Data sources
Data were drawn from the English Longitudinal Study of Ageing (ELSA) and the Survey of Health, Ageing and Retirement in Europe (SHARE). ELSA is a nationally representative cohort study of the English population aged ≥50 years. Data collection began in 2002/03 with follow-up every two years thereafter. SHARE is a nationally representative cohort study of residents aged ≥50 years in 28 European countries and Israel, with follow-up in 2004/05, 2006/07, 2010/11, 2013, 2015, 2017, and 2019/20. These cohorts have a similar survey design and implementation to facilitate cross-country comparisons and harmonization; details are available elsewhere35,36. Both cohorts were granted relevant local ethics approval, with written informed consent given at each interview.
The present analysis included waves 2-8 of ELSA (2004/06-2018/19) and waves 2 and 4-8 (2006/07-2019/20) of SHARE, comprising up to 15 years of follow-up. Waves 2 and 4-8 of SHARE included 13 of the 28 European countries. Wave 3 of SHARE is a life history module that is not part of the main survey. Respondents aged ≥50 years from European countries participating in wave 2 of SHARE or ELSA with at least one round of cognitive testing were eligible for inclusion in the analyzes. ELSA and SHARE data were pooled for analysis.
Lifestyle and behaviors
The four behaviors considered were smoking, alcohol consumption, physical activity, and social contact, which were assessed at the baseline wave of the present study (wave 2 in both SHARE and ELSA) and were not updated during follow-up.
To assess smoking, participants were asked to report whether they currently consider themselves a smoker as well as whether they had smoked previously. Alcohol consumption was assessed by asking participants about their average alcohol consumption in the past six months (SHARE) or maximum alcohol consumption in the past seven days (ELSA). Participants reported the number of drinks they consumed, with examples of what volume constituted one drink for each type of alcoholic beverage provided. The number of alcoholic beverages SHARE participants could report consuming daily was capped at 70; SHARE participants reporting 70 drinks per day were excluded from analyzes. To assess physical activity, ELSA and SHARE participants reported the current frequency of participation in moderate and vigorous physical activity (rarely/never, 1–3 times monthly, weekly, or greater than weekly), with examples given for each physical activity intensity category. This measure of physical activity has been found to be moderately correlated with accelerometer-assessed physical activity in a subset of ELSA participants37; WHO guidelines of 150 min per week of moderate or 75 min per week of vigorous physical activity are also based on self-report measures23. Finally, to examine social contact, participants were asked whether they currently had at least weekly social contact with relatives and/or friends or whether they participated in weekly social activities.
To facilitate interpretation of the results, behaviors were dichotomized into categories corresponding to more recommendation compliant and less recommendation compliant behavior based on WHO and United States guidelines26,38. These dichotomized behaviors made up an individual’s lifestyle, where there were 16 distinct lifestyles corresponding to all possible combinations of the four behaviors.
Smoking was categorized into a current smoker (smoking) and not a current smoker (non-smoking). The dichotomization of alcohol consumption was based on current United States guidelines38. Alcohol consumption was categorized into no-to-moderate consumption (up to 2 drinks per day for men or 1 drink per day for women) or heavy consumption. Physical activity was categorized into those who reported participating in both moderate and vigorous physical activity at least weekly (moderate-plus-vigorous physical activity or weekly MVPA) compared with those who reported less than weekly moderate or vigorous physical activity (less-than-weekly MVPA). We chose to require both weekly moderate and vigorous activity because all categorizations of MVPA requiring less activity resulted in 75–85% of the sample being categorized into the weekly MVPA group, yielding insufficient sample sizes to examine the cognitive decline in lifestyles with less-than-weekly MVPA. Finally, social contact was categorized into those who reported weekly social contact (weekly social contact) compared with those who did not (less-than-weekly social contact). The recommendation-compliant lifestyle refers to participants reporting non-smoking, no-to-moderate alcohol consumption, weekly MVPA, and weekly social contact.
Cognitive function
The cognitive domains examined were episodic memory and verbal fluency. These cognitive domains show decline with ageing39, and dementia40, making them appropriate cognitive domains to examine risk factors for cognitive decline. Tests of episodic memory and verbal fluency were also the same in both cohorts and administered at enough waves to facilitate the examination of long-term cognitive trajectories.
Episodic memory was assessed using the Consortium to Establish a Registry for Alzheimer’s Disease immediate and delayed recall tasks41. Participants were given a 10-word list and asked to recall it immediately and after a delay, with scores on both tests summed to yield an overall recall score. Verbal fluency was assessed using the animal naming task42, in which participants were required to name as many animals as possible within a one-minute period. Memory tasks were administered at every wave; animal naming was not administered during ELSA wave 6. Given absolute differences in cognitive scores between countries, cognitive scores were standardized by country using the mean and standard deviation of each country’s scores at baseline.
To reduce the impact of prodromal dementia on behavior, participants reporting having received a dementia diagnosis at any interview during the follow-up period or who had cognitive scores suggesting cognitive impairment were excluded from analyzes. A cognitive score in either cognitive domain more than 1.5 standard deviations below the mean for an individual’s 5-year age group in their country was considered to be evidence of potential cognitive impairment; this cut-off is widely used to identify cognitive impairment when clinical evaluation is not possible43. We excluded from analyses participants whose scores suggested cognitive impairment for at least two waves of follow-up. We did not exclude participants with just one wave of cognitive impairment unless it was their only wave of participation; this was done to retain participants experiencing transient cognitive impairment as up to 55% of those diagnosed with mild cognitive impairment may revert to normal cognitive function44.
Covariates
Covariates were selected on the basis of previous evidence of associations with lifestyle and cognitive function and were ascertained by self-report. Sociodemographic covariates included gender (man or woman), age at baseline in years, and country (Table S15). Socioeconomic covariates included education (less than upper secondary, upper secondary, or tertiary and above; categorized based on the International Standard Classification of Education 201145) and household non-housing wealth. Wealth was standardized to each country by year and converted into quintiles, with the highest quintile corresponding to the greatest wealth. Chronic conditions were ascertained based on self-report of clinical diagnosis and included high blood pressure, diabetes, cardiovascular conditions (including heart disease and stroke), cancer, lung disease, high cholesterol, and psychiatric conditions. Self-report of clinical diagnosis of chronic conditions has been shown to have good agreement with ascertainment based on medical records46. All covariates except gender and education—which were drawn from the baseline wave—were time-varying and were recorded at each interview or imputed from the closest wave if missing.
Statistical analysis
After imputation of time-varying covariates, participants who were still missing data were dropped from analyzes with one exception: we singly imputed behaviors for 662 participants (2.1% of the analytic sample) missing one behavior out of the four to retain participants who otherwise had a full set of covariates and cognitive scores. This imputation was based on logistic regression models that included country, gender, marital status (married/partnered or not), education level, labor force status (retired, employed, unemployed), age at baseline, all other behaviors, chronic conditions, memory and fluency scores, and body mass index.
We then described participant characteristics in the pooled analytic sample at baseline for each of the four behaviors, with Pearson’s \({x}^{2}\) test used to examine associations with categorical covariates and \(t\)-test with continuous covariates.
Linear mixed models were used to examine associations between lifestyle and 10-year memory and fluency decline. Linear mixed models use all available data regardless of length of follow-up, handle non-monotone missingness patterns, and data missing-at-random47. These models included a random intercept and slope at the individual level with an unstructured covariance matrix to account for the correlation of repeated measurements on each participant and used time since baseline in years as the timescale. We also fitted models that included an additional random intercept on cohort or country but found that these terms explained negligible variation in cognitive trajectory and, therefore, omitted them in the reported results to simplify the model.
We first fitted models adjusted for sociodemographic and socioeconomic covariates and chronic conditions to examine independent associations between each of the behaviors (smoking, alcohol consumption, MVPA, and social contact) and memory and fluency decline over 10 years. These models included time (years since baseline), age at baseline in years (denoted age0), the interaction between time and age0 (denoted time x age0), all covariates, each behavior, and the interaction between each behavior and time (denoted behavior x time). Age0 was centered at 65 years, the mean in the pooled analytic sample. After determining that memory had a non-linear association with time, we also included time2 and behavior x time2 terms in the memory model only.
We then fitted similar models to examine differences in memory and fluency decline over 10 years between the recommendation-compliant lifestyle (non-smoking, no-to-moderate alcohol consumption, weekly MVPA, and weekly social contact) and the other lifestyles. In these models, instead of behavior and behavior x time terms, we included lifestyle, and a lifestyle x time interaction term, with the recommendation-compliant lifestyle as the reference lifestyle. Examination of lifestyle x time x gender and lifestyle x time x age0 terms suggested associations between lifestyle and cognitive decline were similar between men and women and for different ages; as such, we performed analyses in the entire analytic sample without stratifying by gender or age at baseline.
We first used these models to estimate how much memory and fluency scores declined over 10 years from age 65 in each lifestyle; this was done to contextualize differences in cognitive decline between lifestyles. We then determined the difference in memory and fluency decline over 10 years between the reference and the other lifestyles. All analyzes were performed in StataMP 18.0 with a two-sided p <0.05 considered significant.
Additional analyses
As individuals who abstain from alcohol use frequently do so due to poor health30, and as a result abstaining from alcohol consumption has been associated with adverse health outcomes including poor cognitive performance29, we examined how excluding abstainers from analyses impacted the results. As we could not account for changes in behavior during the follow-up period due to the lack of availability of consistent alcohol variables, we also examined the stability of the other three behaviors during the follow-up period to determine whether changes in behavior were likely to have influenced the results.
In further analyses, we examined the impact of adjusting for practice effect, body mass index, and self-reported hearing difficulty on the results as well as independent associations between smoking and cognitive decline when smoking was split into three categories (never smokers, former smokers, and current smokers).
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
