Study design and sample
Participants were drawn from the European Prospective Investigation into Cancer and Nutrition (EPIC)-Norfolk study. The study methodology and sample distribution have been previously reported in detail [12]. In brief, this prospective cohort study recruited members of the general population, aged 40–79 years at the baseline, using general practice age-sex registers from the region of Norfolk – a relatively stable and immobile population in the UK. Participants attended for baseline health assessments between 1993 and 1997 and have since been followed up by way of additional health assessments, questionnaires and national registry data linkage. Ethical approval was obtained from the Norwich Research Ethics Committee.
Prognostic factor of interest
Investigating the determinants of mortality is one of the aims of this cohort. In this specific study, the prognostic factor of interest was fatigue, defined according to the UK version of the Short Form 36 vitality domain (SF36-VT). Although there is no single accepted measure of fatigue, the SF36-VT is commonly used and validated in studies of both diseased and general populations [13, 14]. In fact, it is employed as a gold standard tool to support the development of other fatigue assessment instruments [15]. Collected within the 18-month follow-up study questionnaire, the SF36-VT consists of four questions which refer to the previous 4 weeks: (1) Did you feel full of life? (2) Did you have a lot of energy? (3) Did you feel worn out? (4) Did you feel tired? These were summed and transformed into a 0–100 scale (where a high score represented low fatigue levels) using established scoring algorithms [14]. For the purposes of analysis, the final scores were categorised according to quartiles derived from the whole sample.
Outcome ascertainment
Mortality was ascertained using death certification linkage with the UK Office of National Statistics (ONS) at follow-up (censored in 2013). In addition to data on all-cause mortality, deaths specifically attributed to CVD (defined by International Classification of Disease [ICD], 9:401–448 or ICD 10:I10-I79) or cancer (defined by ICD 9:140–208 or ICD 10:C00-C97) were recorded.
Putative confounders/mechanisms
Multiple putative covariates, commonly considered by clinicians when evaluating the complaint of fatigue, were measured at baseline.
The mechanisms of fatigue are poorly described and some may also be considered confounders. We undertook a conservative approach by assuming that a factor fulfilling the statistical criteria for a confounding variable (i.e. a variable that is significantly associated with both fatigue and mortality) was a confounder. A covariate not fulfilling such criteria was considered to be a possible mechanism.
The covariates were collected by trained staff according to standardised protocols:
Self-reported measures
A survey capturing demographic and lifestyle information was conducted at baseline. This included age, gender and marital status. As with previous EPIC analyses, social class was categorised according to the Registrar General’s occupation classification scheme and further collapsed into manual or non-manual categories [16]. Education status of participants was defined according to whether or not they reached A-level standard (the senior secondary education examination in England). Self-reported total alcohol consumption was quantified as the total units of alcohol (1 unit equates about 8 g) consumed in a week and smoking as current/non-current. In order to ascertain the prevalence of baseline morbidity, participants were asked ‘Has your doctor ever told you have the following?’ followed by a list of illnesses which included cancer, diabetes, depression, heart attack and stroke. The SF36 bodily pain domain (simultaneously collected with the SF36-VT) was scored and transformed to a 0–100 scale employing validated methods [14]. As with SF36-VT, a high score represented a good health state, that is, low pain. Physical activity was assessed using the EPIC-validated questionnaire [17] which rated participants as inactive, moderately inactive, moderately active or active. The consumption of fruit and vegetables (g/day) was determined by the EPIC Food Frequency Questionnaire, also a previously well-described and validated tool [18] which asks participants to quantify their average consumption of a number of food items over the past year. Medication use, including aspirin, was self-reported and captured in the baseline survey.
Clinical measures
Height and weight were assessed and the body mass index subsequently calculated (weight[kg]/(height[m2]). This has been categorised according to World Health Organisation definitions [19].
Non-fasting blood samples were collected. AutoDELFIA time-resolved fluoroimmunoassay kits (Wallac, Finland) were used to assay thyroid-stimulating hormone (TSH) levels (hypothyroid status: >4.0 mU/l) in order to evaluate thyroid status; serum concentrations of the inflammatory marker C-reactive protein (CRP) were measured using a high-sensitivity assay on an Olympus AU640 Clinical Chemistry Analyser (Olympus UK Ltd, Watford, UK); and a Coulter MD18 Haematology Analyzer (Addenbrookes Hospital, Cambridge, UK) measured haemoglobin (Hb) status to enable assessment for anaemia (moderate anaemia defined as Hb < 110 g/L [20]).
Statistical analysis
The cross-sectional relationship between quartiles of fatigue severity and other putative confounders was first described using Chi-squared tests (for categorical confounder data) and ANOVA (for continuous confounder data), as appropriate.
Cox proportional hazards models were then employed to examine the longitudinal relationships between quartiles of fatigue and subsequent all-cause mortality over the follow-up period. First, an unadjusted model was developed (model A). Then, participants with a history of significant CVD (stroke and/or myocardial infarction) or cancer at baseline were excluded in order to avoid major confounding (model B).
Using simple descriptive statistics, as appropriate, the associations between the covariates of interest and fatigue and mortality, respectively, were calculated. Those covariates statistically associated (p < 0.05) with both SF36 vitality and mortality were considered to be confounders and those not were considered as possible mechanisms. Mechanistic factors are on the causal pathway between an exposure and an outcome; if adjusted for in the analysis, they will attenuate the observed exposure–outcome association. We tested our proposed ‘putative mechanisms’ by adjusting for them in the analysis with the expectation that if they are indeed mechanisms then the magnitude of association between fatigue and mortality will be reduced.
Confounders were grouped and introduced (model C) and then the influence of each possible mechanism was evaluated by submitting these, in isolation, to this model. Subsequently, all putative confounders and mechanisms were advanced together in a single model.
To better determine the precise causes of any observed mortality, the modelling process was repeated to examine the longitudinal relationship between quartiles of fatigue and subsequent (1) CVD-related mortality and (2) cancer-related mortality.
Last, models were re-run with the exclusion of all participants who died within 2 years of completing SF36-VT in order to investigate the potential for reverse causality, that is, the possibility of a significant subclinical disease causing fatigue.
Analyses were undertaken using SPSS version 22.0 (IBM, Chicago, IL, US). The proportional hazard assumption was checked by introducing an interaction term of time and vitality and was found not to be violated. All reported p-values are for two-sided significance tests and effect sizes are expressed as hazard ratios (HR) with 95 % confidence intervals (CI).