Data sources

Data for people diagnosed with T2DM in Scotland were obtained from the Scottish Care Information-Diabetes Collaboration (SCI-DC) data set,14 which stores clinical and demographic data on almost all patients with diabetes in Scotland. The study was based on an extract of SCI-DC data performed in 2011, and was restricted to people diagnosed with T2DM at 25 years of age and above during the period from 1 January 2004 to 18 May 2011. The study end date was defined as 18 May 2012, ensuring that 1 year of follow-up time was available for all patients. The SCI-DC data were linked to hospital discharge records, national death records and prescribing data by the Information Services Division (ISD) of National Health Service (NHS) National Services Scotland (NHSNSS). Generation of the linked data set was approved by the responsible ethics committee, Caldicott guardians and the NHSNSS Privacy Application Committee (application no. 33/11). All data were anonymised before release to the researchers.

Comorbidity and SES

Hospital discharge records covering the period 5 years prior to diagnosis with T2DM were used to construct four different measures of comorbidity. Scottish hospital records include six diagnosis fields (one holding the clinical condition mainly responsible for admission, the remaining five storing details of conditions which coexist or develop during the episode of healthcare and affect the management of the patient). The presence of a relevant clinical condition in any of these six diagnosis fields was accepted in constructing the comorbidity measures used in this study. For the period under investigation, all conditions were coded in accordance with the International Classification of Diseases, V.10 (ICD10).15 The first comorbidity measure derived from hospital records was the Charlson Comorbidity Index (CCI)7 as refined by Romano et al.16 This is a single-number (integer) quantity with possible values ranging from 0 (no comorbidity) to 29 (maximum comorbidity). Mapping of the original Charlson groups to ICD10 codes followed the scheme used by ISD to construct hospital standardised mortality ratios.17 The second representation of comorbidity was a set of binary indicators marking the presence or absence of 30 specific disease groups defined by Elixhauser et al.8 In constructing these indicators, the mapping of Elixhauser's original comorbidities to ICD10 provided by Quan et al18 was used. The third comorbidity measure constructed was the number of hospital in-patient bed-days accrued by each person during the 5-year period prior to diagnosis with T2DM. To provide a fourth measure, a group of binary indicators marking the presence or absence of 25 disease groups defined by ISD to guide coding of comorbid clinical conditions on Scottish hospital discharge records19 was created. These groups were developed by condensing 1127 ICD10 codes from 232 categories, and represent conditions deemed to have substantial prognostic significance. A fifth comorbidity measure was derived from patients’ prescribing histories, namely a count of the number of unique drug classes (British National Formulary (BNF) codes20) present in the database of medication prescribed in primary care during the 5 years before T2DM diagnosis.

The measure of SES used was the Scottish Index of Multiple Deprivation (SIMD).21 This is a small area-based ranked measure which combines 38 indicators of deprivation across seven different conceptual domains. Development of SIMD is sponsored by the Scottish Government via a rigorous academic process as the standard method of assessing deprivation in Scotland. It is the only measure of SES for which up-to-date national data covering the whole Scottish population are routinely available, and was thus the natural choice for this population-based study. SIMD is a ranked (as distinct from an absolute) measure and, for this reason, cannot be represented in analysis as a numeric quantity; rather, it must be included as some kind of categorical variable. For the purposes of this investigation, rankings of the 6505 geographical areas recorded in SIMD were expressed as quintiles (the first and fifth quintiles representing the most and least deprived groups in the population, respectively). This reduction to a five-way classification scheme offered a good compromise between sacrificing too much information (too few levels) and ending up with undesirably small numbers in some strata (too many levels).

Statistical analysis

An initial logistic regression model was fitted in which mortality at 1 year after T2DM diagnosis (ie, vital status dead vs alive) was predicted only by age at diagnosis (represented by a restricted cubic spline function to accommodate non-linear effects of age on mortality) and by SES (binary indicators marking quintiles 2–5 of SIMD, with the most deprived quintile 1 defined as the reference category). Separate models were fitted for each sex, to accommodate potential sex-related differences in the prognostic implications of some disease groups (notably cancer, which includes malignancies at sex-specific sites eg, prostate, cervix). From these initial models, the parameter estimates representing the effect of SIMD quintile on 1-year mortality were noted. Modelling was performed using a combination of R software (V.3.0.3; the ‘lrm’ function in package ‘rms’) and SAS (V.9.4; the ‘LOGISTIC’ procedure). The models were then extended by adding each of the five comorbidity measures in turn (only a single measure being included at any one time), and recording the amended estimates for the effect of SIMD on mortality. In fitting the models with comorbidity, certain manipulations of the comorbidity measures were performed in the interests of analytic practicality; these are now summarised for each measure involved. First, the distribution of the CCI was found to be heavily skewed, with high proportions of zero values, that is, no comorbidity (men: 80.5%; women: 80.4%). Reflecting this, the original scores on the CCI were reduced to a four-way categorical scheme: 0, 1, 2–3, and 4 and above (conceptualised as representing respectively zero, mild, moderate and high levels of comorbidity). Values of the CCI discretised in this way were found to exhibit material interactions with age, so terms representing the interaction of CCI with the spline components of age were included in the models. A second set of manipulations was motivated by problems of non-convergence experienced when fitting the logistic model which featured the Elixhauser comorbidity set. These difficulties were attributable to the indicator marking the presence of AIDS, which exhibited extremely low prevalence (<0.01%). Convergence difficulties also arose in connection with a small number of other ‘rare’ Elixhauser comorbidities, for example, blood loss anaemia (prevalence <0.05% in men). To permit successful modelling of the Elixhauser comorbidities, the indicator representing AIDS was omitted from the models: this was justified on the grounds that because the condition was so rare in the cohort, its effect on overall mortality was likely to be extremely small. To permit meaningful parameter estimates to be obtained for other sparsely represented conditions in the Elixhauser set, Firth's bias-reducing penalised likelihood method22 was adopted. A third series of modifications to the original comorbidity measure values was applied in connection with in-patient bed-days. The distribution of this was found to be heavily skewed with a small number of extremely large values, and was therefore reduced for analysis to the four-way ordinal classification used by Parks et al9: 0, 1–4, 5–10, and 11 and above. A fourth manipulation of comorbidity values was performed when fitting models which included the ISD comorbidity suite: here, the indicator representing the presence of AIDS/HIV was again omitted in the interests of permitting the logistic model to converge. Finally, values for the prescribing-based comorbidity measure (number of distinct BNF codes) showed a non-linear relationship with the outcome (1-year mortality), so this measure was represented in analysis as a restricted cubic spline function. In summary, the comorbidity measures as represented in the models consisted of two ordinal schemes (CCI and number of bed-days), two groups of condition-specific binary indicators (Elixhauser with 29 retained indicators and the ISD scheme with 24 retained indicators) and one continuous (more correctly, non-negative integer) quantity (number of unique BNF codes), represented as a spline function.