Thresholds for adjunctive medication usage and probability of initiating injectable naltrexone

Abstract

Aims:

To estimate the effectiveness of different thresholds for administering opioid withdrawal medications (clonidine and clonazepam) on the probability of successfully initiating extended-release naltrexone (XR-NTX) among participants with opioid use disorder (OUD) during medically managed withdrawal.

Design:

Secondary analysis of a multisite clinical trial comparing a rapid vs. standard approach for XR-NTX initiation, 2021–2022.

Setting:

Six community inpatient addiction treatment units in the United States.

Participants:

English-speaking adults seeking treatment for DSM-5 OUD and expressing interest in XR-NTX treatment (N=415).

Measurements(s):

We estimated the extent to which the following thresholds for adjunctive medication administration would affect the probability of initiating XR-NTX over time: 1) where adjunctive medications were given in response to at mild-to-moderate withdrawal symptoms or greater (Clinical Opiate Withdrawal Scale [COWS] score ≥ 5), 2) where adjunctive medications were given in response to minimal withdrawal symptoms or greater (COWS score ≥ 3), and 3) where adjunctive medications were given regardless of withdrawal symptoms. Using a longitudinal sequentially doubly robust estimator we estimated the cumulative probability of XR-NTX initiation under each of these three treatment regimes while accounting for dropout and initiation of other medications as competing events.

Findings:

The estimated probability of initiating XR-NTX by day 14 was 50.4% (95% confidence interval [CI]: 41.8–58.9) under the no-threshold regime, 43.9% (95% CI: 39.1–48.7) under the regime of waiting for minimal withdrawal symptoms, and 38.5% (95% CI: 34.3–42.6) under the regime of waiting for mild-to-moderate withdrawal symptoms. Probability of XR-NTX initiation was a statistically significant 11.9 percentage points higher (95% CI: 3.6, 20.2) under the no-threshold regime versus the moderate threshold regime, and a non-statistically significant 6.4 percentage points (95% CI: −0.8, 13.7) higher under the no-threshold regime versus the mild threshold regime.

Conclusions:

Providing clonidine and clonazepam daily during the first five days of medically managed opioid withdrawal appears to statistically significantly increase the likelihood of initiating extended-release naltrexone treatment compared with waiting for mild-to-moderate withdrawal symptoms to administer adjunctive medications. To improve initiation rates, providers may consider lowering the threshold at which they provide adjunctive medications, giving these medications preemptively or to manage even minimal withdrawal symptoms.

INTRODUCTION

Opioid use disorder (OUD) is estimated to affect between 6–7 million adults in the US (2019)[1] with over 80,000 deaths caused by opioids in 2023.[2] Although successful recovery from OUD may be achieved given proper treatment involving medications for opioid use disorder (MOUD)[3]—extended-release injectable naltrexone (XR-NTX), buprenorphine, and methadone—it is estimated that only 25% of individuals with OUD in the US initiate MOUD.[4]

Initiating XR-NTX is more challenging than initiating buprenorphine or methadone, because participants must be fully withdrawn from opioids prior to the first XR-NTX dose to avoid precipitated withdrawal.[5] As a result, initiation rates for XR-NTX in US are frequently below 75%, even among participants enrolled for treatment in clinical trials (two recent trials reported initiation rates around 50%).[6, 7, 8] The standard approach of managing symptoms of opioid withdrawal before the initial injection includes a 3–5 day agonist taper of buprenorphine or methadone followed by a 7–10 day opioid-free period.[7] An alternative approach involves minimal or no agonist taper and instead manages withdrawal anxiety and other symptoms with adjunctive medications such as clonidine and benzodiazepines. Several studies have shown that such an alternative approach increases the likelihood of XR-NTX initiation.[9, 10, 7]

One of these studies is the Surmounting Withdrawal to Initiate Fast Treatment with Naltrexone (SWIFT) hybrid effectiveness-implementation trial.[7] This trial used a wedge-cluster design to randomly assign, at 14-week intervals, six treatment sites to be trained to provide a rapid initiation approach, which entailed: a single day of buprenorphine, encouragement of more aggressive dosing of clonidine and clonazepam for withdrawal symptoms, and a medically managed withdrawal period of 5–7 days. Prior to the training, sites continued to provide their usual treatment, which in most cases included a buprenorphine taper and a medically managed withdrawal period of 12–14 days, as described above.[7] Training on the rapid approach was found to increase XR-NTX initiation rates.[7]

However, questions remain regarding the extent to which the more proactive use of adjunctive medications to manage withdrawal symptoms, in particular, may increase the likelihood of initiating XR-NTX. Traditionally, clinicians provide these medications as needed in response to withdrawal symptoms, but some have argued that a more aggressive approach, preemptively providing these medications with the goal of preventing withdrawal symptoms, may maximize initiation success.[11, 12] Natural variation in terms of the level of withdrawal symptoms at which participants were given adjunctive medications existed within and across the rapid and standard approaches in the SWIFT trial, which we describe herein, reflects the flexibility the sites were granted in applying the initiation approaches as well as variation in clinician approaches and patient preference.

We harnessed this natural variation to estimate the expected probability of initiating XR-NTX over time under different thresholds for administering adjunctive medications in response to withdrawal symptoms. Our overall aim was to learn the threshold at which adjunctive medications should be administered during medically managed withdrawal to increase XR-NTX initiation rates. Specifically, we considered three approaches to adjunctive medication administration for the first 5 days post-consent: 1) where all participants were given clonidine or clonazepam in response to mild-to-moderate withdrawal symptoms (operationalized as a clinical opiate withdrawal scale (COWS) score ≥5, reflecting at least mild symptoms[13]), 2) where participants were given one of these adjunctive medications in response to minimal withdrawal symptoms or greater (COWS score ≥3, as these scores represented some amount of withdrawal symptoms during a naloxone challenge, as compared to placebo[14]), and 3) where participants were given an adjunctive medication each day regardless of withdrawal symptoms. We contrasted the no-threshold (i.e., always-treat) strategy 3) of giving an adjunctive medication daily with the higher-threshold strategies 1) and 2).

METHODS

Cohort

SWIFT study participants were individuals seeking inpatient treatment for OUD at study sites and expressing interest in XR-NTX treatment (none were currently being treated with either methadone or buprenorphine). As described in the Introduction, study sites were randomly assigned to be trained to provide a rapid initiation approach at 14-week intervals.[7] The trial did not impose a strict protocol for XR-NTX initiation. Clinicians received guidance on the rapid approach (described in the Introduction), but they were free to implement the approach as they would implement any quality improvement initiative. Natural variation in terms of the level of withdrawal symptoms at which participants were given adjunctive medications existed within and across the rapid and standard approaches in the SWIFT trial. For further details on protocols for the trial, refer to the primary study publication.[7] Unlike the primary study, we considered the day of a patient’s completed study consent as day 1, which could occur up to 4 days post-admission, because they would be eligible to be administered study adjunctive medication (clonidine or clonazepam) at this point.

Measures

Exposure

Our time-varying exposure was a daily indicator of receipt (yes/no) of either: clonazepam (≥ 1 mg) and/or clonidine (≥ 0.1 mg) following the maximum COWS score of the day and before the subsequent COWS score, if present, or over the course of the day if no COWS score was recorded that day.

In a secondary analysis, our time-varying exposure was a daily indicator of receipt (yes/no)—at any time that day—of any of the following: clonazepam (≥ 1 mg), clonidine (≥ 0.1 mg), or any other benzodiazepine (> 0 mg). We considered this secondary analysis, because relaxing the temporal ordering requirement of adjunctive medication administration following the maximum COWS score was more data-inclusive and may be reasonable given missing timestamps; for example, no timestamps were recorded for administration of other benzodiazepines. In addition, because the COWS score is likely to decrease following adjunctive medication administration, the maximum score is likely to be temporally prior to medication administration.

We considered time-varying exposures for days 1–5 post-consent, where day 1 was the day of consent. We did not consider the exposure after day 5, because 1) the rapid initiation approach was designed to be 5–7 days long, so some participants would only be treated for 5 days, and 2) beyond day 5, there was little variation in adjunctive medication administration.

Primary Outcome and Competing Events

The outcome was a time-varying binary variable that indicated initiation of XR-NTX by each day up until 14 days post-consent. We chose day 14 post-consent as the primary endpoint for the study, because 14 days is the outer range of the standard initiation approach.

We treated a patient’s dropout from the study or initiation of another MOUD by each day post-consent as a competing event,[15] because leaving the treatment facility or initiating a different MOUD would prevent initiation of XR-NTX.

We conducted a sensitivity analysis where, if the patient initiated XR-NTX or had the competing event on day t but was missing a COWS score and did not receive any of clonidine, clonazepam, or any other benzodiazepine on day t, then we shifted the day of the outcome or competing event to be recorded on day t – 1, because the outcome of interest or competing event likely happened temporally prior to the subsequent day’s exposure.

Covariates

Although we used randomized trial data, our exposure of interest—the threshold at which to administer adjunctive medications—was not randomized and varied across randomization arms. Consequently, we adjusted for the following baseline covariates: number of days between admission and signed consent; age (years); sex; race and ethnicity; alcohol, amphetamine, cannabis, cocaine, or sedative use disorders; history of injection opioid use (both intravenous and non-intravenous); the number of years since first opioid use; non-prescription opioid use; and anxiety, bipolar, and depression diagnoses. We also adjusted for the following time-varying covariates daily: maximum COWS score, an indicator for missing COWS score, an indicator for the patient being administered less than the maximum allowable doses of clonazepam or clonidine (<4 or <1.2 mg, respectively) in the 24-hour period preceding their maximum COWS score, and an indicator for any other benzodiazepines administered (in the case of our primary analysis). If a patient had reached the maximum allowable dose of either clonazepam or clonidine, they would not be eligible to receive additional doses of that medication (but could be eligible to receive additional doses of the other medication).

We did not control for site, as that would have removed much of the natural variation in different thresholds for adjunctive medication usage. However, there could be other site-level practices that may be correlated with these different thresholds and affect XR-NTX initiation, such as non-medical interventions. We tested whether there was a relationship between site and the residuals from our outcome model (such a relationship would be consistent with unobserved confounding by site), and found differences in the residuals for two of the six sites. Thus, as an additional sensitivity analysis, we repeated our primary analysis among the four sites where there was no statistically significant relationship between the site and the residuals (i.e., in a subset where site did not appear to operate as a meaningful confounder).

There were minimal missing data (< 1%) for all baseline covariates except the following: Hispanic ethnicity was missing for 1.2%, race was missing for 3.1%, injection opioid use was missing for 13.5%, years since first opioid use were missing for 13.7%, and non-prescription opioid use was missing for 12.3%. Missing baseline covariates were imputed with the mode for categorical variables, and missing indicator variables were included for injection opioid use, years since first opioid use, and non-prescription opioid use.

Statistical Analysis

We consider baseline and time-varying confounding and time-varying exposure over the first 5 days post-consent, the survival-type outcome of XR-NTX initiation days 5–14 post-consent, and the competing risk of leaving the treatment facility over all 14 days post-consent. We were interested in estimating the longitudinal effects of dynamic treatment regimes with lower vs. higher thresholds for treating withdrawal symptoms on the cumulative probability of initiating XR-NTX. A dynamic treatment regime is a sequence of a priori-specified treatment decisions that may depend on patient baseline and time-varying characteristics.[16] We considered three dynamic treatment regimes, which we call: 1) the mild-to-moderate threshold regime (waiting until a COWS score of ≥ 5 is reached before giving adjunctive medications), 2) the minimal threshold regime (waiting until a COWS score of ≥ 3 is reached before giving adjunctive medications), and 3) the no-threshold regime (giving adjunctive medications daily unless contraindicated), which represent increasingly lower thresholds for using adjunctive medications.

We estimated the following quantities: 1) the probabilities of initiating XR-NTX by each day of days 5–14 post-consent under each of the above-described three dynamic treatment regimes; and 2) the difference in these probabilities between a) the no-threshold regime versus the mild-to-moderate threshold regime and b) the no-threshold regime versus the minimal-threshold regime. In the competing event literature, our estimated risk differences are called “total effects” in that we account for the competing events of leaving the treatment center or initiating a different MOUD, but we do not hypothetically intervene to prevent them, as that would not be realistic.[15]

We used a longitudinal sequentially doubly robust estimator[17] (with 20 cross-fitting folds) with a superlearner[18] (fit using 10 cross-validation folds) consisting of: an intercept-only model, a main-terms generalized linear model, multivariate adaptive regression splines (MARS),[19] eXtreme gradient boosting machines (XGBoost),[20] and random forests.[21] Density ratios (similar to the propensity score of the cumulative time-varying treatment) were truncated at the 97.5th percentile. We conducted the analyses using R (version 4.4.2)[22] with the lmtp[23, 24] and mlr3superlearner[25] packages.

Further technical details of our analysis, including the assumptions under which our estimates may be interpreted causally are given in Section S1 of the Supporting information. All code for the analysis is available at https://github.com/CI-NYC/ctn0097. This analysis was not pre-registered, so the results that follow should be considered exploratory.

RESULTS

The trial consisted of 415 participants (225 enrolled under the rapid approach and 190 under the standard approach) with a median age of 33 years (interquartile range: 28–38). 210 (50.6%) participants were female (Table 1). The median daily maximum COWS score was highest (score of 6) on days 1 and 2, decreased to a score of 5 on day 3, and further decreased to a score of 4 on days 4 and 5, among those still present in the study and not missing a maximum COWS score. Incidence of missing COWS scores was highest on the day of consent (day 1) at 15.7%. By the end of the 14-day period following signed consent, 43.4% of participants initiated XR-NTX while 46.8% had dropped out. We provide the number of participants observed to have followed each of the three dynamic treatment regimes on days 1–5 in Table 2. Maximum cumulative density ratios, which are used to assess the extent of practical violations of the positivity assumptions, across the first five days of treatment for all analyses are provided in Table S1. We did not find evidence of practical positivity violations, meaning that we have support in the data for the treatment regimes we consider.

Table 1:

Descriptive Statistics of Cohort

Variables	N = 4151
Days from admission to consent	1 (1, 2)
Female	210 (51%)
Age (years)	33 (28, 38)
Ethnicity/Race 2
Hispanic	91 (22%)
Ethnicity unknown	5
White	307 (76%)
Black	65 (16%)
Other	53 (13%)
Race unknown	13
Baseline Substance Use
Alcohol use disorder	129 (31%)
Unknown	2
Amphetamine use disorder	150 (36%)
Unknown	2
Cannabis use disorder	165 (40%)
Unknown	2
Cocaine use disorder	147 (36%)
Unknown	2
Sedative use disorder	106 (26%)
Unknown	2
Injection opioid use	177 (49%)
Unknown	56
Years since first opioid use	12 (8, 17)
Unknown	57
Non-prescription opioid use	61 (17%)
Unknown	51
Baseline Psychiatric Conditions
Anxiety	266 (64%)
Unknown	2
Bipolar	114 (28%)
Unknown	2
Depression	157 (38%)
Unknown	2
Maximum COWS score
Day 1	6.0 (2.0, 10.0)
Eligible	350 (100%)
Unknown	65
Day 2	6.0 (3.0, 10.0)
Eligible	368 (100%)
Unknown	40
Day 3	5.0 (3.0, 8.0)
Eligible	348 (99%)
Unknown	26
Day 4	4.0 (1.0, 7.0)
Eligible	302 (99%)
Unknown	37
Day 5	4.0 (1.0, 6.0)
Eligible	237 (99%)
Unknown	31
Time-Varying Medication
Other benzodiazepine use day 1	119 (29%)
Other benzodiazepine use day 2	73 (18%)
Other benzodiazepine use day 3	45 (12%)
Other benzodiazepine use day 4	28 (8.2%)
Other benzodiazepine use day 5	25 (9.3%)
Exposures
Adjunctive medication day 1	248 (60%)
Adjunctive medication day 2	257 (63%)
Adjunctive medication day 3	231 (61%)
Adjunctive medication day 4	184 (54%)
Adjunctive medication day 5	160 (59%)
Outcomes
Dropout by day 14	194 (47%)
Initiation by day 14	180 (43%)

¹Median (Q1, Q3) for numeric measures; N (%) for categorical measures

²Notes: study participants could identify as several races; therefore, counts may not add up to the study total

Table 2:

Number of individuals observed to have followed each dynamic treatment regime and subsequently received treatment

	Mild-to-moderate threshold regime			Minimal threshold regime			No-threshold regime
Day	Followed DTR and received treatment	Followed DTR and did not receive treatment	Did not follow DTR	Followed DTR and received treatment	Followed DTR and did not receive treatment	Did not follow DTR	Followed DTR and received treatment	Followed DTR and did not receive treatment	Did not follow DTR
1	157	73	185	192	52	171	248	0	167
2	162	65	181	208	43	157	257	0	151
3	138	68	172	187	42	149	229	2	147
4	97	98	146	136	58	147	182	0	159
5	73	78	119	108	55	107	158	0	112

Figure 1 shows the results of our primary analysis. The top row depicts the cumulative probability of XR-NTX initiation under each of the three dynamic treatment regimes. The bottom row depicts the difference in estimated probability of initiation comparing the no-threshold regime to the mild-to-moderate threshold regime and to the minimal-threshold regime. Under the no-threshold regime, 50.4% (95% CI: 41.8, 58.9) of participants were estimated to initiate XR-NTX by day 14. Under the minimal-threshold regime, 43.9% (95% CI: 39.1, 48.7) were estimated to initiate by day 14, and under mild-to-moderate threshold regime, 38.5% (95% CI: 34.3%, 42.6%) were estimated to initiate by day 14. The estimated probability of initiation under the no-threshold regime was significantly higher than under the mild-to-moderate regime by 11.9 (95% CI: 3.6, 20.2) percentage points. Similarly, the probability of initiation under no-threshold regime was estimated to be slightly higher compared to minimal-threshold regime by 6.4 (95% CI: −0.8, 13.7) percentage points, though the confidence interval included the null. Estimates by day are provided in Table S2.

Figure 1. Cumulative XR-NTX initiation incidence by period under each dynamic treatment regime.

1. Cumulative XR-NTX Initiation Probability by Day
2. Dynamic Treatment Regime
   1. Mild-to-moderate threshold: adjunctive in response to mild-to-moderate withdrawal symptoms or greater
   2. Minimal threshold: adjunctive in response to mild withdrawal symptoms or greater
   3. No threshold: adjunctive given regardless of withdrawal symptoms
3. Difference in Probability of XR-NTX Initiation by Day

In the secondary analysis, we examined dynamic treatment regimes defined by giving adjunctive medications on the same day that the patient had a recorded COWS score at or above the regime-specific threshold. The estimated probability of XR-NTX initiation under the no-threshold regime (51.3%, 95% CI: 44.2, 58.3%) was relatively similar to the findings in the primary analysis. However, the estimated probabilities of initiation under the minimal and mild-to-moderate threshold regimes were markedly lower than in the primary analysis. In this secondary analysis, we estimated the probability of XR-NTX initiation to be significantly higher under the no-threshold regime as compared to both the mild-to-moderate threshold regime and the minimal-threshold regime, with estimated differences of 20.8 (95% CI: 11.5%, 30.1%) and 19.5 (95% CI: 8.3%, 30.7%) percentage points, respectively, by day 14 (Figure S1). Estimates by day are provided in Table S3.

Both of our sensitivity analyses returned results similar to those of our primary analysis (Figures S2 and S3).

DISCUSSION

We estimated that administering adjunctive medications clonidine and clonazepam daily during the first five days of medically managed withdrawal from opioids (unless clinically contraindicated) significantly increased the likelihood that participants initiated treatment with XR-NTX, as compared to waiting until the patient was experiencing mild-to-moderate withdrawal symptoms to administer these medications. By day 14, we estimated that this “no-threshold” strategy increased the probability of initiating XR-NTX by 11.9 percentage points (95% CI: 3.6, 20.2) over the strategy of waiting for mild-to-moderate symptoms.

To our knowledge, this is the first study to estimate the impact of different thresholds for adjunctive medication administration on XR-NTX initiation. Our findings have practical relevance for OUD treatment, as the requirement for full opioid withdrawal remains a major barrier to XR-NTX initiation. The opioid withdrawal symptoms experienced during XR-NTX initiation can be so uncomfortable that participants abandon treatment and may return to nonprescribed opioid use. Clonidine and clonazepam can alleviate some of those symptoms, and more aggressive treatment with these medications may be clinically appropriate, as risks of relapse and overdose are high if treatment is abandoned.[26, 27, 28, 29] Importantly, despite the risks of clonidine and clonazepam, safety events were rare in this study.[7] Because these medications are most commonly associated with safety concerns such as orthostatic hypotension and falls, site personnel were advised to conduct close monitoring and encourage aggressive oral hydration.[7]

Strengths of this study included the use of recent (2021–2022) data from the SWIFT clinical trial, a relatively large (in terms of sample size), diverse group of participants (Table 1) with ample natural variation in COWS scores and administration of adjunctive medications across the days considered. This natural variation was evidenced by the relative lack of so-called practical positivity violations in our data (Table S1);[30] we can harness this variation to estimate the effects of interest. In addition, as SWIFT was conducted in real-world community treatment settings, the effects we estimate herein may apply to other real-world treatment settings. Another strength is that we treated leaving the treatment facility or starting an agonist medication for OUD as a competing event, which is appropriate given the definition of a competing event as one that precludes the outcome of interest.[15] We estimated the total effect of each dynamic treatment regime on the outcome of initiating XR-NTX; this type of causal estimand essentially treats the competing event as a time-varying confounder with treatment-confounder feedback—appropriately controlling for it but not hypothetically intervening on it, as such an intervention would be impractical in the real world. Finally, the estimator we used represents another strength, as it is doubly robust and flexibly models relationships using an ensemble of machine learning algorithms while retaining theoretically valid inference.[17]

The study was nonetheless limited in several key respects. Missing data—in particular, missing COWS scores and an absence of timestamps on the administration of other benzodiazepines—challenged our primary analysis. Because of these missing data, we conducted the secondary analysis, which represents a cruder analysis but one that is better aligned with the observed variables. In addition, as our time-varying exposure of interest was not randomized, our estimates may be biased due to unobserved confounding. For example, we did not control for site, because the natural variation in our exposure was largely a product of clinician approaches and site practices—controlling for site would have adjusted away much of the exposure of interest, leaving only the within-site variability in thresholds for adjunctive medication administration. However, if there are other site-level practices that are correlated with these same thresholds and affect XR-NTX initiation, such as non-medical interventions, then confounding could result. We did not control for randomization arm in this analysis, because the two key differences between the rapid and standard approaches—adjunctive medication standing doses and timeline—were incorporated into this analysis. Adjunctive medications were captured and included as part of the time-varying exposure or time-varying confounders, as detailed in the Methods, and we conducted a longitudinal analysis, estimating effects for each day. However, it is plausible that if there were other differences between the exposure groups not captured as part of our time-varying exposure or confounding variables (e.g., adjunctive medications used in the days prior to study enrollment), and if these were also related to the outcome, bias could result.

In future work, we plan to conduct a longitudinal mediation analysis using randomization as the exposure, time-varying adjunctive medication usage as the mediators, and initiation of XR-NTX as the time-varying outcome. Such an analysis would quantify the extent to which the rapid protocol increased XR-NTX initiation rates because of increased usage of adjunctive medications.

CONCLUSION

To increase the likelihood of participants initiating treatment with injectionable naltrexone, providers may consider lowering the threshold for administering clonidine and clonazepam, giving it preemptively or to manage even minimal withdrawal symptoms.