The interplay of food insecurity, diet quality and dementia status in their association with all-cause mortality among older US adults in the Health and Retirement Study 2012–2020

Abstract

Background:

All-cause mortality risk and dementia occurrence have been previously hypothesized to be linked with food insecurity and poor dietary quality.

Objective:

The aims of the study were to test mediation and interactions between food insecurity, diet quality and dementia status in relation to all-cause mortality.

Design:

The interplay of food insecurity, diet quality and dementia in their associations with all-cause mortality was studied, in terms of interactions, and mediating effects, using secondary longitudinal data from a sample of older US adults from the Health and Retirement Study (HRS, 2012–2020). Reduced (age, sex, race/ethnicity-adjusted, M1) and fully adjusted (socio-demographic, lifestyle and health-related factor-adjusted, M2) models were tested, and stratification by sex and race/ethnicity was carried out.

Participants/setting:

2,894 US older adults (2012–2014, mean baseline age of 76.4 y) were selected from this national longitudinal sample.

Main outcome measures:

The outcome of interest was all-cause mortality risk for follow-up till end of 2020.

Statistical analyses performed:

Cox proportional hazards, four-way decomposition and generalized structural equations models (GSEM) were utilized.

Results:

Overall, 902 deaths occurred (51 per 1,000 person-years (P-Y). Food insecurity (yes vs. no) was not associated with mortality risk in M1, though inversely related to this outcome in M2, (Cox models and GSEM). Food insecurity was directly related to Ln(dementia odds) in M1 only (β±SE: 0.23±0.05, P<0.001, GSEM). Diet quality as measured by HEI-2015 (z-scored), while inversely related to food insecurity in reduced GSEM (β±SE:−0.18±0.06, P=0.005), was also inversely related to both Ln(dementia odds), z-scored (β±SE:−0.14±0.03, P<0.001) and mortality risk (LnHR±SE:−0.14±0.03, P<0.001, M1). Ln(dementia odds) was strongly associated with mortality risk (HR=1.39, 95% CI: 1.31–1.48, P<0.001, M2). In both four-way decomposition models and GSEM, the total effect of diet quality on mortality risk was partially mediated through Ln(dementia odds) (M1 and M2), explaining 15–21% of this total effect.

Conclusion:

Diet quality-mortality risk association was partially mediated through dementia, with inconsistent findings observed for food insecurity.

INTRODUCTION

Global dementia prevalence is increasing, with over 55 million people living with the condition as of 2020¹. This number is expected to rise to 78 million by 2030 and 139 million by 2050, primarily in low- and middle-income countries¹. The economic impact of the disease is substantial, with costs estimated to reach $2.8 trillion by 2030¹. The World Health Organization (WHO) has formulated a worldwide action plan for dementia management, emphasizing critical health and social system domains^2–4. Notwithstanding recent progress in pharmacotherapy, there exists a paucity of therapeutic options, hence necessitating additional investigation into dementia prevention, risk mitigation, and modifiable lifestyle determinants, that can then be translated into primary preventive interventions^3–5. Dementia risk has been associated with food insecurity ^6–11 and poor diet quality^12–15, among numerous target modifiable risk factors which have also been linked to mortality risk ^{12, 13, 15–21}.

Furthermore, the 2020 Lancet commission, along with its 2024 update, identified other potentially modifiable socio-environmental, cardio-metabolic and psychosocial risk factors for incidence of dementia. Those included education, air pollution, smoking, obesity and its related disorders, depression and social isolation ^{22, 23}. In addition to being linked to dementia risk, including of Alzheimer’s Disease (AD) sub-type^{24, 25}, obesity has been associated with all-cause mortality^{26, 27}, in large part driven by poor dietary choices and quality²⁸, which lead to increased incidence of cardiovascular disease among others²⁹. In turn, diet quality may be the direct consequence of food insecurity, the latter compelling consumers to opt for inexpensive, energy-dense, nutrient-deficient diets, thus increasing the risk of obesity and its associated cardio-metabolic disorders ^30–34.

Moreover, independently of obesity status, and given their neuroprotective properties, certain components of the diet, including antioxidants, certain B-vitamins and omega-3 fatty acids, which are consumed in greater proportionate amounts with higher dietary quality, have been associated with reduced incidence of dementia and a slower pace of age-related cognitive decline^35–43. These components of the diet are also linked to decreased systemic inflammation and oxidative stress, and thus reducing the risk for all-cause mortality ^{29, 35–43}. These dietary components were also found to be consumed to a lesser extent under circumstances of food insecurity, due to their inherent elevated cost particularly in lower income neighborhoods^30–34.

Given the interconnectedness of food insecurity, diet quality, dementia, and mortality, knowing the interplay in terms of mediational pathways and interactions between food insecurity, diet quality and dementia in predicting mortality, presents a valuable subject for innovative research. To our knowledge, no study has explored the interplay of food insecurity, diet quality and dementia status in relation to mortality risk, focusing on mediating and moderating effects among older US persons, by conducting secondary analyses of nationally representative longitudinal data. Furthermore, sex differences might impact health outcomes due to hormonal, genetic, and physiological factors, whereas societal influences affect food accessibility, dietary practices, caregiving responsibilities, and stress levels^{44, 45}. Racial and ethnic differences in access to resources can lead to divergence in relationships among food insecurity, dietary habits, and health consequences across sex and race/ethnicity⁴⁶. Identifying these distinctions is essential for the validity and generalizability of results.

Consequently, the study aimed at investigating whether food insecurity, diet quality and dementia are associated with mortality risk, independently of potential socio-demographic, socio-economic, other lifestyle and health-related factors, across sex and racial/ethnic groups. It was further hypothesized that food insecurity and diet quality’s relationship with all-cause mortality was mediated or moderated by dementia status, whereby food insecurity is inversely related to dietary quality, while dietary quality is inversely related to both dementia status and mortality risk.

MATERIALS AND METHODS

Database

The Health and Retirement Study (HRS) is a longitudinal panel study initiated in 1992 to investigate the health, economic, and social factors affecting older Americans as they age⁴⁷. New cohorts were added over time to maintain a representative sample of adults over the age of 50y⁴⁷.The HRS, funded by the National Institute on Aging (NIA) with grant number U01AG009740 and the Social Security Administration (SSA), collects data from a representative sample of persons aged 50 and older in the United States ^{48, 49}. The sampling strategy of HRS included a multistage area probability sample design to ensure that it represents the U.S. population over 50y of age, with the primary sampling unit (PSU) in the HRS being clusters of geographic areas selected to represent the national population of individuals aged 50 and older, stratified by region and demographic characteristics⁴⁷. It subsequently selects secondary sampling units (SSUs, census blocks or segments) within each PSU, and households within the SSUs, interviews eligible individuals, oversamples minority populations and older age cohorts, and analyses usually apply sampling weights to ensure the final sample reflects the U.S. population composition in terms of age, sex, race/ethnicity, along with other demographics⁴⁷.

The core data includes a long list of variables collected from all HRS participants every two years, covering a wide range of health and retirement-related domains. The Research and Development (RAND) longitudinal dataset used in this study provides curated access to these key data. The Enhanced Face-to-Face Interview (EFTF) for the HRS was introduced in 2006, and it includes physical (performance tests and anthropometric measures), biological (blood and saliva samples), and psychosocial (self-administered psychosocial questionnaires) measures. EFTF interviews were conducted in approximately half of the PSUs that had one or more living responder. More generally, the HRS has broadened its scope to encompass biological, psychological, social, and economic dimensions of aging as of 2006, utilizing wearable sensors to monitor exercise and sleep patterns, while also investigating metabolism and epigenetics. Life-course studies integrate early life experiences with data from midlife and later stages of life⁴⁷. The HRS also gathers biomarkers, cognitive measures, and recently incorporated genetic and metabolomics data⁴⁷.

Despite availability of EFTF data as of 2006, in the present study, only core data were utilized, and those were made available on the entire HRS sample respondents every two years. We also used data from a special survey, known as the 2013 Health Care and Nutrition Study (HCNS) and detailed in a later section, collected in 2013 on 8,035 respondents, which included items on food insecurity and dietary intake. These two data sources were then merged with data on dementia status available every two years on a sub-sample of older adults. The tracker file, which tracks status of respondents over time beyond the dates of the RAND and other files, in this case up to the end of 2020, was used to determine mortality status beyond the year 2014. Respondents that were selected were the ones participating in the HRS for the years 2012 and 2014 waves. Follow-up ended in on December of 2020.

Standard Protocol Approvals, Registrations, and Patient Consents

All procedures complied with the ethical standards set by the institutional or regional human research committees. Approval was obtained from the human subjects committee of the University of Michigan’s Institutional Review Board. Written informed consent was obtained in the initial parent HRS study. For the present prospective longitudinal cohort analysis, approval was obtained by the Institutional Review Board of the National Institutes of Health (NIH IRB), while the ethics board from the parent study determined that participant consent for data use was not required for secondary data analyses and therefore was waived.

Study Sample

The initially utilized RAND file for this present longitudinal study (randhrs1992_2018v2) included data on 43,561 participants in the original and new HRS (2006 onwards), spanning the years 1992 through 2018. Many of these participants were no longer part of the study in part due to loss to follow-up but mainly due to death by the chosen baseline year. According to the participant flowchart (Figure 1), there were 19,719 participants who were still living in 2012 and participated in the 2012 wave of HRS and were 50y or older during that wave of data. Of these participants, 17,312 had complete core data available in 2014 and therefore survived between 2012 and 2014. Those were included in this study with additional exclusions related to availability in outcome, exposure and mediator data. Within this group, we excluded those with no data on dementia status in 2014, resulting in a sample of 6,772 older adults with mean age in 2012 of 76.8 years. In the final sample, we selected those without any missing data on food insecurity, comprising 2,894 participants over the age of 50 in 2012 with dementia status or likelihood available in 2014 and food insecurity data available in 2013, who were followed up from 2014 until 2020 for mortality from all causes using the HRS’s most recent tracker file (trk2020tr_r), with mean age in 2012 of 76.4 years.

Figure 1. Participant Flowchart: Health and Retirement Study 1992–2020.

^aHRS=Health and Retirement Study; ^bRAND=Research and Development; ^cMR=Mortality Rate

All-cause mortality

The present investigation had the outcome variable all-cause mortality between 2014 and 2020. Deaths were discovered in the HRS using data linkage between the population register and interviews⁵⁰. Variables for 2020 (version 2) were obtained from the tracker file. Our final sample consisted of 2,894 people recruited between 2012 and 2014 for various data component. These participants were followed up until 2020 for the main outcome of interest, namely all-cause death. 902 dealths occurred during this period from any reason and had a date of death (month and year) recorded. Time-to-death or censoring was estimated as the number of years elapsed between age in 2012 and age at death or censoring by end of follow-up.

Dementia status and probability

The determination of dementia status is an expensive and labor-intensive diagnostic process, rendering it unfeasible for use in survey and longitudinal investigations^{3, 51}. In this investigation, we employed three established techniques to compute projected probability of dementia^{3, 51}. As of 2006, the data file (hrsdementia_2021_1109.sas7bdat), which is publicly accessible, contains dementia probabilities and status, predicted for HRS participants who were above age 70y between 2000 and 2016 and who self-repoted their race/ethnicity as Non-Hispanic White, Non-Hispanic Black or Hispanic⁵². The algorithms used included the revised version of the Hurd Model, alongside another set of predictions using the novel expert-informed logistic model (Expert Algorithm) and the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm. Training and testing of these models were carried out against all four waves of the Aging, Demographics, and Memory Study (ADAMS; https://hrsdata.isr.umich.edu/data-products/aging-demographics-and-memory-study-adams-wave)⁵³. This ascertainment resulted in a sensitivity of 77–83%, specificity of 92–94%, and overall accuracy of 90–92% in out-of-sample performance. Supplemental Method 1 provides further information concerning these algorithms (available in the GitHub repository: https://github.com/baydounm/HRS_FI_DIET_DEM_MORT). Additional information concerning the algorithms, including the instruments used for cognitive tests, can be found in other sources^{3, 51, 54, 55}.

HCNS 2013

The 2013 HCNS is an ancillary project within the HRS that examines health care utilization, health-related behaviors, and nutritional practices among older persons in the United States. It gathers information on dietary intake, health practices, and healthcare usage, encompassing preventative services, medical consultations, and pharmaceutical utilization. The research enhances comprehension of the effects of diet and healthcare accessibility on aging, chronic illnesses, and general well-being. More details are provided elsewhere^{56, 57}.

Food insecurity

The US Department of Agriculture (USDA) 6-Item Short Form US Household Food Security Survey Module, which was formerly validated, was utilized here to determine the extent of food insecurity ⁵⁸; (URL: https://www.ers.usda.gov/media/8271/hh2012.pdf). Respondents were given a series of questions about their food purchases and consumption during the previous 12 months. Examples of these questions included whether they could afford the food they required or whether they had to cut back on their eating because of budgetary constraints. Positive responses to the six questions (scoring range: 0–6) were added up using USDA criteria, with greater raw scores denoting a higher degrees of food insecurity, yielding the ordinal version of the food insecurity variable. Following the USDA guidelines, the total score is often divided into three categories: food secure (high and marginal statuses, scores: 0–1), low food secure (scores: 2–4), and very low food secure (scores: 5–6). Nevertheless, in this present study, a binary variable was used and coded as ‘food secure (0–1)’ and ‘food insecure (≥2)’ ⁵⁸. More details and the short form of the instrument are provided elsewhere^{59, 60}.

Diet quality

The 2013 HCNS sought to evaluate among others factors the dietary intake of older persons using data collected with a food frequency questionnaire which assessed intake of key nutrients and food groups. Specifically, the 2013 HCNS used a validated Harvard Food Frequency Questionnaire (FFQ) originally developed by Willett and colleagues to assess the intake of over 100 food and beverage items consumed over the previous year^{56, 57, 61}. The study encompassed inquiries regarding the consumption of processed foods and beverage intake, including sugary beverages, alcohol, and coffee. Food patterns can then be derived from food group and nutrient intake data, by estimating total and component scores for diet quality indices such as the such as the Mediterrannean Diet and various versions of the Healthy Eating Index (HEI)^{56, 57}. In our present study, diet quality was measured by HEI-2015. This index measures degree of alignment with the 2015–2020 Dietary Guidelines for Americans, including 13 dietary components^{62, 63}. Component scores ranged from 0 to 5 for total fruits, whole fruits, total vegetables, greens and beans, total protein foods, seafood and plant proteins, and 0 to 10 for the remaining components, which reflected intake of whole grains, dairy, fatty acids, refined grains, sodium, added sugars, and saturated fats. The HEI-2015 has a maximum score of 100 points. A simple HEI scoring algorithm method was used⁶² (See supplementary method 2 on github repository: https://github.com/baydounm/HRS_FI_DIET_DEM_MORT). A three-level version of HEI-2015 was used only in part of the analysis (Kaplan-Meier curves and Log-rank test), using tertiles, while in most analyses, the continuous HEI-2015 total score was included in models as a standardized z-score.

Covariates

Sociodemographic characteristics:

The group of basic covariates included self-reported sex (male, female), age (years) estimated at the end of the 2012 wave using data on date of birth, self-reported race/ethnicity (Non-Hispanic White vs. Non-Hispanic Black/Hispanic), marital status coded as “never married”, “married/partnered”, “separated/divorced” and widowed, education coded as “having no degree”, “GED”,” high school graduate”,” some college”, “college degree or higher”. Work status was re-coded as “working” or “unemployed” and total household wealth was estimated in US dollars, then grouped into “<25,000”, “25,000–124,999”, “125,000–299,000” and 300,000+.

Lifestyle Factors:

Lifestyle factors were also assessed in 2012, including smoking status (never smoker, past smoker, current smoker) and frequency of moderate/vigorous exercise (never, 1–4 times per month, >1 time per week). More specifically, and regarding physical exercise, the HRS Core dataset quantifies the frequency and intensity of physical exercise among individuals. It comprises self-reported inquiries regarding categories of physical activity, frequency of exercise, and intensity levels. The replies are classified as ordinal variables, with elevated values indicating increased frequency or intensity. With respect to smoking status, the HRS includes variables assessing current and past smoking behaviors, such as smoking history and intensity of smoking. The data also includes age at smoking initiation and cessation, duration of smoking, and pack-years of smoking. Current smoking status is determined by whether the respondent currently smokes, while past smoking history is assessed by whether they have ever smoked cigarettes. Combining these two measures yielded the three-level covariate.

Health characteristics:

The following health characteristics were considered in 2012: self-rated health (excellent/very good/good vs. fair/poor), self-reported weight and height measurements from which body mass index computed as weight/height-squared was categorized into (BMI, kg.m⁻², <25: healthy weight/underweight, 25–29.9: overweight, ≥30: obese), as well as existing cardiometabolic risk factors and chronic illness determined by a physician (hypertension, diabetes, heart problems, stroke, the count of which was categorized as 0, 1–2 and 3+), and depressive symptoms measured by an 8-item Centers for Epidemiologic Studies-Depression scale total score⁶⁴.

Statistical analysis

The HRS data were analyzed in a way as to account for sample design complexity by employing PSU, stratifications, and sampling weights, which were applied only on descriptive analyses and regression models (e.g. linear, logistic and Cox models), thus excluding key analyses that do not allow for such adjustment (e.g. four-way decomposition models and generalized structural equations models (GSEM) in imputed data). In the former set of analyses (i.e. excluding GSEM and four-way decomposition models), the following svy setting was used: mi svyset secu [pweight=HCNSWGTR_NT], strata(stratum) vce(linearized) singleunit(missing). The sampling weights used were specific to the 2013 HCNS survey. Covariates, exposures, mediators, moderators, and outcomes were all modeled with multivariable imputation utilizing chained equations ^{65, 66}. In the present analysis, only sociodemographic, lifestyle, and health factors considered as covariates were imputed, using 5 imputations with 10 iterations (Stata mi impute, mi estimate, and mi passive among others). To calculate population means, proportions, and regression coefficients, we utilized the survey (svy) command. Taylor series linearization was employed to correct standard errors.

As a first step, to compare means and proportions of primary variables across sex and racial/ethnic groups, the svy:reg and svy:mlogit commands were used, with sex and racial/ethnic stratification.

Second, we defined time-to-event (in years) as the age at entry >50 years (i.e., delayed entry) until exit age when the event of concern or censoring happened (loss to follow-up or termination of follow-up). Loss to follow-up was defined as the age at which respondents were not present for a wave but were alive and responding in the preceding wave. The final sample analysis included only respondents who were selected in 2012 and remained alive in 2014. The time variable utilized in this research was the number of years between the age at the end of 2014 (age at entrance into the current study cohort) and the age at death, censoring, or by December 31st, 2020. Kaplan-Meier survival curves were employed across food insecurity status (binary), HEI-2015 (three-level version) and dementia status groups, as well as a log-rank test, to assess differences in survival functions.

Third, racial/ethnic- and sex-stratified Cox proportional hazards (PH) models were conducted to assess the link between food insecurity, dementia, and all-cause mortality in both reduced (Model 1) and fully adjusted (Model 2). Standardized Schoenfeld residuals were used to assess the assumptions of proportionality in the hazards in the Cox models. Model 1 included age in 2012, sex, and race/ethnicity, while Model 2 further accounted for education, total wealth, marital status, smoking status, physical activity, self-rated health, body mass index categories, cardiometabolic risk groups, and CES-D total scores. Those covariates were selected for these two models after including them in a directed acyclic graph in addition to the main exposures, mediators and outcomes of interest (See Supplementary methods 3 on github repository: https://github.com/baydounm/HRS_FI_DIET_DEM_MORT). The primary exposures were the food insecurity ordinal metric and each “dementia probability” [Ln odds (Pr/(1-Pr) converted; Hurd, Expert, and LASSO] (Analysis A) for models 1 and 2, an approach used in previous studies^{3, 4}. Food insecurity, as an ordinal variable, and Ln(dementia odds) relationships with mortality risk were tested for heterogeneity of effects by sex and race/ethnicity, through the inclusion of 2-way interaction terms. Note that a 1 unit increase in Ln(odds) is equivalent to going from a probability of 50% to 73.1%. A comparable analysis (Analysis B) was conducted using binary dementia status variables and food insecurity status as primary exposures.

Fourth, to test the relationship between Ln(dementia odds) in Analysis A or dementia status (Analysis B) with mortality risk across food insecurity status groups, an additional set of models (reduced—i.e., Model 1—and fully adjusted—i.e., Model 2) were used, with a similar analytic process. Using Cox PH models, each dementia exposure was entered separately with mortality of all causes as the outcome of interest, while stratifying by food insecurity status group. Food insecurity status group were examined for heterogeneity in an unstratified model with 2-way interaction terms for each separate exposure. Steps 1, 3 and 4 adjusted analyses for sampling design complexity.

Fifth, the present study used four-way decomposition models, applied to the overall sample and stratifying by sex and race/ethnicity, by dividing total effects of food insecurity (or diet quality) on all-cause mortality risk into effects due purely to mediation, purely to interaction, both or neither. Using the method presented by Discacciati et al. (2019), a Stata command Med4way was implemented which facilitates four-way effect decomposition for the analysis of mediation and interaction processes in causal inference. The command disaggregates the overall impact of an exposure on an outcome into four elements: Controlled Direct Effect (CDE), Pure Mediation Effect, Pure Interaction Effect, and Mediated Interaction Effect. This approach enables researchers to evaluate both mediating and interactive effects, facilitating the comprehension of intricate mechanisms when mediation and interaction coexist ⁶⁷, with Stata med4way command available at⁶⁸. The implementation of this method in the context of the present study, included alternating between several potential mediators/moderators, namely the Ln(dementia odds, z-scored) (i.e., Hurd, expert, and LASSO), in the total effect (TE) of food insecurity (or diet quality) on mortality risk from any cause during the follow-up period spanning 2014 through 2020. Exposures, mediators/moderators were transformed into standardized z-scores where applicable, with the possible exception of food insecurity which was entered as a binary variable (0=no, 1=yes). Note that 1 SD increase in each of the algorithmically defined dementia outcomes was equivalent to going from a 50% probability to a ~89%−96% probability, or Ln(odds) increase by 2.1–3.2 (Data not shown). In these four-way decomposition models, Cox PH regression was the final outcome equation, with outcome being all-cause mortality risk. OLS regression for each mediator/moderator as the outcome was conducted to assess the relationship with the main exposure. The full sample was subjected, in the main part of the analysis, to four-way decomposition models with a reduced set of exogenous factors, namely age estimated by end of 2012, sex, and race/ethnicity, with additional models tested stratifying by sex and race/ethnicity, respectively. A secondary analysis was conducted that adjusted for all exogenous covariates, including all socio-economic, lifestyle and health-related factors listed in the Covariates section. When main exposure was food insecurity, diet quality was included among exogenous variables in the fully adjusted model and vice versa.

Finally, the study conducted several generalized structural equations (GSEM) models to examine the mediating effects of diet quality and dementia in the putative association between food insecurity and all-cause mortality, allowing food insecurity to be an antecedent predictor for dietary quality. Specifically, GSEM were tested to explore mediating effects of diet quality and the Hurd dementia odds (Ln transformed) in the putative association between food insecurity and mortality risk, focusing on the reduced model that was adjusted for age, sex and race/ethnicity along with the Hurd algorithm which is more commonly used in various HRS studies^{51, 54}. Weibull models were used for the final equation with mortality risk as the outcome, as was done in a previous study ⁶⁹, while all other equations consisted of linear regression models. Selected mediating effects from the reduced model with two serial mediators were also tested, between food insecurity, diet quality, dementia and mortality, namely: (A) Food insecurity →diet quality → dementia → mortality; (B) food insecurity → dementia → mortality; (C) food insecurity → diet quality → mortality; and (D) Diet quality → dementia → mortality. A type I error of 0.05 was set for all analyses, with the exception of the 2-way interaction terms, which were set to 0.10 ⁷⁰. Data were analyzed with Stata 18.0 (StataCorp, College Station, TX)⁷¹, and the full Stata script along with Output is provided at: https://github.com/baydounm/HRS_FI_DIET_DEM_MORT.

RESULTS

Descriptive findings

Descriptive findings by sex and race/ethnicity are presented in Table 1. Our study analyzed 2,894 participants, with 1,202 male and 1,692 female, 2,367 Non-Hispanic White, while 527 were “Non-Hispanic Black/Hispanic”. The mean age was 76.4 y, and the all-cause mortality rate was estimated at 51 deaths per 1000 person-years. Male older adults were more likely to be married or partnered, currently working, have completed a college degree or higher, and less likely to have very low total wealth ($) <25,000, compared to females. Lifestyle factors’ distributions were inconsistent by sex, with females having a higher proportion of never smokers, alcohol abstinent, and coupled with better overall dietary quality but reduced self-reported physical activity level compared to males. Healthy weight status was a less likely status among male adults compared to females. Male older adults were more likely to report ≥3 cardiometabolic factors and conditions compared to females, while female older adults reported on average more depressive symptoms. No sex differences were found in terms of self-rated health. 1 SD of HEI-2015 within the selected sample was equivalent to 9.42-point increase (Data not shown). Furthermore, 1 SD for each Ln transformed dementia odds was ~3.2, 2.9, and 2.1, for Hurd, Expert and LASSO algorithms, respectively (Data not shown).

Table 1.

Study sample characteristics: overall, by sex and by race/ethnic groups: Health and Retirement Study 2012–2020^a,b

	Overall	Males	Females	Non-Hispanic White	Non-Hispanic Black/Hispanic	Sex Differences P-value	Race/ethnicity Differences P-value
	Mean/%±Standard Error (SE)	Mean/%±SE	Mean/%±SE	Mean/%±SE	Mean/%±SE
	N=2,894	N=1,202	N=1,692	N=2,367	N=527
SOCIODEMOGRAP
HIC, 2012
Sex:
Male	43.3±1.0	100.0±0.0	0.0±0.0	44.0±1.1	39.1±2.0	--	0.049
Female	56.7±1.0	0.0±0.0	100.0±0.0	56.0±1.1	60.9±2.0	--	--
Age (years):
Mean ± SE	76.4±0.2	75.8±0.2	76.8±0.3	76.4±0.3	76.2±0.4	<0.001	0.58
Non-Hispanic	13.9±1.2	12.6±1.2	14.9±1.3	0.0±0.0	100.0±0.0	0.049	--
Black/Hispanic, %
Education:
No degree	16.8±1.0	15.5±1.2	17.8±1.2	12.2±0.9	45.1±3.1	0.77	<0.001
GEDc	4.0±0.4	4.7±0.6	3.4±0.5	3.9±0.5	4.2±1.0	0.010	0.099
High School graduate	35.1±1.1	31.7±1.7	37.6±1.3	36.8±1.3	24.6±2.1	--	--
Some college	20.0±0.7	17.1±1.4	22.1±1.0	20.6±0.8	15.9±1.9	0.54	0.417
College degree or higher	24.2±1.2	31.0±1.7	19.0±1.3	26.5±1.4	10.2±1.5	<0.001	0.009
Marital status:
Never married	3.1±0.4	2.9±0.6	3.3±0.6	2.6±0.6	6.2±1.6	0.009	0.007
Married/ Partnered	57.6±.3	77.3±1.6	42.5±1.4	59.8±1.3	43.4±2.6	--	--
Separated/ Divorced	10.4±0.7	7.6±1.0	12.5±1.0	9.1±0.7	17.9±2.0	<0.001	<0.001
Widowed	28.9±1.4	12.2±1.3	41.7±1.6	28.4±1.5	32.5±2.3	<0.001	<0.001
Working status:
Not Working	81.7±0.9	75.1±1.6	86.7±1.0	81.3±0.9	84.2±2.4	--	--
Working	18.3±0.9	24.8±1.6	13.3±1.0	18.7±0.9	15.8±2.4	<0.001	0.26
Total wealth ($):
< 25,000	32.6±1.3	21.4±1.3	41.2±1.8	27.7±1.2	63.3±3.1	<0.001	<0.001
25,000–124,999	59.2±1.2	66.9±1.4	53.3±1.6	63.1±1.2	34.7±2.9	--	--
125,000–299,999	6.3±0.6	9.1±1.0	4.1±0.6	7.0±0.7	1.9±0.7	<0.001	0.039
300,000–649,999	1.7±0.3	2.3±0.5	1.2±0.3	1.9±0.4	0.0±0.0	0.058	<0.001
≥ 650,000	0.3±0.1	0.3±0.2	0.2±0.1	0.3±0.1	0.0±0.0	0.61	<0.001
LIFESTYLE, 2012
Smoking status:
Never smoker	46.0±1.1	32.9±1.4	56.1±1.5	45.9±1.3	46.9±2.7	<0.001	0.32
Past smoker	46.4±1.1	59.5±1.7	36.5±1.3	47.0±1.2	42.6±2.4	--	--
Current Smoker	7.6±0.7	7.7±1.1	7.5±0.7	7.1±0.7	10.5±1.9	0.013	0.024
Physical activity:
Never	21.8±0.9	16.5±1.4	25.8±1.3	20.8±1.0	28.1±2.2	<0.001	<0.001
1–4 times per month	25.6±0.9	25.6±1.5	25.6±1.2	24.8±1.0	30.3±2.0	0.10	<0.001
>1 times per week	52.6±1.2	57.9±1.6	48.6±1.5	54.4±1.3	41.6±2.2	--	--
Alcohol consumption:
Abstinent	51.2±1.2	42.2±1.8	58.0±1.5	49.2±1.4	63.1±2.5	--	--
1–3 days per month	16.2±0.8	16.0±1.3	16.3±1.0	15.5±0.8	20.5±2.2	0.020	0.82
1–2 days per week	18.0±0.8	21.5±1.1	15.4±1.0	19.0±0.8	12.1±1.5	<0.001	<0.001
≥3 days per week	14.7±0.9	20.3±1.3	10.4±0.9	16.3±1.0	4.3±0.9	<0.001	<0.001
HEALTH RELATED, 2012
Self-rated health:
Excellent/very good/	74.6±1.0	75.0±1.6	74.5±1.4	77.3±1.2	58.7±2.4	--	--
good
Fair/poor	25.3±1.0	25.0±1.6	25.5±1.4	22.7±1.2	41.3±2.4	0.82	<0.001
Body mass index (kg/m2):
<25	33.2±1.0	27.0±1.5	37.9±1.2	34.1±1.0	27.8±2.4	<0.001	0.15
25–29.9	38.4±0.9	44.1±1.5	34.0±1.2	38.5±0.9	37.7±2.6	--	--
≥30	28.4±1.0	29.0±1.8	28.0±1.4	27.4±1.1	34.6±2.9	0.073	0.093
Number of chronic conditionsd
0	22.8±0.8	20.4±1.6	25.6±1.3	23.7±0.9	16.9±1.8	0.17	0.004
1–2	65.9±0.9	66.2±1.7	65.7±1.3	65.2±0.9	70.4±2.2	--	--
≥3	11.3±0.7	13.4±0.9	9.8±0.9	11.1±0.7	12.6±1.6	0.019	0.75
CES-De total score,	1.17±0.05	0.92±0.07	1.35±0.05	1.12±0.05	1.49±0.12	<0.001	0.005
Mean±SE
HEI-2015f (2013 wave),	69.7±0.2	68.9±0.4	70.3±0.2	69.4±0.3	71.0±0.5	<0.001	0.011
Mean±SE
Food insecurityg (2013 wave), %f	9.4±0.8	7.5±0.9	10.8±1.1	6.9±0.7	24.5±2.5	<0.001	<0.001
MORTALITY RATE, 2014–2020:
#death per 1000 P-Y	51	58	47	53	44	--	--
with 95% CI	(48–55)	(52–64)	(43–51)	(49–57)	(37–52)
Dementia Status
Hurd:
No	85.7±0.7	87.3±0.8	84.5±1.1	86.4±0.8	81.6±2.5	--	--
Yes	14.3±0.7	12.7±0.8	15.5±1.1	13.6±0.8	18.3±2.5	0.036	0.062
Expert:
No	84.6±0.8	86.5±1.1	83.1±1.2	85.2±0.8	80.8±2.0	--	--
Yes	15.4±0.8	13.5±1.1	17.0±1.2	14.8±0.8	19.1±2.0	0.053	0.019
Least Absolute Shrinkage and Selection Operator (LASSO):
No	84.7±0.8	87.6±1.0	82.4±1.1	85.9±0.8	77.2±2.5	--	--
Yes	15.3±0.8	12.4±1.0	17.6±1.2	14.1±0.8	22.8±2.5	0.002	<0.001
Dementia Probability
Hurd:
Mean±SE	0.105±0.005	0.092±0.006	0.115±0.008	0.097±0.005	0.157±0.016	0.013	<0.001
Expert:
Mean±SE	0.135±0.006	0.115±0.006	0.150±0.009	0.125±0.005	0.198±0.017	0.003	<0.001
Lasso:
Mean±SE	0.132±0.006	0.112±0.006	0.147±0.009	0.124±0.005	0.179±0.016	0.001	0.001

^aValues are means ± SE column percentages for overall and fixed sample characteristics, across sex and race/ethnicity groups, accounting for sampling design complexity. All covariates are measured in 2012.

^bBased on linear or multinomial logit models using sex or race/ethnicity as main predictors for both continuous and categorical variables, accounting for sampling design complexity. In multinomial logit models, the largest category was automatically chosen as the referent category. P-values marked with – are missing because the category is chosen as the referent in the multinomial logit model.

^cGED=General Education Diploma.

^dNumber of chronic conditions among hypertension, diabetes, heart problems and stroke.

^eCES-D=Center for Epidemiologic Studies-Depression

^fHEI-2015=Healthy Eating Index-2015

^gFood insecurity score had a mean of 0.397 and a standard deviation of 1.17 in the unweighted and unimputed sample. Only the binary version is presented here.

Similarly, food insecurity, algorithmically-defined dementia status and probabilities were higher among females, compared with males. Table 1 also presents findings regarding racial/ethnic disparities in these study characteristics. Most notably, food insecurity was significantly more prevalent among minority groups (i.e. Non-Hispanic Black and Hispanic combined group); (24.5% vs. 6.9%, p<0.001) compared with their Non-Hispanic White counterparts, as were dementia probabilities. In contrast, HEI-2015 was higher among minority groups compared with Non-Hispanic White older adults. In additon, major racial/ethnic disparities were observed with respect to total wealth, educational attainment and several lifestyle and health-related factors, including poor self-rated health.

Food insecurity, dietary quality, dementia status and their association with mortality risk: K-M curve and log-rank tests

Crude associations of food insecurity, dementia status and dietary quality with survival are presented in Figures 2 and 3. Figure 2 depicts survival probabilities by food insecurity status and algorithmically defined dementia status (3 algorithms). Food insecurity was inversely related to mortality, while across all algorithms, dementia was strongly and positively associated with all-cause mortality risk (Log-rank test, P<0.001 for dementia, and P<0.05 for food insecurity). Similarly, Figure 3 illustrates an inverse association between diet quality and mortality risk, using the HEI-2015 tertile groupings, also based on a log-rank test (P<0.001).

Figure 2. Food insecurity, dementia status and their associations with all-cause mortality: Health and Retirement Study 2012–2020, K-M survival curves^a,b.

^a Survival probability on the y-axis plotted against food insecurity (A), dementia status, panels (B)-(D) for Hurd, Expert and LASSO algorithms, respectively and follow-up time (years) on the x-axis. ^b Log-rank test also presented for each analysis. ^c chi2=Chi-square test

Figure 3. Diet quality (Healthy Eating Index-2015 or HEI-2015 total score, tertile version) and its association with all-cause mortality: Health and Retirement Study 2012–2020, Kaplan-Meier survival curves^a,b.

^a HEI-2015 was grouped tertiles using the largest available sample: T1: 28.8–65.59 (mean±SD: 58.6±5.8), T2: 65.6–73.99 (69.7±2.4), T3: 74.0–97.2 (79.8±4.6). 1 standard deviation of HEI-2015 within the selected sample was equivalent to 9.42-point increase.

^b Survival probability on the y-axis plotted against HEI-2015 groupings and follow-up time (years) on the x-axis.

^c Log-rank test (chi2(2)) also presented.

Cox PH models for food insecurity, dementia status and mortality risk

Table 2 shows findings from Cox PH models that included “Food insecurity” as the main exposure and the outcome being all-cause mortality risk, by displaying adjusted Ln(HR) with SE on multiple imputed data. The models indicated that there was generally no direct association between food insecurity and mortality risk in the overall sample (Analysis A, continuous exposure: β±SE: 0.033±0.051, reduced model; β±SE: −0.063±0.052, full model; Analysis B, binary exposure: β±SE: −0.014±0.190, reduced model, with an estimated HR =0.99, 95% CI:0.68–1.43). Nevertheless, an inverse association was detected in the fully adjusted models when food insecurity was entered as a binary variable, overall (β±SE: −0.433±0.178, with an estimated HR=0.65, 95%CI:0.46–0.92, P<0.05), particularly among males (estimated HR=0.57, 95% CI:0.33–0.95, P<0.050) and Non-Hispanic White (estimated HR=0.63, 95% CI: 0.42–0.94, P<0.050) older adults. In reduced models, heterogeneity by sex was noted for the association between Ln(dementia odds, unstandardized) and mortality risk, suggesting a stronger association among males (per unit, HR=1.20, 95% CI:1.16–1.24, P<0.001 for Hurd algorithm, Analysis A, Model 1B), compared to their female counterparts (per unit, HR=1.10, 95% CI:1.07–1.14, P<0.001 for Hurd algorithm, Analysis A, Model 1B). This interaction by sex (P=0.001) was attenuated in the fully adjusted model (P=0.051).

Table 2.

Food insecurity, dementia odds (Ln transformed, Analysis A) or dementia status (Analysis B) and all-cause mortality: Cox PH^a models, overall, by sex and by race/ethnicity, Health and Retirement Study 2012–2020^b,c

	Overall	Male	Female	Non-Hispanic White	Non-Hispanic Black/Hispanic	Psexd	Praced
	β±SEe	β±SE	β±SE	β±SE	β±SE
Unweighted N	N=2,886	N=1,199	N=1,687	N=2,360	N=526
Analysis A, continuous exposures
Reduced models
Model 1A: Food insecurity	0.033±0.051	0.037±0.063	0.034±0.064	0.043±0.061	+0.012±0.074	0.81	0.68
Model 1B: Hurd dementia	0.109±0.015 ***	0.186±0.017 ***	0.097±0.016 * **	0.106±0.015 ***	0.168±0.033 ***	0.001	0.080
Model 1C: Expert dementia	0.185±0.020 ***	0.241±0.028 ***	0.159±0.020 * **	0.185±0.020 ***	0.188±0.062 **	0.003	0.82
Model 1D: LASSO dementia f	0.223±0.020 ***	0.279±0.025 ***	0.194±0.023 * **	0.223±0.021 ***	0.215±0.082 *	0.003	0.84
Full models	N=2,804	N=1,149	N=1,655	N=2,302	N=502
Model 2A: Food insecurity	−0.063±0.052	- 0.081±0.070	−0.051±0.064	- 0.065±0.059	−0.018±0.092	0.82	0.77
Model 2B: Hurd dementia	0.095±0.018 ***	0.152±0.033 ***	0.089±0.021 * **	0.090±0.018 ***	0.205±0.073 ***	0.051	0.22
Model 2C: Expert dementia	0.139±0.024 ***	0.154±0.034 ***	0.134±0.027 * **	0.133±0.025 ***	0.203±0.053 ***	0.15	0.34
Model 2D: LASSO dementia	0.187±0.027 ***	0.205±0.035 ***	0.181±0.033 * **	0.179±0.029 ***	0.251±0.090 **	0.16	0.42
Analysis B, binary exposures
Reduced models	N=2,886	N=1,199	N=1,687	N=2,360	N=526
Model 1A: Food insecurity	−0.014±0.190	−0.100±0.241	0.054±0.235	0.007±0.222	−0.056±0.274	0.72	0.80
Model 1B: Hurd dementia	0.756±0.113 ***	0.798±0.176 ***	0.726±0.124 * **	0.791±0.114 ***	0.553±0.318	0.64	0.38
Model 1C: Expert dementia	0.930±0.103 ***	1.052±0.138 ***	0.835±0.123 * **	0.915±0.108 ***	1.09±0.297 c	0.15	0.84
Model 1D: LASSO dementia	0.869±0.107 ***	1.070±0.139 ***	0.721±0.127 * **	0.894±0.111 ***	0.722±0.304 *	0.013	0.39
Full models	N=2,804	N=1,149	N=1,655	N=2,302	N=502
Model 2A: Food insecurity	−0.433±0.178 *	−0.570±0.267 *	−0.303±0.210	−0.465±0.206 *	−0.290±0.348	0.75	0.86
Model 2B: Hurd dementia	0.481±0.136 **	0.443±0.184 **	0.510±0.145 * **	0.457±0.138 **	0.587±0.335	0.91	0.72
Model 2C: Expert dementia	0.708±0.100 ***	0.673±0.127 ***	0.736±0.143 * **	0.636±0.111 ***	1.216±0.250 ***	0.72	0.070
Model 2D: LASSO dementia	0.624±0.107 ***	0.698±0.118 ***	0.581±0.151 * **	0.585±0.115 ***	0.865±0.308 **	0.12	0.48

^aPH=Proportional hazards.

^bValues are β±SE, with β representing Ln HR from Cox PH model for each exposure-outcome relationship. Cox PH models were conducted overall and across sex or race/ethnicity groups, accounting for sampling weights and sampling design complexity in multiple imputed data. All covariates are measured in 2012 unless stated otherwise. Analysis A included continuous forms of dementia probability (Ln (dementia odds) and ordinal Food insecurity score, while Analysis B included binary dementia status (3 algorithms) and Food insecurity status group. Hazard Ratios with their 95% confidence intervals can be obtained from β±SE through exponentiation as follows: Hazard Ratio (HR) point estimate=exp(β), Lower Confidence Limit of HR=exp(β−1.96×SE), Upper Confidence Limit of HR=exp (β+1.96×SE).

^cReduced models (Models 1A-1D) were adjusted for age in 2012, sex and race/ethnicity. Full models (Model 2A-2D) further adjusted the reduced model by all covariates described under the Covariates section, including socio-demographic, socio-economic, lifestyle (including diet quality), and health-related factors.

^dP-value associated with 2-way interaction term between sex or race/ethnicity and the main exposure (food insecurity score or dementia odds), in a model not stratified by sex or race/ethnicity.

^eSE=Stand Error

^fLASSO= Least Absolute Shrinkage and Selection Operator

^*P<0.05

^**P<0.010

^***P<0.001 for null hypothesis that β=0.

Table 3 displays findings from Cox PH models for the associations of algorithmically-defined dementia status and Ln(dementia odds, unstandardized) with mortality risk, as stratified by food insecurity status. The results indicated that dementia status and Ln(dementia odds, unstandardized) were both associated with mortality risk, with no detectable interaction by food insecurity status (P>0.10 for interaction term in separate model between food insecurity and each Ln(dementia odds) exposure).

Table 3.

Dementia odds (Ln transformed) (Analysis A) or dementia status (Analysis B) and all-cause mortality across food insecurity status: Cox PH^a models, Health and Retirement Study 2012–2020^b,c

	No β±SEe	Yes β±SE	Pfood insecurityd
Analysis A, continuous exposures
Unweighted N
Reduced models	N=2,603	N=283
Model 1A: Hurd dementia	0.107±0.015 ***	0.187±0.067 **	0.90
Model 1B: Expert dementia	0.187±0.021 ***	0.192±0.060 **	0.42
Model 1C: LASSOf dementia	0.221±0.020 ***	0.271±0.084 **	0.79
Full models	N=2,529	N=275
Model 2A: Hurd dementia	0.096±0.018 ***	0.130±0.097	0.12
Model 2B: Expert dementia	0.140±0.025 ***	0.171±0.078 *	0.10
Model 2C: LASSO dementia	0.186±0.028 ***	0.231±0.104 *	0.19
Analysis B, binary exposures
Unweighted N
Reduced models	N=2,603	N=283
Model 1A: Hurd dementia	0.785±0.115 ***	0.624±0.337	0.24
Model 1B: Expert dementia	0.944±0.112 ***	0.996±0.266 ***	0.35
Model 1C: LASSO dementia	0.871±0.117 ***	0.949±0.324 **	0.56
Full models	N=2,529	N=275
Model 2A: Hurd dementia	0.504±0.127 ***	0.487±0.106 ***	0.18
Model 2B: Expert dementia	0.685±0.117 ***	1.228±0.346 **	0.68
Model 2C: LASSO dementia	0.603±0.117 ***	0.845±0.374 *	0.66

^aPH=proportional hazards

^bValues are β±SE, with β representing Ln(HR) from Cox PH model for each exposure-outcome relationship. Cox PH models were conducted on the overall sample, accounting for sampling weights and sampling design complexity in multiple imputed data. All covariates are measured in 2012 unless stated otherwise. Analysis A included continuous forms of dementia probability (Ln (dementia odds); while Analysis B included binary dementia status (3 algorithms).

^cReduced models (Models 1A-1C) were adjusted for age in 2012, sex and race/ethnicity. Full models (Model 2A-2C) further adjusted the reduced model by all covariates described under the Covariates section, including socio-demographic, socio-economic, lifestyle and health-related factors.

^dP-value associated with 2-way interaction term between food insecurity (yes vs. no) and the main exposure Ln(dementia odds), in a model not stratified by food insecurity score status group.

^eSE=Standard Error

^fLASSO=Least Absolute Shrinkage and Selection Operator

^*P<0.05

^**P<0.010

^***P<0.001 for null hypothesis that β=0.

Four-way decomposition model findings: testing moderating and mediating effects

Tables 4 and 5 test moderating and mediating effects using a series of four-way decomposition models with final outcome being all-cause mortality risk. Table 4 shows findings from a four-way decomposition of the TE of food insecurity on mortality risk through dementia odds (Ln transformed, z-scored), displaying both reduced and fully adjusted models, for 3 algorithms, and stratiying by sex and by race/ethnicity. In the reduced models, TE of food insecurity was not detected at type I error of 0.05. Nevertheless, a PIE was found whereby food insecurity was positively associated with mortality risk through Ln(dementia odds, z-scored). This PIE was attenuated in the fully adjsuted model and the TE of food insecurity was <0 in these models. This inverse TE was mainly found among female adults.

Table 4.

Food insecurity and all-cause mortality: four-way decomposition models by dementia odds (Ln transformed, z-scored), overall, by sex and by race/ethnicity: Health and Retirement Study 2012–2020^a,b

	TEc		CDEd		INTREFe		INTMEDf		PIEg
Yi=All-cause mortality risk; Xj=Food insecurity (1=yes, 0=no)	β±SEh	P	β±SE	P	β±SE	P	β±SE	P	β±SE	P
Overall
Reduced Model 1A-C (N=2,894)
Mk=Hurd	+0.008		−0.074	0.57	−0.005		−0.009		+0.096
	±0.127	0.95	±0.129		±0.030	0.88	±0.035	0.79	±0.023	<0.001
M=Expert	−0.0003	1.00	−0.050	0.70	−0.058		−0.076	0.05	+0.183
	±0.1266		±0.126		±0.027	0.032	±0.039	0	±0.040	<0.001
M=LASSOl	−0.0049	0.97	−0.072	0.56	−0.033		−0.040		+0.140
	±0.1260		±0.123		±0.028	0.25	±0.032	0.22	±0.033	<0.001
Full Model 2A-C (N=2,812)
M=Hurd	−0.270	0.011	−0.259	0.016	−0.014		−0.014		+0.017
	±0.106		±0.107		±0.006	0.016	±0.014	0.34	±0.017	0.30
M=Expert	−0.244	0.027	−0.216	0.057	−0.033		−0.012		+0.017
	±0.110		±0.114		±0.011	0.002	±0.016	0.45	±0.023	0.44
M=LASSO	−0.265	0.012	−0.240	0.026	−0.028		−0.006		+0.009
	±0.106		±0.108		±0.010	0.003	±0.015	0.69	±0.022	0.68
Male
Reduced Model 1A-C (N=1,202)
M=Hurd	+0.014		+0.077	0.70	−0.100		−0.096		+0.133
	±0.194	0.94	±0.203		±0.035	0.004	±0.064	0.13	±0.056	0.019
M=Expert	+0.003	0.99	0.080	0.69	−0.127		−0.096		+0.145
	±0.193		±0.197		±0.043	0.003	±0.063	0.12	±0.070	0.037
M=LASSO	−0.0003	1.00	+0.032	0.86	−0.080		−0.067		+0.114
	±0.1924		±0.192		±0.032	0.012	±0.053	0.20	±0.061	0.063
Full model 2A-C (N=1,152)
M=Hurd	−0.124		−0.054	0.78	−0.069		−0.010		0.009
	±0.192	0.52	±0.194		±0.022	0.002	±0.045	0.82	±0.041	0.83
M=Expert	−0.142	0.45	−0.092	0.63	−0.049		−0.001		0.001
	±0.189		±0.191		±0.019	0.008	±0.036	0.99	±0.038	0.99
M=LASSO	−0.176	0.33	−0.152	0.39	−0.021		0.012		−0.015
	±0.179		±0.178		±0.017	0.20	±0.033	0.72	±0.040	0.71
Female
Reduced Model 1A-C (N=1,692)
M=Hurd	+0.0003		−0.157	0.33	+0.032		+0.030		+0.094
	±0.168	1.00	±0.162		±0.048	0.50	±0.049	0.54	±0.028	0.001
M=Expert	−0.022	0.39	−0.140	0.38	−0.025		−0.052		+0.194
	±0.165		±0.162		±0.034	0.46	±0.050	0.30	±0.047	<0.001
M=LASSO	−0.026	0.87	−0.159	0.31	−0.002		−0.013		+0.147
	±0.164		±0.157		±0.040	0.97	±0.044	0.77	±0.039	<0.001
Full model 2A-C (N=1,660)
M=Hurd	−0.329		−0.335	0.012	−0.003		−0.011		+0.020
	±0.130	0.011	±0.133		±0.005	0.63	±0.013	0.41	±0.020	0.33
M=Expert	−0.300	0.030	−0.290	0.042	−0.021	0.12	−0.013	0.42	+0.025	0.39
	±0.138		±0.143		±0.013		±0.017		±0.029
M=LASSO	−0.317	0.018	−0.300	0.030	−0.027		−0.011		+0.020
	±0.134		±0.138		±0.015	0.072	±0.016	0.48	±0.027	0.46
Non-Hispanic White
Reduced Model 1A-C (N=2,367)
M=Hurd	+0.016		−0.032	0.84	−0.017		−0.030		+0.094
	±0.158	0.92	±0.158		±0.023	0.47	±0.047	0.52	±0.029	0.001
M=Expert	−0.001		−0.033	0.83	−0.061		−0.1187	0.04	+0.211
	±0.156	1.00	±0.155		±0.021	0.004	±0.058	0	±0.050	<0.001
M=LASSO	−0.004		−0.038	0.81	−0.045		−0.062		+0.140
	±0.156	0.98	±0.152		±0.025	0.073	±0.044	0.16	±0.042	0.001
Full model 2A-C (N=2,309)
M=Hurd	−0.291		−0.295	0.015	+0.005		−0.007		+0.006
	±0.126	0.021	±0.121		±0.015	0.75	±0.022	0.76	±0.020	0.76
M=Expert	−0.257	0.052	−0.246	0.062	−0.013		−0.016		+0.017
	±0.132		±0.132		±0.012	0.29	±0.025	0.54	±0.028	0.53
M=LASSO	−0.285	0.024	−0.270	0.031	−0.014		+0.002		−0.003
	±0.126		± 0.126		±0.010	0.17	±0.023	0.92	±0.027	0.92
Non-Hispanic Black/Hispanic
Reduced Model 1A-C (N=527)
M=Hurd	−0.030		+0.114	0.56	−0.032		−0.020		+0.136	0.039
	±0.211	0.89	±0.196		±0.099	0.75	±0.051	0.70	±0.066
M=Expert	−0.030		−0.114	0.56	−0.032		−0.020		+0.136	0.039
	±0.210	0.89	±0.196		±0.099	0.75	±0.051	0.70	±0.066
M=LASSO	−0.030		−0.175	0.36	+0.008		−0.0002		+0.138	0.017
	±0.210	0.89	±0.192		±0.080	0.92	±0.057	1.00	±0.058
Full model 2A-C (N=503)
M=Hurd	−0.212		−0.196	0.36	−0.044		−0.013		+0.041
	±0.200	0.29	±0.213		±0.076	0.56	±0.023	0.56	±0.042	0.33
M=Expert	−0.221	0.36	−0.189	0.36	−0.039		−0.001		−0.002
	±0.198		±0.206		±0.092	0.67	±0.012	0.88	±0.012	0.88
M=LASSO	−0.227	0.25	−0.189	0.36	−0.053		+0.008		+0.023
	±0.198		±0.205		±0.066	0.43	±0.018	0.65	±0.044	0.61

^aCox PH regression models with mortality as the main outcome and food insecurity, binary exposure (yes vs. no), as the exposure. Ln[dementia odds), z-scored, using Hurd, expert and LASSO algorithms were potential mediators/moderators allowed to interact with the main exposure, sample size N=2,812 (full model) and 2,894 (reduced model), four-way decomposition analysis. 1 SD for each Ln transformed dementia odds was ~3.2, 2.9, and 2.1, for Hurd, Expert and LASSO algorithms, respectively. 1 SD of HEI-2015 within the selected sample was equivalent to 9.42-point increase. Total effect and controlled direct effects are interpreted as Ln(hazard ratios) comparing food insecure vs. secure.

^bReduced models (Models 1A-1C) were adjusted for age in 2012, sex and race/ethnicity. Full models (Model 2A-2C) further adjusted the reduced model by all covariates described under the Covariates section, including socio-demographic, socio-economic, lifestyle and health-related factors. Diet quality was included among exogenous variables in the full model, as HEI-2015, z-scored.

^cTE=Total Effect

^dCDE=Controlled Direct Effect

^eINTREF=Interaction Referent

^fINTMED=Mediated Interaction

^gPIE=Pure Indirect Effect

^hSE=Standard Error

ⁱY=Outcome

^jX=Exposure

^kM=Mediator/Moderator

^lLASSO= Least Absolute Shrinkage and Selection Operator

Table 5.

Diet quality and all-cause mortality: four-way decomposition models by dementia odds (Ln transformed, z-scored), overall, by sex and by race/ethnicity: HRS 2012–2020^a,b

	TEc		CDEd		INTREFe	INTMEDf	PIEg
Yi=All-cause mortality; Xj=Diet quality (HEI-2015, z-scored)	β±SEh	P	β±SE	P	β±SE	P	β±SE	P	β±SE	P
Overall
Reduced Model 1A-C (N=2,894)
Mk=Hurd	−0.161		−0.131	<0.001	0.001±		+0.002		−0.033
	±0.029	<0.001	±0.029		0.007	0.92	±0.002	0.42	±0.006	<0.001
M=Expert	−0.171	<0.001	−0.130	<0.001	+0.017		−0.001		−0.057
	±0.030		±0.029		±0.012	0.18	±0.004	0.76	±0.009	<0.001
M=LASSOl	−0.164	<0.001	−0.129	<0.001	+0.020		−0.003		−0.053
	±0.030		±0.028		±0.012	0.077	±0.003	0.41	±0.008	<0.001
Full Model 2A-C (N=2,812)
M=Hurd	−0.093	0.006	−0.083	0.016	−0.001		+0.000		−0.012
	±0.034		±0.034		±0.005	0.76	±0.001	0.86	±0.005	0.011
M=Expert	−0.107	0.002	−0.094	0.006	+0.008		−0.001		−0.021
	±0.034		±0.034		±0.007	0.26	±0.002	0.63	±0.007	0.002
M=LASSO	−0.100	0.003	−0.089	0.008	+0.010		−0.001		−0.020
	±0.034		±0.034		±0.007	0.16	±0.002	0.43	±0.006	0.002
Male
Reduced Model 1A-C (N=1,202)
M=Hurd	−0.117		−0.095	0.026	+0.035		−0.006		−0.051
	±0.047	0.014	±0.043		±0.025	0.17	±0.005	0.30	±0.014	<0.001
M=Expert	−0.124	0.009	−0.073	0.10	+0.004		+0.001		−0.054
	±0.047		±0.045		±0.022	0.87	±0.005	0.83	±0.016	<0.001
M=LASSO	−0.121	0.010	−0.082	0.066	+0.021		−0.003		−0.056
	±0.047		±0.045		±0.020	0.30	±0.005	0.55	±0.015	<0.001
Full model 2A-C (N=1,152)
M=Hurd	−0.030		−0.043	0.44	+0.027		−0.002		−0.012
	±0.056	0.59	±0.055		±0.021	0.20	±0.003	0.41	±0.011	0.30
M=Expert	−0.024	0.67	−0.028	0.62	+0.013		−0.001		−0.008
	±0.056		±0.056		±0.015	0.38	±0.002	0.57	±0.010	0.44
M=LASSO	−0.025	0.66	−0.029	0.61	+0.017		−0.002		−0.011
	±0.056		±0.056		±0.013	0.21	±0.002	0.43	±0.011	0.34
Female
Reduced Model 1A-C (N=1,692)
M=Hurd	−0.198		−0.172	<0.001	+0.008		+0.000		−0.033
	±0.037	<0.001	±0.038		±0.009	0.39	±0.004	0.97	±0.007	<0.001
M=Expert	−0.209	0.014	−0.172	<0.001	+0.027		−0.004		−0.060
	±0.038		±0.037		±0.016	0.084	±0.005	0.47	±0.011	<0.001
M=LASSO	−0.200	<0.001	−0.172	<0.001	+0.028		−0.005		+0.052
	±0.038		±0.036		±0.016	0.078	±0.005	0.34	±0.010	<0.001
Full model 2A-C (N=1,660)
M=Hurd	−0.138		+0.127	0.004	+0.003		+0.000		−0.014
	±0.043	0.001	±0.044		±0.005	0.54	±0.002	0.93	±0.006	0.014
M=Expert	−0.160	<0.001	−0.139	0.001	+0.009		−0.001		−0.030
	±0.043		±0.044		±0.009	0.31	±0.003	0.85	±0.008	0.001
M=LASSO	−0.150	<0.001	−0.134	0.002	+0.010		−0.001		−0.026
	±0.042		±0.043		±0.010	0.29	±0.003	0.67	±0.008	0.002
Non-Hispanic White
Reduced Model 1A-C (N=2,367)
M=Hurd	−0.179		−0.148	<0.001	+0.003		+0.002		−0.035
	±0.031	<0.001	±0.031		±0.006	0.65	±0.003	0.47	±0.007	<0.001
M=Expert	−0.191		−0.149	<0.001	+0.020		−0.001		+0.061
	±0.031	<0.001	±0.031		±0.012	0.096	±0.004	0.73	±0.010	<0.001
M=LASSO	−0.182		−0.145	<0.001	+0.022		−0.003		−0.056
	±0.032	<0.001	±0.031		±0.011	0.051	±0.004	0.44	±0.009	<0.001
Full model 2A-C (N=2,309)
M=Hurd	−0.102		−0.093	0.012	+0.004		−0.001		−0.011
	±0.036	0.005	±0.037		±0.004	0.28	±0.002	0.67	±0.005	0.020
M=Expert	−0.118	0.001	−0.108	0.004	+0.012		−0.002		−0.020
	±0.036		±0.037		±0.007	0.074	±0.002	0.29	±0.007	0.004
M=LASSO	−0.109	0.003	−0.099	0.007	+0.012		−0.002		−0.021
	±0.036		± 0.037		±0.007	0.069	±0.002	0.26	±0.007	0.004
Non-Hispanic Black/Hispanic
Reduced Model 1A-C (N=1,179)
M=Hurd	−0.054		+0.001	0.99	−0.031		+0.004		−0.028	0.12
	±0.086	0.53	±0.088		±0.045	0.48	±0.006	0.48	±0.018
M=Expert	−0.059		+0.010	0.91	−0.026		+0.004		−0.046	0.043
	±0.086	0.50	±0.085		±0.053	0.62	±0.008	0.63	±0.023
M=LASSO	−0.054		−0.011	0.89	−0.006		+0.001		−0.037	0.051
	±0.086	0.53	±0.086		±0.044	0.88	±0.007	0.87	±0.019
Full model 2A-C (N=503)
M=Hurd	−0.052		+0.005	0.59	−0.048		+0.004		−0.014
	±0.097	0.59	±0.100		±0.034	0.16	±0.006	0.42	±0.015	0.35
M=Expert	−0.059	0.54	+0.028	0.78	−0.069		+0.008		−0.025
	±0.099		±0.099		±0.043	0.11	±0.008	0.32	±0.021	0.22
M=LASSO	−0.054	0.57	−0.006	0.95	−0.035		+0.004		−0.017
	±0.096		±0.097		±0.004	0.29	±0.005	0.46	±0.017	0.32

^aCox PH regression models with mortality as the main outcome and diet quality measured by HEI-2015, z-scored , as the exposure. Ln[dementia odds), z-scored, using Hurd, expert and LASSO algorithms were potential mediators/moderators allowed to interact with the main exposure, sample size N=2,812 (full model) and 2,894 (reduced model), four-way decomposition analysis. 1 SD for each Ln transformed dementia odds was ~3.2, 2.9, and 2.1, for Hurd, Expert and LASSO algorithms, respectively. 1 SD of Healthy Eating Index-2015 (HEI-2015) within the selected sample was equivalent to 9.42-point increase. Total effect and controlled direct effects are interpreted as Ln(hazard ratios) per SD of HEI-2015 total score.

^bReduced models (Models 1A-1C) were adjusted for age in 2012, sex and race/ethnicity. Full models (Model 2A-2C) further adjusted the reduced model by all covariates described under the Covariates section, including socio-demographic, socio-economic, lifestyle and health-related factors. Food insecurity (yes vs.no) was added as an additional covariate in the full model.

^cTE=Total Effect

^dCDE=Controlled Direct Effect

^eINTREF=Interaction Referent

^fINTMED=Mediated Interaction

^gPIE=Pure Indirect Effect

^hSE=Standard Error

ⁱY=Outcome

^jX=Exposure

^kM=Mediator/Moderator

^lLASSO= Least Absolute Shrinkage and Selection Operator

^mSD=Standard Deviation

Similarly, four-way decomposition of the association between diet quality and mortality through Ln(dmenetia odds) is displayed in Table 5. Food insecurity was adjusted for in all models in addition to other exogenous covariates in the reduced and fully-adjusted models. Overall, among female and Non-Hispanic White older adults, diet quality was inversely related to mortality risk, a TE that was partially mediated through Ln(dementia odds, z-scored), with ~15–35% of the TE being PIE (e.g. Full model, Hurd: TE:−0.093±0.034, P=0.006, PIE: −0.012±0.005, P=0.011). Nevertheless, the CDE was detected at type I error 0.05, despite marked attenuation.

GSEM findings: testing mediational pathways

Going further in depth and testing mediational pathways was carried out and presented in Figure 4 and Table 6 using a series of generalized structural equations models. Similar to four-way decomposition models, in reduced models without mediators, TE of food insecurity on mortality risk was not detected, while there was an inverse association in the fully adjusted Model 2. In both reduced and fully adjusted GSEM models without mediators in which diet quality was the main exposure, there was an inverse relationship between diet quality and mortality risk, with a marked attenuated of the association between Models 1 and 2. In both models, the pathway of diet quality → dementia → mortality indicated that part of the TE of diet quality on mortality risk was explained by Ln(dementia odds), Hurd algorithm, 21% for Model 1 and 15% for Model 2. The reduced model based on Figure 4 and Table 6, indicated that several indirect effects were significant when it comes to food insecurity as the main exposure and those were FI→DQ→DEM→MORT, FI→DEM→MORT, and FI→DQ→MORT, even though the TE of FI was not detected. It is worth noting, that DEM→MORT, DQ→DEM, and DQ→MORT were the path coefficients that remained statistically significant in fully adjusted Model 2 (Figure 4).

Figure 4. Interplay of food insecurity status, z-scored HEI-2015 (diet quality), z-scored Ln(dementia odds), and all-cause mortality: Health and Retirement Study 2012–2020, Generalized structural equations models and mediating pathways^a.

^a Model 1 is adjusted for age, sex, and race/ethnicity (Non-Hispanic Black/Hispanic vs. Non-Hispanic White); Model 2 was further adjusted for other socio-demographic, lifestyle and health-related factors. 1 SD of HEI-2015 within the selected sample was equivalent to 9.42-point increase.

^b FI= food insecurity (yes vs. no)

^c DQ= Healthy Eating Index- 2015, z-scored

^d DEM=Ln(dementia odds), Hurd algorithm, z-scored

^e MORT=all-cause mortality risk, Weibull model, hazards.

^f M1=Model 1 is the reduced model.

^g M2=Model 2 is the fully adjusted model.

Table 6.

Pathways between food insecurity and mortality through diet quality and Ln(dementia odds, Hurd algorithm) using generalized structural equations modeling: Health and Retirement study, 2012–2020^a,b

Pathway	Model 1: Reduced model			Model 2: Full model
	β	SEc	P	β	SE	P
Food insecurity →diet quality → dementia → mortality
	+0.006	0.002	0.013	+0.0005	0.0008	0.55
Food insecurity → dementia → mortality
	+0.086	0.021	<0.001	+0.017	0.016	0.30
Food insecurity → diet quality → mortality
	+0.025	0.011	0.021	+0.003	0.006	0.56
Diet quality → dementia → mortality
	−0.034	0.006	<0.001	−0.012	0.005	0.001
Total effect of food insecurity on mortality	+0.005	0.124	0.97	−0.388	0.134	<0.001
Total effect of diet quality on mortality	−0.160	0.033	<0.001	−0.081	0.036	0.023

^aValues are mediating pathways combining individual path coefficients as non-linear combinations or total effects of each exposure of interest [food insecurity, binary (yes vs. no) or diet quality]. Total effects were obtained from GSEM models without any mediators, adjusted for exogenous covariates. Diet quality was measured using Healthy Eating Index-2015 in continuous form, z-scored. 1 SD of Healthy Eating Index-2015 within the selected sample was equivalent to 9.42-point increase. 1 SD^d for each Ln transformed dementia odds was ~3.2, 2.9, and 2.1, for Hurd, Expert and LASSO algorithms, respectively. All continuous exposures and potential mediators were also z-scored, unless binary (food insecurity).

^bModel 1 adjusted only for age (2012 wave), sex and race (Non-Hispanic Black/Hispanic vs. Non-Hispanic White); Model 2 further adjusted Model 1 for all other socio-demographic, lifestyle and health-related factors. The main models that included mediators (Model 1 and 2) included the following pathway: Food insecurity→diet quality→dementia→mortality (See Figure 3), along with direct effect of food insecurity on dementia and mortality and the direct effect of diet quality on mortality. Exogenous variables were similarly included in all equations.

^cSE=Standard Error

^dSD=Standard Deviation

DISCUSSION

This study is to our knowledge, the first to examine comprehensively the interplay of food insecurity, diet quality and dementia in their association with all-cause mortality risk in a large nationally representative sample of US older adults. Multiple types of statistical analyses were conducted, including Cox proportional hazards models, GSEM and four-way decomposition models. The study found that 902 deaths occurred, with food insecurity not being associated with mortality risk in the reduced socio-demographic only adjusted model or M1 (Analysis B, binary food insecurity: HR=0.99, 95% CI:0.68–1.43, Cox model), but inversely related in the fully adjusted model for Analysis B, M2 (HR=0.65, 95%CI:0.46–0.92, Cox model). A similar pattern was observed in the GSEM model, whereby food insecurity was also found to be directly related to Ln(dementia odds) in M1 only (β±SE: 0.23±0.05, P<0.001). Diet quality, measured by HEI-2015 total score (z-scored), was inversely related to both Ln(dementia odds, z-scored), with a β±SE: −0.04±0.01, P= 0.008 and mortality risk with an estimated HR=0.92, 95% CI:0.85–1.00 P=0.02, in M2. Ln(dementia odds, z-scored) was strongly associated with mortality risk in M2 (HR=1.39, 95% CI: 1.31–1.48, P<0.001). Finally, the total effect of diet quality on mortality risk was partially mediated through Ln(dementia odds) in both M1 (21% of TE) and M2 (15% of TE).

The association of diminished cognitive function at mid- and later life with an elevated mortality risk in later years is not thoroughly comprehended^{3, 4, 72, 73}. Cognitive function is probably affected by genetic and environmental factors during a lifetime, potentially modifying health and mortality pathways^{3, 4}. Cognitive impairments may potentially indicate underlying biological anomalies or hereditary problems outside of neurodegenerative disorders^{3, 4}. A meta-analysis of more than 60 studies revealed that cognitive impairment, encompassing overt dementia, elevates the risk of mortality from all causes⁷⁴. This general conclusion asserts our present findings.

Food insecurity is an economic and social condition that arises from limited access and affordability of healthy food, affecting adults and children ^16–21. According to one study, using National Health and Nutrition Examination Survey, researchers found that underrepresented minorities were disproportionately affected by food insecurity than the general population ¹⁷. In addition to race/ethnicity, using a nationally representative study, researchers also found that the presence of food insecurity among adults is connected to 46% increased risk of overall mortality compared to those without food insecurity after approximately 10.2 years of follow-up ^{17, 19}. Food insecurity was also shown to be connected to 57% increased risk of cardiovascular mortality than those without food insecurity even after contolling for covariates ^{17, 20}. In another study, researchers found that food insecurity moderates the connection between cardiorenal syndrome and overall mortality ¹⁶. As has been previously observed, food insecurity is a social determinant that not only leads to inreased risk of mortality, but also increases the impact of other chronic diseases ¹⁸. In our present study, however, which was restricted to older adjusted aged 70+y, the findings diverged from the existing literature, whereby no association was detected in the reduced models, while being an inverse one in fully adjusted models, mainly in terms of TE. The inverse association may be elucidated by the resilience of food-insecure individuals, compensatory activities, protective social networks, or measurement discrepancies. Subsequent study could investigate if the correlation is context-dependent and whether variables such as the intensity of food insecurity or availability of social support influence the observed relationship.

In terms of food insecurity association with various cognitive outcomes, a study involving 28,508 adults from the 2020 National Health Interview Survey found that food insecurity significantly affects cognitive impairment in the US, particularly among young or middle-aged, females, and Non-Hispanic Black adults ⁶. According to another study by Kim and colleagues, food insecurity was linked to a faster decline in executive function⁷. In another recent analysis of the HRS, low food security was associated with higher dementia odds and lower memory levels ⁹. These patterns of association were also found in other studies in samples from Mexico and India ^{8, 10, 11}. Further research is needed to understand the impact of food insecurity on cognitive function and the need for interventions.

The associations between diet quality and risk of dementia have been inconsistent. A study of 180,532 participants in the UK Biobank found those consuming dietary patterns scoring high on a healthy plant-based index, in contrast to dietary patterns consisting of less healthy plant foods, had lower risk of dementia as well as a lower risk of depression ¹⁵. A systematic review and meta-analysis of the association of diet quality and dementia from prospective global cohort studies revealed individuals with better adherence to a healthy high-quality dietary pattern was associated with lower risk of overall dementia and AD compared with those following a low-quality dietary pattern ¹³. Data from 12 studies were used in the meta-analysis with diet quality assessed by the Mediterranean diet score, Alternate Healthy Eating Index, or MIND diet score. In contrast, higher diet quality, based on scores from a diet screening tool, was not associated with lower risk of all-cause dementia or altered risks of AD in a small sample of US rural adults, ≥80 years ¹².

Previous observational studies have associated diet indices and health outcome; including suboptimal diet to be a modifiable risk factor for the development of non-communicable disease and mortality ⁷⁵. Recent cohort studies have found that decreased dietary acidity and inflammatory food compounds, along with maintaining a high-quality diet over time may reduce cardiovascular disease and mortality ^{76, 77}. In a longitudinal study with repeated measures on diet quality, a plant-based diet index that improved over time, compared to a plant-based diet index that decreased over time showed a lower risk of total mortality by 10% ⁷⁸. However, a healthy diet in combination with positive lifestyle factors such as physical activity has shown to not only reduce risk of mortality, CVD, and cancers, but also be a potential protective factors among prefrail and frail older adults ^{79, 80}. Clinically, poor diet quality that was assessed among adult populations was associated with increased mortality risk ⁸¹ and several common diet indices have shown associations with frailty, mortality, and CVD risk ^{75, 82}; therefore, assessing diet quality using diet indices may not only help assess a healthy diet but also indirectly assess other health factors (e.g., frailty, CVD, and mortality). Our study indicated that diet quality was independently and inversely associated with mortality risk, a total effect that was only partly mediated through dementia odds, Ln transformed (15–21% of the TE).

Our study has various strengths. First, it is possibly the first study to test complex questions related to food insecurity, diet quality and dementia in relation to mortality risk, in nationally representative research of older adults with a 9-year follow-up period. In fact, the study utilized a wealth of HRS data to test various hypotheses and consider potential confounding effects of unrelated factors. It used advanced approaches like multiple imputations, Cox proportional hazards models, 4-way decomposition models, and GSEM to investigate mediation and interaction. However, the study had some drawbacks, including measurement error due to the precision of the date of birth, the use of a short form of the food insecurity questionnaire, and the measurement error in the algorithmically defined dementia status or probability, despite its reliance on the thorough ADAMS sub-study⁵³. Selection biases and residual confounding cannot be excluded due to the observational nature of the study. Alcohol consumption was excluded as a potential exogenous variable due to substantial evidence that drinking is a consequence rather than a cause of food insecurity^{83, 84}. Exposure-induced mediator-outcome confounding cannot be excluded due to the non-empirical testability of several four-way decomposition models’ assumptions⁸⁵.

Conclusions

In sum, the present study suggests an interplay between food insecurity, diet quality and dementia in relation to mortality risk, whereby diet quality was consistently inversely related to mortality risk, a relationship that was partially mediated through dementia. Thus, pending randomized controlled trials, dementia prevention may result in reducing the impact of poor diet quality on mortality risk. Nevertheless, more longitudinal studies are needed to build upon our findings in comparable populations.

Research Snapshot.

Research question

How does the interplay of food insecurity, diet quality and dementia status relate to all-cause mortality among older US adults?

Key findings

Results showed that among older adults participating in the Health and Retirement Study (N =2,894), food insecurity was not associated with mortality risk in the reduced model, but was inversely related in fully adjusted model, while diet quality was inversely related to both Ln(dementia odds) and mortality risk in both models. Greater Ln(dementia odds) was strongly associated with increased mortality risk. Diet quality’s effect on mortality risk was partially mediated through dementia.

Data availability:

While the data is owned by the University of Michigan Ann Arbor and HRS is public use data, this work is owned and funded by the National Institute on Aging at the NIH. Scripts used in this analysis will be made available on a GitHub repository. For additional information please contact the corresponding author by e-mail contact at baydounm@mail.nih.gov.