Participants
The study used data from the English Longitudinal Study of Ageing (ELSA), a representative panel study of adults aged 50 and older living in England. Data collection began in 2002–2003 (wave 1), with follow-up waves biennially [28]. Self-reported questionnaire and interview data are collected at each wave and biological and anthropometric data are collected at alternate waves. Ethical approval for ELSA was obtained from the National Research Ethics Service. All participants provided informed consent.
In the current study, we investigated the association between loneliness measured at wave 2 (2004–2005; the first wave in which loneliness was assessed) and incident type 2 diabetes from wave 3 (2006–2007) to wave 8 (2016–2017). Participants included in the analysis self-reported that they were free of diabetes/high blood sugar at baseline (2004–2005). The median follow-up time was 10 years. A total of 8780 participants took part in wave 2. Participants were included in our study if they had complete data on loneliness and covariates at baseline (2004–2005) and if they provided follow-up data on self-reported type 2 diabetes. Those with HbA1c values in the diabetes range [9] (≥6.5%; 48 mmol/mol) at baseline were excluded. A flowchart of those included and excluded from the study can be found in Fig. 1. Our analytical sample was 4112 participants.
In comparison with those excluded from the analysis (n = 4668), those included were significantly less lonely, and were more likely to be younger, wealthier and of white ethnicity (p < 0.001). They were less likely to smoke, were more physically active and were less likely to have hypertension or CVD at baseline (p < 0.001). They had a lower BMI on average (p < 0.001) and were more likely to consume alcohol regularly than those excluded from the analysis (p = 0.002). No sex differences were evident (p = 0.098).
Measures
Predictor variable: loneliness
We assessed loneliness with the three-item revised University of California, Los Angeles (UCLA) Loneliness Scale [29]. Participants rated items such as ‘How often do you feel you lack companionship?’ with response options of 1, ‘hardly ever/never’; 2, ‘some of the time’; and 3, ‘often’. Ratings were averaged to produce a score ranging from 1 to 3, with higher values indicating greater loneliness [23]. We also assessed loneliness as a continuous score (range 3–9) in supplementary analyses [8, 30]. The Cronbach’s α of the scale was 0.82 in our sample.
Outcome variable: type 2 diabetes incidence
Time to self-reported type 2 diabetes was assessed between wave 3 (2006–2007) and wave 8 (2016–2017). At each wave, participants were asked whether a physician had given them a diagnosis of diabetes or high blood sugar since their last interview. Time of diagnosis was indexed as the wave at which diabetes/high blood sugar was first reported. Time to event was measured in months from wave 2 (2004–2005) to the follow-up wave when diabetes/high blood sugar was mentioned. For those not diagnosed with diabetes by wave 8, time to censoring was the time from wave 2 to drop out.
Covariates
The covariates included in our analyses were measured at baseline (2004–2005). Participants self-reported their age, sex (man/woman) and ethnicity (white/non-white). We controlled for household non-pension wealth, which has been found to be the most relevant indicator of socioeconomic position for this cohort [28]. Wealth was divided into quintiles across the entire wave 2 sample. Participants self-reported whether they smoked (non-smoker/smoker), their frequency of physical activity (light or none weekly/moderate or vigorous once a week/moderate or vigorous more than once a week) and their alcohol consumption (≥5 times a week, <5 times a week). Height (cm) and weight (kg) were objectively measured during the nurse visit at wave 2 and used to calculate BMI (kg/m2). Participants self-reported whether they had received a doctor diagnosis of hypertension and this was combined with the objective nurse measure of blood pressure to create a binary variable (no/yes). We defined hypertension as systolic blood pressure ≥140 mmHg and diastolic blood pressure ≥90 mmHg. Participants self-reported whether they had angina, myocardial infarction or stroke, and we used this information to generate a measure of prevalent CVD (no/yes). HbA1c was objectively measured during the nurse visit and samples were analysed at the Royal Victoria Infirmary laboratory, Newcastle upon Tyne, UK. HbA1c values are reported in Diabetes Control and Complication Trial units (%) and International Federation of Clinical Chemistry units (mmol/mol).
Secondary predictor variables
Depression
Depressive symptoms were measured using the eight-item Centre for Epidemiological Studies Depression Scale (CES-D) [31], where higher scores indicate greater symptoms. Items included statements such as ‘I felt depressed’ and ‘My sleep was restless’. We excluded the CES-D item on loneliness to avoid direct overlap with the loneliness scale. A dichotomous response to each item (0 = ‘no’; 1 = ‘yes’) resulted in a total score ranging from 0 to 7. In line with previous work [23], a score ≥6 was used to define severe depressive symptoms. We also assessed depressive symptoms as a total score in supplementary analyses [30]. The internal consistency of the measure was acceptable (α = 0.76).
Living alone and social isolation
Participants self-reported whether they lived alone (no/yes). Social isolation was measured using an index based on the extent of contact within a person’s social network and their involvement with social organisations [23, 30]. Participants were asked about frequency of contact with their children, other family and friends, with response options of ‘less than once a year/never’, ‘once or twice a year’, ‘every few months’, ‘once or twice a month’, ‘once or twice a week’ and ‘three or more times a week’. Participants received a point if they had less than monthly face-to-face or telephone contact with each of the three categories of social tie. Participants received another point if they did not participate in any social organisation (e.g. social or sports clubs, churches or residents’ groups). Total scores ranged from 0 to 4, with higher scores indicating greater isolation. Few participants received a score of 4, so we combined categories 3 and 4.
Statistical analysis
Descriptive characteristics of the sample are presented as either mean (SD) or number (percentage). The characteristics of those who did and did not develop type 2 diabetes were compared using t tests for continuous variables and χ2 tests for categorical variables. Associations between loneliness and sample characteristics were assessed using Pearson’s correlations for continuous variables and univariate ANOVAs for categorical variables.
We established that the proportional hazards assumption was not violated using log (−log [survival]) vs log (time) graphs. Following this, we used Cox proportional hazards regression to investigate the association between loneliness and type 2 diabetes incidence, controlling for age, sex, wealth, ethnicity, smoking, physical activity, alcohol consumption, BMI, hypertension, CVD and HbA1c (Model 1). Loneliness was inserted as a continuous variable where the HR and 95% CIs represent a 1 U increase.
In secondary analyses, additional covariates were added to the model to test the independent effect of loneliness on diabetes incidence. In Model 2, depression was added. In Model 3, living alone was included. In Model 4, social isolation was added. Model 5 was the final model and included loneliness, all covariates, depression, living alone and social isolation together as predictors of diabetes incidence. We conducted collinearity diagnostic tests to check for collinearity. Variable inflation factors were <1.26, suggesting collinearity was not present. For graphical purposes, total loneliness score (range 3–9) was dichotomised using a median split into low loneliness (scores of 3) and high loneliness (scores 4–9). Incident cases are plotted on a graph to reflect the time to diagnosis for these groups.
We conducted a sensitivity analysis to address the possibility of reverse causality by excluding participants who developed diabetes within 2 years of baseline (wave 3; 2006–2007). In supplementary analyses, we examined whether there was a moderating effect of age, sex or ethnicity on association between loneliness and type 2 diabetes by adding interaction terms to Model 1. Age was entered as a mean-centred interaction term. We also checked whether the pattern of results changed when entering loneliness and depression as continuous scores. Analyses were conducted using IBM SPSS Statistics for Macintosh, version 24 (IBM, Armonk, New York, USA).