Opioid Epidemic or Pain Crisis?

Using the Virginia All Payer Claims Database to Describe Opioid Medication Prescribing Patterns and Potential Harms for Patients With Cancer

Virginia T. LeBaron, PhD; Fabian Camacho, MS; Rajesh Balkrishnan, PhD; Nengliang (Aaron) Yao, PhD; Aaron M. Gilson, PhD


J Oncol Pract. 2019;15(12):e997-e1009. 

In This Article


Purpose and Overview of Study Design

This secondary analysis used multiple statewide data repositories to analyze patterns of POM prescribing and potential harms for patients with cancer living in rural southwest Virginia (Table 1). This research was deemed exempt by the University of Virginia Institutional Review Board. In this article, we (1a) describe POM prescribing patterns at the patient, provider, and insurance claim levels and (1b) explore predictors of POM prescription frequency; (2) explore POM-related harms; and (3) map geospatial patterns related to POM prescribing, cancer incidence, and fatalities.

Data Sources

The primary data source was de-identified claims data from the Commonwealth of Virginia (CV) All Payer Claims Database (APCD). APCDs are state-run insurance claims programs that can be analyzed to evaluate health service delivery.[17–19] At the time of this analysis, the CV-APCD included paid medical and pharmacy claims data for approximately 4.5 million Virginia residents spanning commercial, Medicaid Fee for Service and Medicaid Managed Care, and Medicare Advantage insurance coverage. Additional data sources augmented geospatial analysis (Table 1).

Overview of Data Sampling and Analytic Procedures

We used two data subsets from the CV-APCD that included patients with cancer who had a primary residence in one of the six counties of Health Planning Districts (HPDs) 1 or 2 of southwest Virginia (Appendix Figure A1, online only) between 2011 and 2015: (1) outpatient pharmacy prescription claims and (2) hospitalization medical claims. CV-APCD analysts extracted data using relevant International Classification of Diseases and Diagnosis-Related Group codes after discussion and review by the lead investigator. The study years were selected to reflect an initial date of the Centers for Disease Control and Prevention (CDC) response to the opioid epidemic[20] and CV-APCD data availability. Specific HPDs were selected because of high cancer rates and opioid-related fatalities[7,21] and to better understand health care needs in two of the most rural regions in Virginia. Descriptive and inferential statistical analyses were conducted at the patient, prescriber, and prescription levels (SAS 9.4; SAS Institute, Cary, NC), and geospatial mapping (R package MapGam; ArcGIS 10) identified patterns and predictors of POM prescribing.

Figure A1.

Map of Health Planning Districts (HPDs) and respective counties of the Commonwealth of Virginia. PDC, planning district commissions.

Aim 1a: Describing POM Prescribing Patterns

Patient, prescriber, and claim samples were extracted from the CV-APCD data.

Patients. To focus analysis on adult patients with the highest likelihood of receiving POMs for active cancer-related pain, we included patients 20 years of age or older and excluded patients with diagnosis codes unlikely to require opioid pain management (Data Supplement). To better understand the frequency of POM use, we categorized patients into three groups on the basis of the number of Drug Enforcement Administration (DEA) Controlled Substance Schedule II (C-II) POMprescription claims they received in a calendar year: (1) three or more, (2) one or two, and (3) none.

Prescribers. We categorized prescribers as (1) those likely to prescribe C-II POMs for cancer-related pain (eg, oncologists) and (2) those unlikely to prescribe C-II POMs for cancer-related pain (eg, dentists). Opioid prescriptions written by providers in category 2 were not considered cancer-related opioid prescriptions in the analysis. We also excluded claims with missing prescriber specialty details, because we could not determine if opioid prescriptions written by these providers were cancer related.

Pharmacy Claims. We reviewed all POMs in the CV-APCD data set by National Coverage Determination code and excluded from analysis those opioid products used for nonanalgesic indications (eg, cough suppressants). We categorized the remaining opioids by their respective DEA Controlled Substance Schedule and accounted for hydrocodone upscheduling from C-III to C-II in October 2014. We focused our analysis on C-II POMs (Data Supplement), because they are indicated for treating moderate to severe cancer pain.[22,23]

Aim 1b: Exploring Predictors of POM Prescription Frequency

Two logistic regression models were fit by sex (to account for sex-specific malignancies [eg, prostate cancer]), with lung cancer as the comparator group, to examine factors predicting the number of C-II POM prescriptions received by patients with cancer in a calendar year. On the basis of available data, model predictors included type of cancer, age, county of residence, and Medicaid payer source (as a proxy for income level).

Aim 2: Exploring POM-related Harms

Within the medical claims data, we examined relevant diagnosis codes (Data Supplement) to identify patients hospitalized for an opioid use disorder (OUD). Using the same unique patient identifiers in both the prescription and the medical claims files, we linked the data sets to explore whether the number of POMs received by a patient may be related to subsequent OUD hospitalizations.

Aim 3: Mapping Geospatial Patterns

On the basis of the specificity of available geographical data, choropleth and heat maps were created to explore patterns related to POMs. Choropleth maps provide a visual representation of POM-relevant data at the county level (Table 1). Heat maps show probability rates for (1) a more-than-three-POM category patient receiving a C-II prescription, (2) a provider writing a C-II prescription, and (3) a pharmacy dispensing a C-II prescription. Red areas indicate areas of increased probability (or a hot spot) compared with the median probability in the region; blue areas indicate areas of decreased probability compared with the median.[24,25] (Data Supplement).