Intended for healthcare professionals

CCBYNC Open access

Clinical and healthcare use outcomes after cessation of long term opioid treatment due to prescriber workforce exit: quasi-experimental difference-in-differences study

BMJ 2024; 385 doi: (Published 16 May 2024) Cite this as: BMJ 2024;385:e076509
  1. Adrienne H Sabety, assistant professor1,
  2. Hannah T Neprash, assistant professor2,
  3. Marema Gaye, doctoral student3,
  4. Michael L Barnett, associate professor4
  1. 1Department of Health Policy, Stanford University School of Medicine, Stanford, CA, USA
  2. 2Division of Health Policy and Management, School of Public Health, University of Minnesota, Minneapolis, MN, USA
  3. 3Interfaculty Initiative in Health Policy, Harvard University, Cambridge, MA, USA
  4. 4Department of Health Policy and Management, Harvard T H Chan School of Public Health and Division of General Internal Medicine and Primary Care, Department of Medicine, Brigham and Women’s Hospital
  1. Correspondence to: M L Barnett mbarnett{at} (@ml_barnett on X)
  • Accepted 13 March 2024


Objective To examine the association between prescriber workforce exit, long term opioid treatment discontinuation, and clinical outcomes.

Design Quasi-experimental difference-in-differences study

Setting 20% sample of US Medicare beneficiaries, 2011-18.

Participants People receiving long term opioid treatment whose prescriber stopped providing office based patient care or exited the workforce, as in the case of retirement or death (n=48 079), and people whose prescriber did not exit the workforce (n=48 079).

Main outcomes Discontinuation from long term opioid treatment, drug overdose, mental health crises, admissions to hospital or emergency department visits, and death. Long term opioid treatment was defined as at least 60 days of opioids per quarter for four consecutive quarters, attributed to the plurality opioid prescriber. A difference-in-differences analysis was used to compare individuals who received long term opioid treatment and who had a prescriber leave the workforce to propensity-matched patients on long term opioid treatment who did not lose a prescriber, before and after prescriber exit.

Results Discontinuation of long term opioid treatment increased from 132 to 229 per 10 000 patients who had prescriber exit from the quarter before to the quarter after exit, compared with 97 to 100 for patients who had a continuation of prescriber (adjusted difference 1.22 percentage points, 95% confidence interval 1.02 to 1.42). In the first quarter after provider exit, when discontinuation rates were highest, a transient but significant elevation was noted between the two groups of patients in suicide attempts (adjusted difference 0.05 percentage points (95% confidence interval 0.01 to 0.09)), opioid or alcohol withdrawal (0.14 (0.01 to 0.27)), and admissions to hospital or emergency department visits (0.04 visits (0.01 to 0.06)). These differences receded after one to two quarters. No significant change in rates of overdose was noted. Across all four quarters after prescriber exit, an increase was reported in the rate of mental health crises (0.39 percentage points (95% confidence interval 0.08 to 0.69)) and opioid or alcohol withdrawal (0.31 (0.014 to 0.58)), but no change was seen for drug overdose (−0.12 (−0.41 to 0.18)).

Conclusions The loss of a prescriber was associated with increased occurrences of discontinuation of long term opioid treatment and transient increases in adverse outcomes, such as suicide attempts, but not other outcomes, such as overdoses. Long term opioid treatment discontinuation may be associated with a temporary period of adverse health impacts after accounting for unobserved confounding.


Chronic pain affects more than one in four American adults over 65 years old and is commonly managed using long term opioid treatment (LTOT).123 However, the general shift away from prescribing opioids has meant that millions of patients in chronic pain are also being removed from LTOT, often at faster taper speeds than recommended by guidelines.4563789 Many have raised concerns that poor adherence to guideline-suggested tapering may lead to undertreated pain, mental health crises, and suicide.1011

Work examining the association between tapering or discontinuation of LTOT and health outcomes has important limitations. Systematic reviews on the reduction or discontinuation of LTOT find little high quality evidence, although studies generally indicate improvement in pain and quality of life after discontinuation or tapering.1213 By contrast, large scale observational studies find both increased and decreased risk for death or addiction related adverse events.71415161718192021 These studies typically use statistical techniques adjusting for observable, but not unobservable, differences between users of LTOT who taper or discontinue versus those who do not, populations whose clinical profiles can diverge around the time of LTOT changes.32223 For instance, a patient may be discontinued because prescribers suspect that patients are bordering on behavioral, substance use, or mental health disorders, whereas patients continued on LTOT are maintaining baseline. Without robust evidence that accounts for selection and confounding, clinical knowledge on the association between LTOT discontinuation and patient outcomes is incomplete, impeding both therapeutic management of LTOT and the development of opioid policy.

In this study, we investigated the evidence gap by leveraging prescriber exit from the workforce, a common event,242526 as an external shock to prescribing patterns. Previous work found that primary care physician exit was associated with substantial shifts in patient prescribing patterns.2728 We hypothesized that prescriber market exit would lead to an increase in discontinuation of LTOT unrelated to observed or unobserved patient clinical factors. Patients receiving LTOT who discontinued or tapered from opioids are likely different from those who have not tapered, therefore, we addressed selection bias and confounding by defining the exposure as prescriber workforce exit, an event plausibly not driven by clinical events leading to LTOT discontinuation.18 This method of experimentation enabled testing for the independent effect of discontinuation of LTOT on patient outcomes.


Data source and study population

The cohort study used a 20% random sample of Medicare fee-for-service and Medicare Advantage beneficiaries from 1 January 2011 to 31 December 2018. Medicare is a public insurance program that enrolls 65 million Americans who are age 65 years or older, receiving social security disability income, or diagnosed with amyotrophic lateral sclerosis or end stage renal disease. Our primary study cohort captured all clinical and healthcare use outcomes for fee-for-service beneficiaries enrolled in fee-for service Medicare parts A, B, and D. For outcomes related to prescriptions, we also included Medicare Advantage beneficiaries with Medicare part D prescription claims. We excluded data with missing racial status accounting for less than 1% of the sample. Additionally, we excluded patients diagnosed with cancer at any point over the sample period because the role of LTOT may differ between cancer and non-cancer indications (see appendix figure 1 for cohort flow diagrams).

Identification of prescribers who exited Medicare

The main study exposure occurred when a prescriber stopped providing office based patient care or exited, as in the case of retirement or death. Patients attributed to such an exiting prescriber were considered exposed to this discontinuation. A prescriber’s exit date was defined as the last date the prescriber billed Medicare for an office based service with no subsequent services observed, as defined in prior research.2930 Prescribers were considered exiting if they had at least one office visit 6-12 months before their last observed office visit and had a last office visit between 1 January 2012 and 31 December 2017. These restrictions allowed us to observe patients prescribed an opioid at least four quarters before and after prescriber workforce exit (fig 1). Patients attributed to prescribers with an exit date were considered exposed while those without prescriber exit were considered unexposed.

Fig 1
Fig 1

Matching algorithm. The dashed line indicates the period between the patients’ long term opioid treatment (LTOT) episode and quarter (Q) −5 relative to prescriber exit. The first quarter before exit was excluded when capturing eligible LTOT episodes to avoid bias from anticipation of prescriber exit that could affect patterns of LTOT and subsequent outcomes. Prescriber exit is denoted as quarter 0, with minus numbers indicating quarters before prescriber exit. Patients were matched on calendar year the initial LTOT episode began (2011, 2012, 2013, 2014, 2015, or 2016/2017), patient age, gender (female v male), race (white v not white), state of residence, whether the patient resided in a rural area, Medicare eligible due to disability, Medicare eligible due to end stage renal disease, dual eligibility for Medicare and Medicaid, Medicare Advantage enrollment, diagnosis of chronic non-cancer pain, total number of chronic conditions, average daily morphine milligram equivalents in initial LTOT episode. Information on dual eligibility, Medicare Advantage enrollment, diagnosis of chronic non-cancer pain, and average daily morphine milligram equivalents for the entire duration of the initial LTOT episode were also collected. All other covariate information was obtained from the calendar year in which the initial LTOT episode began

Study sample and LTOT definition

The study sample included Medicare beneficiaries receiving LTOT who were exposed and unexposed to prescriber exit. To be included, the beneficiary had to be at least 18 years old and continuously enrolled in Medicare. LTOT was defined as receipt of at least 60 days’ supply of opioids at a dosage of 25 daily morphine milligram equivalents or more on average per quarter for at least four consecutive quarters. The first four or more quarter period meeting this definition was the initial LTOT episode for patients.

Patients in the exposed group were limited to those with an LTOT episode beginning at least five quarters before prescriber exit (meaning the first quarter before exit was excluded when capturing eligible LTOT episodes) to avoid bias from anticipation of prescriber exit that could affect patterns of LTOT and subsequent outcomes (fig 1).3 Unexposed patients were comprised of those meeting the definition for having an LTOT episode. Patients were attributed to the prescriber providing the plurality of opioid prescriptions over the initial four quarter LTOT episode.

Matching and exit date assignment

To control for observed patient differences, we used propensity score matching to match patients of exiting prescribers (exposed patients) to patients who did not lose their prescriber (unexposed patients). The propensity score, which estimates the likelihood that a given patient would be in the exposed group, was estimated using patient covariates measured the first year they began LTOT. We only matched patients who were unexposed and exposed with an initial LTOT episode in the same calendar year to account for nationwide changes in opioid prescribing over time (appendix methods 1).31

After propensity score matching, we assigned the patient who was unexposed to the same prescriber exit date of the matched patient who was exposed. For example, consider patient A who was exposed to LTOT in January 2011, subsequently losing their prescriber in August 2013. The closest propensity score match is patient B who was not exposed and began LTOT in January 2011 but did not lose their prescriber. We therefore assigned the exit date of August 2013 from exposed patient A’s prescriber to be the synthetic exit date for unexposed patient B. This assignment of exit dates to unexposed patients enabled us to model changes in exposed patients’ outcomes in response to the loss of a prescriber compared with observably similar unexposed patients before and after prescriber exit in a difference-in-differences design (fig 1).

Defining opioid discontinuation

Discontinuation from opioid treatment occurred when a patient had no resumption of opioid treatment at the end of the prescription’s days supplied for at least one year after discontinuation (see appendix figure 2 for the allocation of prescriptions, as well as by provider type). Discontinuation was defined as equal to one in the quarter of discontinuation and zero in the quarters before and after the discontinuation event.

Outcome measures

Our main outcomes were adverse clinical events that could plausibly be a clinical outcome in response to discontinuation of LTOT: all drug overdoses, mental health crises, opioid or alcohol withdrawal, gastrointestinal bleeding or kidney failure due to substitution to non-steroidal anti-inflammatory drugs, or all cause mortality. We measured all cause hospital use as admission to hospital or emergency department visits identified by claims in the inpatient file or outpatient claims. We categorized emergency department visits for pain by using claims with a primary diagnosis of pain (appendix table 1). Mental health crises included hospital visits with a primary diagnosis of depression, anxiety, or a suicide attempt (appendix table 1). We captured mortality using the Medicare beneficiary summary file. We also quantified beneficiaries’ annual rate of office visits to any type of provider, including specialist and primary care providers. Medicare spending included all charges listed on beneficiaries’ claims for the study period. For prescription outcomes, we used prescription data from Medicare part D (which also captures patients using Medicare Advantage) to estimate all filled prescription quarterly as well as mutually exclusive groups of opioid, buprenorphine, naloxone, and non-opioid prescription drugs.


We collected information on patients’ age, gender, race or ethnicity, state of residence, dual eligibility for Medicaid and Medicare coverage, Medicare Advantage enrollment, disability as the original reason for Medicare enrollment, diagnosis of chronic non-cancer pain (appendix table 1), and morphine milligram equivalent daily dose in the period before prescriber exit, so none of the covariates was causally influenced by future exposure or outcomes.32 We quantified patients’ opioid total morphine milligram equivalent, average daily dose, and the total number of days supplied during the first four consecutive quarters a patient met the definition of LTOT. We calculated the total morphine milligram equivalent and days supplied from prescriptions’ generic ingredients for opioid compounds.33 We calculated the average daily dose of opioids as the total morphine milligram equivalent of opioids supplied divided by the total number of days supplied during the initial LTOT episode.

We classified whether the patient’s county of residence was rural or urban,34 and included 27 chronic conditions classified following prior work: acquired hyperthyroidism, acute myocardial infarction, Alzheimer’s disease, Alzheimer’s disease and related senile dementia disorders, anemia, asthma, atrial fibrillation, benign prostatic hyperplasia, cataract, chronic kidney disease, chronic obstructive pulmonary disease, diabetes, depression, heart failure, glaucoma, hip or pelvic fracture, hyperlipidemia, hypertension, ischemic heart disease, osteoporosis, rheumatoid arthritis or osteoarthritis, stroke or transient ischemic attack, breast cancer, colorectal cancer, prostate cancer, lung cancer, endometrial cancer.3

Statistical analysis

We compared outcomes for matched exposed and unexposed patients receiving LTOT before and after prescriber exit in a difference-in-differences framework. The framework allowed us to estimate the average treatment effect of physician exit on exposed patients. The difference-in-differences design required two key assumptions in our context. We first assumed that outcomes would trend similarly for unexposed and exposed patients in the absence of treatment. We visually tested this assumption in figure 1 and figure 2 by observing whether outcomes for exposed and unexposed patients moved in parallel before treatment (visual inspection of pre-trends is another advantage of propensity score matching exposed and unexposed groups).We then assumed the exit of a patient’s assigned prescriber was independent of baseline patient outcomes. Table 1 shows that matching unexposed and exposed patients through a propensity score match effectively limited baseline differences between exposed and unexposed patients, supporting this assumption.

Fig 2
Fig 2

Unadjusted change in quarterly rates of long term opioid treatment (LTOT) discontinuation, clinical, and healthcare use outcomes for patients receiving LTOT who had a prescriber exit versus those who did not have a prescriber exit. Prescriber exit is denoted as quarter 0, with minus numbers indicating quarters before prescriber exit. The vertical dashed line delineates the periods before exit (left) and after exit (right). Appendix Table 4 shows adjusted quarterly point estimates. Appendix Figure 1a and 1b diagrams the sample construction. ED=emergency department

Table 1

Patient characteristics for matched patients whose prescriber exited the workforce versus those whose prescriber remained the same

View this table:

We used linear regression at the level of quarter per patient to estimate a set of interaction terms between indicators for being an exposed patient and indicators for eight quarters relative to prescriber exit (four before exit and four after exit; details in appendix methods 3). The interaction terms describe the mean differential change in the outcome between unexposed and exposed patients by quarter relative to prescriber exit, using quarter −5 as the baseline period. All regression models also included patient and prescriber fixed effects (except for the outcome of mortality, which only contained prescriber fixed effects; appendix methods 3)2935 and clustered standard errors at the matched level of the prescriber pair.30 The use of fixed effects controlled for time invariant differences among patients and prescribers, such as baseline age, race, sex, living in a rural area, reason and type of Medicare enrollment, and baseline chronic conditions.

In a separate set of models, we used the same regression approach, but estimated the differential change between unexposed and exposed patients by year relative to prescriber exit (one year before exit, one year after exit). All regression analyses at the year level defined the period before prescribe exit as quarters −5 to −2 before a prescriber’s exit, excluding quarter −1 to account for potential anticipation (appendix methods 2)

We assessed the robustness of findings with several alternative specifications, including examining patients qualifying for Medicare because of social security disability income as a separate subgroup; repeating adjusted analyses over the entire sample but excluding Medicare Advantage patients; and including patient with cancer. We also examined changes in outcomes attributable to prescriber exit among patients not receiving LTOT with the same prescribers as patients who were receiving LTOT to determine the effect of the exit alone.30 Additionally, we compared treatment effects by whether the patient lost a primary care physician or specialist. We tested the sensitivity of results by adjusting for multiple comparisons and compared treated patients who discontinued to treated patients who did not discontinue to quantify how the provider’s exit alone affected estimates. The 95% confidence intervals (CI) reflected 0.025 in each tail or P≤0.05. Analyses were performed in Stata, version 16 (StataCorp LLC).

Patient and public involvement

No patients were involved in setting the research question or the outcome measures, nor were they involved in developing plans to design or implement the study. No patients advised the interpretation or writing up of results. We used previously collected, de-identified data purchased from the Centers for Medicare and Medicaid Services that is restricted use. The institutional review board at the Harvard TH Chan School of Public Health approved the study, waived informed consent, and did not require us to involve patients and the public in the research process.


Study sample

Before matching, the full study sample consisted of 80 158 exposed and 322 970 unexposed patients who received LTOT. Propensity score matching led to the exclusion of 32 079 exposed patients and 274 891 unexposed patients, leaving 48 079 patients assigned to 15 713 exiting prescribers (exposed) and 48 079 patients assigned to 28 150 stable prescribers (unexposed (appendix figure 1)). Propensity score matching improved balance on observable characteristics (appendix figures 3 and 4). After matching, patients in both exposed and not exposed groups had similar demographic and clinical characteristics, with almost all standardized mean differences of 0.05 or less (table 1). Comparisons of exiting versus stable prescribers and patients receiving versus not receiving LTOT are in appendix tables 2 and 3.

LTOT discontinuation and prescription outcomes

In the first quarter after prescriber exit, the opioid discontinuation rate for exposed patients receiving LTOT increased from 132 to 229 per 10 000 patients per quarter, compared with 97 to 100 per 10 000 unexposed patients (fig 2; adjusted difference of 1.22 percentage points ((95% CI 1.02 to 1.42), 160% increase from a baseline of 0.77%, appendix table 4). The adjusted yearly rate of discontinuation differentially increased 2.08 percentage points (1.66 to 2.50), or a 56% increase from the baseline 3.70% rate of discontinuation, for exposed patients relative to unexposed patients (table 2). In the overall post-exit period, the yearly number of opioid prescriptions declined by 1.01 prescriptions ((95% CI −1.11 to −0.91) or −6% off the baseline mean of 15.71 prescriptions), total days’ supply of opioids declined by 29 days ((95% CI −31 to −26), or −7% off the baseline mean of 414 days’ supply), and total morphine milligram equivalent of opioids declined by 5311 morphine milligram equivalent ((95% CI −5759 to −4864), or −15% off baseline mean of 35 336 morphine milligram equivalent). Further, non-opioid prescriptions declined by 0.66 ((95% CI −1.03 to −0.3) or −1% off baseline mean of 59.68 prescriptions), and buprenorphine prescriptions increased by 0.03 ((0.01 to 0.05), or 25% off baseline mean of 0.12 prescriptions). Naloxone prescriptions increased by 0.0025 ((0.0002 to 0.0047), or 40% off baseline mean of 0.0047 prescriptions).

Table 2

Adjusted differential change in annual rates of LTOT discontinuation, clinical, and healthcare use outcomes for LTOT patients receiving LTOT who were exposed versus not exposed to prescriber exit

View this table:

Clinical and healthcare use outcomes

From quarter −5 to −1 before provider exit, unadjusted trends in outcomes were similar between exposed and unexposed patients receiving LTOT, supporting the parallel trends assumption needed for the differences-in-difference research design (fig 3). Some outcomes showed potential anticipation of an upcoming prescriber exit, most notably mortality, which motivated our exclusion of quarter −1 from year-level regressions.

Fig 3
Fig 3

Adjusted differential change in quarterly rates of long term opioid treatment (LTOT) discontinuation, clinical, and healthcare use outcomes for patients receiving LTOT who had a prescriber exit versus those who did not have a prescriber exit. Prescriber exit is denoted as quarter 0, with minus numbers indicating quarters before prescriber exit. The vertical dashed line delineates the periods before exit (left) and after exit (right). Absolute risk difference estimates are from the matched difference-in-differences model described in the methods, with an indicator for outcomes for each quarter relative to exit. All cause mortality is modeled similarly but without patient fixed effects. Point estimates are relative to quarter −5 (ie, five quarters before prescriber exit). Regressions patient and prescriber fixed effects, and cluster at the prescriber level. Outer lines show the boundaries of the 95% confidence interval for each quarterly estimate. Appendix Figure 1a and 1b diagrams the sample construction. ED=emergency department

In the first quarter after prescriber exit (denoted quarter 0), when discontinuation rates were highest, a significant increase was noted in the rate of suicide attempts (0.05 absolute percentage points (95% CI 0.01 to 0.09); 122% increase off baseline 0.04% suicide attempt), opioid or alcohol withdrawals (0.14 absolute percentage points (0.004 to 0.28); 50% increase off baseline 0.28% withdrawal rate), and emergency department visits or admissions to hospital (0.04 visits (0.01 to 0.06); 9% increase off baseline 0.45 visits), including emergency department visits with a pain diagnosis (0.011 visits (0.002 to 0.02); 10% increase off baseline 0.11 visits with a pain diagnosis), compared with patients who had a prescriber that exited the workforce and those who had a continuous prescriber (fig 3 and appendix table 4). Mortality declined by −0.15 percentage points ((95% CI −0.29 to −0.02); 52% decline off base of 0.5%). Significant differences receded by quarter two except for emergency department visits and admissions to hospital, which were 0.02 visits ((0.001 to 0.04); increase of 4.4% from baseline mean) higher among patients with a provider exit but then receded by quarter three. No significant change in the rate of overdose was noted across all quarters of the study period.

In adjusted analyses averaging across the whole post prescriber exit period, a significant increase was noted in opioid or alcohol withdrawal (0.31 percentage points (95% CI 0.041 to 0.58), or 31% increase from baseline mean of 0.99%) and mental health crises (0.39 (0.08 to 0.69), or a 24% increase from a 1.6% baseline mean) comparing patients who had prescriber workforce exit versus those who did not. Additionally, a significant decrease in mortality was recorded (−0.50 (−0.77 to −0.23), or −23% from baseline mean of 2.18%) (table 2). No significant difference in percentage points was noted between patients who were exposed and unexposed in annual rates of drug overdose (−0.12 (−0.41 to 0.18), or 9% decrease from 1.37% baseline rate), suicide attempt (0.02 (−0.07 to 0.11), or 13% from 0.15% increase from baseline rate), gastrointestinal bleeding (0.33 (−0.07 to 0.73), or 11% increase from 2.87% baseline rate), or kidney failure (0.16 (−0.40 to 0.71), or 2% increase from 7.61% baseline rate).

To quantify the effect of only losing a prescriber on results, we examined changes in outcomes associated with prescriber exit among all patients who were not receiving LTOT but lost the same prescriber (appendix table 3, appendix table 5). The unadjusted differences reflect differences in the main results in table 2 that may be attributable to prescriber exit instead of to LTOT discontinuation. After the loss of a prescriber, increases among patients not receiving LTOT were noted for overdose (0.12 percentage points, or a 27% increase from baseline mean of 0.45%), anxiety (0.10, or a 59% increase from baseline mean of 0.17%), opioid or alcohol withdrawal (0.21, or a 47% increase from baseline mean of 0.45%), and mortality (0.90, or 16% increase from baseline rate of 5.68%). Additionally, differences in mental health crises (0.08, or 15% increase from the baseline mean of 0.55%) were small among patients not receiving LTOT when compared with estimates among patients receiving LTOT (table 2).

Additional analyses

In sensitivity analyses, we focused on: beneficiaries qualifying for social security disability income (people with disabilities); excluding patients in Medicare Advantage; including patients with cancer; and separately, patients above or below median morphine milligram equivalent (median 54.36 average daily), which were all similar to the main analysis (appendix tables 6 and 7). Appendix table 8 replicates the main results focusing on patients who were alive after a prescriber’s exit. Appendix table 9 compares treatment effects modeled in table 2 by whether the patient’s main prescriber was a primary care physician or specialist, showing that effects are similar across the two groups. Appendix table 10 indicates that effects maintain significance when adjusting for multiple comparisons. Appendix table 11 compares exposed LTOT patients who did versus did not discontinue LTOT in response to prescriber exit, showing standardized mean differences of 0.12 or less.


Principal findings

The loss of a prescriber was associated with increased discontinuation of LTOT and transient, but significant, increases in adverse outcomes among patients, including suicide attempts, withdrawal, and admissions to hospital or emergency department visits. Rates of adverse outcomes among patients reverted to baseline rates within four to seven months after prescriber exit. However, a significant increase in mental health crises and opioid or alcohol withdrawal was noted on average across the full four quarters after the exit period. Despite these outcomes, drug overdose rates did not change. Our findings suggest that discontinuation of LTOT may be associated with a temporary period of negative health effects, not including overdose or mortality, after accounting for unobserved confounding.

While a small proportion of patients discontinued LTOT in both groups, a substantial increase in discontinuation was reported in the quarter after prescriber exit. As seen in figure 2, the increase in the first quarter after a prescriber’s exit was not associated with any visible or regression estimated change in overdose rates, despite large increases that might be expected based on standard observational models.71521 The difference in results is likely because discontinuation is a clinical event associated with other health changes, and patients with discontinuation differ from others, preventing the estimation of a causal effect.32223 The small or null results over the post prescriber exit period for outcomes such as overdose suggest that despite potential harm, discontinuation of LTOT may have counteracting benefits, such as reduced overdose risk, for some patients. The overall reduction in mortality among patients receiving LTOT who had a prescriber exit supports the potential longer term benefit of shorter duration and lower dosage to LTOT, although we interpret our mortality results with caution given pre-exit period trends that diverged prior to prescriber exit (fig 2).

The harm we do observe associated with discontinuation of LTOT could be related to low quality management of transitioning patients across prescribers. Most discontinuations have excessively rapid tapers,3 and patients receiving LTOT are a population at high risk with many comorbidities who must frequently navigate substantial stigma in the health care system.36 For instance, discontinuation may accompany distressing clinician abandonment of patients with LTOT and cause opioid withdrawal, emotional harm, and undertreated pain, events that were unobservable in claims data unless they result in diagnoses that we captured.

One question is whether the observed associations were attributable to prescriber exit rather than the accompanying rise in overdose. To address this, we focused on patients receiving LTOT and those who were not on this treatment but who lost the same prescriber. If the prescriber exit explains the effects, patients not receiving LTOT should be impacted similarly to those who were receiving LTOT after prescriber exit. Instead, analyses show that clinical outcomes for patients not on LTOT were either null or opposite to those observed for patients receiving LTOT after the loss of a prescriber. Therefore, the loss of the physician is unlikely to explain the observed results among patients on LTOT and, if anything, may lead to an underestimation of the effects of discontinuation.

Comparison with other studies

One clear conclusion is that the observed clinical effect of LTOT discontinuation is highly dependent on the methods used. Prior research on tapering or discontinuation of LTOT finds a doubling in the rate of overdose and mental health crises comparing populations with discontinuation directly to those without.715161719 By contrast, the absence of change in overdose rates in our analysis is closer to other observational studies using techniques to control for unobserved confounding, which have found opioid discontinuation to have a small or null impact on rates of addiction related adverse events.141822 This discrepancy suggests that overdose risk is more likely to be misestimated in conventional approaches.

Policy implications

While the choice of methods is a technical issue, it has great relevance to patients and policymakers struggling with the clinical and public health challenges of LTOT. Some mechanisms are plausible by which LTOT discontinuation could lead to either benefit, harm, or a mixture of both. Both observational research37 and randomized trials38 describe that LTOT, especially at a high dose, is associated with a multitude of adverse outcomes. On the one hand, discontinuing LTOT without replacing the clinical role of LTOT could lead to untreated pain, withdrawal, or worsening of mental health issues as described previously. On the other hand, discontinuing LTOT could mitigate these risks and promote patient safety in some circumstances. The clinical and scientific uncertainty around this question suggests clinical equipoise to justify ongoing394041 and future randomized interventions that promote patient centered, clinically appropriate LTOT discontinuation or tapering to investigate how tapering LTOT can be done safely and respectfully while prioritizing quality of life.


This study has several limitations. Firstly, the findings may not generalize to the entire Medicare population or to populations outside of Medicare beneficiaries.32 Our sample was a younger, Medicare qualifying population: the average patient in our sample was 58.0 years old, and only 33% of the population was over age 65 years (table 1). Additionally, the study may not generalize to discontinuations outside of those caused by losing a prescriber. Secondly, we cannot observe the reason for a prescriber’s exit, which may be associated with patient outcomes in certain circumstances. We attempt to circumvent this issue by controlling for prescriber specific factors by comparing patients receiving LTOT with patients not receiving LTOT with the same assigned, exiting prescriber, showing that effects are directionally opposed between the two groups. Thirdly, our findings may not apply to patients on LTOT not meeting the threshold of the restrictive definition used in this study, such as those receiving lower doses of opioids or those on LTOT for less than a year. Fourthly, our statistical power, as assessed by the size of confidence intervals in the adjusted results, does not enable us to rule out a low magnitude of harm in response to discontinuation. However, across multiple outcomes, point estimates were consistently close to zero, and they did not change in any consistent pattern with the timing of a large change in discontinuation associated with prescriber exit. Fifthly, we chose what we believe to be the best strategy to handle confounding, but other strategies might be equally valid.

An additional limitation is that, while our analysis overcomes the confounding in prior work, our statistical strategy leverages a specific group of patients: those who have their opioid prescriptions discontinued in response to the loss of a prescriber. For instance, the increase in buprenorphine and naloxone prescriptions may suggest that patients receive replacement prescribers who initiated treatment of opioid use disorder in response to the loss of their prescriber. Also, not all patients had their opioid prescriptions discontinued after losing their main prescriber, suggesting that our results do not generalize to all patients who are discontinued or tapered from opioids. Since we followed up patient outcomes for only four quarters, our study does not consider the effects on patients losing a prescriber beyond four quarters after prescriber exit. We also do not observe prescriptions not billed to Medicare. Finally, our definition of LTOT discontinuation follows prior work to support comparability across studies. However, our results might not generalize to all alternative definitions of LTOT and discontinuation.


This study finds a complex association between the discontinuation of LTOT coinciding with prescriber exit and subsequent health effects. The cessation of LTOT was linked to a short term increase in negative health events, such as suicide attempts and admissions to hospital, indicating a potential need for heightened mental health support during the transition. Despite this, we found no effect of discontinuation on overdose rates or mortality. These findings differ from prior evidence that did not control unobserved confounding, implying that the observed consequences of LTOT discontinuation may vary considerably depending on the methods used. This variation underscores the importance of randomized interventions to better understand how LTOT discontinuation can be managed safely and effectively.

What is known on the topic

  • Much research shows that discontinued versus continued long term opioid treatment (LTOT) is associated with an increased rate of overdoses and mental health crises

  • Uncertainty remains because studies used observational models comparing individuals that discontinue LTOT to those that do not, populations whose clinical profiles diverge around the time of LTOT changes

What this study adds

  • Unobserved confounding was accounted for by leveraging prescriber workforce exit as an external shock increasing LTOT discontinuation and quantify outcomes in a difference-in-differences analysis

  • Prescriber workforce exit significantly changed opioid prescriptions and short term increases in adverse events of opioid or alcohol withdrawal, suicide attempts, and admission to hospital, but overdose rates changed little

  • LTOT discontinuation may be associated with a temporary period of adverse health impacts after accounting for unobserved confounding

Ethics statements

Ethical approval

The study was approved by the institutional review board at the Harvard TH Chan School of Public Heath.

Data availability statement

No additional data available.


  • Contributors: All authors contributed to the design and conduct of the study; management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript. MLB supervised the study and is the guarantor. The corresponding author attests that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted.

  • Funding: Supported by grants from the Retirement Research Foundation for Aging and the National Institute on Aging (K23 AG058806, MLB). The National Institute of Aging had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. The funders had no role in considering the study design or in the collection, analysis, interpretation of data, writing of the report, or decision to submit the article for publication.

  • Competing interests: All authors have completed the ICMJE uniform disclosure form at and declare: funding from Supported by grants from the Retirement Research Foundation for Aging and the National Institute on Aging; no support from any organization for the submitted work; no financial relationships with any organizations that might have an interest in the submitted work in the previous three years; no other relationships or activities that could appear to have influenced the submitted work.

  • Transparency: MLB affirms that the manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies have been disclosed.

  • Dissemination to participants and related patient and public communities: Results will be shared through the dissemination teams at Harvard University, University of Minnesota, and Stanford University. Typical mediums include press releases, social media posts (Twitter, Instagram, Facebook, and LinkedIn), and emails sent directly to journalists representing outlets such as the New York Times, San Francisco Chronicle, CNN, Bloomberg, and Becker’s Hospital Review, which condense the study into its main findings.

  • Provenance and peer review: Not commissioned; externally peer reviewed.

This is an Open Access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited and the use is non-commercial. See: