Participants and procedures
Fully anonymised data were drawn from the English Longitudinal Study of Aging (ELSA), a multi-disciplinary prospective cohort study that began in 2002. The sample includes nationally representative men and women aged 50 years and older.31 Data collection is performed in participants’ homes, through computer-assisted personal interviews (CAPI) and self-completion questionnaires biennially, then nurse visits quadrennially for biological samples. Cross-sectional data and longitudinal exposures were taken from wave 4 (baseline; 2008) and longitudinal outcomes from wave 6 (follow-up; 2012). Using imputed data, 10,749 participants aged ≥50 had measures on exposures and covariates at baseline. The sample was reduced by death, study exit, declined consent or ineligibility (e.g., anticoagulant medication; haematological disorders; a history of convulsions). 5,841 participants had complete data on biomarkers at baseline and 3,562 also had complete data at follow-up. Each biomarker was analysed independently. After exclusions on CRP values >20mg/L (n=116), the analytic sample for CRP was 3,968 (36.92%), 3,932 (36.58%) for Fb, 4,022 (37.42%) for WBCC, and 4,056 (37.73%) for IGF-1. There were no significant differences between participants included and excluded from analyses. Participants provided written consent and ethical approval was granted by the National Research Ethics Service (London Multicentre Research Ethics Committee).
Exposures | Wave 4
Contextual (Neighbourhood-level) Socioeconomic Indicators
The 2004 index of multiple deprivation (i.e., neighbourhood deprivation) for England is a relative measure of deprivation that combines multiple area-level socioeconomic indicators into a single deprivation score. It is predicated on 38 indicators, across seven domains: education; employment; income; skills and training deprivation; barriers to housing and services; living environment deprivation and crime; health and disability. Neighbourhood deprivation was demarcated into tertiles; the first representing the most deprived on a gradient to the third that represents the least deprived (reference category).
Compositional (Individual-level) Socioeconomic Indicators
Wealth. Calculated by summating total household wealth, as determined by net wealth from property, possessions, housing, liquid assets; cash, savings, investments, artwork, and jewellery, net of debt, exclusive of pension wealth. Wealth was divided into tertiles; the first representing the least wealth and the third representing the greatest wealth (reference category).
Education. Categorised into higher education (i.e., degree or equivalent; reference category); primary/secondary/tertiary education (i.e., A-level, higher education below degree, GCSE or equivalent); and alternative/none (i.e., foreign or no qualifications).
Occupational Social Class. A three-category version of the National Statistics Socio-Economic Classification:32 managerial and professional (reference category); intermediate; routine and manual.
Outcomes | Wave 6
Immune and Neuroendocrine Biomarkers
High-sensitivity plasma C-reactive protein (CRP; mg/L), plasma fibrinogen (Fb; g/L), leukocytes (white blood cell counts [WBCC]; 109/L), and serum insulin-like growth factor-1 (IGF-1; mmol/L) were dispatched to the Royal Victoria Infirmary (Newcastle-upon-Tyne, UK) for processing and analysis. Blood samples deemed insufficient or unsuitable (e.g., haemolysed; received >5 days post-collection) were discarded. Exclusion criteria included coagulation, haematological disorders, being on anticoagulant medication or having a history of convulsions.
C-reactive Protein. High-sensitivity plasma CRP (mg/L) was assayed using the N Latex CRP mono Immunoassay on the Behring Nephelometer II analyser (Dade Behring, Milton Keynes, UK). Intra and inter-assay coefficients of variation were <2%. The lower detection limit of the assay was 0.2 mg/L. CRP values >20 mg/L were excluded from analyses (n=116), as these were taken to reflect acute inflammatory processes rather than chronic inflammation.30 CRP was treated as continuous, with higher values indicating greater levels of inflammation.
Fibrinogen. Plasma Fb (g/L) was analysed using a modification of the Clauss thrombin clotting method on the Organon Teknika MDA 180 coagulation analyser (Organon Teknika, Durham, USA). Intra and inter-assay coefficients of variation were <7%. The lower detection limit of the assay was 0.5 g/L. Fb was treated as continuous, with higher values indicating greater levels of inflammation.
Leukocytes (White Blood Cell Counts). WBCC was analysed as continuous counts per 109/L; measured on a haematology-automated analyser (Abbott Diagnostics Cell-Dyn 4000 and Sysmex XE), with higher values indicating greater levels of inflammation.
Insulin-like Growth Factor-1. Serum IGF-1 (nmol/L) was measured using the DPC Immulite 2000 method, by an electrochemiluminescent immunoassay on IDS ISYS Analyser. Inter and intra-assay coefficients of variation were <14%. IGF-1 was treated as continuous, with lower values indicating greater neuroendocrine activity.
Factors likely to confound analyses were selected a priori, including demographic variables: age (≥50 years); sex (male; female); clinical variables: body mass index (BMI; calculated as weight in kilograms divided by height in meters squared [underweight:≤18.5; normal:18.6-24.9; overweight:25-29.9; obese:≥30kg/m2]); limiting longstanding illness (binary:- any long-term illness, disability, or infirmity that limits activity); and mobility difficulties (binary:- one or more difficulties mobilising [walking 100 yards; sitting 2-hours; rising from chairs after sitting long periods; climbing stairs; stooping, kneeling, crouching; reaching or extending arms above shoulders; pulling or pushing large objects; lifting or carrying objects over 10 pounds; picking-up a 5p coin]); lifestyle variables: smoking status (binary:- non-smokers/ex-smokers or smokers); alcohol consumption (binary:- low <3 or high ≥3 day weekly); physical activity (binary:- sedentary or moderate/vigorous weekly activity). Reference categories were being male, of normal weight, not having a limiting longstanding illness, being fully mobile, a non-smoker/ex-smoker, having low alcohol consumption, and being physically active.
Imputation. Missingness ranged from 0.00-52.33% (Supplementary [S] Table 4). Given the possibility of bias in complete case analyses,33,34 missing values on exposures and covariates were imputed using missForest based on Random Forests, an iterative imputation method, in RStudio v.1.4.1717. In the presence of nonlinearity and interactions missForest outperformed prominent imputation methods, such as multivariate imputation by chained equations and k-nearest neighbours.35 In ELSA, socioeconomic variables are the main drivers of attrition,31 so the assumption that missingness was at random (MAR) was likely to be met. The imputation of the missing values yielded a minimal error for continuous variables (Normalized Root Mean Squared Error=0.02%) and categorical variables (proportion of falsely classified=0.20%). Imputed and observed data were homogenous (Table S4).
Baseline characteristics were expressed as means and proportions. Logarithmic transformation was performed on CRP, WBCC, and IGF-1 values because of their originally skewed distribution. Fb was normally distributed. Cross-sectional analyses used multiple linear regressions to assess associations between exposures and outcomes at wave 4 (2008). Longitudinal analyses extended this to outcomes at wave 6 (2012). Results were presented as unstandardised (B) regression coefficients with standard errors (SE). Analyses were two-tailed. The basic model for the analysis can be expressed as: (Ŷi = B0 + B1X1i + B2X2i + ... + BpXpi + ui where Ŷ is the predicted value of the outcome; B0 is the value of Ŷ when all exposures equal zero; B1 through Bp are the estimated regression coefficients, X1-Xp are distinct covariates, and u is the error term). Each regression coefficient represents the change in Ŷ relative to a one-unit change in the respective exposure. Independent multivariate models were fitted to understand the role of different sets of covariates on associations. Biomarkers were modelled independently as CRP was linearly correlated with Fb (r=0.310), WBCC (r=0.262), and IGF-1 (r=0.158) at p<0.001. No further issues existed with collinearity and all models met regression assumptions. The unadjusted model (1), that conditioned on the baseline biomarker being measured, was included in all models. Model 2 adjusted for age and sex (demographic variables). Model 3 adjusted for BMI, limiting longstanding illness, and mobility difficulties (clinical variables); Model 4 adjusted for smoking status, alcohol consumption, and physical activity (lifestyle variables); Model 5 adjusted for all covariates. To test the extent to which different models explained associations, the B for outcomes were calculated using the percentage of the protective association explained (PPAE); a well-established epidemiological method36 using the formula: (where X is the model tested)PPAE = (B [crude model 1 and model X] – B [crude model 1] / (1-B [crude model 1]). Data analyses were conducted in Stata 17.1 (StataCorp, TX, USA).
Four sensitivity analyses were carried out on longitudinal associations. First, sets of covariates were added sequentially rather than independently. Second, due to the potentially confounding effects of inflammaging and somatopause, the moderating effect of age was tested (dichotomised by mean age [≥64.25 years]). Third, the exclusion of CRP values thought to represent acute inflammatory processes (≥20 mg/L) was reassessed on the basis of arguments put forward by Giollabhui et al. (2020)37, so regressions were repeated including those values. Fourth, analyses used complete cases to compare the efficiency and coverage of confidence intervals for the estimated coefficients and to ensure results were not an artefact of the imputed data. Association analyses replicated that in imputed data. The analytical sample formation is illustrated in Figure S1.