During the first decade of the 21st century, the United States experienced a dramatic increase in the number of controlled substances prescribed and dispensed, particularly opioids. The number of opioid prescriptions dispensed from retail pharmacies increased from 174.1 million in 2000 to 256.9 million in 2009 (
1,
2). Over the same period, prescription opioid misuse increased dramatically (
2–
7), along with a simultaneous rise in rates of overdoses due to these substances (
8). As dispensing rates began to decrease in the early 2010s (
9), rates of fatal overdoses due to prescription opioids plateaued. As this plateau began, heroin overdoses rose from 2010 to 2016 before also plateauing, followed by increases in overdoses from illicit synthetic opioids such as fentanyl (
8). These overdose trends suggest that dispensing trends affected prescription opioid misuse, and potentially the rise of heroin and fentanyl use. Yet, little research directly links local-level dispensing rates to individual-level use and dependence using nationally representative data, despite connections between dispensing and patterns of misuse and dependence as an often hypothesized pathway to overdose trends. In this study, we examined directly whether county-level dispensing rates affected individual-level prescription opioid and heroin misuse, frequency, and dependence in a nationally representative data set.
A body of research demonstrates a relationship between opioid dispensing patterns and adverse outcomes from opioid consumption. According to administrative and clinical records, for individuals with valid prescriptions for pain, patients with higher prescribed doses and increased days of supply are more likely to develop an opioid use disorder (
10) or die of an overdose (
11–
14). However, studies show that a substantial proportion of prescription opioid overdose deaths are associated with nonmedical use of opioids obtained via diversion (
15,
16), making evidence from patient behaviors less well extendable to nonmedical use. In the aggregate, increases and decreases in opioid dispensing are associated with concomitant deaths from prescription opioids (
17–
19). Although few national studies directly consider the relationship between dispensing and prescription opioid misuse, we can draw inferences from studies on prescription drug monitoring programs (PDMPs) as an indirect measure of opioid dispensing, as these systems of surveillance are intended to reduce unnecessary prescribing and dispensing. The evidence on whether PDMPs reduce prescriptions or reduce overdoses is mixed (
20–
25). A study of use patterns using this measure found little evidence that PDMPs affect nonmedical opioid use and dependence (
26). Studies with broader geographic scope have tended to use state-level measures, despite evidence that the quantity of prescriptions dispensed and acute outcomes such as overdose are more localized (
22,
27). For example, it is not clear that high rates of dispensing in the western counties of New York would significantly affect pharmaceutical opioid consumption in relatively low-dispensing counties of Queens and Brooklyn several hundred miles away, particularly since opioid diversion commonly occurs with medication diverted from friends and family (
28). In sum, past studies tend to focus on the effect of dispensing on pain patient outcomes, aggregate patterns of overdoses, or indirect measures such as state-level PDMPs and also obscure trends occurring at a more local level. Despite considerable attention given to the opioid crisis, analyses that build on this earlier work to examine the nuances of the relationship between opioid dispensing and opioid misuse and dependence remain much needed. Here, we consider the alternative hypothesis that dispensing rates are associated with changes in prescription opioid misuse outcomes; as we expect an increase in misuse and dependence when the dispensing rate was increasing and a decrease as it declined, this represents a two-sided hypothesis and a test against a null hypothesis of no association.
An additional concern about decreasing the number of available prescriptions is that reductions in the circulation of prescription opioids could push individuals with opioid dependence to illicit opioids such as heroin. While there is some evidence that many heroin users transitioned from prescription opioids (
29–
33), the relationship between this transition and the availability of prescription opioids has not been thoroughly explored using national data. In fact, Schuchat et al., from the Centers for Disease Control and Prevention (CDC), have argued that there is no evidence that prescribing practices have resulted in increased heroin use (
1). However, studies making this claim often rely on indirect evidence of the effect of dispensing rates on overdoses (
17) or a measure of dispensing such as PDMPs (
26). As Compton et al. have remarked, “Some persons certainly use heroin when they are unable to obtain their preferred prescription opioid; however, whether the increases in heroin trends in the overall population are driven by changes in policies and practices regarding prescription opioids is much less clear” (
33, p. 155). Again, the central piece of the argument, namely, whether changes in dispensing rates affected patterns of heroin use, remains uncertain. On the one hand, reductions in dispensing may have had a direct effect on transitions to heroin use. In such circumstances, we would expect to see growth in heroin use in locations that reduced opioid dispensing. On the other hand, the growth in heroin use may be an ancillary component of the wider opioid crisis, but not directly attributable to patterns of opioid dispensing. In examining heroin use outcomes, then, our alternative hypothesis is that decreases in county-level opioid dispensing are associated with increases in heroin use outcomes, while the null hypothesis is that of no relationship.
Relying on national data from the CDC and the Substance Abuse and Mental Health Services Administration (SAMHSA), this study builds on previous research to examine directly the relationship between opioid dispensing rates at the county level and individual-level pharmaceutical opioid misuse and heroin use as well as dependence. The findings may have implications not only for prevention and intervention but also for clinical practice.
Methods
Individual-Level Data
We used data from the National Survey on Drug Use and Health (NSDUH), a national survey on substance use conducted annually by SAMHSA since 1971. In a stratified, multistage area probability sampling design to produce representative national data, approximately 70,000 individuals age 12 and older are surveyed annually. Via a U.S. Census Bureau Research Data Center, we used restricted NSDUH data, which contain state and county Federal Information Processing Standard codes, from 2006 to 2016 to examine individual-level use, frequency of use, and dependence related to prescription opioids and heroin as dispensing rates shifted. The included years represent those for which all dispensing data and covariates were available.
First, we examined a binary indicator (created by the NSDUH) of any prescription opioid misuse or heroin use in the past year. Second, for both substances, given that the sparsity of data for specific counts precluded the use of count models, we used an ordinal measure of past-year frequency of use. The categories coded from the number of days are “none,” “infrequent” (1–11 days), “intermittent” (12–59 days), “regular but nondaily” (60–300 days), and “daily” (>300 days). Because the measure of prescription opioid frequency changed in 2015 from a past-year to a past-month reference period, the model specific to prescription opioid frequency does not include 2015 and 2016 data as the other models do. Finally, we examined a binary indicator of dependence related to both substances, which the NSDUH defines by DSM-IV criteria.
Within the NSDUH, we included appropriate individual-level covariates within the models described below. Specifically, we accounted for birth year (including a squared term to account for nonlinearity), race/ethnicity, gender, income, educational attainment, health insurance status, employment status, marital status, and U.S. nativity.
Dispensing Data
We used county-level opioid dispensing data from the CDC, including years 2006 through 2016. Our key variable was retail opioid prescriptions dispensed for 100 people per county (
34), which does not include administration of opioids in hospitals or other treatment settings. Although there were a total of 3,149 unique U.S. counties as of 2016, the number of counties in our analysis ranges from 2,637 in 2012 to 2,851 in 2015, with the number in other years falling in between. Missing counties primarily result from incomplete dispensing data within the CDC database, which contains data for 87.6% to 94.0% of counties in any given year. According to the CDC (
34), missing data “may indicate that the county had no retail pharmacies and/or prescribers, the county had no retail pharmacies and/or prescribers sampled, or the prescription volume was erroneously attributed to an adjacent, more populous county according to the sampling rules used.”
Policy Data
To determine the effect of county-level dispensing rates independent of the effect of opioid-related policies placed into effect during the period of observation, we used the Prescription Drug Abuse Policy System for a comprehensive listing of policy passage in each state. These measures included state-level variables for any PDMP, any expanded naloxone access to the public, Good Samaritan laws absolving criminal or civil liability in reporting an overdose, any pain clinic prescribing restrictions, and presence of medical marijuana laws. We note that medical marijuana laws—any laws that permit the use of cannabis for medical purposes—were not passed because of the opioid crisis but remain an important policy covariate pertinent to pain management and general patterns of substance use (
35).
County- and State-Level Covariates
The U.S. Census Bureau’s American Community Survey (ACS) and decennial censuses provided county-level time-varying covariates—the unemployment rate, the median household income, and percentages of households in poverty, foreign born, female-headed households, non-Hispanic Black, Hispanic, and over age 25 with a bachelor’s degree. We also accounted for urban/rural location using the National Center for Health Statistics urban-rural classification scheme. We used the 5-year ACS estimates because estimates for shorter time frames are available only for larger counties. Because these begin in 2009, we linearly interpolated between the 2000 census and 2009, using the interpolated values for 2006 to 2008.
Statistical Analysis
We used logistic regression models with fixed effects for state and year to determine the effect of dispensing rates (
36). The strength of including fixed effects is the elimination of unobserved heterogeneity. Fixed-effects estimators are robust to any observed or unobserved time-invariant omitted variables, which removes any constant state-level effects (
36). We included a fixed effect for state to account for unobserved differences across locales over time. We also included a standard error cluster correction for county to account for dependencies between individuals within counties. Finally, to provide estimates of the effect of dispensing rates independent of changes over time, all models included year fixed effects. All analyses included statistical weights to account for the stratified multistage probability sampling design. Since we use logistic regression, all model results are presented as odds ratios with 95% confidence intervals. Given the nature of the variable, we used ordinal logistic regression to model frequency of use. The variance inflation factor indicated no issues with collinearity.
To provide additional evidence for findings, we also present Bayes factors (
37). In particular, Bayes factors are useful for providing evidence for null hypotheses, given that frequentist techniques (i.e., the parameter tests in the fixed-effects models) can only conclude that the null cannot be rejected, rather than providing evidence for the null. A Bayes factor of less than 1/3 indicates evidence for the null hypothesis, with the following levels: 1/3–1/10=moderate; 1/10–1/30=strong; 1/30–1/100=very strong; <1/100=extreme (
37). Bayes factors also support the alternative hypothesis when the value exceeds 3. When a theoretical prior distribution is unknown, the Bayes factor can be computed from a uniform distribution with a plausible maximum (
37). We take this approach here by using a uniform[0,1] prior distribution, but also note that our conclusions were identical using normal(0,1).
Across the years of observation, the entire NSDUH sample considered had about 748,800 respondents. All analytic sample sizes, in accordance with the restricted data agreement, are rounded to the nearest 100.
Results
Table 1 presents descriptive statistics. For the outcomes, 4.6% of the sample in all years reported prescription opioid misuse. Regarding frequency, 95.6% reported no misuse, 2.3% infrequent misuse, 1.3% intermittent misuse, 0.5% regular but nondaily misuse, and 0.3% daily misuse. Across all years, 0.5% of respondents reported dependence on prescription opioids. For heroin, only 0.3% of respondents reported any use. Not surprisingly, among heroin users the rates for each of the frequency categories was below 0.1%, and only 0.2% reported dependence on heroin. While several of these percentages are quite low, the substantial size of the NSDUH ensures that there are still no small cell sizes. For our main predictor of interest, the average county-level opioid dispensing rate across all respondents and years was 78.4 per 100 persons. Temporally, average county-level opioid dispensing rates rose from 80.5 per 100 persons in 2006 to a high of 96.1 per 100 persons in 2012, at which point rates began to steadily decline.
Table 2 summarizes the results of our logistic regression models. Models 1 through 3 show the outcomes for prescription opioid misuse. The county-level dispensing rate was significant and positively associated with all three outcomes. From model 1, a one-unit increase in the county-level opioid dispensing rate per 100 persons is associated with a 0.2% increase in the odds of individual-level past-year prescription opioid misuse (p<0.001), net of opioid-related policies, county-level sociodemographic measures, individual-level correlates, and state and year fixed effects. While this magnitude appears low, it only reflects the magnitude of a one-unit increase. For example, we can consider an increase of a standard deviation in the dispensing rate (SD=34.6; see
Table 1). A one-standard-deviation increase in the dispensing rate is associated with a 7.2% increase in the odds of past-year prescription opioid misuse. Model 2 considers frequency of prescription opioid misuse. A one-unit increase in the county-level opioid dispensing rate is associated with a 0.1% increase in the odds of being in a higher frequency category (p<0.01). For a one-standard-deviation increase in the dispensing rate, the associated odds are 3.5% higher of being in a higher frequency category. Finally, according to model 3, a one-unit increase in the county-level opioid dispensing rate is associated with a 0.3% increase in the odds of individual-level prescription opioid dependence (p<0.001). For a one-standard-deviation increase in the dispensing rate, the associated odds of dependence are 10.4% higher. While we interpret the increase given this expected positive relationship, we may also consider a one-standard-deviation decrease given that dispensing rates began to decline. In this case, a one-standard-deviation decrease in dispensing rates is associated with a 6.7% decrease in the odds of past-year prescription opioid misuse, a 3.4% decrease in the odds of being in a higher frequency category, and a 9.6% decrease in the odds of dependence. We also note that the dispensing effects presented are nearly identical in models without the policy variables (not presented here). Further, although we have strong support using a frequentist approach, we also note that Bayes factors for the effect of dispensing on each of these three outcomes indicate support for the alternative.
Models 4 through 6 in
Table 2 show the same three outcomes for heroin. There were no significant associations between the county-level opioid dispensing rate and these individual-level heroin outcomes. Using the alternative hypothesis that decreases in opioid dispensing increases the heroin outcomes, Bayes factors from the three models were 0.0050, 0.0084, and 0.0052, respectively, demonstrating very strong support for the null hypothesis.
Discussion
The substantial rise and subsequent decrease in prescription opioid misuse and dependence is often attributed to changes in opioid dispensing practices (
1–
9). This study builds on past studies to move beyond indirect outcomes such as overdose (
17–
19) and predictors such as PDMPs (
20–
26), geographically broad measures of dispensing, and administrative and clinical data on patient populations (
10–
16). In using a direct measure of county-level opioid dispensing linked to a nationally representative individual-level data set in the NSDUH, we find evidence of an association between higher levels of opioid dispensing and increased odds of past-year prescription opioid misuse, past-year frequency of use, and dependence. Thus, increases in dispensing at the local level enabled growth in misuse early on, while efforts to curb opioid prescriptions during more recent years appear to have had an effect on reducing prescription opioid misuse and dependence. The results thus support the conclusion that efforts to reduce opioid prescriptions have directly affected nonmedical use of pharmaceutical opioids.
Beyond the linkages of dispensing patterns to prescription opioid misuse, we find no evidence that shifts in local-level opioid dispensing affected odds of heroin use, frequency of heroin use, or heroin dependence, with the frequentist fixed-effects models confirmed by Bayesian techniques. This suggests that trends in heroin use may be an ancillary component of the opioid crisis rather than directly attributable to patterns of opioid dispensing. Given that many heroin users have transitioned from prescription opioid misuse (
29–
33), there has been a reasonable fear that well-meaning attempts to curb opioid prescriptions could result in additional opioid misusers making this transition. However, we found no evidence that this occurred at the population level, such that efforts should continue to reduce opioid prescriptions to levels required for patient care without fears that reductions may drive up heroin use. These results speak directly to Compton and colleagues’ contention that increases in heroin trends in the overall population may not have been driven by changes in policies and practices regarding prescription opioids (
33). These findings also cohere with research on discontinuation of opioids, as recent work shows that discontinuation without tapering was the norm for long-term opioid therapies and that changes in prescribing due to PDMPs do not increase abrupt discontinuation (
38). Nonetheless, as reductions in prescribing and dispensing continue, expansion of substance abuse treatment programs should occur, and such programs should remain available for those who misuse prescription opioids to prevent transitions to heroin (or synthetic opioids), which have been noted in the literature (
29–
33).
Notably, since an effect of dispensing on prescription opioid misuse remains after accounting for opioid-related policies, there are likely other mechanisms operating in the link between dispensing and opioid misuse beyond state policy implementation. We contend that these effects are attributable to more general shifts within the institution of medicine beyond those defined by public policy. General attention to the opioid and overdose crisis may have made providers more cognizant of their own prescribing behaviors (
1). For example, physicians may alter their own prescribing behavior after becoming aware of overdoses within the community (
39). Moreover, as the crisis became apparent, emphasis on physicians considering nonopioid pain management alternatives increased (
40). Thus, although other research has shown that policies such as PDMPs and pain clinic regulations reduce opioid dispensing (
17,
26), our results suggest that the effect of changes in dispensing on patterns of opioid misuse are not solely attributable to such policies and suggest broader changes in prescriber behavior.
Although this study has numerous strengths, we must consider some limitations. First, counties are an imperfect measure of geographic space, and measuring dispensing rates at this level can obfuscate important within-county differentiation. County size and number also vary considerably across states. However, use of county-level data constitutes a substantial improvement over measuring dispensing at the state level given the local nature of opioid prescription dispensing. Second, while we included a large battery of individual-, county-, and state-level controls as well as state and year fixed effects, we recognize that other factors may affect patterns of use and dependence. The use of fixed effects provides results robust to observable and unobservable static state-level factors, increasing confidence in our results. However, unmeasured time-varying factors, including those related to survey methodology, remain a possible source of confounding given the observational nature of the data. Third, with repeated cross-sectional data, we cannot examine within-person change. National-population geocoded individual longitudinal data on prescription opioid misuse and heroin, particularly of a sample size that allows the examination of these uncommon outcomes, remain unavailable. Lastly, we recognize that changes in opioid dispensing may also affect illicit fentanyl use; however, we cannot make any conjectures about fentanyl or fentanyl-adulterated heroin. NSDUH only recently began collecting this information, which limits examinations of fentanyl’s relationship with opioid dispensing. The results for heroin provide some promise that such a shift for fentanyl might not accompany dispensing changes, but this remains an open question.
Conclusions
Our findings indicate that county-level rates of opioid dispensing had a direct effect on individual-level opioid misuse and dependence, but reductions in dispensing did not have any adverse effects on heroin use. Institutionally driven changes among prescribers, potentially shaped by both professional recognition of the problem and policy implementation, may have helped curb the prescription opioid crisis; however, these changes do not appear to have altered heroin use (in either direction) following shifts in dispensing at the county level. We recommend that medical providers continue to monitor patterns of prescribing and dispensing and that states continue to pursue policies that temper unnecessary opioid prescriptions.