Design
This retrospective study used a Generalized Linear Model to estimate the association between a patient’s health literacy and VHA medical costs, adjusting for covariates in a regional population of 112,417 veterans within the VHA. Because this methodology is capturing health literacy at the population level, there is no sample selection bias, thus we are isolating the effect of health literacy without introducing pre-selection bias. As such, we selected a controlled regression analysis to assess the cost profile with controlled variables of interest.
Population description
A total of 112,417 veterans with a health literacy screening between 2007 and 2009 were identified for the study in the North Florida/South Georgia region. This regional study population is representative of the southeast regions of VHA including both rural and non-rural dwelling veterans, with similar age and gender characteristics of not only VISN 8 but the national VHA patient population, being predominantly older white males. To investigate the impact of health literacy on VHA utilization (i.e., composed of: inpatient, outpatient and/or pharmacy), we analyzed data for patients who had utilization of services in each of the three years, for a final study population of 92,749.
Data sources
Veterans within the North Florida/South Georgia sub-region of the VHA are screened for health literacy once every five years (unless re-assessment is needed due to trauma or cognitive decline). Electronic health records within the North Florida/South Georgia sub-region of the VHA (VISN 8 region) were accessed to leverage this regional population of veterans that were routinely screened for health literacy between fiscal years 2007 to 2009. This sub-region is unique in that it is an early adopter for routinely assessing and documenting health literacy for their veteran population. The health literacy screening is conducted along with routine “clinical reminders” by clinicians and/or clinical staff during routine clinical care visits. The health literacy screening instrument is composed of a four item tool known as the BRIEF health literacy screening tool: (1) How often do you have someone help you read hospital materials? (2) How confident are you filling out medical forms by yourself? (3) How often do you have problems learning about your medical condition because of difficulty understanding written information? and (4) How often do you have a problem understanding what is told to you about your medical condition? These items have been found to effectively identify individuals with inadequate/marginal health literacy skills [13]. Response options are scored on five-point Likert-type scales for each of the items [items 1, 3 & 4 (1 = always to 5 = never); and item 2 (1 = not at all to 5 = extremely)]. The potential summative score from the BRIEF ranges from 4–20 with scores categorized as: (1) inadequate (4–12); (2) marginal (13–16); and adequate (17–20).
Data on workload was extracted from the centralized national Medical SAS datasets and cost data was extracted from the Decision Support System National Data Extracts (NDE). The national Diagnostic Cost Group (DCG) SAS datasets provided the DCG information. The computerized patient records system and other administrative systems store data collected during patient care encounters. These data are assembled at national levels into the VHA Medical SAS (MedSAS) Datasets, [14] Decision Support System - National Data Extracts (DSS-NDE) [15] and Diagnostic Cost Group (DCG) SAS datasets [16], stored on the Austin Information Technology Center mainframe. These files provided expenses, encounters and diagnoses for the study’s subset of screened patients, who also had used VHA care – outpatient, inpatient or pharmacy services – for fiscal years 2007 through 2009.
Procedure
Health literacy assessment data and corresponding identifiers for the study were electronically retrieved from the originating site and transferred via an intra-agency data transfer agreement to the investigators. Identifiers were then used to access and extract demographic, health condition and cost data from other VHA centralized administrative data sources. These identifiable data were extracted and collected with University of South Florida Institutional Review Board approval (CR6_Pro00000120). Data extracts were cleaned, and checked for administrative data entry errors by the study data managers. The patient identifiers were replaced using a de-identification cross-walk to a unique study ID and then all data were imported into a single analytical file for analyses.
The main effect (independent) variable is health literacy group
The primary outcome (dependent variable) was each patient’s annual aggregate VHA medical, fee-basis and pharmacy cost for fiscal year (FY) 2007 through 2009. VA inpatient, outpatient and pharmacy cost values from the national DSS-NDE were summed for each patient, in each year. It should be noted that dollar values are nominal, (not adjusted for inflation), and are taken from this VHA’s internal cost allocation system, which are more meaningful to VHA policymakers and administrators than externally valid expense values. Since we seek only to detect the relative within-VHA difference between health literacy groups, these values are sufficient, and do not require further adjustment for inflation or wage differences between facilities. Such adjustments would be necessary if we were comparing outcomes across years or facilities, respectively.
In the VA, cost estimates are recorded for each outpatient unit of care, (i.e., physician, practitioner, labs, and imaging) delivered in both Department of Veterans’ Affairs (VA) Medical Centers and in VA Community-based Outpatient Centers. Expenses for all inpatient care (e.g., acute, extended, observation and non-VA or contractual care) are also recorded. Pharmacy cost data for both outpatient or inpatient care are separately included. Thus, costs for each healthcare encounter (e.g. hospital discharge, specialty service, or outpatient visit) are accounted for in the VHA’s national MedSAS and DSS-NDE datasets for each patient across the national delivery system. Consequently, when aggregated to annual per patient levels, the total resources consumed for that individual whether in local or distant sites (e.g., while vacationing) is captured. We analyzed and report on the annual aggregate of these comprehensive dollar values labeling them as “medical and pharmacy” values to be clear.
Covariates
Guided by Paasche-Orlow and Wolfs causal model [12], demographics and constructed covariates were extracted for each member of this dataset from national VHA data systems. Diagnostic and procedural codes were processed to provide a risk-adjustment in each year at the patient level, i.e., DCG Hierarchical Condition Category (HCC). The DCG HCC score identifies the relative cost categorization by health status, calculated by accessing the clinical complexity and use of care. This analysis approach serves as a patient level measure of medical co-morbidities [17]. HCC categories were grouped into 5 major chronic disease groups for risk adjustment: Cerebral (HCC 097–099), Heart (HCC 083–084), Vascular (HCC 104–105), Cancer (HCC 007–010), Diabetes (HCC 015–020) and Arrhythmias (HCC 092–093). These conditions are common among veterans and are associated with higher costs of care [18–21]. A patient might be found in multiple conditions, if more than one was documented.
To capture relative levels of priority to access VA care for each veteran, Service Priority levels of 1–8 are routinely assigned to all VHA enrollees generally based on their service-connected disability and income levels. The priority level identifies the amount of copayment for VA services and the relative priority to be seen for non-service connected conditions. We created 4 groups for our analysis: priority 1 and 4 (“catastrophically disabled”), priority 2, 3, and 6 (“moderate disability”), priority 5 (“Medicaid assistance/low income”), and priority 7 and 8 (“no service-connected disability”), which served as the reference group. As a composite measure of access and income, this metric is well understood within the VHA system.
Multiple imputations
A multiple imputation method was used to replace (12-15 %) missing race and ethnicity in the analytical data. The two step process included creating 5 imputed datasets using the Markov Chain Monte Carlo method to make the data monotone in nature; and then using the monotone regression approach to impute the remaining missing variables [22]. There were no significant differences in the results from the imputed data and the original data.
Analysis
Patient characteristics and VA medical and pharmacy costs are reported with descriptive statistics. Analyses of bivariate categorical and continuous data were tested with chi-square or Kruskal-Wallis Test. Generalized Linear Model procedures with a log link function and Gamma distribution measured the association of health literacy with annual (2007–2009) patient-level VA medical, fee-basis and pharmacy costs, while other factors are controlled. Generalized Linear Model was used to account for the skewed distribution of the cost data [23]. Guided by the literature, final Generalized Linear Models were adjusted for race, gender, age, marital status, veteran’s service priority level, clinical complexity (co-morbidities and DCG score), and disease categories. We examined the potential for a more complex relationship to exist through Generalized Linear Models that included all covariates and interactions of health literacy with patient demographics and co-morbidity categories, for each year. Coefficients in the interaction models differed only slightly from the more simplified models, which are presented for parsimony [24]. P-values less than 0.05 were considered to be statistically significant. All analyses were conducted with Statistical Analysis Software (SAS, Version 9.3, Cary, NC). Adjusted risk ratios are reported with 95 % confidence intervals (CI).