Design, setting and data source
We undertook a longitudinal cohort study of newly diagnosed patients with type 1 diabetes using data from three pediatric diabetes clinics that are part of the same Healthcare Trust (Barts Health NHS Trust) located in East London, UK. The three clinics largely capture patients living in surrounding areas of East London, where up to 56% of the local population belongs to an ethnic minority, with around 50% of South Asian origin (primarily of Bangladeshi origin) and 40% of Black origin (primarily of Somali origin).
The study was restricted to children <19 years of age who received a diagnosis of type 1 diabetes between 1 January 2005 and 31 December 2015 and attended any one of the three clinics during the same period. Extensive clinical and sociodemographic data were collected prospectively, both at the time of diagnosis and during routine clinic visits. As per recommendations from the National Institute of Health and Care Excellence (NICE), a child with type 1 diabetes is offered an integrated package of care by a multidisciplinary team at a pediatric diabetes clinic four times per year. The team consists of pediatric endocrinologists/diabetologists, diabetes specialist nurses, dieticians, psychologists and interpreters. HbA1c levels are recorded at each visit. All demographical and clinical parameters are systematically measured and electronically documented across all three clinics enabling comparison. Out of 596 children diagnosed with type 1 diabetes during the study period, 571 (96%) children had data on sex, age at diagnosis, duration of diabetes, ethnicity and SES and were eligible to be included in the analysis.
Primary outcome, exposures and covariates
The outcome of interest was glycemic control measured by HbA1c levels. HbA1c was measured at each visit using the point of care Siemens/Bayer DCA 2000+ Analyzer. HbA1c values recorded as percentages were converted to mmol/mol using the formula: (HbA1c value in %−2.15)×10.929.
The main exposures of interest were ethnicity and SES. Participants (or their parents) were asked to self-identify their ethnicity when they visited a clinic and we used the first recorded entries for ethnicity at the time of diagnosis. They were given the option to choose 1 of 15 categories or the option to decline identifying their ethnicity. For this study, the 15 ethnic categories were collapsed into six broad groups: White, mixed ethnicity (any mixed ethnicity combination), Black, African-Somali, Bangladeshi and Asian-Other (any Asian origin excluding Bangladeshi) which reflects the ethnic distribution of the study area in East London. The latter group included CYP mostly of Indian or Pakistani origin and a much smaller proportion originating from other Asian countries. The pH value (blood capillary samples) measured closest to initial presentation was used in the analysis.
SES was derived from postcode of residence using indices of multiple deprivation (IMD) 2010 for England. The IMD is a small geographical area measure of deprivation. It is multidimensional and scores are derived from a weighted combination of several indicators across seven distinct measures of deprivation including: income, employment, education skills and training, health, barriers to housing and services, living environment and crime.13 It captures the ‘relative’ deprivation experienced by an individual living in an area. IMD scores are calculated at the level of lower-layer super output areas, with each area comprising 1500 individuals on average. IMD rank scores were grouped into quartiles for the analysis, with the first and fourth quartiles corresponding to the most and least deprived, respectively.
Other covariates adjusted for in the analysis include: sex, age at diagnosis calculated by subtracting date of diagnosis from date of birth, age at clinic visit calculated by subtracting date of clinic visit from date of birth, duration of diabetes calculated in months by subtracting the date at first visit in the audit year from the date of diagnosis of type 1 diabetes, which of the three Pediatric Diabetes clinics the child attended and pH levels recorded at diagnosis—used as an indicator of diabetic ketoacidosis severity at presentation, measured in a subgroup of patients.
Statistical analysis
Baseline characteristics were compared across all ethnic groups. Categorical variables were compared as frequencies using χ2 or Fisher's Exact test. Mean differences in baseline continuous variables by ethnicity were analyzed using simple linear regression.
Stabilization of glycemic control during the first 6 months postdiagnosis was assessed using linear mixed-effects models (ie, a random intercept and random slope model), which allow comparison of population average HbA1c levels and change over time for the different ethnic categories while controlling for potential covariates. We approximated time trends using a quadratic model for time since diagnosis as this provided a better statistical fit than a linear model. Ethnicity, SES, age at diagnosis, sex and diabetes clinic were entered as time-invariant predictors. We constructed a series of models using the ‘mixed’ commands in Stata V.14 (StataCorp, College Station, Texas, USA). The first model (Model 1) was an unadjusted growth model using the quadratic function of time since diagnosis (disease duration in months) as the time metameter. Subsequent models were additionally adjusted for our hypothesized predictors: sex and age at diagnosis in years (Model 2), ethnicity (Model 3), SES (Model 4) and which of the three diabetes clinic the child attended (Model 5). We tested for a potential interaction between ethnicity and duration to assess whether HbA1c trajectories differed by ethnic group. We estimated all model parameters by maximum likelihood. We used generalized likelihood ratio statistics, −2 log-likelihood (−2LL), Aikake information criterion (AIC) and sample-adjusted Bayesian information criterion (BIC) to compare model fit between subsequent nested models and Wald statistics to test hypotheses about model parameters. We plotted quadratic growth curves at the group level (ie, ethnicity) to visualize model fit. Analyses were run in Stata V.14. In addition, we ran Models 3 and 5 above in a subgroup of patients with data on pH levels at diagnosis to assess any change to the observed ethnicity–HbA1c associations.