The Impact of Opioid Prescribing Limits on Drug Usage in South Carolina: A Novel Geospatial and Time Series Data Analysis

This study developed a classification system based on the location of and distance traveled by stakeholders in prescription records to identify patterns and trends and detect potentially illicit activity in the opioid supply chain. We then used an interrupted time series quasi-experimental design to evaluate the impact of a prescription limit policy on opioid prescribing practices in South Carolina. This approach allowed us to examine the policy’s effects on different subgroups within the population, such as the location of each player (patient, prescriber, and dispenser), and to understand the overall impact on opioid prescribing practices in the state. To control for temporal trends in prescribing of similar medications, we also examined the volume of incident benzodiazepine prescriptions as a controlled substance that is not expected to be affected by the opioid policies [38].

In this study, data obtained from South Carolina Reporting and Identification Prescription Tracking System (SCRIPTS) is used [36]. The database contains 46 million prescription records from 2014 to 2022, including information about patients, prescribers, and dispensers for each particular opioid prescription. The interactions between them are analyzed using a network representation in Figure 1.

The primary focus of this study is the daily dosage of MME per prescription as a means of quantifying drug volume. MME is a unit of measurement that translates the total daily dosage of opioid medications into an equivalent dose of morphine for the purpose of comparing the relative strength of different opioids. This calculation is performed using a standard formula which converts the daily dose of each opioid into a morphine equivalent and adds these amounts together to determine the total MME. This allows healthcare providers to evaluate the potential risk of overdose or other adverse effects associated with opioid use. As recommended by CDC, prescriptions with daily MME values exceeding 50 MME require careful assessment, while those exceeding 90 MME are generally considered to be high risk and should be avoided [18, 17]. In order to express MME in daily terms, the total MME is divided by the number of days that the prescription is intended to last, as indicated by the ”days supply” field in the SCRIPTS data set. Prescriptions with missing days supply data were excluded from analysis in this study.

NarxCare: South Carolina has implemented several policies to reduce opioid misuse and abuse. One of these policies, called NarxCare, was implemented in May 2018 and limits opioid prescriptions to a five-day supply or 90 morphine milligram equivalents daily, with exceptions for severe illnesses. [16, 21, 37] We will focus on studying the impact of this policy because it is the most recent and directly affects prescription dosage. In order to implement the policy impact in the model, we use a binary indicator variable (0 = pre-implementation, 1 = post-implementation).

Internal control series: A dataset of benzodiazepine prescriptions was used as a control for comparison in the study because there is similar pressure to decrease benzodiazepine prescriptions but they were not specifically addressed in the NarxCare policy. All benzodiazepines dispensed from pharmacies in South Carolina were included in the dataset, including alprazolam, chlordiazepoxide, and others. These were used as a control because benzodiazepines are not specifically addressed in the NarxCare policy, despite the CDC recommending avoiding their use with opioid medications and in addition, they have a lower potential for addiction, which makes them a suitable candidate for comparison. [24].

In this section, we demonstrate the efficiency of our proposed classification model in section 2.4.1 using two hypothesis tests. We test if the consumption pattern is significantly correlated with the distance traveled by patients and the relative location of other players.

We conduct a cross-sectional analysis of prescription records from the state of South Carolina. To ensure the validity of our findings, data prior to 2014 was excluded from the analysis due to high levels of noise. Additionally, prescriptions with unusual MME exceeding $10^{5}$ were removed from the dataset. The data was further consolidated at the monthly level to enable time series analysis in the following section. This was done because any other level of consolidation would result in intermittent and non-stationary data, which is not compatible with ARIMA models. Table 2 provides a summary of key statistics, including the days of supply, MME, MME per day (calculated as MME divided by days of supply), and the class percentage (represented as the total number of records within a specific group, divided by the total number of records in the dataset) for opioid drugs.

The first observation is that most of the records (more than 90% of total transactions) have less than 250 miles of total traveled distance (class “0X”). It means most of the transactions are not suspicious in terms of distance. Any other distance categories (“1X”, “2X”, and “3X”) have an almost equal share of the data ( 3% of the transactions).

In terms of MME/day, as stated by CDC the Hazard Ratio (HR) for 1-19 MME/day (level 1), 20-49 MME/day (level 2), 50-99 MME/day (level 3) and any number above 100 MME/day (level 4) defines as 1, 1.44, 3.73, and 8.87 respectively [18]. Overall, according to CDC guidelines, any prescription with more than 90 MME/day should be avoided while the ones with more than 50 MME/day are required to be carefully assessed. In our data set, MME/day average is 46.85 with a standard deviation of 9.29 which falls in the second and third CDC classes and is over 50 for many records. Among different defined classes, five of them (classes 12, 22, 23, 32, and 33) have MME/day of more than 50 on average (Table 2). The higher MMEs are mostly in groups where patients travel more than 500 miles (probably to nearby states) to get their drugs from long-distance pharmacies which increases the probability of illicit activities. We state these observations as hypotheses 1 and 2:

$H_{1}$: The total distance traveled by the patient and the relative location of the prescribers and the dispensers is correlated to the patient’s daily MME.

$H_{2}$: Patients who travel more than 250 miles to obtain their opioid prescription from a distant pharmacy have a significantly higher usage rate compared to other patients.

The first hypothesis tests the suitability of the introduced classification system. If we do not observe any differences between classes, then the proposed system is not meaningful anymore. To evaluate its effectiveness, we utilize the one-way ANOVA test[31]. One-way ANOVA is not sensitive to deviations from normality and works well in the case of a large sample size [13]. As a result, ANOVA is a valid approach in this context and there is no need to perform other tests like Kruskal-Wallis (more statistical result are provided in supplementary material).

While there is no definitive evidence as to the source of diverted or misused opioids, the study of diversion of medically prescribed opioids is most robust in the acute opioid prescription phase [4]. In this regard, the second hypothesis aims to identify classes with a consumption level of more than 50 MME per day as a potential disorder of consumption behavior. According to guidelines from the CDC, any dosage above 50 MME per day is considered risky and should be carefully evaluated [18]. A one-sided t-test with 95% confidence intervals was used, with the alternative hypothesis being MME per day above 50 or 90, and the null hypothesis being MME per day less than 50 or 90. Table 2 shows that groups 12, 22, 23, 32, and 33 have 95% confidence intervals above 50 MME per day, while all other classes are less than 50 MME per day. Additionally, class 32 has MME per day above 90, which is highly risky. These results validate the second hypothesis.

The State of South Carolina implemented a policy called NarxCare on May 1, 2018, with the aim of addressing the opioid epidemic by limiting the number of opioid prescriptions that can be prescribed to patients [16, 37]. The policy’s impact was evaluated by analyzing data from SCRIPTS and comparing the volume of opioid prescriptions before and after the policy’s implementation. A simple pre-post policy comparison shows that on average, the MME per day was reduced by 2.59 (from 53.68 (95% CI, [53.33,54.02]) to 51.09 (95% CI [50.74,51.44])) and the majority of groups in the proposed classification system experienced a significant reduction in the volume of prescribed opioids, except for groups 22 and 31 (as shown in Table 3). To study this observation rigorously, a combination of ARIMA and ARIMAX models, as a class of interrupted time series models, were used to further analyze the policy’s impact over time and how it has affected different groups.

As a first step, an ARIMA model was applied to the pre-intervention data and used to predict the post-intervention prescription volume. The forecast results of the model, along with the 95% prediction confidence interval, are illustrated in Figure 2 (a detailed explanation of the models and parameters can be found in the supplementary materials). The discrepancy between the predicted values and actual ones in the opioid data is an indication of the intervention’s impact. In the absence of any intervention effect, it would be expected that the pre-policy and post-policy data would follow the same pattern. Notably, the discrepancy is not present in the benzodiazepines, which serve as the control group.

The same analysis was conducted at the class level to evaluate the impact of the policy on the groups of the proposed classification system, as shown in Figure 3. The desired mismatch, which is a reduction in actual values but no reduction in predicted values, is apparent for transactions with a total traveled distance less than 250 miles (first row) and patient-isolated classes (first column). However, when examining the prescriber-isolated classes (second column), it seems that the mismatch is occurring in the undesired direction, specifically in classes 11 and 31. This could be an indication of unintended consequences of the policy. The dispenser-isolated classes (third column) tend to have a higher MME per day, but there is no indication of reduction in post-policy predicted or actual values. In the last column, we can see the reduction in actual values and mismatch in groups 03 and 13. A similar analysis was also conducted for benzodiazepine prescriptions, and the results are presented in Figure 4. The policy does not appear to have a significant impact on the volume of benzodiazepine prescriptions, and the mismatch between predicted and actual values does not show any specific trend in many groups. (The details of the pre-policy fitted ARIMA models for each group are provided in the supplementary materials)

To validate the observations and evaluate the policy’s impact mathematically, an ARIMAX model was used, considering the intervention as event inputs in the form of level-shift, ramp, and inverse trend to the previously fitted ARIMA models. As shown in Table 4, it was found that the intervention has a significant impact on reducing opioid consumption in groups 03, 30, 11, 31, 13, and 23, while there is no impact on other opioid classes or benzodiazepines. These are important results since the desired impact in classes 03 and 30, which make up 40% of the total transactions, may lead to the desired overall impact for opioid drugs. Moreover, there is still no desired policy effect over classes 22 and 32 with the highest volumes of prescriptions. Furthermore, there is an unintended impact of the policy over dispenser-isolated groups (class 11 and 31).