Setting
We studied adult members of Kaiser Permanente Colorado (KPCO). KPCO is a group model, closed panel, non-profit HMO providing integrated healthcare services to over 410,000 (covered enrollees), approximately 15 percent of the insured population, in the Denver/Boulder, Colorado metropolitan area. KPCO has over 550 physicians in seventeen separate ambulatory medical offices spread geographically across the greater metropolitan area. Kaiser Permanente Colorado's Institutional Review Board approved this project and the associated analyses of data derived from the administrative databases.
Study population
We identified 16,567 patients from KPCO's inpatient and outpatient based facility and professional claims, and internal outpatient primary care and specialty care databases for years 1996 through 2001. All patients were ≥18 years old at the time first index visit with a LBP diagnosis during 1997 or 1998.
Low Back Pain diagnosis and index visit identification
A diagnosis of LBP was based on algorithm using International Classification of Diseases, Ninth Revision (ICD-9) codes that was developed and validated at the University of Washington and has been used in several previous studies [20, 22, 23]. This algorithm consists of 66 ICD9 codes that include a broad array of diagnoses deemed consistent with mechanical LBP. This algorithm excludes patients if the back pain diagnosis was secondary to major existing conditions: e.g. neoplasms, osteomyelitis, spinal abscess, pregnancy, fracture, dislocation or vehicular accident. The index visit was defined as the first contact with a healthcare provider in either an ambulatory or hospital setting, resulting in a diagnosis, either primary or secondary, of LBP that was preceded by a 12 month period of continuous enrollment with no evidence of LBP. Patients were required to be continuously enrolled for a minimum of 24 months after the index LBP event.
We created a proxy variable to capture the variation in the number of back pain episodes across the study population. This variable was defined as the number of 30-day periods where a patient had one or more healthcare events for LBP. Given the distribution of this variable (67.37% had 1, 17.43% had 2, 10.72% had 3–5, and 4.8% had 6 to 22 episodes), for analytic purposes patients were grouped into LBP episode categories of 1, 2, 3–5, and 6+. In order to better understand the distribution of the diagnoses associated with the patients' index LBP visit, we used the four diagnostic categories noted in Vogt et al (2005) [20]. These categories are (I) LBP without neurological involvement, (II) LBP with neurological involvement, (IIIa) LBP caused by congenital lumbar spinal structural disorders, (IIIb) LBP caused by acquired lumbar spinal structural disorders, (IV) and LBP due to other causes including postoperative issues. For analytic purposes, categories IIIa and IIIb were grouped together.
Utilization measures
We captured the following measures of utilization for 12 months prior, and 24-months subsequent to the index back pain diagnosis: 1) outpatient care including primary care, specialty care, physical therapy, and mental and behavioral health; 2) all hospital based care including inpatient stays, emergency department (ED) visits, and observation stays of less than 24 hours in duration; 3) all pharmaceutical dispenses; 4) major spinal-related radiology procedures including CT and MRI.
We examined the variation in utilization patterns including outpatient primary and specialty care, mental health, physical therapy, imaging, pharmaceutical use, and hospital use by initially stratifying patients by the number of LBP episodes. Given that it is often difficult to isolate back pain specific healthcare resource use, we identified outpatient and hospital based care that was more likely to be related to back pain or musculoskeletal disorders. Outpatient care provided in orthopedics, neurology, neurosurgery, rheumatology, and physiatry departments was aggregated into an outpatient back pain specialty care category. Other specialty care included cardiology, endocrinology, gastroenterology, ophthalmology, etc. We classified hospital admissions that would fall into Major Diagnostic Category (MDC) using Diagnosis Related Groups (DRG). MDC 8 includes inpatient admissions related "to diseases and disorders of the musculoskeletal system and connective tissue" [24]. While this MCD in not specific to LBP, most, if not all, LBP related admissions would be captured in this diagnostic category.
Demographic and co-morbidity measures
Gender and age at the time of the index LBP event were available from the electronic membership records. To adjust for variation in health status and prevalence of chronic conditions that could influence both healthcare service utilization and costs during the observation period, a pharmacy based risk adjustment system, called RxRisk, was used to identify comorbidities [25]. The RxRisk model, also referred to as the Chronic Disease Score, is a clinically validated algorithm that classifies patients into chronic disease categories based on prescription drug fills [26]. The RxRisk (or CDS) system is a valid and reliable predictor of future health services use and future costs. Studies using this system demonstrate that it performs as well as instruments based on International Classification of Disease, 9th Revision, Clinical Modification (ICD-9-CM) inpatient and outpatient diagnoses [25, 27]. Using the RxRisk algorithm, dichotomous variables were created in order to assess the contribution and association of various comorbidities to utilization and cost estimates. We estimated the prevalence of RxRisk based comorbidities for the period 12 months prior to the initial back pain index date.
We used American Hospital Formulary Service groups and the National Drug Code system to identify and classify physician prescribed pharmacy dispenses for the LBP population into categories of non-steroidal anti-inflammatory drugs (NSAIDs) and opioids. We were not able to capture over-the-counter dispenses fro NSAIDs. We estimated the prevalence of use of these products for the 12 months prior and the 12 months after the index LBP visit.
Cost measurements
Using KPCO's cost management information system, we estimated annualized costs of care for services provided after the index back pain visit by type of utilization resource including outpatient, inpatient, hospital, pharmacy, and CT and MRI radiology procedures for the 24 months following the LBP index visit. This system allocated health care cost for all internal services provided directly by KPCO as well as claims for covered services enrollees receive from contracted providers. Internal costs are allocated by resource intensity weights (by service department and procedure) using KPCO's general ledger. Pharmacy costs are estimated using actual acquisition costs and KPCO specific pharmacy dispensing costs. All costs are reported in 1999 constant dollars. As a proxy for total annualized costs in 2005 dollars, we used data from the Medical care services component of the Consumer Price Index to inflate the 1999 cost estimates into 2005 dollars [28].
Statistical analyses
Using chi-square tests of proportion, we compared patterns of age, gender, co-morbidities, and utilization after the initial index visit, categorized by the number of LBP episodes. Given that hospital care may be the most costly component of health service use, we employed logistic regression analyses to examine the likelihood of an inpatient admission based on age, gender and pre LBP index visit comorbidities. Because the hospital admission event may be correlated with the number of LBP episodes, we did not include this variable in the final models. Separate models predicting MDC 08 admissions were also estimated.
In order to adjust for variation in health risk that may influence cost estimates, cost is estimated as a function of age, gender, comorbidities identified prior the back pain index event, and as an extension of a study conducted by Vogt et al, we also included variables capturing a dispense for opioids or NSAIDs in the 12-month baseline period prior to the index LBP event [20]. To avoid potential confounding with the dependent variable, the number of LBP episodes were not included in the cost models. Consistent with other cost studies, the dependent variable was the total cost over the two year post index period, annualized to avoid observations with zero charges and smooth individual year-to-year random variation and to minimize the effect of outlier cost events on the overall cost distribution [29, 30]. The mean and standard deviation of annualized costs, in 1999 dollars, for individuals in this cohort was $2,780 and $6,008, respectively. Given the skewness of the dependent variable, we used a weighted least squares model instead of a standard linear model [30, 31]. This method involves creating a weight from the residuals from the ordinary least-squares regression to adjust for heteroskedastic error terms. This weighted least-squares regression analysis provides unbiased regression coefficients and asymptotically efficient standard errors. In addition, use of the weighted least-squares regression analysis permitted health care costs associated with the covariates in the model to remain in nominal values. The use of nominal values eliminated the need to apply a variance-stabilizing transformation to the dependent variable and subsequently retransform regression results to obtain dollar values. Parameter estimates for each variable may then be interpreted as the marginal or incremental costs associated with patient falling into that particular category.