Adverse childhood experiences and domain-specific cognitive function in a population-based study of older adults in rural South Africa

Abstract

Research on early-life adversity and later-life cognitive function is conflicting, with little evidence from low-income settings. We investigated associations between adverse childhood experiences and cognitive function in an older population who grew up under racial segregation during South African apartheid. Data were from 1,871 adults aged 40–79 in the population-representative “Health and Ageing in Africa: A Longitudinal Study of an INDEPTH Community in South Africa” (HAALSI) in 2015. The adverse childhood experiences were having a parent unemployed for >6 months, having parents who argued or fought often, having a parent who drank excessively, used drugs, or had mental health problems, and, physical abuse from parents. Executive function, language, visuospatial ability, and memory were assessed with the Oxford Cognitive Screen-Plus, a validated cognitive assessment designed for low-income, low-literacy settings. We estimated associations between adverse childhood experiences and latent cognitive domain z-scores using multiple-indicator, multiple-cause structural equation models. Childhood adversities were reported by 15% (parental unemployment for >6 months), 25% (parents argued or fought often), 25% (a parent drank excessively, used drugs, or had mental health problems), and 35% (physical abuse from parent) of respondents. They were not associated with cognition, except having a parent who drank excessively, used drugs, or had mental health problems was associated with lower memory z-scores (−0.07; 95% CI: −0.13, −0.01). This is one of the first investigations into later-life cognitive outcomes associated with early adversity in a population with a historical context of pervasive trauma, and suggests that later-life memory may be vulnerable to early adversity.

Adverse childhood experiences (ACEs) are stressful or traumatic events that are experienced during childhood, defined here as prior to age 16 (Felliti et al., 1998). The most commonly studied examples are physical, social, or emotional abuse, neglect, exposure to violence, substance use or mental illness in the household, and parental divorce (Dong et al., 2004; Felliti et al., 1998). ACEs are documented to have relationships with a range of health outcomes throughout the life course, including increased risks of diabetes, asthma, stroke, myocardial infarction, and all-cause mortality (Young & Widom, 2014; Dong et al., 2004; Dube, Felliti, Dong, Giles, & Anda, 2003; Edwards, Holden, Felitti, & Anda, 2003; Felliti et al., 1998; Geoffroy, Pereira, Li, & Power, 2016; Gilbert et al., 2015; Nikulina, Widom, & Brzustowicz, 2012). ACEs are thought to affect risk of these later-life health outcomes via adverse effects on social, emotional, and cognitive development in childhood and adolescence, and some longitudinal research indicates that adulthood mental health, risk behaviors, and socioeconomic status mediate the effects of ACEs on later-life health outcomes (Young & Widom, 2014; Edwards et al., 2003; Richards & Hatch, 2011).

Although exposures to various ACEs have been associated with impaired cognitive development during childhood and adolescence (Bremner & Narayan, 1998; Roos, Kim, Schnabler, & Fisher, 2016; van der Kolk, 2003), it is surprisingly unclear whether these relationships persist with respect to cognitive health outcomes during aging. Longitudinal evidence from the 1958 British Birth Cohort study demonstrated that childhood exposure to neglect or abuse was negatively associated with domain-specific cognitive functions through adolescence and at age 50 (Geoffroy et al., 2016), while the Chicago Health and Aging Project showed protective associations between childhood adversity and cognitive decline in older African Americans, but no associations in whites (Barnes et al., 2012). Other evidence is limited to studies of small, cross-sectional designs restricted mostly to high-income country settings (Burri, Maercker, Krammer, & Simmen-Janevska, 2013; Donley, Lönnroos, Tuomainen, & Kauhanen, 2018; Gould et al., 2012; Radford et al., 2017).

Very little research has investigated the later-life health consequences of ACEs among older populations in low- and middle-income country (LMIC) contexts. This is a major gap in the literature, as the types, number, and severity of ACEs may differ from those observed in higher-income settings, and the populations of many LMICs are rapidly aging (Tollman, Norris, & Berkman, 2016). Further, the broader social, political, and economic structures which shape individuals’ risks of these experiences and their abilities to recover from them also vary widely. Here, we are interested in South Africa, where legislated racial segregation during apartheid from 1948 to 1994 exposed members of the population who were deemed to be non-white to institutional racism, forced relocation to rural areas, exclusion from quality education and employment, and day-to-day violence and discrimination (Abdi, 2002; United Nations, 1963). In rural northeast South Africa, the region of interest in the present study, a sizable proportion of the population came to the region as refugees from a civil war in neighboring Mozambique between 1975 and 1992 (Kahn et al., 2012; Sartorius et al., 2013). Together, these two long-running and traumatic events have likely shaped the range of early life course exposures to adverse and traumatic events experienced by the current older population of rural northeast South Africa. However, relationships between childhood adversities experienced during the period of South African apartheid and subsequent later-life cognitive health outcomes among older adults in South Africa are unknown.

We thus aimed to estimate the associations between four self-reported adverse childhood experiences related to parental behavior and employment that occurred prior to age 16 and domain-specific cognitive function in a population-based study of older adults living in rural northeast South Africa (Gómez-Olivé et al., 2018). A consideration in this study population is that the simple translation and adaption of cognitive screening batteries developed for high-income, Western populations may be inappropriate in this population, due to cultural differences as well as low literacy levels caused by lack of access to high-quality education. We thus assessed domain-specific cognitive function using the Oxford Cognitive Screen (OCS)-Plus, a measure designed for use in low-literacy, low-income settings and previously validated in our study population (Humphreys et al., 2017). We hypothesized that there would be negative relationships between each of the four ACEs and later-life performance in cognitive domains of executive function, language, visuospatial ability, and memory. The rationale underlying this hypothesis is that these associations may reflect direct neurobiological effects of adversity that impede cognitive development in childhood and/or life course pathways through which childhood adversity persists to affect the rate of cognitive decline during aging up to the time of the cognitive assessment.

Method

Study design and population

Data were from the 2015 baseline wave of “Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa” (HAALSI), a population-representative cohort study of 5,059 adults aged ≥40 years (Gómez-Olivé et al., 2018). HAALSI is the first U.S. Health and Retirement Study (HRS) International Partner Study in Africa, and is representative of its underlying census-based sampling frame of ~110,000 people in Agincourt sub-district, Mpumalanga province, South Africa (response rate: 86%) (Gómez-Olivé et al., 2018; Kahn et al., 2012). Agincourt is part of a former “homeland” region of forced racial segregation during South African apartheid, where black population members were resettled according to their ethnic group affiliation (Kahn et al., 2012). The racial and ethnic composition of this region has remained predominantly Shaangan since the end of apartheid in 1994 (Kahn et al., 2012). Since the end of apartheid in 1994, economic development in the region has improved, but it remains a low-income setting with gaps in public services and persistently high rates of unemployment.

Here, we use data from a sub-sample aged 40–79 years who were randomly selected within age- and sex-strata to complete an additional laboratory visit that included an extended study interview and biomarker assessments (N=2,498). The younger age range of this study population compared to most studies of aging in high-income countries was selected partly because life expectancy is relatively low in South Africa, at 61.1 years at birth for men and 67.3 years at birth for women in 2018 (Statistics South Africa, 2018). Thus, chronological age in the HAALSI cohort may not be equivalent to chronological age in most other studies of aging in settings where average life expectancy is much longer, such as the United States. Within the laboratory sub-sample, a total of 2,114 (85%) participants consented to and completed the Oxford Cognitive Screen Plus (OCS-Plus) tablet measure designed to assess domain-specific cognitive function in low-income and low-literacy settings (Humphreys et al., 2017). The OCS-Plus is described in more detail below. As the childhood adversities under study depended on parental behavior and circumstances, we excluded respondents who reported that one or both of their parents died before they were aged 16 (N=243/2,114), for a total of 1,871 eligible study participants. All participants gave informed consent to participate in the study, and the study was conducted according to the principles embodied in the Declaration of Helsinki. Ethical approvals for HAALSI were obtained from the University of the Witwatersrand Human Research Ethics Committee (#M141159), the Harvard T.H. Chan School of Public Health Office of Human Research Administration (#13–1608), and the Mpumalanga Provincial Research and Ethics Committee.

Measures

All study measures were assessed during in-person interviews in participants’ homes with local, trained fieldworkers (Gómez-Olivé et al., 2018). Most of the HAALSI study measures are harmonized with those used in the HRS and its other International Partner Studies of aging, and have been adapted where appropriate for cultural and linguistic relevance. All study materials were translated and back-translated from English to the local Shaangan language (xiTsonga), and were pilot tested with a random sample of adults who did not become part of the HAALSI sample to ensure appropriateness, comprehension, and reliability of the interview items.

Adverse childhood experiences (ACEs).

Four ACEs experienced prior to age 16 were assessed during a life history module in the study interview: 1) having either parent unemployed for six months or more when they wanted to be working; 2) having parents who fought or argued often; 3) having a parent who drank excessively, used drugs, or had mental health problems; 4) physical abuse from a parent. All four items had binary response options (yes; no), and were adapted from a life history module in the English Longitudinal Study of Ageing (ELSA) (Scholes et al., 2009). Each item was examined individually, and in summation to create a measure of total cumulative ACE exposures.

Domain-specific cognitive function.

Domain-specific cognitive function was assessed with OCS-Plus, a tablet-based measure designed for use in low-literacy and low-income settings, such as Agincourt sub-district (Humphreys et al., 2017). The OCS-Plus uses image-based and auditory cues with no literacy or numeracy skills required to complete the cognitive assessment tasks. Standard cognitive battery assessments that are used in high-income countries ask the respondents to, for example, count backwards from 100 in counts of 7 (serial 7’s) or to connect numbers and letters of the alphabet (trails A and B tasks). Performance on these tests would be influenced by learned literacy and numeracy, even after conditioning on true underlying level of cognitive function. In place of tests such as these, the OCS-Plus asks the respondents, for example, to connect circles and squares and to remember images rather than words. The OCS-Plus assesses domain-specific (language; memory; visuospatial abilities) and domain-general (executive function; auditory attention) cognitive functions. A previous validation study in our study population established the construct validity, external validity, a lack of floor or ceiling effects, and strong psychometric properties of the OCS-Plus battery (Humphreys et al., 2017). The individual cognitive test items are described in detail in Supplemental Material A.

Covariates.

Any common causes of ACEs and later-life cognitive function that could induce confounding bias in our models would have to arise early in life. We thus included in modeling as potential confounders: age (continuous), sex (male; female), country of birth (South Africa; Mozambique or other), and skill level from the International Standard Classification of Occupations (ISCO-08) of father’s main job, as a marker of household socioeconomic conditions during childhood (level 1: unskilled manual labor; level 2: skilled manual labor and service sector labor; level 3: traditional healers and small business assistants; level 4: professional jobs) (Kobayashi et al., 2017). To further contextualize the study population, we additionally described the sample according to the following characteristics, overall and according to the four ACEs: literacy (can read and write; cannot read and/or write), educational attainment (0 years; 1–7 years; 8–11 years; 12+ years), asset-based household wealth (in quintiles), employment status (currently employed part- or full-time; unemployed; homemaker), marital status (married or living as married; not married), living alone (yes; no), score on the 7-item short post-traumatic stress disorder symptom scale (low [<4 symptoms]; high [≥4 or more symptoms]) (Breslau, Peterson, Kessler, & Schultz, 1999), felt depressed in the past week (no; yes), one or more limitations to activities of daily living (no; yes), and, self-rated health (very good/good; moderate/bad/very bad).

Statistical analysis

Characteristics of the sample were described according to the above covariates, overall and according to ACEs. We estimated tetrachoric correlations and associated p-values between the four ACEs. Guided by the neuropsychological theory underlying the domains measured in the OCS-Plus tablet (Humphreys et al., 2017), latent factor scores were derived for the cognitive domains of executive function, language, visuospatial ability, and memory using maximum likelihood estimation within a multiple indicator multiple cause (MIMIC) model (Jones, 2006; Muthén & Muthén, 2017). The MIMIC model is a two-level system of equations estimated simultaneously. The first level is a measurement model for the cognitive domain factor scores, equivalent to a confirmatory factor analysis, where individual i’s response to cognitive test item j is modeled by:

$$y_{ij}={\beta }_j{\eta }_i+{\epsilon }_j$$ 1

where $\beta$ is a measurement slope (factor loading) that relates the underlying latent factor score $\eta$ to the cognitive test item response and $\epsilon$ is the residual cognitive test item variance that does not contribute to $\eta .\eta$ is standardized to a mean of 0 and variance of 1. This latent variable approach to deriving a neuropsychological battery summary score allows the individual cognitive tests to vary in their weight of contribution to underlying cognitive function and utilizes their shared variation in deriving the underlying latent factor score, preventing measurement error that may be present in any single test item from being absorbed into the factor score (Gross et al., 2015; Muthén & Muthén, 2017). Residual error variances on closely related individual cognitive tests (e.g. immediate and delayed recall) were allowed to correlate. All available data were utilized in estimating the four latent cognitive function domains, even if some observations were missing (Muthén & Muthén, 2017). The second level of the MIMIC model is a structural model, which relates the underlying latent factor scores to exposure variables x:

$${\eta }_i=\alpha +{\gamma }_k{x}_{k,i}+{\lambda }_{k}Cov{s}_{k,i}+\zeta$$ 2

The parameter $\alpha$ is the structural model intercept, with direct relationships between each adverse childhood experience $\left(x_k\right)$ represented by ${\gamma }_k$, with ${\lambda }_k$ representing the effects of k confounders (age, sex, country of birth, father’s main occupation during childhood), and the residuals are captured in $\zeta$. We tested for non-linear age slopes with a quadratic age term, but did not include it in the final model as its inclusion worsened model fit. Model fit was iteratively assessed using standard thresholds: root mean square error of approximation (RMSEA) <0.06; comparative fit index (CFI) >0.95, and Tucker-Lewis index (TLI) >0.95 (Hu & Bentler, 1999).

We conducted a secondary analysis to estimate the associations between total cumulative ACEs and domain-specific cognitive function, in order to assess dose-response relationships between the number of adverse experiences in childhood and performance in each cognitive function domain in later-life. Analyses were conducted using MPlus 7.2 (Los Angeles, CA) and Stata 15.1 (College Station, TX).

Sensitivity analyses

Imputation of missing data.

Data on ACEs were missing for 144 (8%; parental unemployment), 121 (6%; parents argued or fought often), 89 (5%; parent drank excessively, used drugs, or had mental health problems), and 14 (<1%; physically abused by parent) study participants. Characteristics of participants missing data on ACEs are shown in Supplemental Table B1. Missing covariate data were rare (Table 1). We conducted a sensitivity analysis by imputing values of missing ACEs using sequential chained equations to generate 10 imputed datasets with a burn-in period of 10 imputations (White, Royston, & Wood, 2011). We then estimated ACE exposure prevalence and associations with cognitive domain factor scores in the imputed datasets (please see full methods in Supplemental Material B).

Table 1.

Sample Characteristics According to Four Adverse Childhood Experiences, “Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa” (HAALSI), Agincourt, South Africa, 2015

Characteristic	Total		Parent unemployed for 6+ months (n = 1727)		Parents argued or fought often (n = 1750)		Parent drank, did drugs, or mental health problems (n = 1782)		Physically abused by parent (n = 1857)
	n	%	n	%	n	%	n	%	n	%
Total	1871	100	252	15	438	25	448	72	667	35
Age (n = 1871)
Mean (SD)	58.9	11.0	56.9	10.0*	58.3	10.7	59.3	11.3	59.5	12.4
Sex (n = 1871)
Male	769	41	99	39	177	40	190	42	314	47*
Female	1101	59	153	61	261	60	258	58	353	53
Country of birth (n = 1870)
South Africa	1320	71	175	69	328	75	320	71	468	70
Mozambique or other	550	29	77	31	110	25	128	29	199	30
Father’s main job (n = 1869)
Skilled labor	933	50	117	47	237	54	217	48	326	49
Unskilled labor	547	29	83	33	125	29	137	31	204	31
Other job	213	11	25	10	47	11	50	11	73	11
Doesn’t know father’s job	176	9	26	10	29	7	44	10	64	10
Literacy level (n = 1870)
Cannot read or write	723	39	88	35	157	36	183	41	251	38
Can read or write	1147	61	164	65	281	64	265	59	416	62
Educational attainment (n = 1866)
No education	800	43	101	40	171	39	199	44	285	43*
Some primary (1–7 years)	682	37	102	41	178	41	156	35	227	34
Some secondary (8–11 years)	227	12	28	11	48	11	51	11	81	12
Secondary or more (12+ years)	157	8	20	8	41	9	42	9	73	11
Household wealth quintile (n = 1871)
1 (poorest)	403	21	53	21	91	21	99	22	151	23
2	362	19	55	22	76	17	97	22	133	20
3	366	20	48	19	84	19	79	18	118	18
4	373	20	49	19	100	23	99	22	138	21
5 (richest)	367	20	47	19	87	20	74	17	127	19
Employment status (n = 1866)
Employed part- or full-time	279	15	42	17	71	16	67	15	123	19*
Not working	1363	73	184	73	314	72	333	75	475	71
Homemaker	224	12	25	10	52	12	46	10	67	10
Marital status (n = 1871)
Married	962	51	131	52	228	52	206	46*	338	51
Unmarried	909	49	121	48	210	48	242	54	329	49
Lives alone (n = 1870)
No	1692	90	231	92	401	92	399	89	600	90
Yes	178	10	21	8	37	22	49	11	67	10
One or more limitations to activities of daily living (n = 1870)
No	1752	94	239	95	411	94	424	95	623	93
Yes	118	6	13	13	27	6	24	5	44	6
Post-traumatic stress disorder (PTSD) symptom score (n = 1852)
Low	1803	97	243	96	424	97	432	97	635	97
High	49	3	9	4	12	3	13	3	22	3
Felt depressed in the past week (n = 1852)
No	1648	89	220	87	375	86*	383	86*	577	88
Yes	204	11	32	13	62	14	63	14	78	12
Self-rated health (n = 1870)
Very good/good	1307	70	180	71	307	70	315	70	461	69
Moderate/bad/very bad	563	30	72	14	131	30	133	30	206	31

^•p < .05 for t-test (age) comparing the variable mean according to reported ACE exposure (yes vs. no), or chi-squared test (all other variables) comparing the frequencies of reported ACE exposure (yes vs. no) across the variable categories.

Quantitative bias analysis to evaluate under-reporting of ACEs.

The four ACEs of interest in this study dealt with sensitive topics related to parental behavior and circumstances that participants may not always wish to disclose. Previous research has demonstrated that ACEs may often be underreported when assessed retrospectively among adult samples (Baldwin, Reuben, Newbury & Danese, 2019; Hardt & Rutter, 2004). We thus conducted a quantitative bias analysis to evaluate the potential impact that under-reporting of ACEs would have on our results (Lash, Fox, & Fink, 2009). We used the Stata package ‘episensi’, which uses binary exposure and outcome measures to generate study estimates that account for systematic and random error due to misclassification of the exposure variable, according to user-specified probability distributions for the sensitivity and specificity of the exposure measure (Lash et al., 2009; Orsini, Bellocco, Bottai, Wolk, & Greenland, 2008). Because binary measures are required to generate values for sensitivity and specificity, we dichotomized cognitive function as cognitive impairment, defined as scoring 1.5 SD or more below the mean for two or more cognitive domains, within age (per decade) and educational degree-level strata.

In the absence of gold-standard validation data for ACE exposures, we based our quantitative bias analysis on previous evidence indicating that self-reported ACEs in adult samples are associated with a substantial degree of under-reporting (low sensitivity), but few false-positive reports (high specificity) (Hardt & Rutter, 2004). We performed two series of bias analyses: the first assumed that under-reporting of ACEs was non-differential with respect to cognitive impairment status, and the second assumed that under-reporting of ACEs was more frequent among cognitively impaired older adults. This allowed for the possibility that the remembering and reporting of ACEs differed based on later-life cognitive status. In both analyses, we varied the degree of potential under-reporting from moderate (probability distribution of sensitivity ranging from 60% to 80%) to substantial (40% to 60%), while consistently holding specificity within a probability distribution ranging from 90% to 100%. We performed Monte Carlo simulations of 2,000 datasets for each of the four ACEs under study, which produced bias-adjusted estimates described by the 2.5^th, 50^th, and 97.5^th percentiles of the distributions of simulated associations. These estimates are analogous to the point estimate and 95% CI of a conventional logistic regression analysis, and incorporate both systematic and random error due to ACE misclassification.

Results

Adverse childhood experiences (ACEs)

The most commonly reported ACE prior to age 16 was physical abuse from parents (35%), followed by having parents who fought or argued often (25%), having at least one parent who drank, used drugs, or had a mental health problem (25%), and having at least one parent unemployed for 6 months or more (15%; Table 1). Forty-five percent of the sample reported experiencing no ACEs, 30% reported one adverse experience, 13% reported two adverse experiences, 9% reported three experiences, and 2% reported four experiences.

Characteristics of the sample, overall and according to ACEs

Characteristics of the sample are shown in Table 1, overall and according to reported ACEs (Table 1). Mean age of the sample was 58.9 (standard deviation 11.0) years, 59% (1101/1871) were female, 71% (1320/1871) were born in South Africa, 39% (723/1871) could not read or write, 43% (800/1871) had no formal education and 8% (157/1871) had secondary education or higher, 15% (279/1871) were employed part- or full-time, and 51% (962/1871) were married (Table 1). The frequencies of all four ACEs were generally similar according to age, sex, country of birth, father’s main job, education, literacy, and current household wealth, employment status, marital status, and living conditions (Table 1). The only exceptions were that participants who reported having a parent unemployed for 6 months or more were slightly younger, on average (56.9 vs. 58.9 years); those who reported that their parents drank, used drugs, or had a mental health problem were less likely to be married; and, those who reported experiencing physical abuse from their parents were more likely to be male, to have secondary education or higher, and to be employed (Table 1). Similarly, indicators of current physical and mental health varied little according to the four ACEs, although those who reported that their parents argued or fought often, and, whose parent(s) drank, used drugs, or had mental health problems were more likely to have reported feeling depressed in the past week (Table 1). Please also see Supplemental Table C1 in Supplemental Material C for the polychoric correlations between the four ACEs and participant characteristics.

Correlations between ACEs

As shown in Table 2, the correlations between ACEs ranged in strength, from .11 (parent unemployed for 6 months or more, with physical abuse from parents) to .78 (parents who argued or fought often, with having at least one parent who drank, used drugs, or had a mental health problem), indicating that the four ACEs are related, but unique dimensions of childhood adversity related to parental behavior.

Table 2.

Tetrachoric Correlations Between the Four Adverse Childhood Experiences, “Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa” (HAALSI), Agincourt, South Africa, 2015

Adverse childhood experience	Parent unemployed for 6+ months		Parents argued or fought often		Parent drank, did drugs, or mental health problems		Physically abused by parent
	Rho	p	Rho	p	Rho	(p)	Rho	(p)
Parent unemployed for 6+ months	1.00
Parents argued or fought often	0.34	<.0001	1.00
Parent drank, did drugs, or had mental health problems	0.26	<.0001	0.78	<.0001	1.00
Physically abused by parent	0.11	0.02	0.33	<.0001	0.40	<.0001	1.00

ACEs and domain-specific cognitive function

Figure 1 presents the factor structure of the latent variables for executive function, visuospatial ability, language, and memory as estimated in the measurement model (first level of the MIMIC model), omitting all covariate paths for simplicity of presentation. The standardized factor loadings, which represent the strength of relationship between each individual cognitive test and the underlying latent cognitive factor, ranged from .420 (semantics on language factor) to .844 (delayed recall – recognition of target words from a multiple-choice list, if the respondent did not identify them during a free recall trial). The mean values for each cognitive domain latent factor score were similar and close to the standardized grand mean of zero across all study participants who did and did not report experiencing each of the four ACEs (Table 3).

Figure 1.

Factor structure of the OCS-Plus tablet assessment of cognitive function. Individual cognitive test items are depicted within rectangles on the right-hand side of the figure, and the latent cognitive domains which they reflect are depicted within circles on the left-hand side of the figure. The values along the paths from the latent domains to the individual items are the factor loadings from a MIMIC model, which regresses the cognitive domains on the four adverse childhood events, adjusting for age, sex, and country of birth, and father’s occupation during childhood (covariate paths not shown for simplicity). The model fit statistics are: RMSEA = .03 (95% CI: .027–.035); CFI = .961; TLI = .952.

Table 3.

Mean (SD) Performance in Four Cognitive Function Domains According to Each Adverse Childhood Experience, “Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa” (HAALSI), Agincourt, South Africa, 2015

Adverse childhood experience	Executive function		Language		Visuospatial ability		Memory
	Mean	SD	Mean	SD	Mean	SD	Mean	SD
Parent unemployed for 6+ months
No (n = 1,474)	0.03	0.86	0.04	0.87	0.04	0.87	0.02	0.85
Yes (n = 252)	0.05	0.86	0.06	0.89	0.05	0.89	0.00	0.90
Declined response (n = 144)	−0.37	0.91	−0.48	1.05	−0.45	1.05	−0.20	1.00
Parents argued or fought often
No (n = 1,311)	0.02	0.85	0.05	0.86	0.05	0.86	0.04	0.83
Yes (n = 438)	0.08	0.86	0.05	0.89	0.04	0.89	0.00	0.91
Declined response (n = 121)	−0.48	0.91	−0.67	1.08	−0.65	1.07	−0.38	1.07
Parent drank excessively, took drugs, or had mental health problems
No (n = 1,333)	0.03	0.86	0.06	0.86	0.06	0.87	0.05	0.84
Yes (n = 448)	0.02	0.83	−0.02	0.87	−0.03	0.89	−0.09	0.93
Declined response (n = 89)	−0.56	0.95	−0.73	1.16	−0.67	1.16	−0.25	1.02
Physically abused by parent
No (n = 1,189)	−0.01	0.86	0.00	0.87	0.00	0.88	0.01	0.85
Yes (n = 667)	0.01	0.89	0.00	0.93	−0.01	0.94	−0.02	0.90
Declined response (n = 14)	0.18	0.97	−0.04	1.22	0.00	1.20	0.08	0.93

The standardized parameter estimates from the MIMIC model for each of the four ACEs in relation to each cognitive domain were close to the null with 95% confidence intervals that included the null (Table 4). There was one exception: respondents who reported that their parents drank excessively, took drugs, or had mental health problems had memory factor scores that were, on average, .07 SD lower than those not reporting this adverse experience (95% CI: −.13, −.01; Table 4). Mean differences in cognitive domain factor scores per additional ACE (total cumulative ACEs) were .01 SD (95% CI: −.02, .05) for executive function, −.01 SD (95% CI: −.04, .02) for language, −.01 SD (95% CI: −.04, .02) for visuospatial ability, and −.03 SD (95% CI: −.07, .00) for memory.

Table 4.

Standardized Parameter Estimates from Multiple Indicator Multiple Cause (MIMIC) Structural Equation Model, “Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa” (HAALSI), South Africa, 2015

Adverse childhood experience	Executive function			Language			Visuospatial ability			Memory
	Estimate	95% CI		Estimate	95% CI		Estimate	95% CI		Estimate	95% CI
		LL	UL		LL	UL		LL	UL		LL	UL
Parent unemployed for 6+ months	−.015	−.066	.036	−.004	−.053	.045	−.003	−.054	.048	−.024	−.073	.025
Parents argued or fought often	.037	−.024	.098	−.010	−.069	.049	−.026	−.087	.035	.002	−.059	.063
Parent drank excessively, took drugs, or had mental health problems	−.002	−.065	.061	−.028	−.087	.031	−.012	−.074	.051	−.067	−.128	−.010
Physically abused by parent	.017	−.034	.068	.018	−.031	.067	−.007	0.058	.044	.019	−.032	.070

Note: All estimates adjusted for age, sex, country of birth, and father’s occupation during childhood. The MIMIC model was of good fit to the data, with RMSEA = .03 (95% CI: .027 – .034); CFI = .961; TLI = .952.

Sensitivity analysis for missing data

When missing values for ACEs were imputed based on covariate data, the prevalence of each ACE was nearly identical to the prevalence according to the observed values, and the parameter estimates negligibly changed in adjusted regressions predicting the cognitive domain factor scores (Supplemental Tables B2 and B3, Supplemental Material B). The exception was the effect estimate for memory in relation to having parents who drank excessively, took drugs, or had mental problems, which was nearly doubled in magnitude in the imputed data (−.11 SD; 95% CI: −.19, −.02; Supplemental Table B3, Supplemental Material B).

Quantitative bias analysis to evaluate under-reporting of ACEs

We conducted a quantitative bias analysis that set sensitivity and specificity parameters for the ACE measures according to probability distributions that corresponded to moderate and substantial under-reporting of ACEs, in scenarios that were both non-differential and differential with respect to participant cognitive impairment status (Table 5). The conventional (non-bias-adjusted) analysis showed increased odds of cognitive impairment associated with exposure to three of the four ACEs under study, although the ORs were of small magnitude and not statistically significant (Table 5). The quantitative bias analysis indicated that if ACEs were under-reported to a degree consistent with the specified probability distributions of sensitivity and specificity, then our results may be biased to the null. The ‘true’ associations between cognitive impairment and each ACE under study became stronger in magnitude as the degree of under-reporting increased (i.e. as sensitivity decreased), and as the degree of differential under-reporting among cognitively impaired individuals increased (Table 5). The width of the confidence intervals in the bias analysis indicates uncertainty due to both systematic and random error due to potential under-reporting of ACEs. Overall, the quantitative bias analysis indicated that even moderate under-reporting of ACEs (60–80% sensitivity of ACE interview measures) has the potential to obscure true associations with later-life cognitive function.

Table 5.

Summary of Odds Ratios (ORs) and Associated 95% Confidence Intervals (CIs) for the Conventional Analysis of ACE Exposures in Relation to Cognitive Impairment, and the Probabilistic Bias Analysis Accounting for Systematic and Random Error due to ACE Misclassification, with Varying Probability Distributions for the Sensitivity of the ACE Measures, “Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa” (HAALSI), South Africa, 2015

Adverse childhood experience	No misclassification			Non-differential ACE misclassification						Differential ACE misclassification
	Conventional			Moderate misclassificationa			Substantial misclassificationb			Moderate misclassificationc			Substantial misclassificationd
	OR	95% CI		OR	95% CI		OR	95% CI		OR	95% CI		OR	95% CI
		LL	UL		LL	UL		LL	UL		LL	UL		LL	UL
Parent unemployed for 6+ months	1.42	.89	2.54	1.70	1.06	2.85	1.81	1.12	3.00	2.01	1.25	3.35	3.32	1.97	5.64
Parents argued or fought often	1.24	.83	1.86	1.35	.92	2.05	1.49	1.00	2.27	1.66	1.11	2.52	3.28	2.05	5.44
Parent drank excessively, took drugs, or had mental health problems	1.11	.74	1.67	1.15	.78	1.74	1.20	.81	1.81	1.41	.95	2.15	2.67	1.70	4.29
Physically abused by parent	0.93	.64	1.34	0.90	.64	1.30	0.83	.58	1.20	1.17	.81	1.69	2.69	1.68	4.91

Note: Cognitive impairment is defined as scoring 1.5 SD or more below the mean for two or more cognitive domains, within age (per decade) and educational degree-level strata. Specificity is held at range of 90% to 100% for all analyses, with distributional modes at 93% and 97%.

^aSensitivity ranges from 60–80% (modes: 65–75%), regardless of cognitive impairment status

^bSensitivity ranges from 40–60% (modes: 45–55%), regardless of cognitive impairment status

^cSensitivity ranges from 60–80% (modes: 65–75%) for cognitively impaired individuals, and 70–90% (modes: 75–85%) for non-impaired individuals

^dSensitivity ranges from 40–60% (modes: 45–55%) for cognitively impaired individuals, and 70–90% (modes: 75–85%) for non-impaired individuals

Discussion

In this population-based study of older adults in rural South Africa, we generally observed no associations between four self-reported ACEs and domain-specific cognitive function, as measured by a novel tablet assessment designed for use in low-income and low-literacy settings. The exception was lower mean memory z-scores among those who reported that their parents drank excessively, used drugs, or had mental health problems. Given the life course literature on the ‘long arm’ of childhood for later-life health (Ben-Shlomo, 2002; Young & Widom, 2014; Dong et al., 2004; Dube et al., 2003; Edwards et al., 2003; Felliti et al., 1998; Geoffroy et al., 2016; Gilbert et al., 2015; Hayward & Gorman, 2004; Nikulina et al., 2012), and the specific social, economic, and political context of South African apartheid (Abdi, 2002), we hypothesized that ACEs would be associated with worse later-life cognitive function across all domains in this study. We did not expect to observe generally null associations. A potential explanation for our findings could be under-reporting of ACEs in the study interview, as indicated by our quantitative bias analysis. However, there are several plausible mechanisms that could underlie our results, each of which deserves investigation in its own right. Future research should corroborate our findings in order to clarify the causal role of ACEs in later-life cognitive function in low-income populations with historical legacies of institutional trauma.

Comparison with other literature

Previous research has demonstrated that exposure to ACEs is associated with impaired cognitive development in early-life, which is thought to be through direct neurobiological effects of trauma (Bremner & Narayan, 1998; Pears, Kim, & Fisher, 2008; van der Kolk, 2003). Our findings are inconsistent with cross-sectional and longitudinal studies among older adults that have observed reporting of adverse childhood experiences to be associated with greater likelihood of poor cognitive health outcomes (Burri et al., 2013; Donley et al., 2018; Geoffroy et al., 2016; Gould et al., 2012; Radford et al., 2017). An exception is the work of Barnes and colleagues, who observed that not having enough food and being thinner than average were protectively associated with rate of cognitive decline in older African Americans but not with rate of cognitive decline in older whites in the United States (Barnes et al., 2012). Barnes and colleagues also observed that financial adversity during childhood was associated with lower baseline cognitive function in their study, but not with rate of cognitive decline over the follow-up, in either African Americans or whites, a finding of unclear consistency with ours (Barnes et al., 2012).

Most existing studies have been conducted in high-income countries, where the population exposure distributions, types, and severities of childhood adversities, and individual vulnerabilities to lasting effects of adversity may vary from our study setting. The prevalence of ACEs in this study were generally, but not always similar to those observed in in the original Kaiser Permanente ACE Study of adults in California from 1995 to 1997 (Dube et al, 2003). For example, 25% of participants in our study reported having a parent who drank heavily, was mentally ill, or used drugs, compared to the 27% of Kaiser ACE Study participants who reported alcohol or drug use of a household member and 19% who reported a mental illness of a household member. However, 35% of participants in our study reported experiencing physical abuse from a parent, compared to 28% of Kaiser ACE Study participants (Dube et al, 2003). In our study, 45% of participants reported 0 ACEs, while 36% of Kaiser ACE Study participants reported 0 ACEs (Dube et al, 2003). In a more recent study of the older English population aged 50 and over, which used the same ACE measures that we had in addition to those regarding other parental behaviors, exposure to ACEs was much less common than in our study, with 65% of older English adults reporting 0 ACEs (Iob, Lacey, & Steptoe, 2019).

In the community-based World Mental Health Surveys conducted by the World Health Organization (WHO), 51,945 adults from 21 high-, middle-, and low-income countries completed an assessment of dichotomous ACE items relating to parental behavior and circumstances, similar to our study (Kessler et al., 2010). The overall prevalence of ACEs was much lower in the world surveys than in our study, with remarkably similar ACE prevalence across high-income, upper-middle-/middle-income, and lower-middle/low-income countries (Kessler et al., 2010). Overall, 61.2% of WHO World Mental Health Survey respondents reported 0 ACEs. The prevalence of each individual ACE was lower than in our study, with 4.0% reporting parental substance disorder, 6.2% reporting parental mental illness, 6.5% reporting family violence, 3.4% reporting economic adversity, and 8.0% reporting physical abuse (Kessler et al., 2010). The authors of study did not report on the demographic, economic, or social patterning of ACEs in the World Mental Health Surveys, so we cannot ascertain whether these ACEs were patterned by these factors, or not, as in our study (Kessler et al., 2010; McLaughlin et al., 2017). It is difficult to speculate on whether similarities and differences in overall prevalence of ACEs across these studies in low-, middle- and high-income contexts are due to true population differences in prevalence of ACEs, heterogeneous methods of ascertaining ACE exposure in different study interviews, or differential propensity to remember and report ACEs across populations.

In the present study, the ACEs related to parental behavior and employment occurred in a context of institutionalized racial segregation and discrimination under South African apartheid. These ACEs showed relatively little patterning by later-life demographic, economic, social, physical health, and mental health-related factors, in contrast to findings from high-income settings (Dube et al., 2003; Barnes et al., 2012; Young & Widom, 2014; Nurius, Green, Logan-Greene, & Borja, 2015). It might be that, in this context, the ubiquitous and insidious day-to-day trauma of living through South African apartheid overwhelmed the effects of the parental behavior-related ACEs that we measured in this study (Abdi, 2002; Hirschowitz & Orkin, 1997; United Nations, 1963). Consistent with our findings, where over half of the sample reported experiencing at least one of four ACEs, other research in rural South Africa has shown that exposure to violence, abuse, and physical hardship are highly prevalent prior to age 18 (Gibbs et al., 2019; Jewkes, Dunkle, Nduna, Jama, & Puren, 2010; Sherr et al., 2016). The prevalence of PTSD in this study was low, at 3% of our sample, and this was in line with results from the nationally representative South African Stress and Health Study (SASH), which identified PTSD prevalence of 2.3% in the adult population according to DSM-IV criteria, and prevalence of 3.5% among those with self-reported exposure to trauma (Atwoli et al., 2013). Our results should be corroborated in other low-income settings, especially those with a history of structural discrimination and violence, using a range of ACE exposure measures and study designs with differing sources of potential study bias in order to triangulate inference across findings (Glymour & Whitmer, 2019).

Potential Mechanisms

In this older South African population, we did not expect to see generally no associations between ACEs and domain-specific cognitive function. There are several possible explanations for our findings. First, there may truly be no association between the ACEs that we measured and later-life cognitive function in this study context. If this is true, then evidence demonstrating direct neurobiological effects of ACEs on cognitive development may not apply in this context, which would be surprising, and would require further inquiry given the strong link between cognitive development in early-life and cognitive decline in later-life (Glymour & Manly, 2008; Richards & Deary, 2005). It might be that the specific ACEs we studied, which relate to the family environment and parental behavior, are not the relevant types of adversities that would have long-lasting impact for later-life health in this setting. The individuals in our study who did not report experiencing the four ACEs we studied may have experiences of other, potentially more relevant forms of adversity or trauma, which were not assessed in our study. Future research should consider the different types of childhood adversity that relate more specifically to the trauma involved in South African apartheid, such as the various forms of interpersonal and institutional racism and violence that were and still are experienced by Black South Africans.

Another possible explanation for our findings, which has experienced relatively little attention in the literature, is selective survival prior to the start of the study. A substantial proportion of the HAALSI cohort has outlived their life expectancy (Statistics South Africa, 2018). If individuals who experienced ACEs and had poor cognitive function differentially died prior to the time of study entry, then no association may be apparent in the surviving older population, making the relationships under study age-dependent (Mayeda, Filshtein, Tripodis, Glymour, & Gross, 2018). This would be true whether poor cognitive function was caused by poor cognitive development in early-life or rapid cognitive decline in later-life. As longitudinal studies of aging such as HAALSI and other HRS International Partner Studies have now been launched in low- and middle-income countries with varying life expectancies, survival mechanisms that occur prior to study baseline should be a consideration for life course research using these data resources.

Limitations and strengths

A major limitation of this study is that we relied on retrospective self-reports of ACEs that occurred many years prior to the time of the study interview, with a range of at least 24 years prior for those aged 40 at the time of interview, to at least 63 years prior to those aged 79 at the time of interview. Retrospective self-reports, especially of potentially sensitive topics such as ACEs, are known to have a high degree of error from various sources. These include forgetting or misremembering, which has been reported in adult survivors of childhood sexual and non-sexual abuse (Williams, 1994; Feldman-Summers & Pope, 1994). Under-reporting (false negative responses) may also occur, if people do not wish to disclose their remembered experiences (Baldwin, Reuben, Newbury & Danese, 2019; Hardt & Rutter, 2004). Previous research has indicated that retrospectively reported ACEs may be a representation of reconstructed memories that can be shaped over the life course by both internal and external social factors (Susser & Widom, 2012; Widom, 2019). The construction of memories during childhood may be dependent on information given from parents, other family, or caregivers to children, and may even be reconstructed later on by an adult in the child’s life (Susser & Widom, 2012). Whether the responses we recorded for the four ACEs under study are an accurate representation of the truth – as opposed to (re)constructed memories – is impossible to know. The gold standard method to assess the later-life health effects of ACEs would be to enroll a cohort of children into a study with contemporaneous assessment of ACEs, followed by multi-decade, longitudinal follow-ups for assessment of later-life health outcomes. In the absence of these data being available, we had to rely on imperfect retrospective reports of ACEs after many years had elapsed.

Imperfect retrospective reporting of ACEs, if non-differential with respect to the cognitive outcomes under study, would likely result in our results being biased towards the null. However, differential retrospective reporting of ACEs according to cognitive status is plausible, and could severely bias our results in a direction that would be difficult to predict a priori. This limitation prompted us to quantify the potential impacts that various degrees of non-differential and differential under-reporting of ACEs could have had on our results. In the quantitative bias analysis, we allowed the degree of under-reporting of ACEs to differ according to cognitive impairment status of the study participants, as (re)constructed memories as well as decisions to disclose memories of these experiences may by influenced by cognitive health in later-life. Quantitative bias analyses often use gold standard validation data from alternative measurement instruments (Lash, Fox, & Fink, 2009), but these were unavailable for our study as it is difficult to validate self-reports of childhood trauma reported in older age. As recommended by Lash and colleagues (2009), we thus based our bias analysis on previous evidence, which has indicated a substantial degree of reporting error and under-reporting of ACEs, with false positive reporting indicated to be much less common (Hardt & Rutter, 2004).

Another limitation is that our ACE exposure measures had binary response options, which reduced a range of adversity dimensions such as frequency, severity, duration, ages when it occurred, and degree of psychological effects at the time. Our ACE measures were adapted from an existing questionnaire and were selected to allow comparability of our findings with the existing body of research on ACEs (Scholes et al., 2009). However, our measures focused only on aspects of parental behavior and employment, and did not capture the range of traumatic exposures specific to the historical context of South African apartheid, namely various forms of interpersonal and institutional racism and violence. Improvement of the measurement of ACEs not only in this specific study context, but also in other low-literate and low-income contexts with aging populations is an area where in-depth qualitative research could inform the development of higher quality interview measures for use in quantitative studies. Strengths of this study include its large sample size and population representativeness, our use of a multi-domain cognitive assessment designed for low-literate and low-income settings, and our inclusion of a sub-Saharan African population that is rapidly aging yet under-represented in cognitive aging research.

Conclusion

In this population-representative study of older adults in rural South Africa, we generally observed no associations between four brief measures of adverse childhood experiences and domain-specific cognitive function in later-life. The exception was lower mean memory z-scores among those who reported that their parents drank excessively, used drugs, or had mental health problems, indicating that later-life memory may be particularly vulnerable to early adversity. This study makes a novel contribution to research on ACEs and later-life cognitive health in a low-literate and low-income setting, with the methodological contribution of presenting data from a novel cognitive screening assessment designed for such settings. This study raises questions about the long-run health effects of early-life adversities experienced during South African apartheid. We demonstrated that under-reporting of ACEs in retrospective study interviews has the potential to bias results for the relationships between ACEs and cognitive function to the null. This potential form of study bias should be more frequently quantified, where possible, as life course studies of aging are often interested in early-life exposures that are self-reported and occurred long before the initiation of data collection. Future research should refine measures of ACEs for use in low-income settings, and corroborate the current findings in a variety of settings and study designs.