Data‐driven lifestyle patterns and risk of dementia in older Australian women

Abstract

INTRODUCTION

Many lifestyle factors have been associated with dementia, but there is limited evidence of how these group together. The aim of this study was to examine the clustering of lifestyle behaviors and associations with dementia.

METHODS

This population‐based study included 9947 older Australian women. Latent class analysis was employed to identify distinct lifestyle classes, and Cox proportional hazard regression compared these with incident dementia over 17 years.

RESULTS

Three classes were identified: (1) “highly social and non‐smokers” (54.9%), (2) “highly social, smokers, and drinkers” (25.1%), and (3) “inactive and low socializers” (20.0%). Women in Class 3 exhibited a higher risk of dementia compared to both Class 1 (hazard ratio [HR] = 1.19, 95% confidence interval [CI]: 1.08 to 1.30) and Class 2 (HR = 1.12, 95% CI: 1.00 to 1.25).

DISCUSSION

A lifestyle pattern characterized by physical inactivity and low social engagement may be particularly detrimental for dementia risk in older women and should be prioritized in preventive strategies.

Highlights

• Latent class analysis was employed to identify distinct lifestyle clusters.

• Three lifestyle‐related clusters were differentially associated with dementia risk.

• Inactive and low socializers exhibited the greatest risk of dementia.

• Targeting physical inactivity and low social engagement in prevention is vital.

1. BACKGROUND

The World Health Organization (WHO) has named dementia a global health priority, with more than 55 million people currently living with the condition. ¹ Dementia is a particular concern in Australia, where it is currently the second leading cause of death overall and the leading cause in women. ² There is currently no cure or effective treatment for dementia, so primary prevention is of utmost importance. ¹ The strongest known dementia risk factor is age. However, several modifiable risk factors have been proposed as targets for preventive efforts. ¹ , ³ For instance, physical inactivity, smoking, excessive alcohol consumption, and social isolation have all been identified as important modifiable risk factors for dementia. ⁴ , ⁵

The importance of participating in regular physical activity and reducing inactivity and sedentary behaviors for dementia and cognitive function has been widely demonstrated ³ , ⁴ , ⁵ and may be attributed to mechanisms such as minimizing vascular and cardiometabolic risk factors, reducing psychological stress, and increasing brain volume, cognitive reserve, amyloid clearance, and brain‐derived neurotrophic factor (BDNF). ⁶ Smoking ³ , ⁷ , ⁸ may be tied to increased oxidative stress and detrimental effects on vascular, inflammatory, and degenerative processes. ⁹ The link between alcohol and neurocognitive disorders is often described as following a J or inverse U‐shaped association. ¹⁰ , ¹¹ Possible mechanisms involved in the role of excessive alcohol consumption on neurodegeneration include neurotoxicity, neuroinflammation, nutritional deficiencies, amyloid aggregation, and neurotransmitter changes ¹² . In contrast, potential protective effects of low to moderate alcohol consumption may be attributed to reducing stress and improving measures of positive affect, ¹³ with consumption of wine also associated with resultant intake of flavonoids that have been associated with cardioprotective effects. ⁶ Moderate alcohol intake may also be correlated with higher levels of social activity and there is an increasing body of evidence to support engaging in social activities and avoiding social isolation as an important modifiable risk factor for late‐life dementia. ³ , ¹⁴ Potential mechanisms involved may include bolstering brain resilience by increasing cognitive reserve and reducing stress‐induced brain damage. ³ However, the existing literature has focused on the association between these and other modifiable risk factors individually on dementia risk, whereas it is more likely that these behaviors co‐occur and/or cluster to influence dementia risk. ¹⁵ , ¹⁶ , ¹⁷

A growing body of evidence has recognized the importance of studying multiple modifiable risk factors in combination and how they might impact the risk of dementia. ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² , ²³ , ²⁴ A previous study examined the clustering of diet, physical activity, smoking, alcohol consumption, social interaction, and church attendance, using the person‐centered approach of latent class analysis (LCA) and subsequent associations with dementia risk in 2491 healthy US adults aged ≥65 years. ²⁵ Four clusters were identified, and those characterized as “healthy‐moderately religious,” “healthy‐very religious,” and “unhealthy‐nonreligious” all exhibited significantly lower risk for dementia than an “unhealthy‐religious” subgroup. ²⁵ A more recent study also applied LCA in a sample of US older adults to explore person‐centered clustering of lifestyle activities and associations between identified clusters and incident dementia. ²⁶ This study identified four distinct subgroups (“variety,” “intellectual,” “social,” and “least active”), with those in the “intellectual” and “variety” classes exhibiting significantly lower risk of incident dementia compared to the “least active” class when cases of mild cognitive impairment (MCI) were excluded.

How such behaviors might cluster together, using a person‐centered data‐driven approach, in an Australian population context, and how this relates to the risk of dementia remains unclear. While similarities in lifestyle patterns may indeed be present across different subgroups/countries, the person‐centered nature of approaches such as LCA prohibits such generalizations from being made, ²⁷ hence the importance of carrying out similar work in broader population groups. As such, this study aimed to examine the associations between clustered modifiable lifestyle behaviors (physical activity, smoking, alcohol consumption, and social engagement) and incident dementia in older Australian women over a 17‐year period.

RESEARCH IN CONTEXT

1. Systematic review: We searched several databases (MEDLINE Complete, EMBASE, PsychINFO, Global Health, CINAHL Complete, AgeLine, Academic Search Complete, and Scopus) for observational studies examining multiple lifestyle‐related factors and associations with cognitive outcomes in adults. Most studies employed index‐based aggregation approaches, with a paucity of evidence employing advanced data‐driven methods.

2. Interpretation: Our findings contribute to a more detailed understanding of the links between combined lifestyle factors and neurocognitive decline, specifically highlighting distinct patterns that may predispose older women to a greater risk of dementia.

3. Future directions: Further research employing advanced data‐driven methods is needed to confirm whether a lifestyle pattern characterized by low social engagement and physical inactivity is particularly detrimental for brain health in broader population groups and in other stages of the lifespan, for example, midlife. These findings can assist in guiding policy messaging, identification of at‐risk subgroups, and tailoring of intervention strategies.

2. METHODS

2.1. Data sources and study design

This study analyzed data from a prospective cohort study, the Australian Longitudinal Study on Women's Health (ALSWH). The ALSWH objectives, design, participants, and survey methods are described elsewhere. ²⁸ Briefly, three cohorts of women (born during the years 1921 to 1926, 1946 to 1951, and 1973 to 1978) were recruited to participate in a long‐term study of women's health in Australia, examining the social, psychological, physical, and environmental factors associated with health across the life course. A random sample of women was initially selected from the national health and insurance database (Medicare), which theoretically covers all Australian citizens and residents. Women in rural and remote areas were intentionally oversampled. For the 1921 to 1926 cohort, baseline data collection commenced in 1996 (Survey 1), and follow‐up surveys were carried out every 3 years (Surveys 2 to 6). Surveys were conducted by mail, with additional efforts to contact participants by telephone to encourage participation. The current study utilized survey data (Survey 2, 1999) for the 1921 to 1926 cohort, along with linked dementia data over a 17‐year follow‐up period. This cohort is broadly representative of women of similar ages in the Australian population. ²⁹ , ³⁰ To be eligible for inclusion in the current study, women from the 1921 to 1926 cohort were required to have not opted out of health record linkage, be alive at Survey 2 (1999), and be free from dementia at Survey 2 (1999).

ALSWH is funded by the Australian Federal Department of Health and is conducted jointly by the University of Newcastle and the University of Queensland. Ethics approval was granted by both the University of Newcastle and the University of Queensland. The ALSWH has also gained approval to access national and state‐based datasets such as the Medicare Benefits Scheme, Pharmaceutical Benefits Scheme (PBS), National Death Index (NDI), Aged Care Datasets, Cancer Registry, emergency department data, and Admitted Patients Data Collections. Ethics approval for access to linked data was provided through the Australian Institute of Health and Welfare (AIHW) and local human research ethics committees. This study was performed in accordance with the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments.

2.2. Lifestyle behaviours

2.2.1. Physical Activity

Physical activity data were collected via the Active Australia survey, which is validated for use in older populations (age ≥65 years). ³¹ Participants recorded total weekly hours in 10‐min bouts for the prior 7 days, categorized as light‐intensity activities (eg, walking briskly), moderate‐intensity leisure‐time activities (eg, social tennis, moderate exercise classes, recreational swimming, and dancing), and vigorous‐intensity leisure‐time activities (which make you breathe harder or puff and pant). Minutes per week spent in each activity were multiplied by a metabolic equivalent (MET) score: 3.0 for walking, 4.0 for moderate activities, and 7.5 for vigorous activities. ³² MET minute values were summed to derive a dichotomous variable: (i) low physical activity—equivalent to ≤150 min/week moderate activity, <800 kcal/week, or <600 MET min/week and (ii) moderate to high physical activity—equivalent to >150 min/week moderate activity, ≥800 kcal/week, or ≥600 MET min/week.

2.2.2. Social engagement

Social engagement was self‐reported through the social interaction subscale from the 11‐item Duke Social Support Index (DSSI). ³³ The social interaction subscale contains four items: (1) “Other than members of your family, how many persons in your local area do you feel you can depend on or feel very close to?” (2) “How many times during the past week did you spend time with someone who does not live with you, that is, you went to see them or they came to visit you or you went out together?” (3) “How many times did you talk to someone (friends, relatives, or others) on the telephone in the past week (either they called you or you called them)?” (4) “About how often did you go to meetings of clubs, religious meetings, or other groups that you belong to in the past week?” Responses were summed to derive a final total score ranging from 4 to 12, with higher scores indicative of greater social interaction.

2.2.3. Alcohol consumption

Participants reported how regularly they drank alcohol from the following response options: “I have never drunk alcohol in my life,” “I never drink alcohol, but I have in the past,” “I drink rarely,” “less than once a week,” “1 or 2 days a week,” “3 or 4 days a week,” “5 or 6 days a week,” and “every day.” Those who consumed alcohol were asked to indicate the volume consumed per day: “1 or 2 drinks per day,” “3 or 4 drinks per day,” “5 to 8 drinks per day,” and “9 or more drinks per day.” Based on the foregoing information, a dichotomous variable was created: (i) meeting current alcohol recommendations, <10 standard drinks per week or (ii) exceeding alcohol recommendations, >10 standard drinks per week, based on the 2020 National Health and Medical Research Council (NHMRC) Australian guidelines to reduce health risks from drinking alcohol. ³⁴

2.2.4. Smoking

Participants self‐reported smoking status by answering the following questions: “How often do you currently smoke cigarettes or any tobacco products” and “In your lifetime, would you have smoked at least 100 cigarettes (or equivalent)?” For the first question response options were (i) daily, (ii) at least weekly (but not daily), (iii) less often than weekly, and (iv) not at all. The second question involved a binary yes or no response. From these data a categorical variable was created: (a) current smoker (ie, smokes cigarettes daily, weekly, or less than weekly), (b) ex‐smoker (ie, had smoked previously but not currently), or (c) never smoker.

2.2.5. Other health‐related variables

Further health‐related variables included in the analyses were body mass index (BMI), stress, number of general practitioner (GP) visits in the past 12 months, and recent chronic conditions. Women self‐reported their current weight and height (“How much do you weigh?” and “How tall are you without shoes?”), from which BMI was calculated (ie, weight [kg] $÷$ height [m]²) and individuals were categorized according to the following WHO classifications ³⁵ : underweight (<18.5 kg/m²), a healthy or normal weight (18.5 to 24.9 kg/m²), overweight (25.0 to 29.9 kg/m²), and obese (≥30.0 kg/m²). Stress was assessed using a perceived stress scale. ³⁶ The scale was made up of seven items: own health, living arrangements, money, health of family members, relationship with spouse/partner, relationship with children, and relationship of other family members. Responses were recodeded with scores of 0 to 4 assigned for each item. The mean score across all items was then calculated. The possible range of mean stress scores was 0 to 4, with higher scores indicative of higher levels of perceived stress. Participants were asked how many times they had consulted a doctor/GP in the last 12 months, and responses were categorized as (i) low: <5 visits per year or (ii) high: ≥5 visits per year, as previously reported for this dataset. ³⁷ The presence of chronic conditions (diabetes, heart disease, hypertension, stroke, and depression) was assessed via self‐report, with participants asked if they had been told by a doctor that they have had any of the listed conditions in the last 3 years.

2.3. Ascertainment of dementia

Aged care records (Aged Care Assessment Program, Aged Care Funding Instrument, events, and home care packages), PBS data, hospital admissions data, hospital emergency data, and self‐reported survey data (from participant or proxy), and cause‐of‐death records from the National Death Index and National Mortality database were integrated to comprehensively identify women with dementia (File S1 for a complete list of individual datasets incorporated), as previously described. ³⁸ A dichotomous (yes/no) variable for the presence of dementia was coded for each participant. As many cases were identified by more than one data source, a variable for the date of first diagnosis was also generated, which was incorporated in the survival analyses.

2.4. Covariates

Age, educational attainment, marital status, income, and area of residence were included in analyses as covariates. Age was self‐reported in Survey 2 (1999). Educational qualifications were obtained from the baseline ALSWH survey (1996) and categorized as university degree (university degree/high education degree), trade certificate/diploma, and high school or below. Marital status was self‐reported in Survey 2 (1999) and dichotomized as partnered (married and de‐facto) versus non‐partnered (separated, divorced, widowed, and never married). Financial position data, as a measure of financial stress, was collected via self‐report in Survey 2 (1999) with participants being asked, “How do you manage on the income you have available?” and included the following response options: “impossible,” “always difficult,” “sometimes difficult,” “not too easy,” amd “easy.” Responses were collapsed into three categories: (i) difficult, (ii) not too bad, and (iii) easy. Area of residence was self‐reported in Survey 2 (1999) and classified as one of (i) capital city/metropolitan, (ii) rural, or (iii) remote.

2.5. Statistical analyses

Descriptive statistics are presented as means and standard deviation (SD) for continuous variables and counts/frequencies for categorical variables. LCA was used to derive subgroups of lifestyle‐based behavior patterns using four modifiable lifestyle behaviors assessed in Survey 2: (i) smoking, (ii) alcohol, (iii) physical activity, and (iv) social interaction. LCA was performed using full information maximum likelihood estimation (FIML) to manage missing data and a robust maximum likelihood estimator for management of non‐normality. ³⁹ , ⁴⁰ Class solutions 1 to 5 were tested, and the selection of the optimal number of latent classes was based on interpretability, class size and proportions (where the relative size of profiles was no lower than approximately 5% of the total sample ⁴¹ ), and maximization of model fit statistics. The following model fit statistics were examined: Akaike's information criterion (AIC), Bayesian Information Criterion (BIC) and adjusted BIC (lower values indicating better model fit), ⁴² , ⁴³ entropy (range 0 to 1, with values closer to one indicating better class separation), ⁴¹ Lo‐Mendell‐Rubin likelihood ratio test (LMR‐LRT), and bootstrapped likelihood ratio test (BLRT) value. ⁴¹

Further characterization of the identified classes was undertaken by comparing several key demographic and health‐related variables across classes, using chi‐squared tests for categorical variables and one‐way analysis of variance (ANOVA) with multiple pairwise comparisons using a Bonferroni adjustment for continuous variables. Cox proportional hazards models tested associations between class membership in Survey 2 (1999) and incident dementia over follow‐up. Participants were followed up until the first diagnosis of dementia, death, other loss to follow‐up, or the end of the follow‐up period, whichever came first. The end of the follow‐up period was set as February 15, 2016. The equality of survival functions across classes was tested using the log‐rank test, and Kaplan‐Meier curves were employed to plot a graph of survival estimates. Both unadjusted and adjusted models were run, with case‐wise exclusion for missing data. Adjusted models included age, education, income, area of residence, and marital status. The inclusion of these variables was informed by background literature, employing a directed acyclic graph (DAG) to model assumed directionality between covariates, confounders, exposure, and outcome (File S2). ⁴⁴ Additional risk factors such as BMI, depression, heart disease, and diabetes that might lie along the causal pathway were omitted from the adjusted model. These are considered intermediate variables in our DAG, and overadjustment for such variables may pose issues that will result in overadjustment bias. ⁴⁵ The proportional hazards assumption was tested by including time‐dependent covariates in the model, running Schoenfeld and scaled Schoenfeld residuals, visual inspection of Schoenfeld residuals plots, and goodness‐of‐fit testing of the final model. Data are presented as hazard ratios (HRs) with 95% confidence intervals (CIs). LCA analyses were carried out in MPlus version 8.6 (Muthen & Muthen, Los Angeles, CA, USA), and all other analyses were performed using Stata/SE 17.0 (Stata Corp., College Station, TX, USA). Statistical significance was set at a p value of α < .05. 

TABLE 1.

Demographic and health characteristics of analytic sample in Survey 2 (1999).

	Descriptive statistics a
n	9947
Age (years)	75.4 ± 1.5
BMI (kg/m2)	25.3 ± 4.5
BMI classification b
Underweight (BMI < 18.5 kg/m2)	302 (3.0%)
Healthy weight (18.5 kg/m2 ≤ BMI < 25 kg/m2)	4365 (43.9%)
Overweight (25 kg/m2 ≤ BMI < 30 kg/m2)	2933 (29.5%)
Obese (BMI ≥ 30 kg/m2)	1196 (12.0%)
Missing	1151 (11.6%)
DSSI Social Interaction subscore c	9.0 ± 1.5
Education
High school or below	7977 (80.2%)
Trade certificate/diploma	1109 (11.1%)
Higher (university degree/high university degree)	365 (3.7%)
Missing	496 (5.0%)
Marital status
Partnered	5080 (51.1%)
Non‐partnered	4380 (48.6%)
Missing	37 (0.4%)
Area of residence
Capital city/metropolitan	4.080 (41.0%)
Regional/rural	5617 (56.5%)
Remote	230 (2.3%)
Missing	20 (0.2%)
Income management
Difficult	2310 (23.2%)
Not too bad	4706 (47.3%)
Easy	2105 (21.2%)
Missing	826 (8.3%)
Smoking status
Non‐smoker	6023 (60.5%)
Ex‐smoker	2899 (29.1%)
Current smoker	455 (4.6%)
Missing	570 (5.7%)
Physical activity
Low	5214 (52.4%)
Moderate‐high	3576 (35.9%)
Missing	1157 (11.6%)
Alcohol consumption d
Meeting recommendations (< 10 standard drinks per week)	7340 (73.8%)
Exceeding recommendations (≥ 10 standard drinks per week)	1243 (12.5%)
Missing	1364 (13.7%)
GP visits
Low (<5 visits per year)	4137 (41.6%)
High (≥5 visits per year)	5686 (57.2%)
Missing	124 (1.2%)

Abbreviations: BMI, body mass index; DSSI, Duke Social Support Index; GP, general practitioner.

^a Descriptive statistics are presented as mean ± standard deviation for continuous variables and frequency (%) for categorical variables.

^b WHO classifications for adult BMI. ³²

^c Social interaction subscale score ranges from 4 to 12, with higher scores indicative of increased social interaction.

^d Categorical variable for alcohol consumption coded to align with the 2020 NHMRC Australian guidelines to reduce health risks from drinking alcohol. ³¹

3. RESULTS

At baseline, 12 432 women participated in the study, but after the exclusion of individuals who did not consent to data linkage (n = 353), women who died or were lost to follow‐up between baseline (1996) and Survey 2 (1999) (n = 1946), and women who were missing all lifestyle behavior variables in Survey 2 (n = 186), the analytic sample for this study comprised 9947 women. Comparing baseline (1996) data for the analytic sample to those excluded/lost to follow‐up, the former had a slightly higher mean BMI, a higher proportion of partnered (versus non‐partnered) women, and a slightly higher proportion of women from regional/rural areas compared to capital city/metropolitan areas (Table S1). The mean age of the analytic sample in Survey 2 (1999) was 75.4 years, with the majority (80.2%) having a low level of formal education (high school or below) and a higher proportion (56.5%) living in regional/rural areas, while just over half (51.1%) were partnered (Table 1). Approximately half (47.3%) of the analytic sample felt it was “not too bad” to manage on their available income. The mean BMI was 25.3 kg/m2, with 43.9% of participants falling into the healthy BMI classification. The mean stress score was 0.4 (range: 0 to 4), and the mean DSSI social interaction score was 9.0 (range: 4 to 12). The majority (60.5%) were non‐smokers; over half were classified into the low PA group (52.4%); the majority (73.8%) met current alcohol consumption guidelines, and over half (57.2%) reported a high number of GP visits in the past 12 months. 

FIGURE 1.

Kaplan‐Meier curves of dementia‐free probabilities across follow‐up by lifestyle class. Class 1: “Highly social and non‐smokers” (54.9%); Class 2: “Highly social, smokers and drinkers” (25.1%); Class 3: “Inactive and low socializers” (20.0%).

Model fit indices are summarised in Table S2. Log‐likelihood increased as class solutions increased, and AIC, BIC, and sample‐size adjusted BIC (SSABIC) all declined across the same class solutions. However, the BLRT for class solutions 4 and 5 indicated model overfit and derived classes with a very low probability of membership relative to the overall sample size (four‐class approach: 5.1%, and five‐class approach: 3.3%) and as such was considered too small to be a meaningful profile. After consideration of model fit indices, class proportions, and class interpretability, the three‐class solution was considered the most optimal. Class 1 made up over half of the study sample (54.9%), followed by Class 2 (25.1%), then Class 3 (20.0%). Based on response probabilities (for categorical input variables: alcohol, smoking, and PA) and means (for continuous: social interaction score) obtained from LCA (Table 2), Class 1 was labeled “highly social and non‐smokers,” Class 2 “highly social, smokers and drinkers,” and Class 3 “inactive and low socializers.”

TABLE 2.

Distribution of lifestyle‐related behaviors in each class (with smoking, exercise, and alcohol expressed as a percentage of class membership and social interaction as a mean for class membership).

Class	N	Percentage	Within alcohol guidelines	Exceeds alcohol guidelines	Low PA	Moderate‐High PA	Never smoker	Ex‐smoker	Current smoker	DSSISI
1. Highly social and non‐smokers	5458	54.9	91.1%	8.9%	54.9%	45.1%	91.1%	8.9%	0%	9.7
2. Highly social, smokers, and drinkers	2496	25.1	69.7%	30.3%	53.7%	46.3%	0%	84.3%	15.6%	9.5
3. Inactive and low socializers	1993	20.0	87.3%	12.7%	73.4%	26.6%	60.1%	33.8%	6.1%	7.1

Abbreviations: DSSISI, Duke Social Support Index Social Interaction; PA, physical activity.

Differences in other key demographic and health‐related variables across classes are presented in Table 3. There were little to no differences in age and the highest mean perceived stress scores across the classes. However, several other notable differences were observed. Class 3 had the lowest educational levels, followed by Class 2. Class 1 had a higher proportion of women with education levels above high school attainment. Class 3 also exhibited the most difficulty with managing available income. Similar proportions resided in remote regions across classes. However, Class 3 had the highest proportion living in capital city/metropolitan regions, followed by Class 2, and, lastly, Class 1. Class 3 had the greatest proportion of partnered women, and Class 2 had the lowest. Class 3 exhibited the highest proportion of women reporting diabetes, heart disease, stroke, and depression. Class 1 was slightly above Class 3 in relation to hypertension, with Class 1 exhibiting the lowest proportion of women reporting the condition (Table 3).

TABLE 3.

Sample characteristics within each lifestyle class (using final three‐class approach) ^{A .}

	Class 1: “Highly social and non‐smokers”	Class 2: “Highly social, smokers, and drinkers”	Class 3: “Inactive and low socializers”
	n = 5458 (54.9%)	n = 2496 (25.1%)	n = 1993 (20.0%)	p value B
Age in years, mean (SD)	75.3 (1.4)	75.3 (1.4)	75.4 (1.5)	.93
Stress score, mean (SD) C	0.3 (0.4)	0.4 (0.4)	0.5 (0.5)	<.001
Highest level of education, n (%)				<0.001
High school or below	4380 (84.8%)	1930 (81.1%)†	1652 (87.4%)†*
Trade certificate/diploma	594 (11.8%)	316 (13.3%)†	197 (10.4%)†*
Higher (university degree/high university degree)	190 (3.7%)	134 (5.6%)†	41 (2.2%)†*
Income management				<0.001
Difficult	1131 (23.2%)	592 (26.1%)†	582 (29.6%)†*
Not too bad	2547 (52.3%)	1163 (51.2%)†	987 (50.1%)†*
Easy	1188 (24.4%)	515 (22.7%)†	399 (20.3%)†*
Area of residence				<0.01
Capital city/metropolitan	2150 (39.6%)	1032 (41.5%)	886 (44.6%)†
Regional/rural	3159 (58.1%)	1395 (56.1%)	1055 (53.1%)†
Remote	126 (2.3%)	59 (2.4%)	45 (2.3%)†
Marital status				<0.001
Partnered	2778 (51.2%)	1139 (45.9%)†	1153 (58.3%)†*
Non‐partnered	2652 (48.8%)	1343 (54.1%)†	825 (41.7%)†*
BMI (WHO classifications)				<0.001
Underweight, BMI < 18.5	126 (2.6%)	103 (4.6%)†	72 (4.2%)†*
Healthy weight, 18.5 ≤ BMI < 25	2415 (50.0%)	1101 (49.2%)†	841 (49.1%)†*
Overweight, 25 ≤ BMI < 30	1645 (34.0%)	767 (34.3%)†	518 (30.3%)†*
Obese, BMI ≥ 30	645 (13.4%)	268 (12.0%)†	281 (16.4%)†*
GP visits per year				<0.001
Low (<5 visits per year)	2354 (43.7%)	975 (39.6%)	806 (41.1%)
High (≥5 visits per year)	3028 (56.3%)	1487 (60.5%)	1153 (58.9%)
Diabetes				<0.01
No	4962 (92.4%)	2272 (93.5%)	1783 (91.8%)
Yes	406 (7.6%)	159 (6.5%)	159 (8.2%)
Heart disease				<0.05
No	4653 (87.3%)	2103 (86.5%)	1644 (84.6%)†
Yes	679 (12.7%)	328 (13.5%)	298 (15.4%)†
Hypertension				<0.01
No	3464 (65.0%)	1671 (68.7%)†	1270 (65.4%)*
Yes	1868 (35.0%)	760 (31.3%)†	672 (34.6%)*
Stroke				<0.001
No	5206 (97.6%)	2369 (97.4%)†	1857 (95.6%)*
Yes	126 (2.4%)	62 (2.6%)†	85 (4.4%)*
Depression				<0.001
No	5054 (94.8%)	2259 (92.9%)†	1754 (90.3%)†*
Yes	278 (5.2%)	172 (7.1%)†	188 (9.7%)†*

Bolded values indicate significant differences at level of p < .05.

^A Results are presented based on final three‐class approach based on highest‐probability class assignment.

^B P values were obtained through using chi‐squared tests for categorical variables and one‐way analysis of variance (ANOVA) with multiple pairwise comparisons using a Bonferroni adjustment for continuous variables

^C Stress scores range from 0‐4 with higher scores indicating increased levels of perceived stress.

Post hoc testing: †Significant difference from Class 1 (p < .05), *Significant difference from Class 2 (p < .05).

Cox proportional hazards modeling was employed to examine associations between class membership for the final three‐class solution and dementia incidence over the 17‐year follow‐up period. The proportional hazards assumption was tested using multiple methods and showed no significant evidence of violation of the assumption. The log‐rank test indicated inequality in survival estimates across classes (p = .017), which can be visually observed in the separation of class survival curves across the latter half of the analysis period (Figure 1). In unadjusted Cox proportional hazards regression models, there was a significantly higher risk of dementia for Class 3 compared to Class 1 (HR: 1.14, 95% CI: 1.04 to 1.25, p < .05) (Table 4). No differences in dementia risk were observed between Classes 1 and 2. In models adjusted for age, education, area, income, and marital status, dementia risk for Class 3 compared to Class 1 remained statistically higher (HR: 1.19, 95% CI: 1.08 to 1.30, p < .001). In adjusted models, Class 3 also exhibited significantly higher risk than Class 2 (HR: 1.12, 95% CI: 1.00 to 1.25, p < .05) (Table 4).

TABLE 4.

Cox proportional hazards regression examining associations between class membership and risk of dementia, over 17‐year follow‐up period.

	Comparison group	Unadjusted HR [95% CI]	p value	Adjusted HR [95% CI]	p value
Class 1 (reference group)	Class 2	1.05 [0.97, 1.15]	.222	1.06 [0.96, 1.16]	.238
	Class 3	1.14 [1.04, 1.25]	.005	1.19 [1.08, 1.30]	<.001
Class 2 (reference group)	Class 1	0.95 [0.87, 1.03]	.222	0.95 [0.86, 1.04]	.238
	Class 3	1.08 [0.97, 1.20]	.147	1.12 [1.00, 1.25]	.041

Abbreviations: CI, confidence interval; HR, hazard ratio

P values < .05 are bolded to highlight statistical significance.

Class 1: “Highly social and non‐smokers” (54.9%); Class 2: “Highly social, smokers, and drinkers” (25.1%); Class 3: “Inactive and low socializers” (20.0%).

Adjusted models included the following: age, education (high school or below, trade certificate/diploma, higher, that is, university degree/high university, area of residence [capital city/metropolitan, rural, or remote], income management, and marital status (partnered/non‐partnered).

4. DISCUSSION

This study employed the data‐driven approach of LCA to identify three distinct lifestyle‐based classes in a population of older Australian women. The most prevalent class was characterized as (1) “highly social and non‐smokers,” followed by (2) “highly social, smokers and drinkers” and (3) “inactive and low socializers.” The “inactive and low socializers” (Class 3) exhibited significantly higher risk of dementia compared to the “highly social and non‐smokers” (Class 1) and “highly social, smokers and drinkers” (Class 2).

The first novel component of this study was the identification of distinct lifestyle‐based classes, derived through the sophisticated person‐centered approach of LCA, in a cohort of older Australian women. Similar classes have been identified in other populations. Liao et al. explored the clustering of the same four lifestyle behaviors as the current study (smoking, alcohol consumption, PA, and social activity) in a US and English sample of men and women aged ≥50 years, with models stratified by sex. ⁴⁶ Their “multi‐health‐related‐behaviour” cluster is comparable to the “highly social and non‐smokers” (Class 1) of the current study, both of which are characterized by moderate drinking, ex‐/never smoking, and being highly social and physically active. Furthermore, their “inactive” cluster can be likened to the “inactive and low socializers” (Class 3), with both demarcated by low engagement in physical and social activities. Their final cluster, that is, “(ex‐)smoking,” is also similar to the “highly social, smokers and drinkers” (Class 2) of the current study with respect to being characterized by a high proportion of ex‐smokers and the highest likelihood of exceeding alcohol recommendations. However, the coupling with social and physical inactivity in this cluster found by Liao et al. ⁴⁶ was not shown in our study. This may be partly attributed to the relatively high social engagement scores exhibited by the current sample of older Australian women (overall mean = 9.0 out of a maximum score of 12), resulting in two classes clustering around the high end of engagement, rather than only one in the Liao et al. study. The presence of similar classes emerging across different population groups lends support to the generalizability of these findings and, thereby, the implementation of multinational behavioral recommendations that consider the key classes observed, such as strategies focused on joint targeting of physical activity and social engagement. It is important to note that the sample of women in the Liao study was younger (mean age 65 years for US women and mean age 67 for English women) compared to the current sample (mean age 75.4). This may be indicative of relative stability in lifestyle patterns in women aged 65+. However, further research is needed to explore longitudinal clustering trajectories over time to determine whether this is the case.

Another key finding was that a class of older Australian women characterized as “inactive and low socializers” (Class 3) was at increased risk of dementia compared to the two other identified classes. Specifically, the “inactive and low socializers” had a 1.19 times greater dementia risk than the “highly social and non‐smokers” (Class 1) and 1.12 times greater risk than the “highly social, smokers and drinkers” (Class 2). There is sound rationale to support these findings. For instance, a 2018 meta‐analysis examining the association between social engagement or loneliness and dementia risk highlighted the importance of targeting social isolation and disengagement in dementia prevention strategies, with poor social engagement resulting in a 1.28 times greater risk of dementia. ⁴⁷ Social isolation has also been highlighted as a key late‐life modifiable risk factor for dementia, estimated to account for approximately 4% of total population risk. ¹⁹ Mechanistically, social interaction may act to build or preserve brain reserves and, hence, ensure the mind is more resilient to neurodegenerative effects, especially in late life. ⁴⁸ , ⁴⁹ , ⁵⁰ Furthermore, social engagement may increase positive emotional factors, such as self‐esteem, and may also have an effect on improving immune system functions, contributing to delayed dementia progression. ⁴⁸ Similarly, physical inactivity is an established late‐life risk factor for dementia. ¹⁹ The 2019 WHO risk reduction guidelines for cognitive decline and dementia strongly recommend participation in regular physical activity for adults with normal cognition to reduce neurodegenerative risk. ⁵¹ The underlying mechanisms may be direct, for example, beneficial effects on brain structures such as increasing brain volume, ⁵² but also likely indirect, for example, enhancing immune function, reducing inflammation, increasing neurotrophic factors, and optimizing cardiovascular risk factors. ⁵¹ Based on research focused on social engagement and physical activity as individual predictors of dementia risk, it is logical that a population class characterized by low engagement in these behaviors may be at increased risk for dementia. An added benefit from this study that explored the clustering of multiple health‐related behaviors is that we can elucidate which unhealthy behaviors cluster together and, hence, should be considered for joint targeting in preventive strategies.

An unexpected finding from this study was that a class (“highly social, smokers and drinkers” [Class 2]) characterized by the highest proportion of current and ex‐smokers and the highest proportion exceeding alcohol guidelines had a lower risk of dementia compared to the “inactive and low socializers” (Class 3) in adjusted models. However, it is important to highlight that the “highly social, smokers and drinkers” (Class 2) also exhibited high social engagement and the highest proportion of women attributed to the moderate‐high PA classification (albeit very closely followed by the “highly social and non‐smokers” [Class 1]). This finding indicates that perhaps, despite being more likely to have histories of smoking and drinking, individuals in Class 2 may not be at as high a risk for dementia as those in Class 3 because they remain both physically and socially active. This lends further support to the prioritization and joint targeting of these types of behaviors in preventive strategies.

This is the first study to explore the clustering of alcohol, smoking, physical activity, and social interaction in older Australian women and compare resultant classes to incident dementia risk. A key strength of this study is the large sample size and the long‐term follow‐up period. Furthermore, incident dementia has been comprehensively identified using multiple linked administrative health records and capture‐recapture methodology, ensuring that the number of “unidentified” women with dementia is minimized. ³⁸ Another strength and novel aspect is the use of LCA to identify person‐centered subgroups based on similar patterns of health‐related behaviors.

Despite these strengths, several limitations of this study need to be recognzsed. A key limitation is the cross‐sectional capture of lifestyle‐related behaviors in Survey 2 (1999). Over time, some of the current smokers are likely to quit, and there is likely to be fluctuations in levels of physical activity over time. ⁵³ Future research exploring longitudinal changes in class membership and any resultant differences in associations with the risk of dementia is warranted. We also used a highest probability classification/classify‐analyze approach to class assignment, which may have introduced bias due to classification errors. Future work should endeavor to employ a flexible‐model‐based approach in order to preserve uncertainty in latent class asssignments. ⁵⁴ Another important consideration is that the separation across classes with respect to dementia risk does appear to be quite subtle. However, this is focused on population‐level (based on class membership) rather than individual‐level risk, and hence even small reductions in risk can be attributed to a reduction in overall dementia prevalence and dementia‐attributed mortality. Lastly, we did not model death as a competing risk in our analyses, and hence there is some potential bias through possible overestimation of dementia incidence.

The analyses in this study can be considered exploratory in nature, and it is important that further research be carried out to confirm the validity of these findings. Another limitation is the limited generalizability to wider population groups. ²⁷ Therefore, the findings from this work represent the lifestyle patterns of older Australian women specifically. The reliance on self‐reported health behavior data is also a limitation. As such, these data may be subject to bias from over‐ or underreporting. Another consideration is that this study was focused on older women (mean age of 75.4 years in Survey 2, 1999). Several modifiable lifestyle factors appear to have the greatest impact on dementia risk at midlife (45 to 65 years) rather than in later life (>65 years). ¹⁹ For example, excess alcohol consumption has been proposed as a midlife risk factor for dementia. ¹⁹ In future research, it may be important to explore the longitudinal clustering of multiple health‐related behaviors across the life course and consider whether specific life stages pose greater intervention potential than others.

4.1. Conclusion

In conclusion, this study indicates that a lifestyle pattern characterized by inactivity and low social engagement may be particularly detrimental for dementia risk in older Australian women and could be the target for future intervention efforts. However, further research employing data‐driven approaches is warranted to confirm these findings.

CONFLICT OF INTEREST STATEMENT

Author disclosures are available in the supporting information.

FUNDING INFORMATION

SED is funded by a Deakin University Postgraduate Research Scholarship. The sponsors had no role in the design and conduct of the study; in the collection, analysis, and interpretation of data; in the preparation of the manuscript; or in the review or approval of the manuscript.

CONSENT STATEMENT

All participants provided written informed consent, and participants are free to withdraw or suspend their participation at any time with no need to provide a reason.

Supporting information

ICMJE DISCLOSURE FORM

Supplementary File 1

Supplementary Figure 1

Supplementary Table 1

Supplementary Table 2

Dingle SE, Bowe SJ, Bujtor M, et al. Data‐driven lifestyle patterns and risk of dementia in older Australian women. Alzheimer's Dement. 2024;20:798–808. 10.1002/alz.13467