Table 2

Outcomes and analysis methods for 22 studies included in systematic review

Study	Cognitive outcomes	Risk factor aggregation, classification of risk factor exposure measure	Analysis methods
Betula Study 99	11 episodic recall tasks, 3 recognition tasks, 4 fluency tasks (2 semantic, 2 phonemic) and a spatial ability task Component scores. Cognitive scores converted to z scores and combined for each cognitive domain	Six factors were obtained from the 14 health variables using principal components analyses. Metabolic component (systolic blood pressure (BP), diastolic BP and body mass index (BMI) Glycaemic component (glucose, haemoglobin A1c) Lipid component (triglycerides, total cholesterol) Inflammatory component (erythrocyte sedimentation rate, haemoglobin, albumin) Nutritional component (B12, folate) Thyroid component (thyroid-stimulating hormone, thyroxine) Residual change scores computed for health and cognitive change between waves 1 and 2 and cognitive change between waves 1 and 3 and waves 2 and 3	Three sets of longitudinal analysis: 1. Health factors at baseline predicting cognitive change between waves 1 and 3. 2. Change in health factors between waves 1 and 2 associated with cognitive change between waves 1 and 2. 3. Change in health factors between waves 1 and 2 predicting cognitive change between waves 2 and 3.
Cache County Study 95	Incident dementia (Diagnostic Statistical Manual IIIR (DSMIIIR)) and Alzheimer’s disease (AD) (National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA))	Latent class analysis to identify patterns in the six lifestyle behaviours	Relationship between latent classes and incident dementia examined by proportional hazards regression using years to dementia starting at age 65 years. Four lifestyle classes were identified.
Cardiovascular Risk Factors, Ageing and Dementia study101	Incident dementia (DSMIV) and AD (NINCDS-ADRDA)	Dichotomised then summed three midlife risk factors (from surveys in 72, 77, 82 and 87)	Logistic regression comparing those with 1, 2 or 3 risk factors with those with no risk factors for incident dementia at re-examination as the outcome. Were there >3?
Coronary Artery Risk Development in Young Adults90	Digit Symbol Substitution Test, Stroop test, Rey Auditory Verbal Learning Test (long delay free recall)	Seven health factors were categorised as ideal, intermediate or poor heath using a slightly modified version of the American Heart Association (AHA) criteria (online supplement text 1). The total number of health components at ideal levels was calculated based on the average level of each across the 0, 7 and 25 year examinations as well as the number present at years 0 and 25. The score ranged from 0 (none at ideal levels) to 7 (all ideal). Also used a cut point of ≥5 ideal health metrics at 0, 1, 2 or all 3 examinations and a score of 1–14 where poor health scored 0, intermediate health 1 and ideal health 2 also based on average exposure	Multivariable linear regression was used to estimate the association between health components and each cognitive function. Multiple imputation used to impute missing values using data from all eight examinations and resulting in complete year 0, 7 and 25 data for 2932 individuals. Additional sensitivity analysis performed on 1753 participants who had complete information on all health behaviours and factors across all three examinations.
Framingham Study 96	Kaplan Albert neuropsychological battery—includes logical memory, visual reproduction, paired associates, digit span forwards and backwards, similarities, word fluency, delayed memory	Score used as the independent variable, 0 for neither risk factor, 1 for each risk factor and 2 for presence of both risk factors	Linear regression to examine the combined effect of obesity and hypertension on cognitive measures.
Honolulu Asia Ageing Study88	Incident dementia (DSMIIIR), AD (NINCDS-ADRDA), vascular dementia (criteria provided by the California Alzheimer’s Disease and treatment centres)	Risk factor measures converted to z scores. Those with skewed distributions were transformed prior to conversion and the relationships with the outcome checked to ensure they were linear. Z scores were summed over the seven risk factors thus ensuring a contribution from each risk factor	Logistic regression used to evaluate the relationship between the z score sum and dementia outcome. Results are reported as relative risks (RR) under the rare disease assumption that OR can be considered an approximation of RR.
Hoorn Study102	Neuropsychological battery including memory, working memory, immediate memory, delayed memory, attention and executive function, processing speed, visuoconstruction, language and abstract reasoning test z scores Z scores were adjusted on an individual basis for age, sex, IQ Impairment defined as z score≤1.5	Weighted risk score included additional weights for older age, lower education and sex. The four modifiable risks were scored as 0 or 1 and summed. See online supplementary table 1 for details of cut points.	Logistic regression to evaluate risk of impairment. Reanalysed excluding non-modifiable risk factors from the score.
Intervention project on cerebrovascular disease and dementia in the district of Ebersberg104	Dementia diagnoses retrieved from health insurance claims data, diagnosis was required to be recorded on at least two quarterly records	Six risk factors scored as ideal 2 points, moderate 1 point and poor 0 points	Proportional hazard regression used to evaluate the relationship between baseline score and incident dementia. Time from baseline until date of diagnosis in the health insurance records.
Kaiser Permanente Medical Care Program93	Dementia diagnoses entered by treating physician. Dementia ascertainment from 1994 to 2003	A composite cardiovascular risk score was created using midlife hypertension, diabetes, high cholesterol, smoking. Each risk factor scored 1 if present and summed to a maximum of 4	Proportional hazard regression used to examine the relationship between baseline risk factors and dementia outcomes.
Kungsholmen project100	Dementia assessment at each visit, DSMIII. Diagnoses: dementia and AD	Created vascular risk profiles by scoring vascular risk factors. Overall vascular risk profile included high systolic and low diastolic BP, low pulse pressure, diabetes or pre-diabetes, prior stroke and diagnosis of heart failure. Atherosclerotic risk profile included high systolic BP, diabetes or pre-diabetes and stroke; hypoperfusion risk profile included low diastolic BP, low pulse pressure and heart failure	Cox proportional hazard models used to examine various vascular profiles in association with risk of dementia and AD.
Maastricht Ageing Study103	Dementia diagnosis by consensus committee (neuropsychologist and neuropsychiatrist) based on DSMIV. Cognitive testing, verbal memory, executive function, processing speed. Incident cognitive impairment defined as <1.5 SD below the mean on any of the cognitive tests at 6-year or 12-year assessments	Risk score created by taking the natural logarithm of the RR for each risk factor, standardised by taking the result from the lowest natural log of the RRs as a reference value and dividing the other values by this value. Then summing the resulting scores assigned to each risk factor to create a risk score	Proportional hazard regression used to examine relationships between risk score and incident cognitive decline and dementia. Linear mixed models used to examine relationships between risk score and cognitive change.
Northern Manhattan Study91	Neuropsychological battery combined into z scores for episodic memory, processing speed, semantic memory and executive function (based on exploratory factor analysis of the full battery and prior work) For change in cognitive score, composite scores in the four cognitive domains were calculated using regression-based reliable change indices of the corresponding individual test adjusted for age, education years and the time between the two tests	The seven health factors were categorised as ideal or not ideal based on the AHA definitions and summed to reach a score between 0 and 7	Multivariable linear regression models used to examine the association between baseline health factor score and z scores at neuropsychological testing wave 1 and change in z scores between neuropsychological testing between waves 1 and 2. Scores were examined continuously and divided into four categories: 0–1 (reference), 2, 3, 4–7 health factors.
Personality and Total Health, Path through life study108	Neuropsychological battery including measures of verbal ability, processing speed, delayed and immediate recall, working memory and reaction time. A global score was calculated by summing standardised test scores for the six individual items and dividing by 6	A risk score (PATHrisk) was constructed from six individual risk factors (each risk factor contributed one point to a total of six)	Multivariable models were used to examine the relationship between baseline PATHrisk score and cognitive function across all three waves of the study. Two main models were used, the first included gender, time, PATHrisktime and PATHrisk. The second included gender, time, education, timePATHrisk and education*PATHrisk. Individual risk factors were also examined.
San Luis Valley Health ad Aging Study 92	Incident cognitive decline defined as a fall in Mini-Mental State Examination (MMSE)≥2 points at follow-up. Incident executive dysfunction defined as a decline≥0.5 of SD (2.5–3 points) in the executive control behavioural dyscontrol scale	The three individual risk factors were dichotomised as present/absent and summed to create a score	Logistic regression was used to examine the relationship between risk factors and cognitive decline.
Supplementation en vitamines et mineraux antioxydants study105	Several standard neuropsychological tests administered and two summary measures based on executive function and verbal memory plus an overall composite cognitive score derived	To identify latent unhealthy lifestyle factors related to cognition used structural equation models	Used analysis of covariance to estimate associations between individual and combined unhealthy behaviours (as categories and on a continuous scale). Also created and modelled a score of 7 dichotomised unhealthy variables.
Suwon Longitudinal Ageing Study109	Korean MMSE. Change over follow-up	Dichotomised then summed positive four health behaviours to form a protective score	Used multivariable linear regression to examine influence of risk/protective factors on cognitive change.
Uppsala Longitudinal Study of Adult Men98	Expert panel review of medical records up to 1 January 2010. Dementia (DSMIV criteria), AD (NINCDS-ADRDA), vascular dementia (Alzheimer’s Disease Diagnosis and Treatment Centre, mixed dementia (AD and cerebrovascular contribution)	Five risk factors scored 1 if present (smoking) or above a defined cut-off for BMI, systolic BP, fasting plasma glucose, serum cholesterol, maximum score 5	Cox proportional hazard regression used to evaluate risk of dementia calculated for individual and summed risk factors present at age 50 and at age 70 years.
Washington Heights cohort89	Consensus conference to diagnose dementia. Diagnosis of AD based on NINCDS-ADRDA	Four risk factors were dichotomised and treated as time dependent covariates where follow-up date was date of diagnosis. Median and quartiles used for BMI and low-density lipoprotein (LDL) cholesterol. Retained variables were summed to create a score. Date of event was age of onset of dementia	Proportional hazards regression. Risk factors entered into univariate analyses, those achieving significance values of ≤0.1 were retained in multifactorial regression.
Washington Heights cohort97	Consensus conference to diagnose dementia based on DSMIIIR. Diagnosis of AD based on NINCDS-ADRDA	Diet score non-binary, range 0–9, higher is better, physical activity dichotomised into low and high. Risk evaluated for combinations of physical activity and diet score.	Proportional hazard regression time to AD (first visit with AD diagnosis).
Washington Heights cohort94	A global composite, executive function composite and memory composite score from factor analysis of data from a neuropsychological battery	Four risk factors were dichotomised and treated as time-dependent covariates where follow-up date was date of diagnosis. Median and quartiles used for BMI and LDL cholesterol. Retained variables were summed to create a score. Date of event was age of onset of dementia	Multiple group parallel process random effects regression using data from all follow-up evaluations adjusted for retest effects.
Whitehall II study106	Global cognitive score combining z scores from tests of inductive reasoning, short-term verbal memory, verbal fluency. Cognitive function assessed at baseline, in 2002–2004 and 2007–2009	Examined association between smokers, never and ex-smokers, abstinent, moderate and heavy alcohol users and their interactions and global cognition score	Latent growth curve models (allowing correlation between repeated measures) to examine the association between smokers, never and ex-smokers, abstinent, moderate and heavy alcohol users and their interactions and global cognition score. Sensitivity analyses: analyses repeated for those with an MMSE≥24 in 2002–2004 and 2007–2009.
Whitehall II study107	Memory an executive function. The latter derived from a composite of three neuropsychological tests. Memory was assessed using a verbal memory free recall test. Poor executive function defined as the lowest sex specific quintile. Poor memory as <5/20 words correctly recalled	Summed (dichotomised) scores of 4 health behaviours at each phase and across all three phases	Univariate logistic regression relating individual health behaviours to cognitive outcomes at phase I, V and VII (cross-sectional) followed by summed (dichotomised) scores of health behaviours at each phase and across all three phases.