Data for this analysis came from the Health Core Managed Care Database. This limited, Health Insurance Portability and Accountability Act (HIPAA) compliant database contains approximately 2.1 million individuals with medical, pharmacy, and eligibility data in the health plan lines of business with complete capture of healthcare provider encounters. Data for the present study were obtained from the Southeastern health plan of the database and covered the time period from October 1, 1998 to April 30, 2003.
We examined the costs associated with a diagnosis of type 2 diabetes by focusing the analysis on individuals with continuous insurance coverage who were diagnosed with type 2 diabetes. Individuals were identified as having type 2 diabetes if they received an oral glucose lowering medication, or both insulin and a diagnosis of type 2 diabetes (ICD-9 of 250.00, 250.10, 250.20, 250.30, 250.40, 250.50, 250.60, 250.70, 250.80, 250.90 (e.g., 250.x0) or 250.02, 250.12, 250.22, 250.32, 250.42, 250.52, 250.62, 250.72, 250.82, 250.92 (e.g. 250.x2)), or at least two diagnoses of type 2 diabetes (ICD-9 of 250.x0 or 250.x2) between the period of October 1, 1998 and April 30, 2001 (e.g. the identification period). We identified an individual's index date as first date of diagnosis of type 2 diabetes or receipt of an antidiabetic agent (a sulphonylurea, amino acid derivative, biguanide, meglitinide, alpha-glucose inhibitor, insulin sensitizing agent, or antidiabetic combination) during the identification period and required that individuals have continuous insurance coverage, at a minimum, from twelve months prior to twenty-four months post-index date. An individual's post-period was allowed to be as long as forty-three months, provided they had continuous insurance coverage during the time-period of interest. We also required that each individual have at least two HbA1c values in the post-period. We included in the analyses the 10,780 individuals who fit the above criteria. Individuals were stratified into groups of good (N = 6,069), fair (N = 3,586), and poor (N = 1,125) glycemic control based upon HbA1c values of ≤ 7, >7 and ≤ 9, or >9, respectively, where HbA1c control groups were based upon the mean HbA1c level in the last year of the post-period.
The analysis focused on differences in costs for individuals with different levels of glycemic control. Costs were measured as the direct medical payments associated an inpatient or outpatient claims with a corresponding diagnosis of type 2 diabetes as well as receipt of any outpatient antidiabetic prescription medication. Such payments include both payments by insurance companies as well as payments by patients. All costs were converted into 2003 dollars using the medical care component of the consumer price index (Source: bls.gov). Costs are presented per member per year in order to account for the differences in follow-up periods.
Descriptive statistics included mean (± standard deviation [SD]) and median values for continuous data and relative frequencies for categorical data. Continuous variables were compared with analysis-of-variance (ANOVA) with Scheffe test for multiple-comparisons. This post-hoc test was chosen because it is more stringent (i.e., less likely to make an alpha error) as compared to other tests. Categorical variables were compared based on Pearson chi-squared tests. The length of follow-up was not uniform for all patients.
To determine if costs for post-index events were different between index diagnoses, multivariate generalized linear model techniques were used. Model covariates included demographic characteristics, patient severity, complications and comorbidities of clinical relevance, and HbA1c values. Patient demographic characteristics consisted of the individual's age, sex, and type of insurance coverage. Patient severity was proxied by previous resource healthcare use (e.g. prior year costs) as well as a count of distinct medications prescribed during the year prior to index date. In addition, disease prevalence was proxied by an indicator variable equal to one if the individual was diagnosed with type 2 diabetes (ICD-9 of 250.x0 or 250.x2) in the 12 months prior to the index period. Since an individual's index date was based upon first diagnosis of type 2 diabetes or receipt of an antidiabetic agent during the identification period, it is possible to be diagnosed with diabetes prior to the index date. Indicator variables were also constructed for the diabetic complications of nephropathy, retinopathy, foot ulcer, and amputation as well as for the comorbid conditions of myocardial infarction, stroke, coronary artery bypass surgery, and angioplasty. Identification of such variables were dependent upon receipt of a diagnosis or procdure in the one year prior to the index date. The diagnostic and procedure codes used to identify each of the above complications or comorbidities are given in Table . Finally, indicator variables were constructed for different rules of identifying individuals with type 2 diabetes. Specifically, indicator variables were set equal to one if an individual received a diagnosis of type 2 diabetes only or if an individual received insulin plus a diagnosis of type 2 diabetes.
International classification of diseases and current procedural terminology codes for complications and comorbidities
In estimating the multivariate regressions, costs associated with post-index episodes were assumed to follow a gamma-distribution. In addition, costs of care were empirically determined to be linear on the natural logarithm scale. Patient demographic characteristics, indicators of patient severity, type 2 diabetes classification and level of HbA1c
control were included in the model, while complications and comorbidites covariates with Wald p-values of < 0.15 were considered significant and included in the final model if they did not significantly change the Bayesian Information Criteria scores. The final estimated regression therefore included variables of clinical relevance and best model fit. These models were used to predict estimated costs per member per year with covariates held constant at their sample means. Statistical significance was defined a priori
at an alpha of less than 0.05. All statistical analyses were performed using STATA version 8.2. [18