Data were derived retrospectively from a large de-identified claims database covering more than 12 million employees, retirees, spouses and dependents from self-insured companies in the USA. The database includes enrolment data, medical service claims (classified as inpatient, emergency department or outpatient services), associated International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnosis codes and prescription drug claims.
People with a recorded diagnosis of type 2 diabetes who had filled at least two prescriptions for an antidiabetes drug, but had no evidence of insulin use, were identified between 1 January 1998 and 31 March 2010. Type 2 diabetes was identified as the presence of ICD-9-CM code 250.xx (diabetes mellitus) without 250.x1 or 250.x3 (type 1 diabetes). Antidiabetes drugs were identified using generic product identifier codes. The first prescription fill for an antidiabetes drug was defined as the index date. Selected people were required to have continuous health plan enrolment for the year preceding the index date, which was defined as the baseline period. Selected people were followed from their index date until health plan disenrollment or the end of data availability (31 March 2010). People with insulin use during the baseline period or any time following the index date were excluded from the analysis. Selected people were classified as either having or not having evidence of medical care for a hypoglycaemic event during the study period, based on ICD-9-CM codes for hypoglycaemia (Table S6
) at any place of service [22
]. As in previous studies of hypoglycaemia, these observed hypoglycaemia events are considered as a proxy for greater underlying hypoglycaemia risk throughout the study period [5
]. To achieve a computationally manageable study population, a random subsample of people without hypoglycaemia was chosen to achieve a 5 : 1 ratio to the sample with hypoglycaemia.
Baseline characteristics assessed before the index date included demographics, the Charlson comorbidity index (CCI) [25
], common comorbidities of diabetes (obesity, mental disorders, neurological disorders, cardiovascular disorders, endocrine disorders and renal disorders) [4
], health conditions associated with accident risk (epilepsy, stroke and substance abuse) [28
], healthcare resource use [inpatient, outpatient and emergency room (ER) utilization] and type of index antidiabetes drug (see Table S6
for ICD-9-CM codes associated with each comorbidity). Baseline characteristics were compared between people with and without hypoglycaemia during the study period using chi-squared tests for categorical variables and Wilcoxon rank-sum tests for continuous variables. Accidents resulting in hospital visits during the study period were identified from inpatient and ER claims based on ICD-9-CM codes and were then grouped into three categories: accidental falls, motor vehicle accidents and other accidents (which included accidents occurring either at home, such as during housework or in relation to occupation, physical exertion, striking or being struck by an object, suffocation, foreign body entering eye/orifice, explosion and unspecified) (see Table S6
for ICD-9-CM codes).
For each accident outcome, a multivariable Cox proportional hazard model was used to assess the association between hypoglycaemia and occurrence of a first accident during the study period. Models were adjusted for baseline age 65 or older, gender, comorbidities of diabetes, conditions associated with accident risk, CCI, history of inpatient admission, index prescription for sulphonylurea and index prescription for thiazolidinedione. Adjusted hazard ratios (HRs) and corresponding 95% confidence intervals were estimated to compare people with, versus those without, a hypoglycaemia-related claim. Age group-specific HRs, for those aged ≥ or <65 years, were also estimated for the association between hypoglycaemia and each accident type by including an age-by-hypoglycaemia interaction term in the multivariable model. Annual incidence rates for accidents among people with, and without, hypoglycaemia were predicted based on the multivariable Cox model for hypothetical people with average baseline characteristics. A p-value of less than 0.05 was considered statistically significant.
As accidents could hypothetically increase the risk of hypoglycaemic events, especially if a person's eating routine is disrupted, a sensitivity analysis was conducted by excluding accidents that occurred within the 2 weeks before a hypoglycaemia claim. The time frame of 2 weeks was chosen for this sensitivity analysis based on the observation that three quarters of hypoglycaemia events among people with diabetes observing Ramadan occurred within the first 2 weeks of daytime fasting [29
All analyses were conducted using SAS statistical software version 9.2 (SAS Institute Inc, Cary, NC, USA).