Marital status has been linked with cognitive aging through proposed mechanisms such as sustained social engagement, emotional support, and shared cognitive activities. Prior literature reports inconsistent associations: some large studies report higher cognitive impairment among never-married older adults, while others find lower risk after adjustment for socioeconomic and health factors. A potential source of inconsistency is the cognitive assessment modality—performance-based tests versus informant-capable questionnaires—which capture different constructs and may introduce differential reporting bias. This study examined whether associations between baseline marital status and incident screen-positive cognitive impairment (hereafter referred to as MCI) differ across two screening instruments within the same Taiwanese cohort and whether effects vary by population subgroups.
This prospective cohort analysis used the Taipei City Elderly Health Examination Database covering 2005–2016 and included adults aged ≥65 years attending government-sponsored annual health examinations. The analytic approach compared two non-overlapping assessment periods defined by the cognitive screening instrument used: SPMSQ during 2005–2011 and AD8 during 2012–2016. Separate Cox proportional hazards models with time-updated covariates were fit for each period, adjusting for demographic, socioeconomic, health, and behavioral factors.
From 184,403 individuals initially identified, exclusions for single-examination records, baseline dementia medication use, and incomplete questionnaires produced a final sample of 103,863 participants: 61,304 in the SPMSQ period and 42,559 in the AD8 period. Mean follow-up was 3.7 years for SPMSQ and 2.8 years for AD8. Marital status at baseline was self-reported and categorized as married (reference), divorced, widowed, or never-married and treated as fixed in the primary analyses. The study was approved by the Taipei City Hospital IRB with data anonymized prior to analysis.
Screen-positive cognitive impairment (MCI) was defined using validated Taiwanese traditional Chinese versions of the two screening tools. The Short Portable Mental Status Questionnaire (SPMSQ) is a 10-item performance-based test scored by number of incorrect responses; scores ≥3 were classified as MCI. The 8-item Ascertain Dementia (AD8) asks about change in cognitive/functional abilities with binary items and uses a cutoff of ≥2 for MCI. The AD8 can be completed by an informant or by self-report; the database did not record respondent type, preventing stratification by informant.
Baseline covariates included age, sex, education (<6, 7–9, ≥10 years), low-income household status, and living arrangement (alone vs with others). Time-dependent covariates updated annually included BMI category, medication-based chronic disease indicators (hypertension, diabetes, hyperlipidemia, heart disease, stroke), psychiatric medication history, psychological distress measured by BSRS-5 (cutoff ≥6), current smoking, and alcohol consumption. Low-income status was defined using administrative criteria linked to household registration.
Cox proportional hazards regression with time-dependent covariates estimated hazard ratios (HRs) and 95% confidence intervals (CIs) for incident MCI with married participants as the reference. Analyses were conducted separately for the SPMSQ (2005–2011) and AD8 (2012–2016) periods. Model 2 adjusted for baseline demographic and socioeconomic factors and the time-updated health and behavioral covariates. Participants who did not screen positive were censored at last attendance; death was treated as censoring. Complete-case analyses were conducted. Subgroup analyses examined effect modification by sex, age group (65–79 vs ≥80), and education (<10 vs ≥10 years). Sensitivity analyses included stratification by median follow-up duration and exclusion of MCI cases identified within the first 2 years to address reverse causality. In the AD8 period, marital status was also tested as a time-varying covariate as a robustness check.
Approximately 75% of participants were married in both periods. Widowed participants were the oldest group (mean ages 75.9–77.9 years), predominantly female (76.6–79.9%), and had higher prevalence of hypertension and heart disease but lower smoking and alcohol consumption. Never-married participants exhibited the highest low-income prevalence (16.0% decreasing to 7.4%) and had higher smoking prevalence relative to other groups. Divorced participants had higher rates of living alone, psychological distress, and psychotropic medication use.
Findings differed by assessment period. During the SPMSQ period (2005–2011; n = 61,304), never-married status was associated with an increased hazard of incident screen-positive MCI compared with married participants (HR 1.37, 95% CI 1.08–1.74). In contrast, during the AD8 period (2012–2016; n = 42,559), never-married status was associated with a lower hazard of incident MCI versus married (HR 0.60, 95% CI 0.46–0.77). Effect modification by sex and age was observed across the two periods; detailed subgroup estimates are reported in the original tables and figures.
Sensitivity analyses—stratifying by follow-up duration, excluding early incident cases within two years to minimize reverse causality, and modeling marital status as time-varying in the AD8 period—yielded materially consistent patterns with the primary results. Interaction tests for effect modification were adjusted for multiple comparisons using the Benjamini–Hochberg false discovery rate.
The opposing associations for never-married participants across the two measurement periods suggest that measurement heterogeneity—differences between performance-based vs. informant-capable screening, variation in follow-up duration, and changes in population composition—may account for discrepant findings. Informant-based tools like the AD8 may be subject to differential detection depending on co-residing informants; married individuals with a cohabiting spouse may have subtle declines detected earlier, potentially raising apparent risk in married groups when using informant-report measures. Conversely, performance-based instruments like the SPMSQ assess current cognitive capacity and in this cohort were associated with higher MCI risk among never-married participants. The authors caution that these differences should not be interpreted as definitive temporal changes in risk without accounting for measurement and sampling differences.
Strengths include a very large community-based sample, longitudinal design, standardized annual data collection, use of validated Chinese versions of two common screening instruments, and time-updated covariate adjustment. Limitations explicitly reported are reliance on screening tools rather than clinical diagnoses, inability to record who completed the AD8 (informant vs self), potential reverse causality given baseline marital status ascertainment, complete-case analysis (exclusions for incomplete data), and incomplete linked mortality data after 2012 preventing competing-risks modeling.
In this large Taiwanese cohort aged ≥65 years, associations between baseline marital status and incident screen-positive cognitive impairment varied by the screening instrument and assessment period. The contrasting associations observed for never-married participants likely reflect measurement heterogeneity and study-period differences rather than a clear temporal change in marital status effects. The authors recommend that interpretations of marital status–cognition relationships account for the cognitive assessment modality and study context when informing research and policy.