An opioid preferred drug list was not associated with an increase in emergency department visits or hospitalizations in a Medicaid population.
Medicaid provides prescription drugs for certain low-income individuals and families who meet state and federal eligibility requirements. State budgetary constraints coupled with rising healthcare costs have placed considerable pressure on state Medicaid programs to enact significant cost-containment policies. States have made attempts to reduce pharmacy costs by using a variety of strategies, ranging from cost-sharing to prior authorization policies. Preferred drug lists (PDLs) are used in many states and are associated with reducing costs by guiding prescribers to economically preferred agents. However, there is considerable heterogeneity in state management of pharmacy benefits. Currently, there is little evidence regarding the effects on patient outcomes—including the magnitude of potential risks—associated with different cost-containment policies.
In 2001, Oregon developed a PDL as part of the Practitioner Managed Prescription Drug Plan (PMPDP) for its fee-for-service (FFS) Medicaid program. The PMPDP used a systematic, transparent, and evidence-based approach to make the preferred drug selections and has been described previously.1,2 Initial implementation of the Oregon DL required the prescriber to indicate “medically necessary” on the prescription blank for nonpreferred agents. In May 2003, this method was phased out in favor of a quasi—prior authorization program that required prescribers to listen to an educational message summarizing Oregon’s evidence-based drug class review, with no submission of justification for request required.
Drugs were selected for inclusion on Oregon’s PDL based on extensive systematic reviews of comparative effectiveness and safety conducted by the Oregon Health and Science University’s Evidence-Based Practice Center. If the clinical evidence was insufficient to distinguish between drugs within a given class, the PMPDP legislation allowed for cost to be considered. Long-acting opioids were 1 of the 4 drug classes initially listed on the PDL. Oregon found insufficient evidence to determine that any specific long-acting opioid was safer or more effective than other available alternatives.3 Therefore, the preferred selections of long-acting opioid for the PDL were determined by price and included morphine sulfate long-acting (generic, Kadian, Oramorph SR), methadone (generic), and levorphanol (generic).4
Because Medicaid patients are a more vulnerable subset of the population, they also may be more sensitive to drug use limitations. Previous studies suggest that limitations on Medicaid prescriptions for chronic pain in vulnerable populations can result in a 35% reduction in the use of clinically essential drugs and increased use of health services related to a reduction in pain management that exceeded the costs of the discontinued drugs.5 Market share trends indicated that the PMPDP policy effectively increased the use of preferred agents and decreased total opioids dispensed, translating into reduced costs to the state.2 However, the unintended health outcomes of these policies remain unclear. Observers suggested that the reduced use of medications for chronic pain management might impact other sectors of healthcare utilization such as emergency department (ED) encounters and hospitalizations.
The goal of this study was to test the hypothesis that persons who switched opioids coincident with this policy would have higher inpatient and ED utilization rates than individuals not affected by the policy.
METHODSStudy Design and Data Sources
This study had a retrospective cohort design utilizing administrative claims from the Oregon Medicaid program. Data were abstracted from pharmacy and medical encounter claims from the Oregon FFS Medicaid program. The Oregon FFS population was a pool of approximately 100,000 high-risk, low-income beneficiaries not eligible for managed care coverage.1
We specified the following time periods for analyses: 1 year before policy implementation (August 2001-August 2002), during active policy enforcement (September 2002-October 2003), and 1 year after policy implementation (November 2003-November 2004).
The study sample included categorically eligible FFS Medicaid beneficiaries who were prescribed a long-acting opioid during the analysis period. We selected adult (age >18 years) users of long-acting opioids during the year before implementation of the PMPDP prescription practice guidelines who were continuously enrolled for the entire study period. Persons with a diagnosis of cancer pain (338.3) were excluded from the analyses using the International Classification of Diseases, Ninth Edition, Clinical Modification (ICD-9-CM) coding. The purpose of this study was to understand the effects of prescribing changes on persons using opioids for chronic pain. The indication for opioid usage in cancer patients is acute pain associated with the progression of their illness. Therefore, persons with a diagnosis of cancer pain were excluded from analyses. In addition, persons with a diagnosis of an opioid abuse condition (opioid dependence [304.0X, 304.7X, and 304.9X] or abuse [305.5]) were excluded. Substance abuse by itself is an indicator for increased health utilization. Persons diagnosed with an opioid abuse condition were excluded to minimize the potential confounding effects of these conditions.
We also restricted our study sample to beneficiaries who sustained prescription use for the prepolicy and active policy period. Sustained prescription use was defined as 1 prescription fill per quarter for any long-acting opioid during the prepolicy period. Individuals who discontinued use during the entire active policy period were excluded. Although discontinuation is an accepted adverse outcome of a prohibitory policy change, the number of persons who ceased filling prescriptions during the study period was insufficient for analysis of this group.
Finally, the study cohort was categorized into 1 of 2 groups depending on whether they changed their fill pattern during the policy period. Switchers were defined as those persons who continuously filled prescriptions for any nonpreferred long-acting opioid during the year preceding PDL implementation and were switched to a PDL long-acting opioid during the active policy period. Nonswitchers were defined as those persons who continuously filled prescriptions for any long-acting opioid during the year preceding PDL implementation and who either continued to have these medications dispensed or were switched to a nonpreferred long-acting opioid during the active policy period.
We examined 2 outcomes: ED visits and hospitalizations. Emergency department visits that did not lead to an admission were identified by procedure codes and revenue center codes. An ED encounter was defined using Current Procedural Terminology (CPT) codes 99281 through 99285, or as any visit that resulted in advanced life support (CPT code 99288) or encounters generating claims with a revenue center code of 45x or 981. Hospitalizations were identified through the Diagnosis-Related Group coding system.6 All claims submitted with a Diagnosis-Related Group payment were considered hospitalizations. The number of ED encounters and hospitalizations per person for each study period was used for analyses.
In addition to the primary predictor variables, the following baseline demographic variables were included for analysis: age, race, and sex, as defined by the first recorded medical encounter data. The following origin of pain indicators was included: osteoarthritis, low back pain, peripheral nervous system disorders, and fibromyalgia, defined by ICD-9-CM codes abstracted from medical record pain diagnoses. These are the most common chronic pain diagnoses for this population and were expected to identify a majority of appropriate persons suffering from noncancer chronic pain.
The Charlson Comorbidity Index was used as a measure of disease severity. This index serves as a proxy measure for mortality prediction based on prevalence of comorbid conditions and has been validated using administrative data.7-10
A polypharmacy variable was constructed to estimate the number of unique medications a beneficiary was exposed to in the prepolicy period. The presence of concurrent drug classes known to have pharmacodynamic interactions with long-acting opioids such as benzodiazepines, skeletal muscle relaxants, barbiturates, sedative hypnotics, and short-acting narcotics were quantified as a binary variable. Polypharmacy was quantified as number of unique drugs used per person during the prepolicy period.
A “difference-in-differences” approach was utilized to assess the changes in ED visits and hospitalizations before and after the Oregon Health Plan PMPDP policy change. In this approach, the first “difference” refers to health service event changes in those affected (switchers) 1 year before and 1 year after policy implementation. The second “difference” was determined through use of a comparison group (nonswitchers) that was observed during the same period but was not subject to prescription changes as a result of the policy change. This group provided baseline data on any secular changes that could be driving changes in our outcome not related to the policy change. Remaining significant differences may be considered attributable to the policy change.
We used a negative binomial model to estimate the number of ED visits or hospitalizations. In addition to the covariates described above, our model also included 3 dummy variables. The first dummy variable took a value of 1 if the beneficiary was a switcher and 0 otherwise. The second dummy variable took a value of 1 in the postpolicy period, and 0 in the prepolicy period. The third dummy variable was an interaction of these dummy variables, representing the difference-in-difference.
In nonlinear models (eg, negative binomial models), it is insufficient to interpret the coefficients and their standard errors as a true measure of the interaction.11 To provide an appropriate interpretation of the effect of the policy change, we estimated the model and then empirically simulated the outcomes under different scenarios (ie, with and without the PDL). We developed interpretable confidence intervals (CIs) by using bootstrapping methods. Specifically, we estimated the negative binomial model and saved the coefficients. Then, we estimated utilization (hospitalizations and ED visits) for 4 quantities of interest: (Q1) estimated utilization before the policy change, among nonswitchers; (Q2) estimated utilization before the 2003 policy change, among switchers; (Q3) estimated utilization following the policy change, among nonswitchers; and (Q4) estimated utilization following the policy change, among switchers. Our difference-in-differences estimate of interest is the mean of Q4 — Q2 – (Q3 – Q1). This process provided a single point estimate. We derived 95% bias-corrected CIs of our estimates through bootstrapping with 500 replications. To account for the multiple observations for each individual, we used block bootstrapping, with each individual defining a single block. We used Stata version 9.2 (StataCorp LP, College Station, TX) for all analyses. A more detailed algorithm and generalizable Stata code are available from the authors. As part of our results, we also display the relative risks (RRs) for all included variables in our negative binomial model, which we estimated with generalized estimating equations to account for the repeated-measures feature of the data.
Demographics and covariates were compared between switchers and nonswitchers with parametric and nonparametric statistical tests as appropriate. Statistical significance was considered to be P <.05. Sensitivity analyses for polypharmacy and comorbidity were conducted to investigate the impact of extreme outliers, excluding persons with more than 50 concurrent medications or persons with an index score of greater than 5. Results were not qualitatively different from those for full sample models. Therefore, outliers were not excluded in interpreted models.
In total, 947 Medicaid beneficiaries met our inclusion and exclusion criteria. The average age of study subjects was approximately 57 years, and the majority were female and white. The only statistically significant difference between switchers and nonswitchers was in low back pain. A greater proportion of switchers were diagnosed with low back pain than nonswitchers (P = .001). The characteristics of the study sample are shown in
displays simple statistics for outcomes by treatment group. Switchers and nonswitchers had similar mean ± SD values for prepolicy and postpolicy ED utilization (1.22 ± 2.69 visits vs 1.54 ± 3.56 visits prepolicy; 0.99 ± 1.46 visits vs 1.23 ± 3.53 visits postpolicy). Switchers and nonswitchers also had similar mean ± SD values for prepolicy and postpolicy number of hospitalizations (0.42 ± 1.22 vs 0.44 ± 0.97 prepolicy; 0.34 ± 0.84 vs 0.35 ± 0.82 postpolicy). The prepolicy and postpolicy differences were significant for both outcomes between groups. The mean number of ED visits for switchers significantly decreased compared with the number for nonswitchers (P = .001). The mean number of hospitalizations for switchers also significantly decreased compared with the number for nonswitchers (P = .005).
The interaction term, presented in
, suggests that the policy change did not have a differential effect on switchers’ utilization of the ED (RR = 1.05 [95% CI = 0.78, 1.40]) or the hospital (RR = 1.02 [95% CI = 0.64, 1.61]). Because the interaction terms are difficult to interpret in a nonlinear model, we present bootstrapped estimates of the effect of the policy change in
. We found no statistically significant changes in ED visits (0.13 [95% CI = −0.36, 0.56]) or hospitalizations (0.02 [95% CI = −0.12, 0.18]).
Table 3 also reports covariate risk ratios associated with ED utilization based on the multivariate generalized estimating equation negative binomial regression model. The following variables were determined to be significant predictors of ED utilization: sex (P <.001), low back pain (P <.001), number of concurrent medications (P <.001), and comorbidity score (P <.001). The ED utilization rate was 34% lower for men than for women (RR = 0.66 [95% CI = 0.55, 0.78]). Diagnosis with low back pain was associated with a 62% increase in the rate of ED utilization (RR = 1.62 [95% CI = 1.38, 1.90]). Each additional concurrent medication increased the rate of ED utilization by a factor of 1.01 (RR = 1.01 [95% CI = 1.01, 1.02]). Each additional comorbidity index unit increased the rate of ED utilization by 24% (RR = 1.24 [95% CI = 1.20, 1.29]).
Table 3 also reports the covariate risk ratios associated with hospitalizations based on the multivariate generalized estimating equation negative binomial regression model. The following variables were determined to be significant predictors of hospitalization: sex (P = .002), low back pain (P <.001), fibromyalgia (P = .05), number of concurrent medications (P = .001), and comorbidity score (P <.001). The hospitalization rate for men was 31% lower for men than for women (RR = 0.69 [95% CI = 0.55, 0.87]). Disability status was associated with a 30% increase in hospitalizations (RR = 1.30 [95% CI = 1.04, 1.63]). Diagnosis of low back pain was associated with a 47% increase in hospitalizations (RR = 1.47 [95% CI = 1.19, 1.82]). Each additional concurrent medication increased the rate of hospitalizations by a factor of 1.01 (RR = 1.01 [95% CI = 1.00, 1.01]). Each additional comorbidity index unit increased the rate of hospitalizations by 29% (RR = 1.29 [95% CI = 1.23, 1.35]).
Previous research indicated that the PDL effectively increased the use of preferred agents and decreased total opioids dispensed.2 Results from our study indicate that ED visits and hospitalizations were not different in individuals subject to long-acting (opioid) prescription changes due to PDL policy implementation compared with individuals not affected by the PDL. Our findings suggest that policy restrictions did not increase health service utilization as a result of less effective pain management for patients with chronic, noncancer pain. Our findings are in contrast to those of previous studies that documented the potential for PDLs to increase Medicaid patients’ use of other services that are potentially much more expensive.12,13
The validity of this finding is strengthened by the significance of multiple clinically relevant variable estimates. For example, the number of concurrent medications and the disease status of a beneficiary significantly predicted ED utilization and hospitalizations. Because these variables are clinically known risk factors, our finding of significance demonstrates the sensitivity of our study design to detect differences in utilization.
This observational study was complicated by factors not measured or accounted for in an administrative database. Comparison group specification and analysis methods were carefully chosen to minimize this bias. Specifically, a control group was chosen within the Medicaid FFS population to reduce the amount of between-subject variance. An alternative design could have designated a comparison group outside of the FFS population during the same time period. In addition, the difference-in-differences analysis was chosen because it is considered a valid quasi-experimental design to estimate intervention effects in nonrandomized settings. Difference-in-differences analyses control for secular trends in the outcome measure and quantify changes in response to an intervention in a nonrandomized setting.
A principal concern in the estimation of treatment effects is selection bias, as medication switchers could systematically differ from nonswitchers for reasons other than group status. The observational retrospective study design did not allow for control of all relevant factors, and bias may exist from unobserved and uncontrolled differences between the treatment groups. For example, mental health status was not quantified and controlled for by the Charlson Comorbidity Index in our analysis. If the degree of mental illness was related to cohort specification, this differential distribution of a relevant factor could have biased the results.
Our use of the Medicaid administrative claims database also is a potential limitation. The purpose of our study was to determine whether a reduction in pain management due to a restrictive drug formulary translated into increased health utilization. However, outcomes are only included in the database if a utilized service was reimbursed by the state and do not necessarily account for all services rendered to a beneficiary; therefore, these outcomes represent a minimum level of utilization for pain management. In addition, diagnosis misclassification bias can occur in administrative claims databases. Medical diagnosis coding could be influenced by reimbursement incentives and may not reflect the true medical condition. Finally, medical encounter and claims forms only permit use of proxy measures (eg, hospitalizations, ED utilization) for changes in pain management. These outcomes could lack specificity, and interpretations of direct causality are not possible. Moreover, the outcome measures available in an administrative database are markers for more serious outcomes, but are indirect and likely insensitive measures for poorer pain management or other clinical outcomes.
Finally, our study could have limited statistical power to detect differences in clinical outcomes. To ensure validity, the study cohort was created with very restrictive inclusion and exclusion criteria. Although approximately 8000 FFS beneficiaries used long-acting opioids for noncancer pain, our study cohort included only 947 persons. This small sample size could have provided limited statistical power to detect a difference between insensitive outcomes such as hospitalizations and ED visits. However, point estimates were clinically irrelevant, and increased sample size would not have led to different conclusions, assuming that results from a sample with enhanced statistical power would be similar.
Follow-up studies that included the incorporation of clinical outcomes would strengthen the interpretation of our results. Supplementing data from an administrative database with clinical data would enhance our ability to evaluate more specific patient-centered outcomes (eg, decrease in pain, improvement in function) and our ability to analyze potential confounders. Selecting a small sample of beneficiaries to create a detailed clinical picture would be a more appropriate design for addressing fluctuations in pain management. Outcome measures such as number of provider visits per month, adherence to prescribed opioid regimen, and self-reported adverse effects and pain management would be necessary to directly address the impact of the restrictive drug policy on patient-centered outcomes.
In our study, health utilization was predicted by factors such as disability status and specific pain diagnoses. The increased risk for ED visits and hospitalizations due to disability status warrants attention. Perhaps this increase indicates unmet needs in the population. Future studies should address the additional access barriers that result in higher utilization rates of the ED and hospital by disabled persons.
Legislative reversal of the restrictive policy prohibited lengthening the postpolicy period. It is possible that the policy effect was underestimated with this abbreviated observational period. This is a general limitation of policy research. Future studies should attempt to acquire data with longer periods of follow-up time when beneficiaries were subject to one stable, consistent policy.
Determining the implications of cost-containment policies in vulnerable Medicaid populations is pertinent for the current debate about the relative merits of drug policy in the public and private sector. Because of increased healthcare costs and a growing number of uninsured individuals, cost-containment policies continue to be attractive to payers. Understanding the magnitude of risks associated with these interventions is essential for prioritizing policies and improving population health. Public health professionals have a responsibility to create policies based on best-evidence practices and to design policies that cause the least harm to beneficiaries. Results of our study suggest drug policies targeting long-acting opioids may allow for cost savings to state Medicaid programs, while not increasing utilization of the ED or hospital by beneficiaries.