Subjective cognitive impairment (SCI) and future probable dementia risk in the National Health and Aging Trends Study (NHATS) – 2012-2019

Abstract

Background:

Subjective cognitive impairment (SCI) is routinely assessed in national surveys, offering potential for dementia monitoring and early detection. The association between SCI and dementia has not been evaluated using population-based data and longitudinal methods. Estimating the causal effect of SCI on dementia risk requires clear research questions and appropriate statistical methods to account for competing risks.

Objectives:

To evaluate whether SCI causally affects dementia risk in U.S. older adults (≥65 years), considering the competing risk of all-cause mortality.

Methods:

We analyzed data from the National Health and Aging Trends Study (NHATS) for participants who completed the U.S. Census Bureau’s disability questionnaire in 2012 and were dementia-free at baseline (2011-2012). We estimated total, direct, and separable effects of SCI on dementia and mortality risks.

Results:

Among 1,622 dementia-free older adults, 7.6% reported SCI at baseline. SCI was associated with nearly a twofold increased risk of dementia over 8 years (risk ratio [RR]: 1.95, 95% confidence interval [CI]: 1.07,3.07) and a lower risk of mortality (RR: 0.31, CI: 0.10,0.67) while free of dementia. Direct effect analysis indicated a potential causal link between SCI and dementia (RR: 1.98, CI: 1.02,3.37), with the effect being mainly direct rather than mediated through mortality.

Conclusions:

SCI predicted dementia onset and was inversely related to mortality in older adults over 8 years. These findings highlight the complex relationship between cognitive decline and mortality, emphasizing the need for early detection and intervention in SCI, and precise analytic methods and questions to accurately interpret aging research.

INTRODUCTION

Subjective cognitive impairment (SCI) is recognized as a potential early indicator of preclinical Alzheimer’s Disease (AD) and other related dementias (ADRD) (Reisberg et al., 2010; Ritchie & Touchon, 2000) and is a core component of the diagnostic criteria for mild cognitive impairment (MCI), one of the early manifestations of dementia (Jessen, Wolfsgruber, et al., 2014; Schwarz et al., 2021). SCI refers to self-perceived impaired cognitive function and performance, and is conceptually independent of normal aging, performance on a cognitive test, or clinical diagnosis (Jessen, Amariglio, et al., 2014). The Medicare Annual Wellness Visit (AWV), introduced in 2011, involves completion of a health risk assessment with self-reported concerns similar to SCI, followed by direct observation, potentially incorporating SCI into screening (Jacobson et al., 2020). Additionally, the updated AD research framework of the National Institute on Aging and Alzheimer’s Association (NIA-AA) now recognizes SCI within the cognitively unimpaired stage of the cognitive continuum, implying that SCI may even be a manifestation and stage of dementia (Jack et al., 2024; Petersen et al., 2021). Thus, there is a growing interest in understanding whether SCI is causally associated with dementia.

A recent cross-sectional analysis of the NHATS data found a strong association between SCI and dementia (Chyr et al., 2024). A Framingham Heart Study analysis also linked SCI to a higher dementia risk (Kang et al., 2024). Other studies using similar methodologies and in various settings and populations, reached similar conclusions (Chapman et al., 2023). However, these studies are limited by their cross-sectional designs, or reliance on hazard rates or ratios, and/or failure to account for competing risks, making it difficult to establish a clear causal relationship between SCI and dementia. Structural changes and disrupted functional connectivity in the rostral middle frontal gyrus (RMFG) may contribute to increased dementia risk in SCI (Kang et al., 2024). Hyperperfusion in this area, along with asymmetries in the hippocampus and amygdala, as well as atrophy and glucose hypometabolism in AD-signature regions, are linked to cognitive decline in SCI (Jessen, Amariglio, et al., 2014). Increased cerebral beta-amyloid deposition further supports the role of RMFG alterations and amyloid accumulation as key SCI pathways contributing to dementia risk (Pike et al., 2022; Turner et al., 2023). At the same time, SCI indirectly affect dementia risk through its effect on mortality before dementia onset. SCI is associated with all-cause mortality, independent of baseline function, with contributing factors including poor adherence to medical care, limited mobility, falls, social isolation, and difficulty managing daily tasks. Shared risk factors with cardiovascular disease also suggest a complex relationship between SCI and mortality, which remains, however, inadequately understood and warrants further study (Roehr et al., 2016).

Understanding whether SCI increases dementia risk (in a causal sense) could enhance efforts to identify early persons who may benefit from early interventions. While no pharmacological treatments for SCI exist, lifestyle changes, such as increased physical activity, have shown protective effects against progression to dementia (D’Elia et al., 2024; Iso-Markku et al., 2024). Surveillance could efficiently target those with SCI and at high risk of dementia, ensuring timely access to diagnostic technologies and care (Palmqvist et al., 2024). Precise risk assessments can also inform public health strategies, and guide resource allocation, awareness campaigns, and preventive measures. Additionally, improved risk information can shape research priorities and advance our understanding of dementia mechanisms – including those by which SCI may cause dementia – and the effectiveness of interventions.

Yet, establishing the causal effect of SCI on dementia risk is challenging due to competing risks like mortality, which prevent the primary outcome—dementia—from occurring. This issue is particularly relevant in studies of older adults, where comorbidities and mortality are common. Traditional time-to-event analyses often censor participants at study end without distinguishing between independent censoring and censoring due to events like death (Schober & Vetter, 2018), leading to biased estimates. Participants who die before study completion are removed from the risk set, violating the assumption of noninformative censoring and complicating interpretation (Andersen et al., 2012; Efron, 1967; Tsiatis, 1975). Failing to account for competing risks such as mortality can distort estimates and obscure the true relationship between SCI and dementia risk. Historically, mortality has been considered in survival analysis a “semicompeting event,” distinct from competing events (Andersen et al., 2002; Fine et al., 2001), but recent arguments suggest this distinction may be unnecessary when addressing causal questions (Rojas-Saunero et al., 2023).

With that backdrop, a robust literature has emerged that offers statistical estimands in the presence of competing events. But until recently, the vast majority of this literature described these estimands without the use of a formal causal framework, thus making the interpretation of the estimated effects difficult and, at times, confusing (Austin et al., 2016). In recent work, Young et al. placed the historical estimands from the survival analysis literature in competing events settings within a formal counterfactual framework for causal inference (Young et al., 2020). In subsequent studies, they formalized the interpretation, conditions for identification in real-world studies, and some corresponding statistical methods for estimation (Rojas-Saunero et al., 2023; Stensrud et al., 2021; Stensrud et al., 2022). This more recent literature introduced several estimands and corresponding estimators that may be of interest for estimating causal effects in the presence of competing events, depending on the particular research question: total, direct, and separable effects.

In this study, we use a one-third random sample of Medicare beneficiaries (≥65 years) from the 2011-2019 NHATS dataset and apply novel estimands to assess the potential causal impact of SCI on dementia risk, with mortality as a competing event, and avoiding mathematical equations and proofs already covered in recent theoretical papers by Young et al. and Stensrud et al. (Stensrud et al., 2021; Young et al., 2020).

METHODS

Study Design, Population, and Data Source

We analyzed data from Rounds 1-9 (2011-2019) of the National Health and Aging Trends Study (NHATS), a nationally representative survey of Medicare beneficiaries aged 65 and older (Freedman & Kasper, 2019; Kasper & Freedman, 2014). We focused on 1,622 participants without dementia in 2011-2012 who completed the 2012 Disability Questionnaire (Kasper & Freedman, 2020). The cohort was followed annually until the first occurrence of dementia, death, loss to follow-up, or the study’s end in 2019.

Primary Outcomes – Events

During follow-up, participants could experience one of four mutually exclusive events: 1) develop dementia, 2) die from any cause, 3) be lost to follow-up, or 4) remain under observation without dementia or death (administrative censoring).

Dementia Onset

The primary event was dementia onset, determined using the NHATS probable dementia algorithm, which classifies participants as having probable, possible, or no dementia based on a validated composite measure (Kasper et al., 2013). Participants (or proxies) reported physician diagnoses, and proxies completed the AD8 when needed (Kasper et al., 2013). Participants also underwent cognitive tests assessing orientation, memory, and executive function (Crimmins et al., 2011; Kasper et al., 2013; Langa et al., 2005; Schretlen et al., 2010). Dementia onset was defined based on the first round after Round 2 when participants were categorized as having probable dementia. Time to dementia onset was calculated from Round 2 to the first dementia diagnosis.

All-cause Mortality, Loss to Follow-up, and Administrative Censoring

Death was defined as a binary indicator for all-cause mortality, recorded in the NHATS Sample Person Sensitive Demographic files. The round of death was determined from the month and year of death. Since assessments occurred annually, a participant could develop dementia and then die or be lost to follow-up within the same interval. Thus, we used the time between NHATS interviews and the date of death to determine event sequences. Participants who did not develop dementia or die were considered censored due to loss to follow-up or administrative censoring.

Exposure: Subjective Cognitive Impairment (SCI)

SCI status was defined from responses to the cognitive disability item of the Disability Questionnaire, administered to a random one-third sample of NHATS participants in Round 2 (2012). This item (version administered in the American Community Survey) asks participants the following: “Because of a physical, mental, or emotional problem, does this person have difficulty remembering, concentrating, or making decisions?” (Adams et al., 2020; Taylor et al., 2018) Respondents who affirmatively answered this question were categorized as having SCI; those who answered “No” were classified as not having SCI; all other respondents (e.g., those with missing answers, or who answered “Don’t know/not sure”) were classified as missing responses and excluded from the sample.

Other Covariates

Other covariates (measured in Round 2) included known risk factors for dementia (i.e., age, biological sex, race and ethnicity, education) (Alzheimer’s Disease Association, 2023), as well as other socioeconomic, sociodemographic, and health characteristics of the population: living arrangement, enrollment in Medicaid, household income as a percentage of the federal poverty guideline (FPL), number of impairments with activities of daily living (ADL) and instrumental activities of daily living (IADL), body-mass-index (BMI), comorbid conditions, mental health symptoms, residential setting (community vs non-nursing home residential care setting), respondent type (self vs proxy), and death status after Round 2 (Table 1).

Table 1.

Characteristics of NHATS Respondents who Participated to the U.S. Census Bureau’s Disability Questionnaire in 2012, Stratified by SCI Status.

		Subjective Cognitive Impairment (SCI)
Characteristic	Overall, N = 1,622*	No, N = 1,507*	Yes, N = 115*
Age (y), Mean (SD)	78.04 (7.42)	77.94 (7.39)	79.33 (7.62)
Age groups (y), %
65-74	40.5	41.1	32.2
75-84	39.8	39.5	42.6
≥85	19.7	19.3	25.2
Female, %	58.1	57.6	64.3
Race and ethnicity, %
White, non-Hispanic	72.0	72.4	67.0
Black, non-Hispanic	19.5	19.2	24.3
Hispanic	5.3	5.1	7.8
Other, non-Hispanic	3.1	3.3	0.9
Educational level, %
Less than high school	21.3	20.2	35.7
High school graduate	34.7	35.0	31.3
Some college	18.9	19.0	16.5
College graduate and beyond	25.1	25.7	16.5
Marital status, %	4.8	4.9	3.5
Living arrangement, %
Alone	31.3	30.5	41.7
With spouse or partner	50.2	51.0	40.0
With others	18.5	18.5	18.3
Income as % of FPL, %
<100% FPL	21.9	21.0	33.9
100-199% FPL	24.5	24.2	28.7
200% FPL and above	53.6	54.8	37.4
Medicaid enrollment, %	14.2	13.5	24.3
ADL impairments, %	21.8	19.8	47.8
IADL impairments, %	65.7	64.0	87.0
Body-mass-index, Mean (SD)	27.66 (5.69)	27.62 (5.62)	28.15 (6.49)
Stroke, %	11.2	10.8	15.7
Heart Disease, %	28.1	27.1	41.7
Diabetes, %	25.2	24.7	32.2
Hypertension, %	68.4	67.9	74.8
Depression, %	34.6	31.6	74.8
Anxiety, %	29.8	27.1	66.1
Proxy respondent, %	0.9	0.8	1.7
In residential setting ‡ , %	4.8	4.4	10.4

Abbreviations: NHATS, National health and aging trends study; SCI, subjective cognitive impairment; FPL, federal poverty line; ADL, activities of daily living; IADL, instrumental activities of daily living.

Notes: Participants with dementia (as assessed by the NHATS probable dementia definition) or residing in a nursing home facility in Round 2 (2012) or earlier were excluded. All estimates are unweighted, unless noted otherwise.

^*Mean (SD), or frequency (%)

^‡Non-nursing home facility.

Statistical Analyses

We first compared characteristics of NHATS participants assessed for SCI in 2012 using Pearson’s chi-square for categorical variables and t-tests for continuous variables. Next, we estimated the total, direct, and separable effects of SCI on probable dementia and all-cause mortality risks, following recent methodological advances (Rojas-Saunero et al., 2023; Stensrud et al., 2021; Stensrud et al., 2022; Young et al., 2020). We used cumulative incidences rather than hazards due to challenges in interpreting hazard ratios as causal effects (Hernán, 2010; Young et al., 2020). SCI was treated as a time-fixed exposure, as it was measured only once (Round 2) for each participant.

Since participants with SCI at baseline might have differed from those without SCI, the exposure could not be considered random, leading to potential confounding. Thus, inverse-probability-of-treatment weights (IPTW) were used to adjust for confounding (Hernán & Robins, 2023), and inverse-probability-of-censoring weights (IPCW) accounted for bias due to loss to follow-up and removed mortality under direct effects (Howe et al., 2016). These weights create a pseudo-population, simulating random exposure, loss to follow-up, and mortality (Hernán & Robins, 2023; Howe et al., 2016). All weights were stabilized for efficiency and truncated at the 1st and 99th percentiles to reduce variance (Cole & Hernán, 2008; Robins et al., 2000; Xiao et al., 2013). Estimands are defined in Table 2.

Table 2.

Research Questions and Corresponding Estimands for Different Counterfactual Risk Contrasts for Dementia and All-cause Mortality, As Well As Their Interpretation And Identifying Assumptions, When All-cause Mortality Acts as a Competing Risk for Dementia.

Research Question*	Estimand†	Interpretation†	Expression for Risk Difference‡	Identifying Assumptions‡	Weights for Confounder Adjustments‡	Terminology in Literature**
RQ1: What would be the average difference in dementia risk, by some time t of follow-up, had no study participant experienced SCI versus had they all experienced SCI at baseline?	Total effect of SCI on dementia risk	Average causal effect of SCI on dementia risk: Captures all pathways by which SCI affects dementia, which may include both direct and indirect exposure effects outside and through the exposure’s effect on mortality	$\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a}=1}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a}=0}=1\right]$	Exchangeability, positivity, or consistency not needed for competing event (mortality)	IPTW and IPCW for loss to follow-up	Subdistribution function, cause-specific cumulative incidence, crude risk
RQ2: What would be the average difference in all-cause mortality risk before the onset of dementia, by some time t of follow-up, had no study participant experienced SCI versus had they all experienced SCI at baseline?	Total effect of SCI on mortality risk before dementia onset	Average causal effect of SCI on all-cause mortality risk before dementia onset: Captures all pathways by which SCI affects mortality before dementia onset	$\mathrm{P}\mathrm{r}\left[{\mathrm{D}}_{\mathrm{k}+1}^{\mathrm{a}=1}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{D}}_{\mathrm{k}+1}^{\mathrm{a}=0}=1\right]$	Exchangeability, positivity, or consistency not needed for competing event (mortality)	IPTW and IPCW for loss to follow-up	Subdistribution function, cause-specific cumulative incidence, crude risk, expected loss in life due to a cause of death
RQ3: What would be the difference in dementia risk, by time t of follow-up, had no study participant experienced SCI at baseline compared to had they all experienced SCI at baseline if there was no competing risk of death (i.e., under a scenario in which an intervention existed to prevent the competing event) throughout the study period?	Controlled direct effect of SCI on dementia risk	Effect of SCI on dementia risk, under complete elimination of mortality as a competing event: Direct effect on dementia not via death, referring to scenario where death has (somehow) been eliminated	$\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a}=1,\stackrel{-}{\mathrm{d}}=\stackrel{-}{0}}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a}=0,\stackrel{-}{\mathrm{d}}=\stackrel{-}{0}}=1\right]$	Exchangeability, positivity, and consistency	IPTW and IPCW for both loss to follow-up and mortality	Marginal cumulative incidence, net risk, net probability
RQ4: What would be the average difference in dementia risk, by some time t of follow-up, had every study participant experienced exposure to both the SCI dementia and mortality versus had they all been unexposed to both pathways at baseline?	Separable total effect of SCI on dementia Risk	Average effect of SCI on dementia risk under a setting of participants exposed to both SCI dementia pathway and mortality pathway as compared to being unexposed to both dementia and mortality pathways	$\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=1,{\mathrm{a}}_{\mathrm{D}}=1}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=0,{\mathrm{a}}_{\mathrm{D}}=0}=1\right]$	Generalized decomposition, exchangeability, positivity, consistency, dismissible components, and strong positivity	IPTW and IPCW for both loss to follow-up and mortality	***
RQ5: What would be the difference in dementia risk, by some time t of follow-up, had every study participant experienced exposure to the SCI dementia pathway compared to had they all been unexposed to the SCI dementia pathway at baseline, while controlling for the SCI mortality pathway by holding exposure status constant (e.g., ${\mathrm{A}}_{\mathrm{D}}={\mathrm{a}}_{\mathrm{D}}=1$)?	Separable direct effect of SCI on dementia risk	Direct effect of SCI on dementia risk, holding the exposure to mortality pathway constant	$\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=1,{\mathrm{a}}_{\mathrm{D}}=1}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=0,{\mathrm{a}}_{\mathrm{D}}=1}=1\right]$	Generalized decomposition exchangeability, positivity, consistency, dismissible components, and strong positivity	IPTW and IPCW for both loss to follow-up and mortality	***
RQ6: What would be the difference in dementia risk, by some time t of follow-up, had all study participants experienced exposure to the SCI mortality pathway versus had they all been unexposed to the SCI mortality pathway at baseline, while holding exposure status to the SCI dementia risk pathway constant for all participants (e.g., ${\mathrm{A}}_{\mathrm{Y}}={\mathrm{a}}_{\mathrm{Y}}=0$)?	Separable indirect effect of SCI on dementia risk	Indirect effect of SCI on dementia risk through mortality pathway while setting exposure status to dementia risk pathway constant	$\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=0,{\mathrm{a}}_{\mathrm{D}}=1}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=0,{\mathrm{a}}_{\mathrm{D}}=0}=1\right]$	Generalized decomposition exchangeability, positivity, consistency, dismissible components, and strong positivity	IPTW and IPCW for both loss to follow-up and mortality	***

Abbreviations: SCI, subjective cognitive impairment; RQ, research question; IPCW, inverse-probability-of-censoring weights; IPTW, inverse-probability-of-treatment weights.

Notes: ${\mathrm{Y}}_{\mathrm{k}}$ and ${\mathrm{D}}_{\mathrm{k}}$ denote, respectively, the primary outcome of interest (dementia onset), and the competing event (mortality), a terminal state which prevents dementia from onsetting by time $\mathrm{k}$ of follow-up. $\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}}=1\right]$ denotes the probability of outcome ${\mathrm{Y}}_{\mathrm{k}}$ being observed by follow-up time $\mathrm{k}$. ${\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a}}$ and ${\mathrm{D}}_{\mathrm{k}+1}^{\mathrm{a}}$ denote, respectively, the counterfactual dementia onset and counterfactual all-cause mortality outcomes by time $\mathrm{k}+1$ under exposure level $\mathrm{a}$. ${\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a},\stackrel{-}{\mathrm{d}}=\stackrel{-}{0}}$ denotes the counterfactual dementia onset by time $\mathrm{k}+1$ under exposure level $\mathrm{a}$ and under an intervention that eliminates the competing event, such that $\stackrel{-}{\mathrm{d}}=\stackrel{-}{0}$. ${\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}},{\mathrm{a}}_{\mathrm{D}}}$ denotes the counterfactual dementia onset by time $\mathrm{k}+1$ under level ${\mathrm{a}}_{\mathrm{Y}}$ of the exposure pathway ${\mathrm{A}}_{\mathrm{Y}}$ and level ${\mathrm{a}}_{\mathrm{D}}$ of the exposure pathway ${\mathrm{A}}_{\mathrm{D}}$. For details on notation, see Supplementary Materials.

^*Adapted from Table 2 in Rojas-Saunero et al. (Rojas-Saunero et al., 2023).

^†Based on Table 1 of Young et al. and on equations in Stensrud et al. (Stensrud et al., 2021; Young et al., 2020) Young et al. noted that similar descriptions of risk were also reported in earlier work on competing events, albeit without definition of counterfactuals (Chiang, 1961; Young et al., 2020).

^‡Based on equations in Young et al. and Stensrud et al. (Stensrud et al., 2021; Young et al., 2020).

^**Based on Table 1 of Young et al. which itself was based on Table 1.1 of Geskus (Geskus, 2016; Young et al., 2020).

^***Recently described terminology of separable total effect, separable direct effect, and separable indirect effect (Rojas-Saunero et al., 2023; Stensrud et al., 2021; Stensrud et al., 2022).

We calculated and presented estimates for different counterfactual risks (Supplementary Table 1) as well as all effects (i.e., total, direct, and separable effects, as risk differences and risk ratios [RR]; Table 3) at 8 years of follow-up, along with their 95% confidence intervals (CIs), using percentile-based bootstrapping with 500 bootstrap samples. All analyses were conducted in the R programing language, adapting scripts from Young et al., Rojas-Saunero et al., and Stensrud et al. (Rojas-Saunero et al., 2023; Stensrud et al., 2021; Stensrud et al., 2022; Young et al., 2020)

Table 3.

Total Effects, Controlled Direct Effects, Separable Effects of SCI (Compared with No SCI) on Dementia Risk, and the Total Effect on All-cause Mortality Risk, at 8 years of Follow-up, for Participants in the NHATS, 2012-2019.

Risk Difference Estimands	Risk Difference (95% CI)*	Risk Ratio (95% CI)*
Total effects on dementia and mortality
Dementia: $\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a}=1}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a}=0}=1\right]$	0.28 (0.02,0.64)	1.95 (1.07,3.04)
Mortality: $\mathrm{P}\mathrm{r}\left[{\mathrm{D}}_{\mathrm{k}+1}^{\mathrm{a}=1}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{D}}_{\mathrm{k}+1}^{\mathrm{a}=0}=1\right]$	−0.45 (−0.90,−0.11)	0.31 (0.10,0.67)
Controlled direct effect on dementia
$\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a}=1,\stackrel{-}{\mathrm{d}}=\stackrel{-}{0}}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a}=0,\stackrel{-}{\mathrm{d}}=\stackrel{-}{0}}=1\right]$	0.27 (0.01,0.67)	1.98 (1.02,3.37)
Separable effects on dementia
Direct: $\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=1,{\mathrm{a}}_{\mathrm{D}}=1}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=0,{\mathrm{a}}_{\mathrm{D}}=1}=1\right]$	0.13 (−0.13,0.32)	1.38 (0.85,2.25)
Indirect: $\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=0,{\mathrm{a}}_{\mathrm{D}}=1}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=0,{\mathrm{a}}_{\mathrm{D}}=0}=1\right]$	0.20 (−0.08,0.61)	1.73 (0.73,3.25)
Total: $\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=1,{\mathrm{a}}_{\mathrm{D}}=1}=1\right]-\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}}=0,{\mathrm{a}}_{\mathrm{D}}=0}=1\right]$	0.33 (0.16,0.52)	2.17 (1.54,2.94)

Abbreviations: SCI, subjective cognitive impairment; NHATS, national health and aging trends study; CI, confidence interval; IPCW, inverse-probability-of-censoring weights; IPTW, inverse-probability-of-treatment weights.

Notes: Participants with dementia or residing in a nursing home facility in Round 2 (2012) or earlier were excluded. All estimates are unweighted, unless noted otherwise. All estimates are adjusted for IPTW, using baseline covariates. IPCW adjusts for all-cause mortality and loss to follow-up. Weights were calculated adjusting for the baseline characteristics in Table 1, as well as time-varying characteristics, and were stabilized. ${\mathrm{Y}}_{\mathrm{k}}$ and ${\mathrm{D}}_{\mathrm{k}}$ denote, respectively, the primary outcome of interest (dementia onset), and the competing event (mortality), a terminal state which prevents dementia from onsetting by time $\mathrm{k}$ of follow-up. $\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}}=1\right]$ denotes the probability of outcome ${\mathrm{Y}}_{\mathrm{k}}$ being observed by follow-up time $\mathrm{k}$. ${\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a}}$ and ${\mathrm{D}}_{\mathrm{k}+1}^{\mathrm{a}}$ denote, respectively, the counterfactual dementia onset and counterfactual all-cause mortality outcomes by time $\mathrm{k}+1$ under exposure level $\mathrm{a}$. ${\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a},\stackrel{-}{\mathrm{d}}=\stackrel{-}{0}}$ denotes the counterfactual dementia onset by time $\mathrm{k}+1$ under exposure level $\mathrm{a}$ and under an intervention that eliminates the competing event, such that $\stackrel{-}{\mathrm{d}}=\stackrel{-}{0}$. ${\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}},{\mathrm{a}}_{\mathrm{D}}}$ denotes the counterfactual dementia onset by time $\mathrm{k}+1$ under level ${\mathrm{a}}_{\mathrm{Y}}$ of the exposure pathway ${\mathrm{A}}_{\mathrm{Y}}$ and level ${\mathrm{a}}_{\mathrm{D}}$ of the exposure pathway ${\mathrm{A}}_{\mathrm{D}}$. For details on notation, see Supplementary Materials.

^*Percentile-based bootstrap from 500 bootstrap samples.

Total Effects of SCI on Dementia and Mortality Risks

The total effect of SCI on dementia risk can be quantified to answer the following research question (RQ1): “What would be the average difference in dementia risk, by some time t of follow-up, had no study participant experienced SCI versus had they all experienced SCI at baseline?” (Table 2) This estimates the average effect of SCI on dementia risk, through all causal pathways between SCI and dementia onset, including those possibly mediated by mortality, and reflects real-world scenarios where both the primary event of interest and the competing event occur (Figure 1a). Because a competing event is an event that prevents the occurrence of the event of interest (i.e., death prior to dementia prevents dementia from occurring) then one can imagine that the effect of SCI on dementia risk works not only directly but also indirectly by affecting the risk of mortality (i.e., if a risk factor causes death, it also prevents the event of interest). Therefore, the total effect of SCI on dementia measures the effect of how SCI impacts dementia regardless of whether it is due to a direct relationship, an indirect effect on dementia through the competing event, or a combination of both (see Supplementary Materials for technical details).

Figure 1. Illustrative Causal Directed Acyclic Graphs Representing the Relationship Between SCI, Dementia, and Mortality.

Abbreviations: SCI, subjective cognitive impairment’ NHATS, national Health and aging trends study.

Notes: $\mathrm{A}$ denotes the exposure (SCI); ${\mathrm{Y}}_{\mathrm{k}}$ and ${\mathrm{D}}_{\mathrm{k}}$ denote, respectively, the primary outcome of interest (dementia onset), and the competing event (mortality), a terminal state which prevents dementia from onsetting by time $\mathrm{k}$ of follow-up. ${\mathrm{A}}_{\mathrm{Y}}$ denotes the exposure $\mathrm{A}$ dementia pathway, which can assume values; ${\mathrm{A}}_{\mathrm{D}}$ denotes the exposure $\mathrm{A}$ all-cause mortality pathway. ${\mathrm{A}}_{\mathrm{Y}}$ can assume levels ${\mathrm{a}}_{\mathrm{Y}}=0$ or ${\mathrm{a}}_{\mathrm{Y}}=1$. Likewise, ${\mathrm{A}}_{\mathrm{D}}$ can assume levels ${\mathrm{a}}_{\mathrm{D}}=0$ or ${\mathrm{a}}_{\mathrm{D}}=1$. ${\mathrm{L}}_{\mathrm{k}}$ denotes time-varying individual characteristics updated at each period $\mathrm{k}$; ${\mathrm{C}}_{\mathrm{k}+1}$ denotes an indicator for loss to follow-up by period $\mathrm{k}+1$. Lastly, $\mathrm{U}$ denotes unmeasured shared risk factors for the event of interest (dementia), and individual characteristics, as well as the competing event (i.e., mortality for the total effect). In all graphs, some arrows are omitted for simplicity and to reduce clutter.

(a) A causal directed acyclic graph representing observed data generating assumptions under which the identifying assumptions of the total effect hold, permitting identification of the total effect of SCI (exposure $\mathrm{A}$) on dementia (event of interest, ${\mathrm{Y}}_{\mathrm{k}}$ after $\mathrm{k}$ periods of follow-up). Adapted from Young et al. (Young et al., 2020).

(b) A causal directed acyclic graph representing observed data generating assumptions under which the identifying assumptions of the total effect hold, permitting identification of the total effect of SCI (exposure $\mathrm{A}$) on dementia (event of interest, ${\mathrm{Y}}_{\mathrm{k}}$ after $\mathrm{k}$ periods of follow-up). We must include the arrow from ${\mathrm{D}}_{\mathrm{k}}$ to ${\mathrm{Y}}_{\mathrm{k}}$ or from ${\mathrm{D}}_{\mathrm{k}+1}$ to ${\mathrm{Y}}_{\mathrm{k}+1}$, since, by definition, if ${\mathrm{D}}_{\mathrm{k}}=1$ then ${\mathrm{Y}}_{\mathrm{k}}=0$. Adapted from Young et al. (Young et al., 2020)

(c) A causal directed acyclic graph representing observed data generating assumptions under which the identifying assumptions of the separable effects hold, permitting identification of the separable direct, indirect, and total effects of SCI (exposure $\mathrm{A}$) on dementia (event of interest, ${\mathrm{Y}}_{\mathrm{k}}$ after $\mathrm{k}$ periods of follow-up) through the direct effect that SCI exerts via its component affecting dementia (${\mathrm{A}}_{\mathrm{Y}}$), or through SCI’s component affecting mortality (${\mathrm{A}}_{\mathrm{D}}$), or both. Adapted from Stensrud et al. (Stensrud et al., 2021).

It is equally important to examine whether SCI affects mortality risk before the onset of dementia, by additionally examining the cumulative incidence for mortality across exposure groups. The “before the onset of dementia” portion of the preceding sentence is key, as we are interested in isolating the independent effect of SCI on mortality risk, free from any effect of dementia. This motivates the following research question (RQ2): “What would be the average difference in all-cause mortality before dementia onset, by some time t of follow-up, had no study participant experienced SCI versus had they all experienced SCI at baseline?” (Table 2)

To assess the total effects of SCI on dementia and mortality risks, we used a weighted Aalen-Johansen estimator comparing participants with and without SCI at Round 2 of the NHATS. Weights were calculated as the product of the IPTW and IPCW, to adjust for baseline confounders (Table 1) and censoring due to loss to follow-up, respectively (Hernán & Robins, 2023), under the assumption that the same set of measured confounders used to estimate the total effect of SCI on dementia risk were sufficient for addressing confounding of the total effect of SCI on all-cause mortality risk. The IPTW for participants with SCI was the inverse of the probability of having SCI given baseline covariates, and the inverse of the probability of being “SCI-free” for participants without SCI at baseline. These probabilities were estimated using a logistic regression model relating SCI status to the baseline covariates. The time-t IPCW was zero for participants who were lost to follow-up by time t , and the time-specific inverse probability of being lost to follow-up for those who experienced a dementia onset or died. The total effect of SCI on mortality was estimated using similar approach but with the IPTW and the IPCW for loss to follow-up that occurred before death or dementia onset.

Answering RQ1 and RQ2 alone does not clarify whether SCI’s impact on dementia risk is partially mediated by mortality. Conceptually, the indirect effect of SCI on dementia through mortality is necessarily protective, as those who die earlier are “protected” from developing dementia. This “pathological mediation” complicates total effect interpretation (Rojas-Saunero et al., 2023), warranting caution when a competing event influences the exposure’s total effect. This also highlights the need to examine SCI’s total effect on mortality and motivates the introduction of the direct effect estimand to address a different research question, as discussed below.

Controlled Direct Effect of SCI on Dementia Risk

Alternatively, we may be interested in answering the following question (RQ3): “What would be the difference in dementia risk, by time t of follow-up, had no study participant experienced SCI at baseline compared to had they all experienced SCI at baseline if there was no competing risk of death (i.e., under a scenario in which an intervention existed to prevent the competing event) throughout the study period?” To address this question, estimating the direct effect is appropriate. Unlike total effects, the direct effect assumes a scenario where only the primary outcome occurs, with all competing events prevented (Figure 1b) (Rojas-Saunero et al., 2023; Young et al., 2020). The contrast between outcomes in the exposure groups, known as the controlled direct effect, isolates the impact of exposure on the outcome, independent of the competing event (Table 2). Eliminating the competing event allows the direct effect to more accurately assess the exposure’s impact on the outcome, free from the exposure affecting the competing event which prevents the outcome of interest. This adjustment cannot be done with total effects, even if the total effects on all competing events are considered. The direct effect measure relates to the marginal or “net” probability measures in the historical competing risks literature (Fine & Gray, 1999; Pepe & Mori, 1993).

We note that in this context, death acts as a form of censoring, leading to missing data on dementia onset. While the historical competing risks literature often equates death with censoring, this is misleading because death directly causes missingness in dementia outcomes (Rojas-Saunero et al., 2023). The definition of censoring depends on the research question, especially when outcomes are considered under the hypothetical elimination of death. Identifying the controlled direct effect of SCI on dementia risk requires additional assumptions—exchangeability, positivity, and consistency—with respect to censoring, beyond those needed for the total effect (Hernán & Robins, 2023; Pearl, 1995; Young et al., 2020), and adjustments for shared causes of dementia and mortality are necessary (see Supplementary Materials).

We estimated the controlled direct effect of SCI on probable dementia risk using a weighted Aalen-Johansen estimator, comparing participants with and without SCI starting from their NHATS Round 2 assessment. The weights, which varied over time, were calculated as the product of baseline IPTWs and time-specific IPCWs. For participants alive and uncensored at time t, the censoring weight was the product of inverse survival probabilities from prior periods, adjusted for shared causes of death, loss to follow-up, and dementia onset. IPCWs were zero for those who died or were lost to follow-up by time t. Survival probabilities were estimated using logistic regression, with baseline and time-varying covariates as predictors (Table 1).

Identification of the direct effect requires the hypothetical scenario where all competing risks are eliminated. An alternative measure which does not assume elimination of the competing events – reference-adjusted event probabilities – has been proposed recently, in addition to separable effects. We describe next that latter approach. Readers interested in the reference-adjusted event probabilities can refer to recent work by Rutherford et al. (Rutherford et al., 2022).

Separable Direct, Indirect, and Total Effects of SCI on Dementia Risk

When the exposure affects the primary outcome and competing events through distinct pathways, separable effects can be estimated without conceptually eliminating competing events, a condition known as the generalized decomposition assumption (Stensrud et al., 2021). However, identification of these effects requires that the exposure components be adjustable to different values (Figure 1c). In our application, this requires conceptualizing SCI as having two distinct components: one pathway that directly influences dementia and another that affects dementia indirectly through its impact on all-cause mortality. Therefore, this is denoted as exposure $\mathrm{A}$ (SCI) being decomposed into an exposure to SCI that affects the dementia pathway, ${\mathrm{A}}_{\mathrm{Y}}$, and an exposure to SCI that affects the mortality pathway, ${\mathrm{A}}_{\mathrm{D}}$ (Figure 1c). The counterfactual is linked to the observed outcomes via the assumption that measured confounders (${\mathrm{L}}_{\mathrm{k}}$ in Figure 1c) are sufficient to control for confounding between the treatment and competing events (hence no edges being drawn in Figure 1c between exposure and ${\mathrm{L}}_{\mathrm{k}}$). It is further assumed that, after adjusting for these confounders, the primary event is independent of the treatment component affecting the competing event, and vice versa. Additional structural assumptions required for identification of the separable effects are discussed in Stensrud et al. and summarized in Supplementary Materials.

Separable effects can answer three types of questions. First, one might be simply interested in the average effect of SCI on dementia risk, under the assumption that SCI exerts an effect on dementia and mortality through different pathways, but irrespective of the pathway the effect may be exerted. Hence, the question one might want answered is (RQ4): “What would be the average difference in dementia risk, by some time t of follow-up, had every study participant experienced exposure to both the SCI dementia and mortality pathways versus had they all been unexposed to both pathways at baseline?” In that case, the separable total effect can be estimated as the sum of the separable direct and separable indirect effects.

Alternatively, it is appropriate to estimate the separable direct effect when the question of interest is (RQ5): “What would be the difference in dementia risk, by some time t of follow-up, had every study participant experienced exposure to the SCI dementia pathway compared to had they all been unexposed to the SCI dementia pathway at baseline, while controlling for the SCI mortality pathway by holding exposure status constant (e.g., $a=a_D=1$)?” The separable direct effect assesses the causal effect of SCI on dementia risk, that is, SCI’s effect on dementia risk that is not mediated by the effect it exerts on mortality, under the generalized decomposition assumption.

Lastly, the question of interest may be (RQ6): “What would be the difference in dementia risk, by some time t of follow-up, had all study participants experienced exposure to the SCI mortality pathway versus had they all been unexposed to the SCI mortality pathway at baseline, while holding exposure status to the SCI dementia risk pathway constant for all participants (e.g., $a=a_Y=0$)?” That question can be answered by estimating the separable indirect effect, which quantifies the causal effect of SCI on dementia risk, that is, only through SCI’s effect on mortality, under the generalized decomposition assumption.

To estimate the separable effects of SCI on dementia risk, we used the approach by Stensrud et al. (Stensrud et al., 2021). We created an additional copy of the SCI variable to isolate its effects on dementia and mortality. This allowed us to separately manipulate each component in hazard calculations for both outcomes. We then estimated logistic regression models for dementia and mortality, adjusting for confounders, and created three datasets with the following features: 1) SCI affects dementia through both components, 2) SCI has no effect on dementia (directly or indirectly via its mortality component), and 3) SCI has a direct effect on dementia but not via mortality. We calculated cause-specific hazards and predicted cumulative incidences under hypothetical interventions that either prevent or allow SCI’s effect on dementia through its mortality component. IPTW and IPCW were applied to isolate the separable direct and indirect effects. We summed the separable direct and indirect effects to derive the total effect (see Supplementary Materials for details).

RESULTS

Sample Characteristics

Table 1 summarizes the characteristics of the sample, which included n=1,622 living dementia-free participants in both Rounds 1 and 2 of the NHATS, including n=115 with SCI and n=1,507 without SCI. The mean age was 78.0 (standard deviation [SD], 7.4) years. The sample included a higher proportion of females, whites, and individuals with at most a high school degree, unmarried, but living with a spouse or partner, income above 200% of the FPL, at least one IADL impairment, and with hypertension. About a third of participants had a mental health symptom of depression and/or anxiety. Compared to those without SCI at baseline, participants with SCI were less educated, more likely to live alone, participate in Medicaid, be poor (income below 100% of the FPL), have higher numbers of ADL and IADL impairments, have heart disease, diabetes, depression, or anxiety.

Total, Direct, and Separable Effects of SCI on Dementia and Mortality Risks

Table 2 summarizes the counterfactual risks (cumulative incidence) of dementia and mortality, along with their estimates after 8 years of follow-up, in the presence of the competing event of mortality from all-causes, and by SCI status at baseline. Figure 2 depicts these cumulative incidences as curves under different estimands. Table 3 reports the estimated total, direct, and separable effects of SCI on probable dementia and all-cause mortality, after 8 years of follow-up, as risk difference and risk ratios. We discuss each below.

Figure 2. Risk of probable dementia onset and all-cause mortality over eight years, by experience of subjective cognitive impairment (SCI) in 2012, using data from the National Health and Aging Trends Study (NHATS).

Abbreviations: SCI, subjective cognitive impairment; NHATS, national health and aging trends study.

Notes: ${\mathrm{Y}}_{\mathrm{k}}$ and ${\mathrm{D}}_{\mathrm{k}}$ denote, respectively, the primary outcome of interest (dementia onset), and the competing event (mortality), a terminal state which prevents dementia from onsetting by time $\mathrm{k}$ of follow-up. $\mathrm{P}\mathrm{r}\left[{\mathrm{Y}}_{\mathrm{k}}=1\right]$ denotes the probability of outcome ${\mathrm{Y}}_{\mathrm{k}}$ being observed by follow-up time $\mathrm{k}$. ${\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a}}$ and ${\mathrm{D}}_{\mathrm{k}+1}^{\mathrm{a}}$ denote, respectively, the counterfactual dementia onset and counterfactual all-cause mortality outcomes by time $\mathrm{k}+1$ under exposure level $\mathrm{a}$. ${\mathrm{Y}}_{\mathrm{k}+1}^{\mathrm{a},\stackrel{-}{\mathrm{d}}=\stackrel{-}{0}}$ denotes the counterfactual dementia onset by time $\mathrm{k}+1$ under exposure level $\mathrm{a}$ and under an intervention that eliminates the competing event, such that $\stackrel{-}{\mathrm{d}}=\stackrel{-}{0}$. ${\mathrm{Y}}_{\mathrm{k}+1}^{{\mathrm{a}}_{\mathrm{Y}},{\mathrm{a}}_{\mathrm{D}}}$ denotes the counterfactual dementia onset by time $\mathrm{k}+1$ under level ${\mathrm{a}}_{\mathrm{Y}}$ of the exposure pathway ${\mathrm{A}}_{\mathrm{Y}}$ and level ${\mathrm{a}}_{\mathrm{D}}$ of the exposure pathway ${\mathrm{A}}_{\mathrm{D}}$. For details on notation, see Supplementary Materials.

(a) Total effect of SCI on dementia risk: Curves represent the cause-specific cumulative incidence or crude risk of dementia over 8 years of follow-up, had participants not experienced SCI (dashed dark grey curve) vs had they experienced SCI (solid orange curve) at baseline (Round 2 of NHATS).

(b) Total effect of SCI on all-cause mortality risk before dementia onset: Curves represent the cumulative incidence or risk of mortality before dementia onset over 8 years of follow-up had participants not experienced SCI (dashed dark grey curve) vs had they experienced SCI (solid orange curve) at baseline.

(c) Controlled direct effect of SCI on dementia risk: Curves represent the marginal cumulative incidence or net risk of dementia (had death been eliminated) over 8 years of follow-up, had participants not experienced SCI (dashed dark grey curve) vs had they experienced SCI (solid orange curve) at baseline.

(d) Separable total, direct, and indirect effects of SCI on dementia risk. Curves represent the cumulative incidence of dementia over 8 years of follow-up, had participants not experienced SCI vs had they experienced SCI at baseline (Round 2 of NHATS), based on the methods in Section 7 of Stensrud et al. (Stensrud et al., 2021): With SCI containing both the ${\mathrm{A}}_{\mathrm{Y}}$ and ${\mathrm{A}}_{\mathrm{D}}$ pathways ($\mathrm{a}={\mathrm{a}}_{\mathrm{Y}}={\mathrm{a}}_{\mathrm{D}}=1$; solid orange curve), without SCI, hence no ${\mathrm{A}}_{\mathrm{Y}}$ or ${\mathrm{A}}_{\mathrm{D}}$ pathways ($\mathrm{a}={\mathrm{a}}_{\mathrm{Y}}={\mathrm{a}}_{\mathrm{D}}=0$; dashed dark grey curve), and with a modified SCI exposure under an intervention that removes only the ${\mathrm{A}}_{\mathrm{D}}$ pathway (${\mathrm{a}}_{\mathrm{Y}}=1$, ${\mathrm{a}}_{\mathrm{D}}=0$; solid blue curve) for everyone in the population.

Sample for panels includes n=1,622 older adults aged ≥65 and residing in the community or in a non-nursing home residential setting in rounds 1 and 2 of NHATS, who did not have dementia in either rounds 1 or 2 of the survey. The cohort is followed over the study period (8 years) until dementia onset, death (from all causes), loss to follow-up, or the end of the observation period, whichever occurred first. All-cause mortality was treated as a competing risk for probable dementia onset. Dementia status defined based on the NHATS probable dementia definition. SCI status determined based on the Disability Questionnaire’s Cognitive Disability item.

Total Effects

The answer to research question RQ1 is obtained by analyzing the standardized (i.e., IPTW and IPCW adjusted) Aalen-Johansen estimates of the cumulative incidence of dementia with and without SCI over the 8-year follow-up period (Figure 2a). The standardized cumulative incidence of dementia at 8-year follow-up was 0.58 (CI: 0.33,0.95) with SCI and 0.30 (CI: 0.25,0.35) without SCI (Supplementary Table 1), translating into a difference of 0.28 (CI: 0.02,0.64) and risk ratio of 1.95 (CI: 1.07,3.04) (Table 3). This is depicted over the follow-up time as the difference between the dark blue and orange curves in Figure 2a. The 8-year risk difference represents the estimated total effect of SCI on probable dementia risk over an 8-year follow-up period. Thus, if no study participant had experienced SCI at baseline, the average difference in dementia risk by 8 years of follow-up would have been reduced by half. In other words, SCI leads to a two-fold increase in the cause-specific cumulative incidence, or crude risk, of probable dementia over that period.

To answer RQ2 we assessed whether SCI at baseline affects all-cause mortality risk before dementia onset and calculated the cumulative incidence for all-cause mortality before dementia over time (Figure 2b). The curves of the cumulative incidences with and without SCI almost overlap in the earlier follow-up periods. While the gap between the curves widens over time, confidence intervals continue to overlap, largely due to high variability in the cumulative incidence of mortality without SCI. The standardized cumulative incidence of all-cause mortality before dementia onset at 8-year follow-up was 0.16 (CI: 0.08,0.26) with SCI and 0.61 (CI: 0.30,1.00) without SCI (Supplementary Table 1), translating into a risk difference of −0.45 (CI: −0.89,−0.11) and risk ratio of 0.32 (CI: 0.10,0.67). This is the estimated risk of the competing event (mortality) occurring prior to dementia onset, thus preventing dementia from onsetting. Thus, if no study participant had experienced SCI at baseline, the average difference in mortality risk by 8 years of follow-up would have been higher. This suggests that while SCI does not appear to affect mortality risk over a short period of follow-up, SCI appears to be protective against dementia-free all-cause mortality occurring prior to dementia at 8-years follow-up.

Controlled Direct Effect

Research question RQ3 is answered by analyzing Figure 2c, which depicts the standardized risk of probable dementia with and without SCI over the 8-year follow-up time, if mortality from all-causes were somehow eliminated by some intervention. After 8 years of follow-up, and under this hypothetical intervention where mortality is eliminated as a competing event, the standardized probability of probable dementia was 0.54 (CI: 0.28,0.96) with SCI and 0.27 (CI: 0.23,0.32) without SCI (Supplementary Table 1), resulting in a statistically significant difference of 0.27 (CI: 0.01,0.67) and a risk ratio of 1.98 (CI: 1.02,3.37) (Table 3). This represents the controlled direct effect of SCI on probable dementia risk, or put otherwise, the marginal cumulative incidence or “net” risk of probable dementia over 8 years of follow-up, under the hypothetical elimination of all-cause mortality, and had participants experienced SCI vs had they not experienced SCI at baseline. The result suggests that SCI exerts a direct effect on dementia risk, that is not mediated by its effect on all-cause mortality.

Separable Effects

Analysis of Figure 2d, answers research questions RQ4-RQ6 by depicting the 8-year standardized cumulative incidence of dementia with and without SCI, and under a hypothetical intervention that eliminates the component of SCI affecting mortality, rather than removing mortality entirely as in controlled direct effects. The cumulative incidence of dementia with SCI under no intervention ($\mathrm{a}={\mathrm{a}}_{\mathrm{Y}}={\mathrm{a}}_{\mathrm{D}}=1;$ solid orange curve) was 0.62 (CI: 0.45, 0.79) after 8 years. The cumulative incidence was 0.29 (CI: 0.24, 0.34) without SCI, under the hypothetical intervention removing the mortality-related SCI component ($\mathrm{a}={\mathrm{a}}_{\mathrm{Y}}={\mathrm{a}}_{\mathrm{D}}=0;$ dashed dark grey curve). Lastly, with a modified SCI, under an intervention that removes only the mortality-related SCI component, ${\mathrm{A}}_{\mathrm{D}}$ (${\mathrm{a}}_{\mathrm{Y}}=1$, ${\mathrm{a}}_{\mathrm{D}}=0;$ solid blue curve), the incidence was 0.49 (CI: 0.21, 0.91) after 8 years.

The separable total effect of SCI on dementia risk over time is the difference between the solid orange and dashed dark-grey curves, and answers RQ4. The effect is statistically significant at 8-year follow-up and even earlier – there is overlap in the confidence intervals only before approximately 24 months of follow-up. At 8-year follow-up the risk difference and risk ratio were 0.33 (CI: 0.16,0.52) and 2.17 (CI: 1.54,2.94), respectively (Table 3). Thus, SCI appears to exert an effect on dementia risk, that is comprised of a direct effect (i.e., not mediated by SCI’s effect on mortality), indirect effect (i.e., only through SCI’s effect on mortality), or combination of both direct and indirect effects. To understand why and how, we examine closely the separable direct and indirect effects.

The separable direct effect of SCI on dementia risk over the follow-up period is represented by difference between the solid orange and solid blue curves and answers research question RQ5. While the two curves do not overlap, their confidence bands converge after approximately 24 months of follow-up, suggesting that the differences may not be statistically significant. Indeed, at 8-year follow-up the estimated separable direct effect was 0.13 (CI: −0.13,0.32) in risk difference and 1.38 (CI: 0.85,2.25) in risk ratio (Table 3). Thus, SCI appears to exert a direct effect on dementia risk, that is not mediated by SCI’s effect on mortality but there is a lack of precision in the estimate suggesting that the null relationship cannot be ruled out.

The separable indirect effect of SCI on probable dementia risk is the difference between the solid blue and the dashed dark-grey curves and answers RQ6. The two curves are not that far apart from each other, and their confidence bands do overlap indeed over the entire follow-up period. At 8-year follow-up, the estimated separable effect of SCI on dementia risk was 0.20 (CI: −0.08,0.61), corresponding to a risk ratio of 1.73 (CI: 0.73,3.25). This represents the difference in dementia risk with (a modified) SCI compared to without SCI, that is due to the effect of SCI on mortality. Thus, SCI also appears to exert an effect on dementia risk, that is only through SCI’s effect on mortality but there is again a lack of precision in the estimate to rule out a null relationship.

However, combined, these results suggest that SCI impacts dementia risk through both components affecting dementia and mortality, rather than through one alone. Thus, under the structural assumptions of separable effects theory, the total effect of SCI on dementia is not primarily due to a protective effect of SCI against all-cause mortality that would extend life and increase dementia risk.

DISCUSSION

This study examined the relationship between SCI and the risks of dementia and all-cause mortality over an 8-year period in a sample of older adults (“dementia-free” at baseline) participating in a special module of the NHATS administered in Round 2 (2012). We found that on average, SCI was significantly associated with an increased risk of dementia and a decreased risk of long-term dementia-free all-cause mortality, emphasizing the complex interplay between cognitive decline and mortality.

Older adults with SCI had nearly twice the risk of developing dementia within 8 years compared to those without SCI (RR: 1.95, CI: 1.07,3.07), aligned with prior research that identifies SCI as an early marker of dementia-related neuropathology (Jack et al., 2024; Jessen, Wolfsgruber, et al., 2014; Petersen et al., 2021; Schwarz et al., 2021). The controlled direct effect of SCI on dementia, independent of mortality, further suggests a potential causal relationship (RR: 1.98, CI: 1.02,3.37). Interestingly, SCI did not appear to impact short-term mortality despite being associated with a lower risk of long-term mortality prior to dementia onset (RR: 0.31, CI: 0.10,0.67), possibly due to differences in healthcare utilization or social support, and reflecting the concept of competing risks, where longer survival increases the likelihood of dementia onset. The analysis of the direct and separable effects revealed that SCI likely has a causal impact on dementia, further supporting a direct causal link.

These findings have significant implications for clinical management, dementia surveillance, and public health policy. Identifying SCI as an early indicator and potential cause of dementia can facilitate timely interventions aimed at delaying or preventing dementia onset, consistent with the updated AD research framework of the NIA-AA (Jack et al., 2024; Petersen et al., 2021). Given that SCI is routinely assessed in large national surveys, these findings suggest that monitoring SCI could enhance our ability to measure dementia risk, identify at-risk populations, and allocate resources more effectively.

The study also underscores the importance of clarifying research questions and applying appropriate analytic methods to account for competing risks to properly interpret findings in aging research. Many time-to-event studies involving older adults fail to specify the research question and properly address competing risks, complicating the identification, estimation and interpretation of relationships between factors like SCI and dementia. Applying suitable analytic frameworks is essential for advancing our understanding of these relationships. For example, distinguishing between the “total effect” and “separable total effect” can have implications for clinical and policy decision-making, as the choice of the estimand can influence the bias and precision of the estimates, as evidenced by our results.

This study’s strengths include its large, representative cohort and advanced methods for addressing competing risks. However, limitations exist. Dementia onset was assessed using the NHATS probable dementia definition, which, despite being valid and reliable, may lack sensitivity to education level and struggle to differentiate between mild cognitive impairment and dementia. The absence of biomarker data further limits diagnostic accuracy. The SCI measure may not fully capture distinctions between dementia-related impairment and other cognitive issues like depression or anxiety (Langa et al., 2005). Additionally, SCI was assessed only once, and the study’s observational nature, combined with the omission of genetic information to account for genetic risk factors for dementia, limits causal inference. Furthermore, the variability in estimates, as indicated by wide confidence intervals, underscores the need for further research to validate these findings. This variability could be explained by the relatively low SCI prevalence in the sample. Despite these challenges, the study underscores SCI’s importance as a factor for dementia risk.

CONCLUSION

This study establishes SCI as a significant predictor of increased dementia risk in older adults over an 8-year period. The findings underscore the complex relationship between cognitive decline and mortality and support the need for early detection and intervention strategies in SCI. Additionally, the results highlight the necessity of advanced analytic methods to account for competing risks in aging research.

Supplementary Material

Supplementary Materials: Subjective cognitive impairment (SCI) with future probable dementia risk in the National Health and Aging Trends Study (NHATS) – 2012-2019