to be guided by the life-course perspective on health outcomes.

Overall, our findings emphasize the importance of research on cognition among aging populations living in extreme poverty and have important implications for public health policy in SSA LICs. The staggering prevalence of poor cognitive health, especially among older women, calls for more research and policy interventions. Poor cognition, often combined with poor mental and physical health [7 , 93], suggests even a higher burden of diseases and higher adverse impact on individual wellbeing and family members, especially since the latter are the main caregivers of older adults in this context [94 –97]. As one of the first studies to describe longitudinal patterns of cognition and analyze their predictors in an SSA LIC population, our findings are of public health relevance since they contribute to a better understanding of the needs of older adults in poor contexts and help identify key sub-populations that are particularly affected by poor cognition (i.e., older women). The study emphasizes the importance of prioritizing cognitive health and research on cognition among older individuals in SSA LICs to which relatively few health care resources continue to be allocated.

Footnotes

ACKNOWLEDGMENTS

The authors are thankful for comments on manuscript drafts by members of the MLSFH Research Group.

FUNDING

The authors gratefully acknowledge the support by the Swiss Programme for Research on Global Issues for Development (SNF r4d Grant 400640_160374) and the National Institute of Aging (NIA R03 AG069817) Catalyst grant awarded to Iliana V. Kohler. Fabrice K. was supported by the Swiss National Science foundation (grant number: P2LAP1_187736). The Malawi Longitudinal Study of Families and Health (MLSFH) was also supported by a pilot funding received through the Penn Center for AIDS Research (CFAR), supported by NIAID AI 045008 and the Penn Institute on Aging; the Rockefeller Foundation; the National Institute of Child Health and Human Development (NICHD, Grant Nos. R03 HD05 8976, R21 HD050653, R01 HD044228, R01 HD053781); the National Institute on Aging (NIA, Grant Nos. P30 AG12836 and R21 AG053763); the Boettner Center for Pensions and Retirement Security at the University of Pennsylvania; and the NICHD Population Research Infrastructure Program (Grant Nos. R24 HD-044964), all at the University of Pennsylvania.

CONFLICT OF INTEREST

The authors have no conflict of interest to report.

DATA AVAILABILITY

The data supporting the findings of this study are available on request from the corresponding author. The data will be made publicly available as per commitment to funders later.

The supplementary material is available in the electronic version of this article: .

References

Avan

, Hachinski

, Brain

Health Learn and Act Group

(2022) Brain health: Key to health, productivity, and well-being. Alzheimers Dement 18, 1396–1407.

Alwin

, Hofer

(2011) Health and cognition in aging research. J Gerontol B Psychol Sci Soc Sci 66(Suppl 1), i9–i16.

Shah

, Albanese

, Duggan

, Rudan

, Langa

, Carrillo

, Chan

, Joanette

, Prince

, Rossor

, Saxena

, Snyder

, Sperling

, Varghese

, Wang

, Wortmann

, Dua

(2016) Research priorities to reduce the global burden of dementia by 2025. Lancet Neurol 15, 1285–1294.

GBD 2019 Dementia Forecasting Collaborators (2022) Estimation of the global prevalence of dementia in 2019 and forecasted prevalence in 2050: An analysis for the Global Burden of Disease Study 2019. Lancet Public Health 7, e105–e125.

Feigin

, Vos

, Nichols

, Owolabi

, Carroll

, Dichgans

, Deuschl

, Parmar

, Brainin

, Murray

(2020) The global burden of neurological disorders: Translating evidence into policy. Lancet Neurol 19, 255–265.

Prince

, Guerchet

, Prina

(2013) The global impact of dementia 2013-2050. Alzheimer’s Disease International (ADI), Policy Brief (Available on ADI website). http://www.alz.co.uk/research/G8-policy-brief.

Payne

, Kohler

, Bandawe

, Lawler

, Kohler

(2018) Cognitive health among older adults: Evidence from rural Sub-Saharan Africa. Eur J Popul 34, 637–662.

Livingston

, Sommerlad

, Orgeta

, Costafreda

, Huntley

, Ames

, Ballard

, Banerjee

, Burns

, Cohen-Mansfield

, Cooper

, Fox

, Gitlin

, Howard

, Kales

, Larson

, Ritchie

, Rockwood

, Sampson

, Samus

, Schneider

, Selbæk

, Teri

, Mukadam

(2017) The Lancet Commission on Dementia Prevention, Intervention, and Care. Lancet 390, 2673–2734.

Bloom

, Chen

, Counts

, Han

, Malik

, Nandi

, Seligman

, Vigo

(2021) Addressing Alzheimer’s disease and related dementias to realise the promise of the UN’s Decade of Healthy Ageing. VOX-EU CEPR: Research-based policy analysis and commentary from leading economists. Online available on VOX-EU website, accessed 26 March 2021. https://cepr.org/voxeu/columns/addressing-alzheimers-disease-and-related-dementias-realise-promise-uns-decade.

10.

GBD 2016 Neurology Collaborators (2019) Global, regional, and national burden of neurological disorders, 1990-2016: A systematic analysis for the Global Burden of Disease Study 2016. Lancet Neurol 18, 459–480.

11.

UN Population Division (2019) World Population Prospects, the 2019 revision. United Nations, Department of Economic and Social Affairs, Population Division. https://population.un.org/wpp/.

12.

World Bank (2020) World Development Indicators database. http://data.worldbank.org. Washington, DC.

13.

Fleming

, Zeisel

, Bennett

(2020) World Alzheimer Report 2020. Design, Dignity, Dementia: Dementia-Related Design and the Built Environment. Alzheimer’s Disease International, London, England.

14.

Langa

, Ryan

, McCammon

, Jones

, Manly

, Levine

, Sonnega

, Farron

, Weir

(2020) The Health and Retirement Study Harmonized Cognitive Assessment Protocol project: Study design and method. Neuroepidemiology 54, 64–74.

15.

Kohler

, Bandawe

, Ciancio

, Kämpfen

, Payne

, Mwera

, Mkandawire

, Kohler

(2020) Cohort profile: The Mature Adults Cohort of the Malawi Longitudinal Study of Families and Health (MLSFH-MAC). BMJ Open 10, e038232.

16.

Gurven

, Fuerstenberg

, Trumble

, Stieglitz

, Beheim

, Davis

, Kaplan

(2017) Cognitive performance across the life course of Bolivian forager-farmers with limited schooling. Dev Psychol 53, 160–176.

17.

G2Aging (2021) Gateway for Global Aging Data: A platform for population survey data on aging around the world. G2aging website, Center for Economic and Social Research (CESR), University of Southern California. Accessed June 22, 2022. https://g2aging.org/.

18.

UN and WHO (2021) UN Decade of Healthy Ageing: 2021-2030. United Nations and World Health Organization. https://www.who.int/initiatives/decade-of-healthy-ageing.

19.

Castañeda

, Doan

, Newhouse

, Nguyen

, Uematsu

, Azevedo

, World Bank Data for Goals Group (2018) A new profile of theglobal poor. World Dev 101, 250–267.

20.

Jha

, Mills

, Hanson

, Kumaranayake

, Conteh

, Kurowski

, Nguyen

, Cruz

, Ranson

, Vaz

, Yu

, Morton

, Sachs

(2002) Improving the health of the global poor. Science 295, 2036–2039.

21.

United Nations Development Programme (2020) Human Development Report 2020. United Nations.

22.

Kohler

, Kämpfen

, Ciancio

, Mwera

, Mwapasa

, Kohler

(2022) Curtailing Covid- 19 on a dollar-a-day in Malawi: Role of community leadership for shaping public health and economic responses to the pandemic. World Dev 151, 105753.

23.

UNDP (2022) Uncertain times, unsettled lives: Shaping our future in a transforming world. United Nations Development Programme - Human Development report 2021/2022.

24.

Short

, Baram

(2019) Early-life adversity and neurological disease: Age-old questions and novel answers. Nat Rev Neurol 15, 657–669.

25.

Kobayashi

, Farrell

, Payne

, Mall

, Montana

, Wagner

, Kahn

, Tollman

, Berkman

(2020) Adverse childhood experiences and domain-specific cognitive function in a population-based study of older adults in rural South Africa. Psychol Aging 35, 818.

26.

Kobayashi

, Farrell

, Langa

, Mahlalela

, Wagner

, Berkman

(2021) Incidence of cognitive impairment during aging in rural South Africa: Evidence from HAALSI, 2014 to 2019. Neuroepidemiology 55, 100–108.

27.

Zuelsdorff

, Okonkwo

, Norton

, Barnes

, Graham

, Clark

, Wyman

, Benton

, Gee

, Lambrou

, Johnson

, Gleason

(2020) Stressful life events and racial disparities in cognition among middle-aged and older adults. J Alzheimers Dis 73, 671–682.

28.

Majoka

, Schimming

(2021) Effect of social determinants of health on cognition and risk of Alzheimer disease and related dementias. Clin Ther 43, 922–929.

29.

Zuelsdorff

, Sonnega

, Byrd

, Benton

, Turner

(2020) Lifetime stressful experiences and cognitive performance in African American and white older adults: New evidence from a population-based cohort. Alzheimers Dement 16, e046422.

30.

, Yang

, Wan

, Shen

, Qiu

(2021) The influence of childhood adversities on mid to late cognitive function: From the perspective of life course. PLoS One 16, e0256297.

31.

Farah

(2017) The neuroscience of socioeconomic status: Correlates, causes, and consequences. Neuron 96, 56–71.

32.

Kohler

, Watkins

, Behrman

, Anglewicz

, Kohler

, Thornton

, Mkandawire

, Honde

, Hawara

, Chilima

, Bandawe

, Mwapasa

(2015) Cohort profile: The Malawi Longitudinal Study of Families and Health (MLSFH). Int J Epidemiol 44, 394–404.

33.

Malawi NSO (2019) 2018 Malawi Population and Housing Census: Main Report. Malawi National Statistical Office (Malawi NSO), Zomba, Malawi. https://malawi.unfpa.org/sites/default/files/resource-pdf/2018%20Malawi%20Population%20and%20Housing%20Census%20Main%20Report%20%281%29.pdf.

34.

UNDP (2010) Human Development Report 2010: The Real Wealth of Nations — Pathways to Human Development. United Nations Development Programme (UNDP), New York, NY. http://hdr.undp.org/en/reports/global/hdr2010.

35.

GBD 2019 Disease Collaboration (2020) Global burden of 369 diseases and injuries in 204 countries and territories, 1990-2019: A systematic analysis for the Global Burden of Disease Study 2019. Lancet 396, 1204–1222.

36.

Payne

, Mkandawire

, Kohler

(2013) Disability transitions and health expectancies among adults 45 years and older mature in Malawi: A cohort modeling approach. PLoS Med 10, e1001435.

37.

Kohler

, Payne

, Bandawe

, Kohler

(2017) The demography of mental health among mature adults in a low-income high HIV-prevalence context. Demography 54, 1529–1558.

38.

Kohler

, Hoang

, Amin

, Behrman

, Kohler

(2022) Resilience, accelerated aging and persistently poor health: Diverse trajectories of health among the global poor. University of Pennsylvania Population Center (PSC/PARC) Working Paper #2022-96. https://repository.upenn.edu/psc_publications/96.

39.

Malawi

DHS

(2017) Malawi Demographic and Health Survey 2015-16 (Final Report). National Statistical Office (NSO) and ICF Macro, Zomba, Malawi, and Calverton, Maryland, USA. http://www.measuredhs.com.

40.

Malawi

NSO

(2019) Malawi Poverty Report 2018. Government of Malawi, National Statistical Office (NSO), Lilongwe, Malawi. http://www.nsomalawi.mw/images/stories/data_on_line/economics/poverty/Malawi%20Poverty%20Report%20-%202019%20.pdf.

41.

Ciancio

, Behrman

, Kämpfen

, Kohler

, Maurer

, Mwapasa

, Kohler

(2022) Barker’s hypothesis among the global poor: Positive long-term cardiovascular effects of in-utero famine exposure. University of Pennsylvania Population Center (PSC/PARC) Working Paper #2022-80. https://repository.upenn.edu/psc_publications/80.

42.

Kohler

, Kohler

, Anglewicz

, Behrman

(2012) Intergenerational transfers in the era of HIV/AIDS: Evidence from rural Malawi. Demogr Res 27, 775–834.

43.

Stieglitz

, Schniter

, von Rueden

, Kaplan

, Gurven

(2014) Functional disability and social conflict increase risk of depression in older adulthood among Bolivian forager-farmers. J Gerontol B Psychol Sci Soc Sci 70, 948–956.

44.

Bollyky

, Templin

, Cohen

, Dieleman

(2017) Lower-income countries that face the most rapid shift in noncommunicable disease burden are also the least prepared. Health Aff (Millwood) 36, 1866–1875.

45.

Ciancio

, Kämpfen

, Kohler

(2021) Health screening for emerging non-communicable disease burdens among the global poor: Evidence from sub-Saharan Africa. J Health Econ 75, 102388.

46.

Busch

, Chelune

, Suchy

(2006) Using Norms in Neuropsychological Assessment of the Elderly. Guilford Publications.

47.

Chandra

, Ganguli

, Ratcliff

, Pandav

, Sharma

, Belle

, Ryan

, Baker

, DeKosky

, Nath

(1998) Practical issues in cognitive screening of elderly illiterate populations in developing countries. The Indo-US Cross-National Dementia Epidemiology Study. Aging Clin Exp Res 10, 349–357.

48.

Beath

, Asmal

, Van den Heuvel

, Seedat

(2018) Validation of the Montreal cognitive assessment against the RBANS in a healthy South African cohort. S Afr J Psychiatr 24, 1304.

49.

O’Driscoll

, Shaikh

(2017) Cross-cultural applicability of the Montreal Cognitive Assessment (MoCA): A systematic review. J Alzheimers Dis 58, 789–801.

50.

Wen

, Zhang

, Niu

, Li

(2008) The application of Montreal Cognitive Assessment in urban Chinese residents of Beijing. Zhonghua Nei Ke Za Zhi 47, 36–39.

51.

Saleh

, Alkholy

RSAEHA

, Khalaf

, Sabry

, Amer

, El-Jaafary

, Khalil

MAEF

(2019) Validation of Montreal Cognitive Assessment-Basic in a sample of elderly Egyptians with neurocognitive disorders. Aging Ment Health 23, 551–557.

52.

Harvey

(2019) Domains of cognition and their assessment. Dialog Clin Neurosci 21, 227–237.

53.

Venti

(2011) Economic measurement in the Health and Retirement Study. In Forum for Health Economics & Policy, vol. 14. De Gruyter.

54.

Gross

, Kueider-Paisley

, Sullivan

, Schretlen

, International Neuropsychological Normative Database Initiative (2019) Comparison of approaches for equating different versions of the Mini-Mental State Examination administered in 22 studies. Am J Epidemiol 188, 2202–2212.

55.

Kang

, Park

(2021) Validation of four methods for converting scores on the Montreal Cognitive Assessment to scores on the Mini-Mental State Examination-2. Dement Neurocogn Disord 20, 41–51.

56.

Yang

, Land

(2013) Age-period-cohort analysis: New models, methods, and empirical applications. Taylor & Francis.

57.

MacQueen

(1967) Some methods for classification and analysis of multivariate observations. In Proceedings of the 5th Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Statistics, University of California Press Berkeley, pp. 281–297.

58.

Vora

, Oza

(2013) A survey on k-mean clustering and particle swarm optimization. Int J Innov Sci Mod Eng 1, 24–26.

59.

Kassambara

(2017) Practical guide to cluster analysis in R: Unsupervised machine learning 1^st Edition. CreateSpace Independent Publishing Platform.

60.

Thakare

, Bagal

(2015) Performance evaluation of k-means clustering algorithm with various distance metrics. Int J Comput Appl 110, 12–16.

61.

Dutta

, Bhattacharya

(2015) A short review of clustering techniques. Int J Adv Res Manag Soc Sci 4, 131–139.

62.

Myroniuk

, Payne

(2019) The longitudinal dynamics of household composition and wealth in rural Malawi. J Comp Fam Stud 50, 242–260.

63.

Snyder

, Asthana

, Bain

, Brinton

, Craft

, Dubal

, Espeland

, Gatz

, Mielke

, Raber

, Rapp

, Yaffe

, Carrillo

(2016) Sex biology contributions to vulnerability to Alzheimer’s disease: A think tank convened by the Women’s Alzheimer’s Research Initiative. Alzheimers Dement 12, 1186–1196.

64.

Nebel

, Aggarwal

, Barnes

, Gallagher

, Goldstein

, Kantarci

, Mallampalli

, Mormino

, Scott

, Yu

, Maki

, Mielke

(2018) Understanding the impact of sex and gender in Alzheimer’s disease: A call to action. Alzheimers Dement 14, 1171–1183.

65.

Ferretti

, Martinkova

, Biskup

, Benke

, Gialdini

, Nedelska

, Rauen

, Mantua

, Religa

, Hort

, Santuccione Chadha

, Schmidt

(2020) Sex and gender differences in Alzheimer’s disease: Current challenges and implications for clinical practice: Position paper of the Dementia and Cognitive Disorders Panel of the European Academy of Neurology. Eur J Neurol 27, 928–943.

66.

Islam

KMM

, Asadullah

(2018) Gender stereotypes and education: A comparative content analysis of Malaysian, Indonesian, Pakistani and Bangladeshi school textbooks. PLoS One 13, e0190807.

67.

Asadullah

, Chaudhury

(2010) Religious schools, social values, and economic attitudes: Evidence from Bangladesh. World Dev 38, 205–217.

68.

Jayachandran

(2015) The roots of gender inequality in developing countries. Annu Rev Econ 7, 63–88.

69.

Levine

, Harrati

, Crimmins

(2018) Predictors and implications of accelerated cognitive aging. Biodemography Soc Biol 64, 83–101.

70.

Liu

, Zhang

, Choi

, Langa

(2020) Marital status and dementia: Evidence from the Health and Retirement Study. J Gerontol B Psychol Sci Soc Sci 75, 1783–1795.

71.

Asiimwe

, Farrell

, Kobayashi

, Manne-Goehler

, Kahn

, Tollman

, Kabudula

, Gómez-Olivé

, Wagner

, Montana

, Berkman

, Glymour

, Bärnighausen

(2020) Cognitive differences associated with HIV serostatus and antiretroviral therapy use in a population-based sample of older adults in South Africa. Sci Rep 10, 16625.

72.

Habib

, Yakasai

, Owolabi

, Ibrahim

, Habib

, Gudaji

, Karaye

, Ibrahim

, Nashabaru

(2013) Neurocognitive impairment in HIV-1-infected adults in Sub-Saharan Africa: A systematic review and meta-analysis. Int J Infect Dis 17, e820–e831.

73.

Sacktor

, Nakasujja

, Robertson

, Clifford

(2007) HIV-associated cognitive impairment in sub-Saharan Africa— the potential effect of clade diversity. Nat Clin Pract Neurol 3, 436–443.

74.

Baranov

, Kohler

(2018) The impact of AIDS treatment on savings and human capital investment in Malawi. Am Econ J Appl Econ 10, 266–306.

75.

Verhagen

SJW

, Daniëls

NEM

, Bartels

, Tans

, Borkelmans

KWH

, de Vugt

, Delespaul

PAEG

(2019) Measuring within-day cognitive performance using the experience sampling method: A pilot study in ahealthy population. PLoS One 14, e0226409.

76.

Makles

(2012) Stata tip 110: How to get the optimal k-means cluster solution. Stata J 12, 347–351.

77.

Castaneda

, Doan

, Newhouse

, Nguyen

, Uematsu

, Azevedo

(2018) A new profile of the global poor. World Dev 101, 250–267.

78.

Baulch

, Hoddinott

(2000) Economic mobility and poverty dynamics in developing countries. J Dev Stud 36, 1–24.

79.

NIA (2023) Approved concept: Building neuroscience research infrastructure for AD/ADRD in Africa. National Institute on Aging, Approved Concepts Announcement, January 2023. https://www.nia.nih.gov/approved-concepts#SSAfrica.

80.

Ikanga

, Reyes

, Kaba

, Akilimali

, Mampunza

, Epenge

, Gikelekele

, Kavugho

, Tshengele

, Hammers

, Alonso

(2023) Prevalence of suspected dementia in a sample of adults living in Kinshasa-Democratic Republic of the Congo. Alzheimers Dement, 10.1002/alz.13003.

81.

Mcintosh

, Tureson

, Rotblatt

, Singer

, Thames

(2021) HIV, vascular risk factors, and cognition in the combination antiretroviral therapy era: A systematic review and meta-analysis. J Int Neuropsychol Soc 27, 365–381.

82.

Rubin

, Sundermann

, Moore

(2019) The current understanding of overlap between characteristics of HIV-associated neurocognitive disorders and Alzheimer’s disease. J Neurovirol 25, 661–672.

83.

Scott

, Woods

, Carey

, Weber

, Bondi

, Grant

(2011) Neurocognitive consequences of HIV infection in older adults: An evaluation of the “Cortical” Hypothesis. AIDS Behav 15, 1187–1196.

84.

Kobayashi

, Mateen

, Montana

, Wagner

, Kahn

, Tollman

, Berkman

(2019) Cognitive function and impairment in older, rural South African adults: Evidence from “Health and Aging in Africa: A longitudinal study of an INDEPTH community in rural South Africa”. Neuroepidemiology 52, 32–40.

85.

Ismail

, Creese

, Aarsland

, Kales

, Lyketsos

, Sweet

, Ballard

(2022) Psychosis in Alzheimer disease— Mechanisms, genetics and therapeutic opportunities. Nat Rev Neurol 18, 131–144.

86.

Varney

, Roberts

, Springer

, Connell

, Wood

(1997) Neuropsychiatric sequelae of cerebral malaria in Vietnam veterans. J Nerv Ment Dis 185, 695.

87.

Kihara

, Carter

, Newton

(2006) The effect of plasmodium falciparum on cognition: A systematic review. Trop Med Int Health 11, 386–397.

88.

Kanmogne

, Kuate

, Cysique

, Fonsah

, Eta

, Doh

, Njamnshi

, Nchindap

, Franklin

, Ellis

, McCutchan

, Binam

, Mbanya

, Heaton

, Njamnshi

(2010) HIV-associated neurocognitive disorders in sub-Saharan Africa: A pilot study in Cameroon. BMC Neurol 10, 60.

89.

Robertson

, Liner

, Hakim

, Sankalé

, Grant

, Letendre

, Clifford

, Diop

, Jaye

, Kanmogne

, Njamnshi

, Langford

, Weyessa

, Wood

, Banda

, Hosseinipour

, Sacktor

, Nakasuja

, Bangirana

, Paul

, Joska

, Wong

, Boivin

, Holding

, Kammerer

, Van Rie

, Ive

, Nath

, Lawler

, Adebamowo

, Royal

3rd , Joseph

; NeuroAIDS in Africa Conference Participants (2010) NeuroAIDS in Africa. J Neurovirol 16, 189–202.

90.

Lawler

, Mosepele

, Seloilwe

, Ratcliffe

, Steele

, Nthobatsang

, Steenhoff

(2011) Depression among HIV-positive individuals in Botswana: A behavioral surveillance. AIDS Behav 15, 204–208.

91.

Lawler

, Jeremiah

, Mosepele

, Ratcliffe

, Cherry

, Seloilwe

, Steenhoff

(2011) Neurobehavioral effects in HIV-positive individuals receiving highly active antiretroviral therapy (HAART) in Gaborone, Botswana.6. PLoS One e17233.

92.

Kämpfen

, Wijemunige

, Evangelista

(2018) Aging, non-communicable diseases, and old-age disability in low-and middle-income countries: A challenge for global health. Int J Public Health 63, 1011–1012.

93.

Kohler

, Ciancio

, Kämpfen

, Kohler

, Mwapasa

, Chilima

, Vakhumbo

, Mwera

, Maurer

(2022) Pain is widespread and predicts poor mental health and subjective wellbeing among older adults in rural Malawi. Innov Aging 6, igac008.

94.

Ainamani

, Alele

, Rukundo

, Maling

, Wakida

, Obua

, Tsai

(2020) Caregiving burden and mental health problems among family caregivers of people with dementia in rural Uganda. Glob Ment Health (Camb) 7, e13.

95.

Imarhiagbe

, Asemota

, Oripelaye

, Akpekpe

, Owolabi

, Abidakun

, Akemokwe

, Ogundare

, Azeez

, Osakue

(2017) Burden of informal caregivers of stroke survivors: Validation of the Zarit burden interview in an African population. Ann Afr Med 16, 46.

96.

Prince

; 10/66 Dementia Research Group (2004) Care arrangements for people with dementia in developing countries. Int J Geriatr Psychiatry 19, 170–177.

97.

Addo

, Agyemang

, Tozan

, Nonvignon

(2018) Economic burden of caregiving for persons with severe mental illness in sub-Saharan Africa: A systematic review. PLoS One 13, e0199830.