Penn Medicine spans two states—NJ and Pennsylvania (PA)—within the greater Philadelphia area. For the study, we included new opioid prescriptions written in ambulatory, non-teaching practices within Penn Medicine from specialties represented in both NJ and PA. Specialties represented in both states included adult primary care (family and internal medicine), obstetrics and gynecology (Ob/Gyn), and cardiology. We also excluded practices with other ongoing opioid-related interventions (sports medicine) during the study period.
Study Design and Intervention
We used a difference-in-differences study design to estimate the effect of the combined policy-EMR intervention accounting for background trends in opioid prescribing. We compared practices in NJ exposed to the policy and EMR alert to a control group of unexposed practices in PA. The NJ law mandates that new prescriptions for schedule II controlled substances or any opioid drugs not exceed a 5-day supply.19 The law does not specify what constitutes a 5-day supply but requires the use of immediate release formulations at the lowest dose. The legislation passed in February 2017 and took effect in May 2017. PA also had a state prescribing limit in effect during the study period that limited new opioid prescriptions for adults to 7 days in specific contexts, including emergency departments (EDs), urgent care, and hospital observation settings. These restrictions did not apply elsewhere, including outpatient practices included in our study.14
Penn Medicine instituted an EMR alert in July 2017 to facilitate compliance with the law. The alert employed an “enhanced active choice” design, requiring a prescriber to make an active decision between alternatives while highlighting the advantage of one option.20 In this case, prescribers were notified if a new opioid prescription exceeded the 5-day limit based on the tablet quantity and dosing instructions. Prescribers were required to make an active choice to acknowledge the alert and then could either continue with the original order or adjust the prescribed quantity, with suggested amounts provided that were consistent with the law. More detail on the EMR intervention is available in Online Appendix A.
We included index prescriptions for the six most common opioids: oxycodone, morphine, hydrocodone, hydromorphone, codeine, and tramadol. Index prescriptions were defined as a new opioid prescription (no opioids in the prior year, consistent with NJ law) for adult patients (≥ 18 years) within Penn Medicine ambulatory practices. We excluded prescriptions for liquid formulations and for patients with cancer or enrolled in hospice or palliative care. We also excluded prescriptions written with a total opioid dose of ≥ 4000 MME or with long- and short-acting opioid formulations on the same day. These were unlikely to represent prescriptions for acute pain given the risks of respiratory depression associated with such doses.4 We conducted additional sensitivity analyses where we further restricted the study population to include those prescriptions most likely to represent patients who were opioid-naïve, including those with less than 90 tablets and excluding prescriptions written with refills on the original order (only possible for lower DEA schedule medications).
We used data from the EMR supplied by the Penn Data Analytics Center. Penn Medicine uses the Epic EMR (Hyperspace 2017; Epic Systems Corporation, Verona, WI). We had access to all prescriptions written within Penn Medicine.
We extracted patient demographics, including age, sex, race, ethnicity, and insurance status as well as ICD-10 codes for depression, alcohol use disorder (AUD), and other substance use disorders (SUDs). These diagnoses are associated with greater likelihood of long-term opioid use.21,22,23 We also extracted data on encounters for inpatient, ED, and ambulatory visits and documented patient telephone calls as well as prescribing provider specialty, provider type (MD or advanced practice provider), and practice location.
The primary outcome was total opioid dose per index prescription in morphine milligram equivalent (MME), a measure that reflects both the dose and number of tablets in the entire prescription and standardizes across opioid type.24 Although the outcome MME/day is commonly used in the literature, MME/prescription has been used previously in opioid prescribing policy25, 26 and hospital and health system–based studies27, 28 and allowed us to report change in the total volume of opioids in a new prescription. Further, by using total MME/prescription, we were able to distinguish between true decreases in prescribing and changes in dosing instructions (i.e., increasing the dose or frequency to fit within 5 days). We also assessed quantity of tablets/index prescription as an a priori secondary outcome. Finally, we analyzed the proportion of prescriptions written using prepopulated orders from the EMR preference list among prescribers in NJ.
We also examined secondary outcomes that might serve as proxies for under-prescribing or poor symptom control. These included proportion of prescriptions with refills within 30 days and numbers of telephone encounters, outpatient visits, and ED visits within 30 days of the index prescription. Although additional prescriptions and close follow-up may be appropriate or preferable to excessive prescribing, these could be markers of an overly restrictive policy if initial supplies were inappropriate to meet patient needs.
We included three time periods: (1) a 12-month pre-period (May 15, 2016–May 15, 2017), (2) a transition period during which the prescribing limit law was implemented without the EMR alert (May 16, 2017–July 7, 2017), and (3) a post-period in which the prescribing limit and EMR intervention were implemented (July 8, 2017–March 14, 2018). The 5-day limit passed in February 2017, so 1 year provided sufficient pre-period data from prior to the announcement of the policy that would not be subject to the effects of the announcement itself. The post-period was chosen because the health system implemented a similar EMR intervention in PA practices after that time.
For patients with an index opioid prescription, we compared patient and prescription characteristics by state before and after the law. We used Student’s t tests to compare continuous variables and χ2 test or Fisher’s exact test to compare categorical variables. We also included descriptive analysis of trends in prescription outcomes over time by state.
We then fit a multivariable linear regression model for each outcome variable to estimate the difference-in-differences effects. Opioid prescribing has been declining nationally over time,29 so comparing outcomes before and after the policy and alert may inappropriately estimate the effect. However, the difference-in-differences analysis allowed for comparison of NJ outcomes to a control group in PA within the same health system not subject to interventions, allowing us to account for these background trends. In order to use this approach, several assumptions must be met.30 First, the two sites should be subject to “common shocks,” meaning that other than the exposure being tested, outside events occurring during the study period should impact the intervention and control groups equally.30 In this case, clinics in NJ and PA were subject to the same health system policies and guidelines for opioid stewardship, with the exception of the prescribing limit and EMR intervention in NJ. Second, the outcomes need to meet the assumption of parallel trends prior to the exposure, meaning that trends must be similar in the in the pre-period, which we tested in our analysis.
We fit a linear regression model us to examine the intervention effects on both the absolute change and rate of change in outcomes between the two states. The model included the state, calendar time (prescription date), and indicators for the three time periods (pre-period, transition period with policy, and post-period with policy and EMR alert), and all two-way and three-way interactions between time period, calendar time, and state. We eliminated non-significant interactions, and only the two-way interaction between the time period and state was retained in the final model. The interaction between time period and state can be interpreted as whether the rates of change in outcomes differed between NJ and PA across time periods.
All regression models were adjusted for patient age, sex, race, insurance status, comorbidities (Charlson Comorbidity Index, and history of depression, AUD, OUD, and other SUDs), prescribing specialty and title (MD vs advanced practice provider), and presence of a procedure—all factors that could impact prescribing patterns or patient risks. We also accounted for clustering by prescriber in our model. Analyses were conducted using statistical software (Stata, version 15.1; StataCorp, College Station, TX).
Finally, we performed sensitivity analyses to test the robustness of our findings. First, we excluded patients from one commercial health plan that implemented its own 5-day limit in PA during the study period.13 We also limited our sample to only prescriptions of smaller quantities (< 90 tabs) or prescriptions written with zero refills in case we were capturing some chronic opioid prescriptions in our cohort.
The study was approved by the University of Pennsylvania Institutional Review Board.