Data Source and Patient Selection
A retrospective study was conducted using data extracted from a large nationwide US administrative claims database (PharMetrics Integrated Outcomes Database) dating from July 2004 through July 2006. PharMetrics data represent more than 70 different managed-care organizations across the United States and more than 58 million commercially insured patients. The PharMetrics database is Health Insurance Portability and Accountability Act (HIPAA) compliant, de-identified, commercially available to the public, and widely considered exempt from institutional review board (IRB)/ethics committee approval. Due to full data de-identification on the collected data, IRB approvals were neither needed nor sought. The data encompasses comprehensive records on member demographic characteristics, health plan enrollment, inpatient and outpatient services, and prescriptions.
Diagnostic and prescription data were extracted for 12 months before the date of treatment initiation with duloxetine or venlafaxine XR (index date), between July 1, 2005, and July 30, 2006. The index date was defined as the date of the most recent prescription for duloxetine or venlafaxine XR where no prescription for or use of the same medication was present in the prior 3 months. Patients were included in the study if they were commercially insured, 18 to 64 years of age on the index date, and had 1 or more diagnosis of MDD (International Classification of Diseases, 9th Edition [ICD-9-CM] codes 296.2x for MDD single episode, or 296.3x for MDD recurrent episode) during the 12 months before the index date. Study patients were also required to have continuous enrollment for the 12 months before the index date. Patients were categorized into two mutually exclusive study cohorts based on the most recent index pharmacy claim: either for duloxetine or venlafaxine XR.
For patients in each cohort, we used National Drug Codes (NDC) to identify and categorize prior medications in the 12 months before index SNRI initiation, including antidepressants (SSRIs, monoamine oxidase inhibitors [MAOIs]; tricyclic antidepressants [TCAs]); anxiolytics; other psychotropic medications (e.g., antipsychotics, stimulants, atomoxetine, antimanics); sedatives/hypnotics; anticonvulsants; pain-related medications (analgesics, skeletal-muscle relaxants, antimigraine medications); other medications for gastrointestinal, cardiovascular, and respiratory diseases; diabetes mellitus; or allergies. Two additional variables were created to predict initiation of treatment with duloxetine versus venlafaxine XR based on prior uses of antidepressants and pain medications: prior uses for ≥3 unique antidepressants and ≥3 unique pain medications.
Comorbid diagnostic histories were identified based on ICD-9-CM codes in the 12 months leading up to and including the index visit. Medical conditions considered were other depressive disorders (300.4x for dysthymic disorder, 309.1x for adjustment reaction with prolonged depressive reaction, and 311.x for depressive disorder not elsewhere classified), pain, diabetic neuropathy, fibromyalgia, anxiety (classified as generalized anxiety disorder, panic anxiety, post-traumatic stress disorder, social anxiety, and other anxiety disorder), schizophrenia, bipolar disorder, organic psychosis, alcohol dependence, dyslipidemia, hypertension, sleep disorders, gastrointestinal disorders, diabetes mellitus, asthma, heart disease, attention-deficit/hyperactivity disorder (ADHD), drug dependence, and nondependent drug abuse. Specific pain diagnostic subcategories were also identified, including skeletal muscle, back, head, chest, neuropathic pain, irritable bowel syndrome, and other pain conditions not classified elsewhere. (ICD-9 coding groups available from corresponding author on request.) On the basis of prior diagnoses, a predictive indicator variable for patients with ≥8 unique medical disease classes was derived (see Appendix).
To compare health care utilization between the two SNRI cohorts, we calculated total health care costs based on amounts paid by health plans for medical services and prescription medications for 6 months before (and including the index date) and the 6 months after the index date. Medical costs were classified by place of service into inpatient, emergency room, outpatient, and pharmacy costs.
The following factors were compared for patients in the duloxetine cohort with those in the venlafaxine XR cohort: sociodemographic characteristics (age [mean age and by-group ages 18-35, 36-50, and 51-64 years], gender, plan type, geographic region, and prescriber specialty (psychiatrist or other at the index date), prior medication use, prior medical conditions, and health care claims costs for the patients. Chi-square and Mantel-Haenszel tests were performed for comparisons of categorical variables between cohorts, and 2-sample t tests, Wilcoxon signed-rank, and wilcoxon rank-sum test were performed for comparisons of continuous variables.
To determine predictors of initiation with duloxetine versus venlafaxine XR, a multivariate logistic regression model was used, with initiation of treatment with duloxetine versus venlafaxine XR as a binary dependent variable coded as 1 = duloxetine and 0 = venlafaxine XR. Covariates in the model included patient age (with age 18-35 years as a reference group), gender, prescriber specialty, dummy variables for prior medication use, and medical and diagnostic histories. Additional predictors included 3 derived indicators for prior uses of ≥3 unique antidepressants and ≥3 unique pain medications and patients with ≥8 unique disease classes. Adjusted odds ratios (ORs) and 95% confidence intervals (CIs) are presented to show the strengths of the associations with each significant predictor in the model. The reliability of the model was checked to evaluate predictability values, including receiver-operator characteristic (ROC) curves. All statistical analyses were performed using SAS version 9.1 (SAS Institute, Inc., Cary, NC). Tests were conducted at a two-tailed α = 0.05. Two-tailed p-values are presented unadjusted for multiplicity. However, as a total of 65 different covariates were examined, Hochberg's adjustment was computed and p-values of approximately 0.001 or less met the multiplicity adjusted level of statistical significance [21
]. Thus, p-values of 0.001 or less are denoted as statistically significant in the tables.