To generate the sample for this study, we had to determine which CBOCs (and VAMCs to which they were affiliated) to include, and then which veterans from these facilities to sample. CBOCs were included based on the following criteria: 1) CBOCs had to have congressional approval, 2) the CBOC had to be open by 10/1/98 to ensure that cost data was tracked for three years to ensure stable estimates, 3) the clinic must have enrolled 200+ veterans in FY99 for sufficient power, 4) cost data had to be trackable separately from the affiliated VAMC. There were 744 CBOCs and other outpatient clinics in operation in FY00. Only 315 of these clinics were CBOCs that required Congressional approval, which was the original process for designating CBOCs. Another 429 outpatient clinics that have been retrospectively labeled CBOCs but did not require Congressional approval were excluded to ensure parity between the CBOCs examined in the prior study and this study. Of these 315 CBOCs, 180 were excluded because they did not meet the inclusion criteria of being in operation before 10/1/98, leaving 135 CBOCs. Twenty-two of the 135 CBOCs were excluded because they cared for fewer than 200 veterans in FY99. The 113 CBOCs with sufficient sample size was reduced by five because they couldn't be found in VA administrative databases (n = 3) or were closed in FY01 (n = 2), resulting in 108 CBOCs that ml the inclusion criteria mentioned above. These 108 CBOCs were affiliated with 72 VAMCs because several VAMCs had established multiple CBOCs. Seventy-six of these 108 CBOCs were VA-staffed CBOCs, and the remaining 32 were Contract CBOCs.
Veterans in this study were classified into one of three mutually exclusive groups based upon where the veteran obtained primary care in FY00: VAMC patients, CBOC patients or crossover patients. VAMC patients were defined as veterans who had no primary care visits to a CBOC and at least one primary care visit to a VAMC primary care clinic, including general internal medicine, geriatrics, women's health, or primary care. This definition of primary care stop codes was consistent with the primary care codes defined in previous studies [4, 5]. VAMC patients could have no primary care visits to its affiliated CBOC, but could have visits to other unaffiliated CBOCs or VAMCs. CBOC patients were defined as veterans who had at least one visit to a CBOC and no primary care visits to the affiliated VAMC in FY00, but several CBOC patients had primary care visits at other VAMCs in FY00 not included in our sample. We defined them as CBOC patients for the purpose of this study because veterans may seek care at VAMCs not affiliated with their CBOC during vacations or other reasons. A crossover patient was a veteran who had at least one visit to a CBOC and an affiliated VAMC primary care clinic in FY00. In the prior CBOC analyses, 12% of CBOC patients had VAMC and CBOC primary care but were included in the CBOC group . Crossover patients were separately categorized from CBOC patients to examine whether their utilization and expenditures were different from VAMC and CBOC patients' use. We contrast the experience of CBOC patients as a group against the experience of VAMC patients and crossover patients in this paper, and examine the utilization and expenditure differences between Contract and VA-staffed CBOC patients in a separate paper (available from authors).
A random sample of 250 patients was drawn from each CBOC and VAMC primary care population to obtain sufficient power for all analyses and frequency weights were computed for use in regression analysis. For CBOCs with 200–250 enrolled veterans, all veterans were drawn for the study sample and frequency weights for regression analysis equalled 1.0. All crossover patients receiving primary care in the CBOCs and affiliated VAMCs in our sample were included and were assigned frequency weights of 1.0. Patients were then excluded if they didn't survive until the end of FY01 (n = 6,754), or had missing (n = 74) or excessive (e.g., >200 miles) travel distance from their home to the closest VAMC (n = 4,688), which resulted in 23,894 patients in 108 CBOCs, 26,176 patients from the 72 affiliated VAMCs, and 11,074 crossover patients.
The five datasets used in this study were drawn from administrative files that are maintained at the VA's central data repository, the Austin Automation Center. The first two datasets – the FY00 Outpatient Care File (OPC) and FY00 Patient Treatment File (PTF) – contained demographic information, county of residence, and utilization data that were used to identify the study sample and control for covariates related to health care utilization and expenditures. The outpatient and inpatient utilization files report the location of care (CBOC or VAMC), clinic stop codes that were used to identify the type of care, clinic visit date or admission and discharge dates, and diagnosis and procedure codes.
The second two datasets – the FY00 and FY01 Decision Support System (DSS) Outpatient and Inpatient National Extracts – contained similar utilization and diagnostic data as OPC and PTF, but also have direct, indirect and total expenditures for each VA outpatient visit and inpatient hospitalization. Expenditures were inflation-adjusted using the Medical Component of the Current Price Index. We also obtained 1999 Diagnostic Cost Groups (DCG) risk scores that are routinely generated for all veterans receiving care in a given year to adjust for differential risk between patient groups, because DCG was predictive of utilization and expenditures in the prior CBOC studies [4, 5].
The dependent variables for our analysis included outpatient and inpatient utilization and expenditure variables. Outpatient utilization and expenditures were partitioned into different types of care by clinic stop codes (e.g., primary care, mental health, specialty, pharmacy, radiology, laboratory, other) and summarized into total outpatient visits and expenditures. In the VA, lab space is physically separated from provider clinics, so veterans have to go to VA laboratories to have one or more tests done in a single lab "visit". Radiology and pharmacy "visits" work in the same way.
Inpatient utilization was defined as the number of hospitalizations, and inpatient expenditures were defined as the sum of expenditures across all hospitalizations in the same year. Demographic information used as independent variables regression analysis included age, gender, race, marital status, means test-based eligibility for free care, service-related disability, and 1999 DCG risk score. We also adjusted for distance between veteran's zipcode and the veteran's usual source of care (e.g., VAMC or CBOC) to control for access issues related to VAMC and CBOC travel costs. Differential distance, defined as (Distance between VAMC county and home county of residence) – (Distance between CBOC county and home county of residence), was used in primary care analyses, but distance between VAMC county and home county of residence was used to analyze all other outcomes.
To assess whether CBOCs influence the probability of care and/or the level of care, we estimate three measures of VA resource use – 1) odds of service use, 2) number of outpatient visits or inpatient hospitalizations, and 3) expenditures incurred. Ninety percent of veterans had primary care and outpatient pharmacy visits and 95% of veterans had one or more outpatient visits of all kinds, so the odds of use were not estimated for these two outcomes. Bivariate comparisons of odds of service were assessed via analysis of variance methods and bivariate comparisons of utilization and costs were assessed by the Cuzick extension of the nonparametric Wilcoxon rank-sum test for three or more groups .
In multivariate analysis, one-part models and two-part models of utilization and costs were run. One-part models were run for outcomes that had few observations with zero use (e.g., primary care, total outpatient, and overall care). Two-part models are used to estimate the odds of use and the level of use separately for those that use health care . A hurdle variant of the two-part model was run for outcomes that had many observations with zero values (e.g., mental health, specialty, radiology, laboratory, other outpatient, inpatient) because of the poor numerical properties of count data models (e.g., zero-inflated negative binomial models) that use the same covariates in both parts of the model . The odds of mental health, specialty, and ancillary (radiology, laboratory, and other) outpatient service use, as well as odds of inpatient admission, were estimated using logit regressions. The number of outpatient visits and number of hospitalizations were estimated using count data methods . The number of mental health, specialty, and ancillary visits for those with visits, and the number of hospitalizations for those who were hospitalized, was estimated with negative binomial models. This "modified" two-part model approach allows the conventional, separate modeling of the process for generating zeros and a count data model for users that explicitly accounts for overdispersion. Poisson regression models were estimated for total outpatient visits, primary care visits, and outpatient pharmacy visits because overdispersion due to high utilizers was not a problem .
Expenditure models for overall care, outpatient care, primary care and outpatient pharmacy were estimated using one-part generalized linear models (GLMs) with a normal distribution and log link function to restore log-normality. GLMs have been shown to fit expenditures with significant heteroskedasticity better than ordinary least squares  and have the added benefit of yielding predicted expenditures without having to retransform log costs. In a comparison of various generalized linear models and ordinary least squares from another CBOC study , the GLM with normal distribution and log link fit the middle three quintiles of the expenditure distribution better than OLS, based on predictive ratios . Inpatient, specialty, mental health, radiology, laboratory or other outpatient expenditures for those with one or more visits were also estimated using GLMs with a normal distribution and log link function. Specifications, cluster corrections for repeated measures on each veteran, and sampling weights were standardized across all regressions. Significance was lowered from 0.05 to 0.01 to adjust for multiple comparisons.
Since veterans were not randomly assigned to groups, we used propensity score matching with quintiles to improve the balance in the observed characteristics of veterans seen in CBOCs and VAMCs . To account for the possibility that the clustering of CBOCs within VAMCs and clustering of VAMCs within Veteran Integrated Service Networks (VISNs) might impact our variance estimates, we also tested models that included VAMCs as random effects and VISNs as fixed effects. These propensity score and cluster adjustments obtained similar results to our unadjusted models, so the unadjusted models are presented.