Trends in healthcare expenditures and resource utilization among a nationally representative population with opioids in the United States: a serial cross-sectional study, 2008 to 2017

A serial cross-sectional design was used to compare the economic burden of adult household respondents with and without an opioid using pooled data from the Medical Expenditure Panel Survey (MEPS) between 2008 and 2017. MEPS is a nationally representative sample of the U.S. population and collects data on their use of health services including costs associated with specific services curated by the Agency for Healthcare Research and Quality (AHRQ) [6, 7]. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for a cross-sectional study design [8].

Data from adult MEPS household respondents (18 years old or older) between 2008 and 2017 were pooled. The pooled population was based on the subsample of the National Health Interview Survey households, which is a national representative sample of the non-institutionalized U.S. population. We used the consolidated MEPS Household Component, Prescription Medicines and Medical Conditions files to identify and characterize respondents with and without an opioid on their prescription profile. The Household Component file contains information on responder demographics, socioeconomic information, insurance information, employment information, and health status. The Prescription Medicines file contains information of respondents’ self-reported prescription drug fills. Medical condition file contains information about the respondents’ self-reported diagnoses.

Respondents with an opioid were defined as household respondents who reported having been prescribed an opioid prescription. Information on opioid use was acquired using the Prescription Medicines file, which provides details on the household-reported prescribed medications. Each record describes a unique prescription event (purchased or obtained by the household respondent). The therapeutic classes and subclasses of the prescription medication were based on Multum Lexicon Variables from Cerner Multum, Inc., which was used to identify household respondents with opioid prescriptions defined as narcotic analgesics. Data from the therapeutic classes and subclasses were cross referenced with the names of the medications and grouped into categories based on the number of unique opioid prescriptions acquired during the respective year: household respondents that reported having only one unique opioid in a given year, two unique opioids in a given year, and three or more unique opioids in a given year. Unique opioid represents the mutually exclusive generic name of the opioid.

The outcomes of interest included total healthcare expenditures, prescription expenditures, outpatient expenditures, emergency department expenditures, inpatient expenditures, number of prescriptions filled, number of office-based visits, number of emergency department visits, and number of inpatient night stays. Expenditure estimates in MEPS are based on the Medical Provider Component (MPC) and Pharmacy Component (PC) of the survey, which include payments and not charges [9]. Total healthcare expenditures captured all payments related to healthcare services including direct payments, out-of-pocket payments, and insurance payments (e.g., private, Medicaid, Medicare, and other sources). Prescription expenditures include out-of-pocket payments and insurance payers for prescription drugs. MEPS does not report expenditures for over-the-counter medications or inpatient administration of medications. Outpatient expenditures include all provider visits (e.g., physician and non-physician) in the ambulatory setting. Emergency department expenditures included all visits to the emergency department but does not include any visit that resulted in an inpatient admission to avoid double counting. Inpatient expenditures included all expenses for direct hospital care (e.g., room, board, diagnostic and laboratory work, and imaging); MEPS does not record provider services (e.g., anesthesiologists, radiologists, and other specialists) as part of the inpatient expenditures. All expenditures were adjusted for inflation using the Consumer Price Index to reflect costs in 2019 $US. Missing data for expenditures were imputed using a weighted hot deck procedure where other survey responses were used to input the missing data based on survey-weighted distributions [10].

Number of prescriptions filled was based on the PC and included the name of the medication, the number of times the medication was acquired, and payments associated with the medication. Number of outpatient visits was based on MPC and included encounters in office-based settings. Number of emergency department visits was based on the MPC and included the count of emergency department visits reported. Number of inpatient night stays were based on the MPC and included the total number of nights associated with a discharge event.

Respondent demographics that were collected included age (categorized as 18–24, 25–44, 45–64, and 65 and older), race (White, Black, Native American/Alaskan Native, Asian/Pacific Islander, and Multiple races reported), ethnicity (Hispanic and Non-Hispanic), marital status (Married, Widowed, Divorced, Separated, Never), education level (No degree, GED/High School, Associated or other degree, Bachelor degree, Master/Doctor degree, Not ascertainable, Don’t know, and Refused to answer), region (Northwest, Midwest, South, and West), poverty status (Poor/Negative, Near Poor, Low Income, Middle Income, and High Income), insurance coverage (Any Private, Public, and Uninsured), and comorbidities.

Federal poverty status was categorized based on the federal poverty level (FPL) defined by the Current Population Survey for the respective years: Poor/Negative (less than 100% of FPL), Near Poor (100% to less than 120% of FPL), Low Income (125% to less than 200% of FPL), Middle Income (200% to less than 400% of FPL), and High Income (greater than or equal to 400% of FPL).

Comorbidities included high blood pressure, coronary heart disease, angina, myocardial infarction, other heart diseases, stroke, high cholesterol, cancer, diabetes, and arthritis. Comorbidities were identified using MEPS priority conditions definitions that ask respondents if they were ever diagnosed with these conditions. Priority conditions were selected due to their high prevalence and established standards for clinical care.

A propensity score matching method for complex survey data was used to balance the measurable covariates between the two groups [11]. This allowed us to create weighted matched cohorts that would be generalizable to the original survey population. We applied this method to generate a 1:1 propensity match between respondents with and without an opioid. Propensity scores were generated using a logistic regression by regressing the covariates to the treatment assignment variable (respondents with and without an opioid). We included variables into the propensity score matching based on the Anderson-Newman Behavioral Health Model [12‐14], which provides a framework for the social and individual determinants of health care utilization. These include age, gender, race, ethnicity, marital status, education level, region, poverty status, and insurance coverage, and comorbidities. Moreover, we selected these comorbidities based on their impact on health care utilization and availability from the MEPS data [15, 16]. Matches were made using the nearest neighbor approach with a caliper distance of 0.01 without replacements. Balance between the groups was assessed using standardized differences; a value of 0.1 or less was considered balanced [17].

For the primary aim, we evaluated whether respondents with an opioid prescription have higher healthcare expenditures and resource utilization compared to matched respondents without an opioid. Healthcare expenditures included total, prescription, outpatient, emergency department, and inpatient expenditures. Healthcare resource utilization included number of prescriptions filled, number of office-based visits, number of emergency department visits, and number of inpatient night stays. Additionally, we compared the trends in healthcare expenditures and resource utilization between respondents with and without an opioid across 2008 to 2017. In the secondary aim, we performed a subgroup analysis to evaluate whether a higher number of unique opioids prescribed within a given year was associated with greater healthcare expenditures and resource utilization. Respondents were grouped into three categories based on the number of unique opioid prescriptions they received during the given year: one opioid prescription, two unique opioid prescriptions, and three or more unique opioid prescriptions.

We applied a survey-weighted two-part generalized linear model to compare the annual health expenditures and resource utilization between respondents with and without an opioid prescription adjusting for their characteristics [18, 19]. In the first part, we used a logistic regression model to assess the likelihood of having nonzero healthcare expenditures. In the second part, we used a generalized linear model with gamma distribution to evaluate the association between healthcare expenditures with treatment assignment conditions on whether the respondents had nonzero healthcare expenditures adjusting for respondent characterisitcs [20]. Results were reported as annual mean expenditures and resources utilized with corresponding 95% confidence intervals (CIs). Goodness of fit tests included the Pearson correlation of the predicted values and residuals, Pregibon’s link test, and modified Hosmer-Lemeshow test [21].

Comparison of trends between respondents with and without an opioid prescription were evaluated using linear regression models adjusting for covariates. An interaction term between the respondents with an opioid prescription variable and time was used to estimate the average annual differences in expenditures (differences in trends) between respondents with and without an opioid prescription across 2008 to 2017. These findings were presented as mean annual differences with corresponding 95% CIs.

Statistical significance was defined as a two-tailed alpha < 0.05. Propensity score matching was performed using the MatchIt [22] package for R software version 4.0.3 (The R Foundation for Statistical Computing; http://​www.​r-project.​org) [23]. All other analyses were performed using Stata SE version 15 (Stata Corp, Inc., College Station, TX).

There are several limitations with our study. Although we based our findings on a nationally representative sample of the non-institutionalized U.S. population, respondents are subject to recall bias, in particular when reporting on their healthcare expenditures and resource utilization. MEPS mitigates this problem by cross-referencing self-reports with the Medical Provider Component follow-back surveys collected from medical providers and pharmacies; however, we cannot rule out the possibility of error. Furthermore, we do not have data on illicit opioid use or opioid misuse which have been associated with increased healthcare expenditures and resource utilization. Diagnostic codes in MEPS only include the first three digits of the International Classification of Diseases, Tenth Edition codes due to patient confidentiality protection, which limited our ability to identify respondents with opioid use disorder or dependency. Additionally, we were unable to determine whether responders were using opioids for the first time or chronically using opioids. Moreover, increased expenditures may be driven by increases in the number of prescriptions filled and number of office-based visits; however, we were unable to determine the reasons for these behaviors. Furthermore, propensity score method requires inclusion of relevant potential confounders to balance the groups; however, unobserved confounders or omitted variables can compromise the internal validity of the method [36, 37]. Finally, the results from our findings were based on a nationally representative sample of the civilian, non-institutionalized U.S. population, which may limit the generalizability to other countries. However, other nations have reported similar experiences with the opioid crises and may find these findings useful [38‐41].