Introduction
As the global population ages, the incidence of age-related diseases, particularly dementia, is on the rise, placing increasing pressure on individuals, families and healthcare systems (Patterson, 2018; Livingston et al., Reference Livingston, Huntley, Liu, Costafreda, Selbæk, Alladi, Ames, Banerjee, Burns, Brayne, Fox, Ferri, Gitlin, Howard, Kales, Kivimäki, Larson, Nakasujja, Rockwood, Samus, Shirai, Singh-Manoux, Schneider, Walsh, Yao, Sommerlad and Mukadam2024). Currently, approximately 57 million people are living with dementia, with about two-thirds residing in low- and middle-income countries, and this number is projected to exceed 153 million by 2050 (Nichols et al., Reference Nichols, Steinmetz, Vollset, Fukutaki, Chalek, Abd-allah, Abdoli, Abualhasan, Abu-gharbieh, Akram, Al Hamad, Alahdab, Alanezi, Alipour, Almustanyir, Amu, Ansari, Arabloo, Ashraf, Astell-burt, Ayano, Ayuso-mateos, Baig, Barnett, Barrow, Baune, Béjot, Bezabhe, Bezabih, Bhagavathula, Bhaskar, Bhattacharyya, Bijani, Biswas, Bolla, Boloor, Brayne, Brenner, Burkart, Burns, Cámera, Cao, Carvalho, Castro-de-araujo, Catalá-lópez, Cerin, Chavan, Cherbuin, Chu, Costa, Couto, Dadras, Dai, Dandona, Dandona, De la Cruz-góngora, Dhamnetiya, Dias da Silva, Diaz, Douiri, Edvardsson, Ekholuenetale, El Sayed, El-jaafary, Eskandari, Eskandarieh, Esmaeilnejad, Fares, Faro, Farooque, Feigin, Feng, Fereshtehnejad, Fernandes, Ferrara, Filip, Fillit, Fischer, Gaidhane, Galluzzo, Ghashghaee, Ghith, Gialluisi, Gilani, Glavan, Gnedovskaya, Golechha, Gupta, Gupta, Gupta, Haider, Hall, Hamidi, Hanif, Hankey, Haque, Hartono, Hasaballah and Hasan2022). Cognitive decline, an early hallmark of dementia, plays a crucial role in facilitating early detection and enabling timely intervention. In response to the growing burden of ageing, large-scale ageing studies such as the US Health and Retirement Study (HRS) and its international sister studies have been established, alongside the Harmonized Cognitive Assessment Protocol (HCAP) sub-studies, which standardize cognitive measures across countries to enable cross-national comparisons (Langa et al., Reference Langa, Ryan, McCammon, Jones, Manly, Levine, Sonnega, Farron and Weir2020; Gross et al., Reference Gross, Li, Briceño, Arce Rentería, Jones, Langa, Manly, Nichols, Weir, Wong, Berkman, Lee and Kobayashi2023).
Previous studies have shown that exposure to traumatic life events is associated with an increased risk of dementia (Severs et al., Reference Severs, James, Letrondo, Løvland, Marchant and Mukadam2023), and emerging evidence suggests that the loss of a spouse constitutes a negative late-life stressor with potential cognitive consequences. The stress model (Williams and Umberson, Reference Williams and Umberson2004) posits that spousal loss constitutes a life stressor requiring substantial adaptation, with sustained stress adversely affecting mental health. Meanwhile, the marital resource model (Waite et al., Reference Waite, Gallagher and England2001) emphasizes that marriage provides social, psychological and economic resources that protect physical and mental health. Accordingly, widowed older adults – particularly those with limited socioeconomic resources – are expected to exhibit poorer cognitive function.
Notwithstanding the growing recognition of associations between widowhood and subsequent cognitive changes, much of the existing evidence comes from high-income countries and remains inconsistent. For example, some studies have reported increased risks of cognitive decline and dementia (Håkansson et al., Reference Håkansson, Rovio, Helkala, Vilska, Winblad, Soininen, Nissinen, Mohammed and Kivipelto2009; Shin et al., Reference Shin, Kim and Park2018; Sommerlad et al., Reference Sommerlad, Ruegger, Singh-Manoux, Lewis and Livingston2018; Liu et al., Reference Liu, Zhang, Burgard and Needham2019; Zhang et al., Reference Zhang, Li, Xu and Liu2019, Reference Liu, Zhang, Choi and Langa2020; Singham et al., Reference Singham, Bell, Saunders and Stott2021; Min and Song, Reference Min and Song2023), whereas others have found no significant association (Vidarsdottir et al., Reference Vidarsdottir, Fang, Chang, Aspelund, Fall, Jonsdottir, Jonsson, Cotch, Harris and Launer2014; Lee et al., Reference Lee, Chi and Palinkas2019; Xu et al., Reference Xu, Wei, Cheng, Wang, Li, Li, Sun, Du, Sheng, Liu, Tao and Yang2021; Nerobkova et al., Reference Nerobkova, Park, Shin and Park2022; Lee and Jiang, Reference Lee and Jiang2023). Thus, a comprehensive study that includes both developed and developing countries is essential for improving our understanding of these associations.
Indeed, previous studies on widowhood have largely focused on psychosocial factors like pre-loss marital quality (Chen et al., Reference Chen, Wu, Wang, Zeng, Huang, Lv, Niu, Meng, Cai, Shen, Gang, You, Lv, Ren, Shi and Ji2022), whereas cognitive decline is a multifactorial condition influenced by a variety of risk factors, such as educational attainment and health-related behaviours (Livingston et al., Reference Livingston, Huntley, Liu, Costafreda, Selbæk, Alladi, Ames, Banerjee, Burns, Brayne, Fox, Ferri, Gitlin, Howard, Kales, Kivimäki, Larson, Nakasujja, Rockwood, Samus, Shirai, Singh-Manoux, Schneider, Walsh, Yao, Sommerlad and Mukadam2024; Zhang et al., Reference Zhang, O’shea, Yu, Cho, Zhang, Kler, Langa, Weir, Gross and Kobayashi2024). Therefore, additional research is needed to investigate other potential factors that underlie the association between widowhood and cognitive health.
Using data from six countries, this study aims (1) to examine whether spousal loss is differentially associated with specific cognitive domains, including orientation, memory, executive function and language; (2) to assess whether these associations vary by sex and age; and (3) to identify individual-level characteristics associated with better cognitive performance among widowed individuals.
Methods
Data sources and study design
This observational prospective study included six cohorts from six countries (United States, England, Mexico, China, India and South Africa). These cohorts are all part of the Gateway to Global Aging Data project (https://g2aging.org/), described elsewhere (Lee et al., Reference Lee, Phillips, Wilkens, Gu and Dupre2021). The cohorts were the HRS, English Longitudinal Study of Ageing (ELSA), the Mexican Health and Aging Study (MHAS), the China Health and Retirement Longitudinal Study (CHARLS), the Longitudinal Ageing Study in India (LASI) and the Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa (HAALSI). All cohorts had ethical approvals from respective national or regional ethics committees with varying numbers of waves of data collection. All participants provided written or electronic informed consent. Details of these cohorts are shown in Figure 1 and the Supplementary information.
Study profile and design. A. Geographic distribution of the six cohorts, with participant numbers for each cohort. B. Study design, showing the exposure, outcome, and covariates (modifiable factors). C. Flowchart of the study.
Figure 1 Long description
The composite image consists of three sections: A. Six Prospective Cohort Studies: A world map highlights six countries involved in cohort studies: United States (HRS), England (ELSA), Mexico (MHAS), China (CHARLS), India (LASI) and South Africa (HAALSI). B. Study Design: This section outlines the study design with modifiable factors, baseline and waves 1 and 2. It includes HCAP cognitive tests such as orientation, memory, executive function, language and outcome/domain-specific. The exposure is spousal loss. C. Flowchart: The flowchart illustrates the classification of marital status across waves 1 and 2. Part icipants are classified as married or widowed. The flowchart shows transitions between marital statuses, with a note on error rates: continuously married error is less than 5 percent and baseline to incident widowed error is less than 10 percent.
Measures
Procedures
The six cohorts had varying numbers of waves, ranging from two to four. To minimize reverse causality between the exposure (spousal loss) and outcomes (cognitive function), we used the earlier waves (pre-final waves) to define spousal loss and then used cognitive function data from the final wave. Detailed information on the waves used to define spousal loss and cognitive function in each cohort is shown in Supplementary Figure S1.
Ascertainment of spousal loss (exposure)
We defined the exposure and control groups based on the previous studies (Lee et al., Reference Lee, Chi and Palinkas2019). Spousal loss was defined using self-reported legal marital status. Prior evidence indicates that legal marriage confers greater social and legal stability than cohabitation and is associated with more favourable emotional and cognitive outcomes (Perelli‐harris and Styrc, Reference Perelli‐harris and Styrc2018), therefore, participants who were partnered but not legally married (i.e., cohabiting) were excluded from the analyses. The exposure group included participants who reported being widowed during the observation period (i.e., the pre-final waves), whereas the control group consisted of participants who reported being married throughout the observation period. Participants who remarried after widowhood were excluded due to changes in exposure status.
Ascertainment of cognitive function (outcomes)
Given that previous mixed findings may be due to the different cognitive outcomes examined (e.g., global cognitive function, Alzheimer’s disease and dementia), this study assessed cognitive function using the HCAP neuropsychological battery from six countries. We used standardized z-scores of domain-specific cognitive functions (orientation, memory, executive function and language) as the outcome measures in this study. Detailed information on the measurement of cognitive function is shown in Supplementary Table S1.
Covariates and modifiable factors
Covariates were selected based on established theoretical frameworks, including age (continuous in years), sex (male or female), minority status (ethnicity [White/Caucasian or non-White] or living region [urban community or rural village]) and baseline cognition (≤25%, 25–50%, 50%–75% or >75%). In addition, guided by the modifiable dementia risk factors identified by the 2024 Lancet Commission (Livingston et al., Reference Livingston, Huntley, Liu, Costafreda, Selbæk, Alladi, Ames, Banerjee, Burns, Brayne, Fox, Ferri, Gitlin, Howard, Kales, Kivimäki, Larson, Nakasujja, Rockwood, Samus, Shirai, Singh-Manoux, Schneider, Walsh, Yao, Sommerlad and Mukadam2024), we included educational attainment (less than upper secondary, upper secondary/vocational training or tertiary), household wealth level (low, normal or high), body mass index (BMI) (underweight/normal, overweight or obesity), smoking status (current, former or never), drinking patterns (heavy, moderate, light or never), internet use (no or yes) and comorbidities (mild, moderate or severe) as the modifiable factors of interest in the present study. Detailed information on the measurement of these covariates is provided in Supplementary Table S2.
Statistical analysis
Given the cross-national nature of the study, we adopted a parallel cohort-specific analytic framework, whereby all analyses were conducted separately within each cohort using harmonized variable definitions, exposure classifications, outcome measures and covariate adjustments. This approach allows direct comparison of cohort-specific associations while accounting for contextual differences across populations (Kobayashi et al., Reference Kobayashi, Jones, Briceño, Rentería, Zhang, Meijer, Langa, Lee and Gross2024).
Descriptive analyses
We examined the distribution of sociodemographic and health-related factors between widowed and married groups within each cohort, using means with standard deviations (SDs) for continuous variables and counts with percentages for categorical variables.
Spousal loss and cognitive function
Associations between spousal loss and domain-specific cognitive function were examined using generalized linear models, with results reported as β coefficients and 95% confidence intervals (95% CI). To find out whether these associations varied by sex, analyses were also conducted separately for males and females. All models were adjusted for all covariates (including modifiable factors). We meta-analysed based on effect size that estimated from each cohort with a random-effects model using the metafor package in R to estimate overall β coefficients (Viechtbauer, Reference Viechtbauer2010). Heterogeneity for each overall cognitive outcome was examined using the I 2 statistic (Higgins et al., Reference Higgins, Thompson, Deeks and Altman2003).
Age-related nonlinearity and interaction analyses
To account for potential nonlinear age-related cognitive changes associated with widowhood (Shin et al., Reference Shin, Kim and Park2018; Liu et al., Reference Liu, Zhang, Burgard and Needham2019), we applied restricted cubic spline models with four knots. Effect modification by age was further evaluated by incorporating an age × widowhood interaction term to assess whether the association between spousal loss and cognitive function varied across the age distribution.
Cognitive correlates in widowed individuals
To identify potentially modifiable individual-level characteristics associated with cognitive performance among widowed individuals, we conducted multivariable generalized linear models restricted to this subgroup. The studied modifiable factors included educational attainment, household wealth level, BMI, smoking status, drinking patterns, internet use and comorbidities. All models were adjusted for age, sex, minority status and baseline cognition.
Sensitivity analyses
We performed several sensitivity analyses. First, to assess the effect of missingness on our results, we re-ran our analysis restricted to participants without the missing values. Second, to address potential concerns regarding temporal ordering and the operationalization of exposure status across multiple pre-final waves, we conducted a two-time-point analysis in which exposure was defined using baseline marital status only and cognitive outcomes were assessed at the final follow-up wave. Third, considering the strong influence of ageing on cognitive decline, we further restricted the HCAP sample to participants aged 65 years and older and re-evaluated the analysis. Last, to examine whether the meta-analytic results were disproportionately driven by high-income countries, we conducted an additional sensitivity analysis excluding results estimated from HRS and ELSA and re-ran the meta-analyses using the remaining four cohorts.
Additionally, outliers were identified by the quartile method and subsequently converted into missing values. We performed multiple imputation by chained equations to adequately address missing data in all analyses (Sterne et al., Reference Sterne, White, Carlin, Spratt, Royston, Kenward, Wood and Carpenter2009).
All analyses were conducted in Stata/MP 18.0 and R 4.3.2, with statistical significance defined as a two-sided P value <0.05. Further details are provided in the Supplementary Methods.
Results
Participant characteristics
A total of 18,551 participants were included from six cohorts (Table 1). The baseline mean age (SD) ranged from 61.39 (6.40) to 71.06 (7.22) years. Specifically, 2672 participants were from HRS, 1075 from ELSA, 1553 from MHAS, 8757 from CHARLS, 3924 from LASI and 570 from HAALSI. The proportion of widowed individuals varied substantially, ranging from 14.1% in CHARLS to 53.9% in HAALSI.
Characteristics of participants by countries
Table 1 Long description
Participant characteristics are reported for six cohorts: HRS with 2672 participants, ELSA with 1075, MHAS with 1553, CHARLS with 8757, LASI with 3924, and HAALSI with 570. Widowed group counts and percentages are 851 (31.9) in HRS, 308 (28.7) in ELSA, 407 (26.2) in MHAS, 1231 (14.1) in CHARLS, 1313 (33.5) in LASI, and 307 (53.9) in HAALSI. Mean age in years with standard deviation is 70.57 (7.50) for HRS, 71.06 (7.22) for ELSA, 65.35 (8.38) for MHAS, 61.39 (6.40) for CHARLS, 68.86 (7.43) for LASI, and 64.99 (11.29) for HAALSI. Minority status is 475 (17.8) in HRS, 31 (2.9) in ELSA, 469 (30.2) in MHAS, 5254 (60.0) in CHARLS, 2453 (62.5) in LASI, and listed as NA in HAALSI. Sex is male 1077 (40.3) and female 1595 (59.7) in HRS; male 475 (44.2) and female 600 (55.8) in ELSA; male 656 (42.2) and female 897 (57.8) in MHAS; male 4446 (50.8) and female 4311 (49.2) in CHARLS; male 1794 (45.7) and female 2130 (54.3) in LASI; and male 201 (35.3) and female 369 (64.7) in HAALSI. Educational attainment is less than upper secondary 461 (17.3), upper secondary or vocational training 1543 (57.7), tertiary 667 (25.0), and missing 1 (0.0) in HRS; 380 (35.3), 469 (43.6), 126 (11.7), and missing 100 (9.3) in ELSA; 1367 (88.0), 39 (2.5), 132 (8.5), and missing 15 (1.0) in MHAS; 7839 (89.5), 782 (8.9), 136 (1.6), and missing 0 in CHARLS; 2967 (75.6), 800 (20.4), 157 (4.0), and missing 0 in LASI; and 524 (91.9), 37 (6.5), 9 (1.6), and missing 0 in HAALSI. Household wealth level is low 896 (33.5), normal 892 (33.4), high 884 (33.1), and missing 0 in HRS; low 357 (33.2), normal 357 (33.2), high 357 (33.2), and missing 4 (0.4) in ELSA; low 614 (39.5), normal 569 (36.6), high 370 (23.8), and missing 0 in MHAS; low 3166 (36.2), normal 2567 (29.3), high 2845 (32.5), and missing 179 (2.0) in CHARLS; low 1231 (31.4), normal 1233 (31.4), high 1229 (31.3), and missing 231 (5.9) in LASI; and low 169 (29.6), normal 167 (29.3), high 167 (29.3), and missing 67 (11.8) in HAALSI. BMI categories are underweight or normal 495 (18.5), overweight 871 (32.6), obesity 1090 (40.8), and missing 216 (8.1) in HRS; 244 (22.7), 444 (41.3), 342 (31.8), and missing 45 (4.2) in ELSA; 469 (30.2), 639 (41.1), 389 (25.0), and missing 56 (3.6) in MHAS; 5605 (64.0), 2328 (26.6), 362 (4.1), and missing 462 (5.3) in CHARLS; 2563 (65.3), 783 (20.0), 255 (6.5), and missing 323 (8.2) in LASI; and 206 (36.1), 176 (30.9), 154 (27.0), and missing 34 (6.0) in HAALSI. Smoking status is current 221 (8.3), former 1191 (44.6), never 1242 (46.5), and missing 18 (0.7) in HRS; 95 (8.8), 590 (54.9), 390 (36.3), and missing 0 in ELSA; 169 (10.9), 414 (26.7), 970 (62.5), and missing 0 in MHAS; 2796 (31.9), 870 (9.9), 5066 (57.9), and missing 25 (0.3) in CHARLS; 1317 (33.6), 259 (6.6), 2325 (59.3), and missing 23 (0.6) in LASI; and 32 (5.6), 61 (10.7), 477 (83.7), and missing 0 in HAALSI. Drinking patterns are heavy 299 (11.2), moderate 683 (25.6), light 468 (17.5), never 1222 (45.7), and missing 0 in HRS; 239 (22.2), 408 (38.0), 238 (22.1), 161 (15.0), and missing 29 (2.7) in ELSA; 27 (1.7), 209 (13.5), 134 (8.6), 1183 (76.2), and missing 0 in MHAS; 1402 (16.0), 400 (4.6), 1802 (20.6), 5130 (58.6), and missing 23 (0.3) in CHARLS; 54 (1.4), 58 (1.5), 467 (11.9), 3324 (84.7), and missing 21 (0.5) in LASI; and 27 (4.7), 30 (5.3), 201 (35.3), 312 (54.7), and missing 0 in HAALSI. Internet use is no 1288 (48.2), yes 1384 (51.8), and missing 0 in HRS; no 479 (44.6), yes 567 (52.7), and missing 29 (2.7) in ELSA; no 660 (42.5), yes 893 (57.5), and missing 0 in MHAS; no 7024 (80.2), yes 193 (2.2), and missing 1540 (17.6) in CHARLS; no 3776 (96.2), yes 96 (2.4), and missing 52 (1.3) in LASI; and no 74 (13.0), yes 457 (80.2), and missing 39 (6.8) in HAALSI. Comorbidities are mild 1791 (67.0), moderate 639 (23.9), severe 242 (9.1), and missing 0 in HRS; 816 (75.9), 212 (19.7), 47 (4.4), and missing 0 in ELSA; 1172 (75.5), 309 (19.9), 72 (4.6), and missing 0 in MHAS; 7395 (84.4), 751 (8.6), 151 (1.7), and missing 460 (5.3) in CHARLS; 3313 (84.4), 536 (13.7), 66 (1.7), and missing 9 (0.2) in LASI; and 472 (82.8), 57 (10.0), 7 (1.2), and missing 34 (6.0) in HAALSI. Baseline cognition categories are 856 (32.0) for 25 percent or less, 685 (25.6) for 25 to 50 percent, 595 (22.3) for 50 to 75 percent, 536 (20.1) for greater than 75 percent, and missing 0 in HRS; 339 (31.5), 253 (23.5), 235 (21.9), 248 (23.1), and missing 0 in ELSA; 396 (25.5), 381 (24.5), 411 (26.5), 365 (23.5), and missing 0 in MHAS; 2906 (33.2), 2059 (23.5), 1853 (21.2), 1831 (20.9), and missing 108 (1.2) in CHARLS; 1082 (27.6), 1236 (31.5), 742 (18.9), 820 (20.9), and missing 44 (1.1) in LASI; and 218 (38.2), 75 (13.2), 180 (31.6), 92 (16.1), and missing 5 (0.9) in HAALSI.
Notes: HRS, the Health and Retirement Study; ELSA, the English Longitudinal Study of Ageing; MHAS, the Mexican Health and Aging Study; CHARLS, the China Health Retirement Longitudinal Study; LASI, the Longitudinal Aging Study in India; HAALSI, the Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa; SD, standard deviation; NA, not applicable; BMI, body mass index.
Unless indicated otherwise, data are expressed as No. (%) of participants. Percentages have been rounded and may not total 100.
Spousal loss and cognitive function
Figure 2 presents the β coefficients for the associations between spousal loss and cognitive outcomes in each cohort and in the meta-analysis. The meta-analysis showed that, compared with married individuals, widowhood was associated with poorer memory (multivariable-adjusted β = −0.07, I 2 = 98.27%) and poorer executive function (multivariable-adjusted β = −0.08, I 2 = 98.37%), but not with orientation (multivariable-adjusted β = −0.03, I 2 = 98.05%) or language (multivariable-adjusted β = −0.03, I 2 = 99.30%) (Supplementary Table S3). The association patterns were broadly consistent in five cohorts, although the magnitude of the effect sizes varied. In contrast, data from ELSA showed a divergent pattern, with spousal loss associated with neutral or even better cognitive performance in several domains (orientation: β = 0.10, 95% CI: 0.06–0.13; memory: β = 0.05, 95% CI: 0.02–0.08; language: β = 0.16, 95% CI: 0.12–0.19).
Associations between spousal loss and subsequent domain-specific cognitive function. HRS: US Health and Retirement Study; ELSA: English Longitudinal Study of Ageing; MHAS: Mexican Health and Aging Study; CHARLS: China Health and Retirement Longitudinal Study; LASI: Longitudinal Ageing Study in India; HAALSI: Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa; Overall: Meta-analysis.
Figure 2 Long description
Orientation. The horizontal axis has tick labels negative 0.2, negative 0.1, 0, 0.1, 0.2. A vertical reference line is drawn at 0. Each cohort row shows a point with a horizontal line and a numeric value with parentheses. The vertical axis lists HRS, ELSA, MHAS, CHARLS, LASI, HAALSI, Overall. Values shown: HRS 0.00 left parenthesis negative 0.02, 0.02 right parenthesis; ELSA 0.10 left parenthesis 0.06, 0.13 right parenthesis; MHAS negative 0.08 left parenthesis negative 0.10, negative 0.05 right parenthesis; CHARLS negative 0.08 left parenthesis negative 0.10, negative 0.07 right parenthesis; LASI negative 0.08 left parenthesis negative 0.09, negative 0.07 right parenthesis; HAALSI negative 0.04 left parenthesis negative 0.07, 0.00 right parenthesis; Overall negative 0.03 left parenthesis negative 0.09, 0.03 right parenthesis. The most positive value shown is ELSA 0.10 and the most negative values shown are MHAS negative 0.08, CHARLS negative 0.08 and LASI negative 0.08. The image B showing Memory. The horizontal axis has tick labels negative 0.2, negative 0.1, 0, 0.1, 0.2. A vertical reference line is drawn at 0. Each cohort row shows a point with a horizontal line and a numeric value with parentheses. The vertical axis lists HRS, ELSA, MHAS, CHARLS, LASI, HAALSI, Overall. Values shown: HRS negative 0.06 left parenthesis negative 0.07, negative 0.04 right parenthesis; ELSA 0.05 left parenthesis 0.02, 0.08 right parenthesis; MHAS negative 0.14 left parenthesis negative 0.16, negative 0.12 right parenthesis; CHARLS negative 0.10 left parenthesis negative 0.12, negative 0.09 right parenthesis; LASI negative 0.03 left parenthesis negative 0.04, negative 0.02 right parenthesis; HAALSI negative 0.11 left parenthesis negative 0.15, negative 0.07 right parenthesis; Overall negative 0.07 left parenthesis negative 0.12, negative 0.01 right parenthesis. The most positive value shown is ELSA 0.05. The most negative value shown is MHAS negative 0.14. The image C showing Executive function. The horizontal axis has tick labels negative 0.2, negative 0.1, 0, 0.1, 0.2. A vertical reference line is drawn at 0. Each cohort row shows a point with a horizontal line and a numeric value with parentheses. The vertical axis lists HRS, ELSA, MHAS, CHARLS, LASI, HAALSI, Overall. Values shown: HRS negative 0.09 left parenthesis negative 0.10, negative 0.07 right parenthesis; ELSA 0.01 left parenthesis negative 0.02, 0.03 right parenthesis; MHAS negative 0.07 left parenthesis negative 0.09, negative 0.05 right parenthesis; CHARLS negative 0.14 left parenthesis negative 0.16, negative 0.13 right parenthesis; LASI 0.02 left parenthesis 0.03, 0.01 right parenthesis; HAALSI negative 0.17 left parenthesis negative 0.20, negative 0.13 right parenthesis; Overall negative 0.08 left parenthesis negative 0.13, negative 0.03 right parenthesis. The most positive value shown is LASI 0.02. The most negative value shown is HAALSI negative 0.17. The image D showing Language. The horizontal axis has tick labels negative 0.2, negative 0.1, 0, 0.1, 0.2. A vertical reference line is drawn at 0. Each cohort row shows a point with a horizontal line and a numeric value with parentheses. The vertical axis lists HRS, ELSA, MHAS, CHARLS, LASI, HAALSI, Overall. Values shown: HRS negative 0.05 left parenthesis negative 0.07, negative 0.04 right parenthesis; ELSA 0.16 left parenthesis 0.12, 0.19 right parenthesis; MHAS negative 0.01 left parenthesis negative 0.03, 0.01 right parenthesis; CHARLS negative 0.15 left parenthesis negative 0.16, negative 0.14 right parenthesis; LASI negative 0.02 left parenthesis negative 0.04, negative 0.01 right parenthesis; HAALSI negative 0.11 left parenthesis negative 0.15, negative 0.07 right parenthesis; Overall negative 0.03 left parenthesis negative 0.12, 0.06 right parenthesis. The most positive value shown is ELSA 0.16. The most negative value shown is CHARLS negative 0.15. Across the four images, the same cohort labels appear on the vertical axis in each image and each image uses the same horizontal axis tick labels from negative 0.2 to 0.2 with a vertical reference line at 0.
Figure 3 presents the β coefficients for the associations between spousal loss and cognitive outcomes stratified by sex. Among male participants, widowhood was associated with poorer executive function (multivariable-adjusted β = −0.14, I 2 = 94.31%), but not with orientation (multivariable-adjusted β = 0, I 2 = 95.21%), memory (multivariable-adjusted β = −0.03, I 2 = 93.97%) or language (multivariable-adjusted β = −0.08, I 2 = 98.41%). Among female participants, widowhood was associated with poorer memory (multivariable-adjusted β = −0.07, I 2 = 97.65%), but not with orientation (multivariable-adjusted β = −0.03, I 2 = 97.15%), executive function (multivariable-adjusted β = −0.05, I 2 = 98.12%) or language (multivariable-adjusted β = −0.01, I 2 = 98.38%). Similar to the combined analyses of male and female participants, the ELSA cohort continued to show a divergent pattern in the sex-stratified analyses.
Associations between spousal loss and subsequent domain-specific cognitive function stratified by sex. HRS: US Health and Retirement Study; ELSA: English Longitudinal Study of Ageing; MHAS: Mexican Health and Aging Study; CHARLS: China Health and Retirement Longitudinal Study; LASI: Longitudinal Ageing Study in India; HAALSI: Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa; Overall: Meta-analysis.
Figure 3 Long description
Eight forest plots display beta coefficients for cognitive outcomes related to spousal loss, stratified by sex. The plots are organized into two sections: male and female participants. Each section includes four plots representing orientation, memory, executive function and language. The x-axis is labeled with beta coefficients, ranging approximately from minus 0.3 to 0.3, indicating the strength and direction of associations. The y-axis lists cohorts, including HRS, ELSA, LASI, CHARLS and others, along with an overall estimate. Each plot features horizontal lines with markers showing beta coefficients and confidence intervals for each cohort. The zero line is marked, indicating no effect. For males, executive function shows a notable negative association, while other domains have varied results. For females, memory shows a consistent negative association. Confidence intervals crossing zero suggest non-significant results in some domains. Cohorts are distinguished by different colors and symbols, aiding in comparison across studies. The plots provide insight into how spousal loss impacts cognitive functions differently across sexes and cohorts.
Age-related nonlinearity and interaction analyses
Supplementary Figures S2 and S3 reveal several nonlinear associations between age and cognitive performance, which varied by cohort and widowhood status. Among widowed participants, significant nonlinear age-related patterns were observed for orientation, executive function and language in LASI, as well as for executive function in ELSA. In addition, a borderline nonlinear trend was noted for memory in HAALSI (P = 0.05). However, evidence of effect modification by widowhood status was limited; a statistically significant age × widowhood interaction was detected only for executive function in ELSA (P = 0.04).
Cognitive correlates in widowed individuals
When specifically studying widowed individuals, we found higher educational attainment and internet use were consistently associated with better cognitive outcomes in multiple cohorts (Fig. 4). In middle-income national cohorts such as LASI and MHAS, higher BMI was associated with better performance in executive function and language, with similar trends observed in CHARLS. Certain individual characteristics showed cohort-specific associations with cognitive outcomes among widowed individuals, such as a positive association between household wealth level and cognitive function in HRS and an adverse association between smoking and cognitive performance in HAALSI.
Associations between individual characteristics and domain-specific cognitive function among widowed individuals from six cohorts. HRS: US Health and Retirement Study; ELSA: English Longitudinal Study of Ageing; MHAS: Mexican Health and Aging Study; CHARLS: China Health and Retirement Longitudinal Study; LASI: Longitudinal Ageing Study in India; HAALSI: Health and Aging in Africa: A Longitudinal Study of an INDEPTH Community in South Africa; BMI, body mass index.
Figure 4 Long description
Contains four heatmaps, each representing coefficients for various factors across different cohorts. The row labels include factors such as smoking status, household wealth, educational attainment, drinking patterns and body mass index categories. Column labels represent cohorts including CHARLS, ELSA, HRS, MHAS and LASI. The numerical values within the cells range from negative to positive, indicating the strength and direction of the coefficients. The legend at the bottom shows a gradient from negative 0.5 to positive 0.5, with darker shades indicating higher magnitude coefficients. The first heatmap focuses on smoking status, showing varying coefficients across cohorts, with some positive values indicating a stronger association in certain cohorts. The second heatmap highlights household wealth, where positive coefficients are more prevalent, suggesting a positive association with cognitive outcomes. The third heatmap examines educational attainment, with consistently positive coefficients across cohorts, indicating a strong positive relationship. The fourth heatmap covers body mass index categories, showing mixed coefficients, with some cohorts displaying negative associations. Overall, educational attainment shows consistently positive coefficients across all cohorts, while body mass index categories reveal varied associations. Notable contrasts are observed between cohorts, particularly in smoking status and household wealth factors.
Sensitivity analyses
Associations remained in the same direction, though few reached statistical significance in the complete-case analysis, likely due to the substantially reduced sample size (Supplementary Table S4). The estimates from the sensitivity analyses in which exposure was defined using baseline marital status only (Supplementary Table S5) and those restricted to participants aged 65 years and older in HCAPs (Supplementary Table S6) were largely consistent with the reported results. In sensitivity analyses excluding HRS and ELSA, associations between widowhood and cognitive performance became more consistent. The pooled association for orientation strengthened markedly with minimal heterogeneity (multivariable-adjusted β = −0.08, I 2 = 1.97%), while associations for memory and executive function remained directionally consistent with slightly larger effect sizes (Supplementary Table S7). The association with language became statistically significant after exclusion of the two high-income national cohorts.
Discussion
The present multinational study represents one of the most comprehensive investigations to date examining the association between spousal loss and cognitive changes in diverse regions and socioeconomic contexts. By analysing harmonized data from six prospective cohorts spanning high- and middle-income countries, we found that widowhood was associated with poorer cognitive performance in specific domains, particularly memory and executive function, with clear sex-specific patterns. The substantial cross-national heterogeneity, driven in part by divergent findings in ELSA, highlights the context-dependent nature of the cognitive consequences of spousal loss. In addition, we observed that age-related patterns of cognitive decline differed by widowhood status in a few cohorts, and that certain individual characteristics such as higher education, internet use and BMI may mitigate the cognitive risks associated with spousal loss. Taken together, these findings suggest that the cognitive consequences of spousal loss are not driven by a single mechanism but reflect the interplay of biological vulnerability, psychological adaptation to loss, and social and contextual resources. This integrative framework provides a conceptual basis for developing context-sensitive, biopsychosocial interventions to support cognitive resilience among vulnerable subgroups of widowed older adults, particularly in resource-limited settings.
Our findings add to existing mixed literature by showing that, although spousal loss is associated with cognitive decline in many settings, these effects are domain-specific, sex-dependent and vary across countries. A consistent downward trend in executive function was observed in nearly all regions, with more pronounced among men. Executive function – including working memory, cognitive flexibility and inhibitory control – is critical for attention regulation and decision-making (Guevara et al., Reference Guevara, Gilbert, Murdock, Stowe and Fagundes2019), and appears particularly vulnerable to external stressors (Stroebe and Stroebe, Reference Stroebe and Stroebe1983). Impairment in executive function can lead to difficulties in daily function, which may partly explain the higher rates of functional limitations among widowed individuals compared with their married counterparts (Johnson et al., Reference Johnson, Lui and Yaffe2007; Hughes and Waite, Reference Hughes and Waite2009). Evidence from one preliminary study further shows that bereaved spouses exhibit declines in working memory and cognitive inhibition (Wu-Chung et al., Reference Wu-Chung, Kennedy, Medina, Schulz, Oswald, Heijnen, Leal, Denny and Fagundes2025). The marital resource model (Waite et al., Reference Waite, Gallagher and England2001) may partly explain the observed sex-specific differences, particularly the consistent executive decline among widowed men. Men may rely more on their spouses for emotional regulation and instrumental support in daily life, including health management and household responsibilities (Umberson, Reference Umberson1992; Kalmijn, Reference Kalmijn2007). In contrast, widowed women are more likely to experience memory impairment. Previous evidence indicates that widowed women experience higher levels of grief and depression following spousal death (Chen et al., Reference Chen, Bierhals, Prigerson, Kasl, Mazure and Jacobs1999), which can induce fluctuations in oestrogen levels (Barth et al., Reference Barth, Crestol, De Lange and Galea2023). These hormonal changes may impair memory-related brain regions such as the hippocampus. However, Aartsen et al. (Reference Aartsen, Van Tilburg, Smits, Comijs and Knipscheer2005) found that only widowed men, but not women, had a greater memory decline than their married counterparts. The rate of memory decline over time was similar for widowed men and women. Orientation and language abilities tend to be relatively preserved in normal ageing and are less sensitive to psychosocial stressors such as spousal loss compared with other cognitive domains, reflecting domain-specific patterns in cognitive ageing (Rut et al., Reference Rut, Jose and Antonieta2018).
Consistent with previous studies (Feng et al., Reference Feng, Ng, Yap, Li, Lee, Håkansson, Kua and Ng2014; Liu et al., Reference Liu, Zhang, Choi and Langa2020), widowhood is associated with sex differences in cognitive decline. In the traditional marital patterns, women typically serve as family kin-keepers, providing emotional support to their husbands; therefore, losing a spouse tends to have a greater negative impact on men’s social networks and support systems (Williams and Umberson, Reference Williams and Umberson2004), which are closely linked to cognitive impairment. Also, men tend to exhibit higher pre-bereavement engagement in harmful behaviours like excessive drinking. U.S. CDC data indicate that 25.6% of older men (vs. 12.4% of women) consume ≥4 drinks daily, a known risk factor for accelerated cognitive deterioration (Robin and Michael, Reference Robin and Michael2022). In countries with limited formal support systems and enduring gender inequalities – such as China, India and South Africa – spousal loss may trigger more disruptions in social identity, access to care and economic security. For instance, widowed women in India are often subjected to restrictive norms, including wearing white and dietary limitations, that deepen economic dependency and social exclusion (Singh, Reference Singh2022). In China, spousal loss is significantly associated with loneliness among older adults (Yang and Gu, Reference Yang and Gu2021), supporting that Chinese widows showed severe language decline because of reduced linguistic stimulation. Countries with better welfare systems may offer buffering effects. In England, public healthcare and social support services may facilitate adaptation to bereavement through improved access to medical care, community programmes or volunteer opportunities (Alderwick et al., Reference Alderwick, Hutchings and Mays2022). However, given the elevated mortality risk linked to widowhood (Wang et al., Reference Wang, Wu, Hsu, Lin, Fan, Chen, Hsu and Wu2024), the widowed population in England may reflect ‘healthy survivors’ (Marshall and Donald R, Reference Marshall and Donald1996), potentially leading to an underestimation of widowhood’s adverse impact on cognitive decline. In Mexico, policies like non-contributory pensions have been shown to improve older adults’ well-being and may indirectly preserve cognitive health by alleviating financial strain (Aguila and Casanova, Reference Aguila and Casanova2020). Moreover, multigenerational living arrangements (Aguila et al., Reference Aguila, Park and Vega2020), prevalent in Mexico, may offer additional protection against cognitive decline through increased social interaction. Future cross-national research is warranted to elucidate how cultural contexts and macro policies jointly shape the cognitive consequences of spousal loss.
Our findings do not fully support the hypothesis that the cognitive impact of widowhood intensifies nonlinearly with advancing age (Shin et al., Reference Shin, Kim and Park2018; Liu et al., Reference Liu, Zhang, Burgard and Needham2019). While we observed some nonlinear age-related cognitive trajectories, these patterns were not systematically modified by widowhood status. Instead, the largely null age × widowhood interaction effects suggest that widowhood may act as an independent risk factor for cognitive decline, rather than amplifying age-related deterioration. The only exception was in ELSA, where a significant interaction indicated that the effect of widowhood on executive function varied across the age distribution. Overall, these results suggest that the intersection of age and widowhood is not universally associated with cognitive ageing across cohorts. However, the lack of consistent interaction effects should be interpreted cautiously due to limited sample sizes in some cohorts. Future research should further explore how contextual factors shape the age-dependent consequences of widowhood across different sociocultural environments.
Within the framework of established modifiable risk factors for dementia (Livingston et al., Reference Livingston, Huntley, Liu, Costafreda, Selbæk, Alladi, Ames, Banerjee, Burns, Brayne, Fox, Ferri, Gitlin, Howard, Kales, Kivimäki, Larson, Nakasujja, Rockwood, Samus, Shirai, Singh-Manoux, Schneider, Walsh, Yao, Sommerlad and Mukadam2024), our findings offer additional insights into how individual characteristics relate to cognitive function among widowed older adults. Notably, educational attainment remained a robust protective factor for cognitive function (Zhang et al., Reference Zhang, O’shea, Yu, Cho, Zhang, Kler, Langa, Weir, Gross and Kobayashi2024). The current findings suggest that internet use can decrease the likelihood of cognitive decline. Internet access or smartphone use, a low-cost way, can effectively promote social connections and can serve as a health management platform (Lu et al., Reference Lu, Yao and Jin2022; James et al., Reference James, Basu, Nallamothu and Kullgren2025). Moreover, access to cognitive training apps and virtual reality may further reduce cognitive decline by providing targeted neuroplasticity exercises and immersive cognitive stimulation. A meta-analysis of 23 RCTs (N = 894 stroke patients) found that VR-based therapies significantly improved executive function, memory and visuospatial skills in patients with stroke (Zhang et al., Reference Zhang, Fu, Lu, Zhang, Huang, Yang, Zhang and Li2021). In older populations and resource-limited settings, higher BMI may be a factor that can reduce the risk of cognitive decline among participants who lose a spouse (Lv et al., Reference Lv, Mao, Gao, Ji, Kraus, Yin, Yuan, Chen, Luo, Zhou, Li, Duan, Zhu, Zeng, Wang, Wang and Shi2022). However, this interpretation requires caution, as reverse causation or survival bias may play a role. Further observational studies or clinical trials are needed to confirm our findings, as well as to develop tailored interventions to improve the widowed population’s physical and mental health.
Strengths of this study include an extensive sample derived from nationally representative ageing studies in high- and middle-income countries. This broad representation enhances the generalizability of the cross-cultural findings, facilitating meaningful comparisons between countries at varying levels of economic development. To better assess the cognitive performance of the elderly in diverse cultural backgrounds, we adopted standardized HCAP scores to assess cognitive function. Comprehensive statistical approaches – including generalized linear models, random-effects meta-analysis and restricted cubic splines regressions – allowed us to capture both average effects and nonlinear trends in cognitive changes. Moreover, analyses restricted to widowed individuals offered insights into how individual factors may shape the effects of widowhood. Additionally, we conducted multiple imputations for the missing values and, after performing a series of sensitivity analyses, our results remained robust.
Nevertheless, certain limitations should be noted. First, given the significant increase in mortality risk following spousal loss (Wang et al., Reference Wang, Wu, Hsu, Lin, Fan, Chen, Hsu and Wu2024), this may introduce selection bias, whereby only healthier individuals who survive are included in the study. Relatedly, reverse causation cannot be fully excluded, as pre-clinical cognitive decline may influence survival, participation or marital trajectories. Second, we failed to identify the time-varying effects of being widowed (Hanes and Clouston, Reference Hanes and Clouston2021) and precise timing of spousal loss (Chen et al., Reference Chen, Wu, Wang, Zeng, Huang, Lv, Niu, Meng, Cai, Shen, Gang, You, Lv, Ren, Shi and Ji2022). Additionally, several factors relevant to widowhood research – including pre-loss marital quality (Min and Song, Reference Min and Song2023), caregiver status (Pinquart and Sörensen, Reference Pinquart and Sörensen2003), perceived social support and cause of spousal death (Wu-Chung et al., Reference Wu-Chung, Kennedy, Medina, Schulz, Oswald, Heijnen, Leal, Denny and Fagundes2025) – were not consistently measured in six cohorts and could not be incorporated into the harmonized analyses; future studies with richer psychosocial and environmental data are needed. Third, the exposure definition did not encompass other relationship statuses such as divorce, remarriage or cohabitation, which may limit generalizability beyond these groups. In addition, the lack of harmonized pre-baseline marital histories may have led to exposure misclassification and underestimation of the observed associations. Future studies with more comprehensive relationship histories are needed to address this limitation. Fourth, the minimum age of HCAP ranged from 50 to 65 years, which might have led to the observed differences in associations, as age is a key determinant of cognitive performance. Finally, considering the cross-cultural nature of this study, we endeavoured to define the variables in accordance with common protocols. However, significant heterogeneity was still observed in the pooled results, though our parallel analysis approach mitigated this concern.
Conclusions
Drawing on harmonized data from six longitudinal cohorts, this multinational study demonstrates that widowhood-associated cognitive changes exhibit domain-specific and sex-dependent patterns in diverse regions and socioeconomic contexts, with executive function and memory potentially emerging as the most consistently affected cognitive domains. Our findings illustrate how cognitive outcomes among widowed individuals are associated with individual and contextual factors, such as education, internet use and BMI. These results highlight the importance of identifying protective mechanisms and tailoring public health strategies to vulnerable subgroups. Future studies should integrate biological and psychosocial measures to clarify mechanisms and inform targeted, context-sensitive interventions supporting cognitive resilience among widowed older adults, particularly in resource-limited settings.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S2045796026101048.
Availability of data and materials
Original data from the HRS, ELSA, MHAS, CHARLS, LASI and HAALSI are publicly available. The data can also be obtained on request (fenfeng@hst.aau.dk).
Acknowledgements
We gratefully thank the administrations and data collection teams of the HRS, ELSA, MHAS, CHARLS, LASI and HAALSI for providing publicly accessible datasets.
Author contributions
Conceptualization: C.G., Y.W. and F.G. Data curation: C.G. Formal analysis: C.G. with the help of Y.W. Funding acquisition: X.S. Supervision: X.S. and F.G. Visualization: Y.W. and C.G. Writing – original draft: C.G. Writing – review and editing: F.G. and Y.W.
Financial support
This study was supported by the National Natural Science Foundation of China (grant number 82171524).
Competing interests
The authors have no conflicts of interest to declare.
Ethical standards
HRS was approved by the Health Sciences and Behavioral Sciences Institutional Review Board (IRB-HSBS) at the University of Michigan. ELSA was approved by NHS Research Ethics Committees under the National Research and Ethics Service (NRES). MHAS was approved by the Institutional Review Boards of the University of Texas Medical Branch (UTMB) in the United States, the Instituto Nacional de Estadística y Geografía (INEGI) in Mexico, and the Instituto Nacional de Salud Pública (INSP) in Mexico. CHARLS was approved by the Institutional Review Board at Peking University. LASI obtained ethical approvals from the following collaborating organizations: Indian Council of Medical Research (ICMR), Delhi; IRB, International Institute for Population Sciences (IIPS), Mumbai; IRB, Harvard T.H. Chan School of Public Health (HSPH), Boston; IRB, University of Southern California (USC), Los Angeles; IRB, ICMR National AIDS Research Institute (NARI), Pune; and IRB, Regional Geriatric Centres (RGCs), MoHFW. HAALSI was approved by the University of the Witwatersrand Human Research Ethics Committee, the Harvard T.H. Chan School of Public Health, Office of Human Research Administration and the Mpumalanga Provincial Research and Ethics Committee. All participants provided written informed consent. The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2013.
Open access
