Citalopram for Treatment of Cocaine Use Disorder: A Bayesian Drop-The-Loser Randomized Clinical Trial

Abstract

Background.

Medication development research for cocaine use disorder (CUD) has been a longstanding goal in addiction research, but has not resulted in an FDA-approved treatment. Rising cocaine use rates underscore the need for efficient adaptive designs. This study compared differences between two doses of the selective serotonin reuptake inhibitor (SSRI) citalopram (versus placebo) on duration of cocaine abstinence and applied adaptive decision rules to select the ‘best efficacy’ dose.

Methods.

A double-blind, placebo-controlled, randomized Bayesian drop-the-loser (DTL) trial with three arms compared placebo to citalopram 20mg and 40mg. Adults (N=107) with CUD attended thrice-weekly clinic visits for 9 weeks. The primary outcome was longest duration of abstinence (LDA), based on continuous cocaine-negative urine drug screens (UDS). The secondary outcome was probability of cocaine-negative UDS during treatment. A planned interim analysis performed at approximately 50% of recruitment dropped the least-effective active medication. Bayesian inference was used for all analyses with a pre-specified posterior probability (PP) threshold PP≥95% considered statistically reliable evidence.

Results.

Citalopram 40mg satisfied interim efficacy criteria and was retained for the second half of the trial. For LDA, analyses indicated PP=82% and PP=65% of benefit for 40mg and 20mg, respectively (each relative to placebo). The odds of having cocaine-negative UDS decreased in all groups over 9 weeks but remained higher for 40mg (PP=97.4%).

Conclusions.

Neither dose met the 95% PP threshold for the primary outcome; however, 40mg provided moderate-to-strong evidence for positive effects on LDA and cocaine-negative UDS. The 40mg dose was declared the “winner” in this DTL trial.

1. INTRODUCTION

Currently there are no approved medications for the treatment of cocaine use disorders (CUD) but many plausible targets exist. The mesolimbic dopaminergic system is well known for its critical role in mediating the reinforcing effect of cocaine and other stimulants and, as such, has been a primary focus of medication development for CUD. However, the limited utility of dopaminergic agents to reduce cocaine use has prompted investigation of additional neurochemical targets. Serotonin (5-hydroxytryptamine, 5-HT) neurotransmission has been mechanistically linked to cocaine-seeking in preclinical models, representing another rational target in the search for an effective pharmacotherapy of cocaine addiction (Cunningham et al. 2010; Cunningham and Callahan 1994; Filip et al. 2006; Howell and Cunningham 2015). Cocaine binds to the 5-HT transporter (SERT), inhibits reuptake, and increases dopamine neurotransmission in key brain areas that receive serotonergic innervation. These cocaine-induced alterations of 5-HT have been shown to amplify cocaine’s rewarding effects while causing impairment in the regulation of impulse control (Matsui and Alvarez 2018; Wright et al. 2017). Indeed, selective antagonists for 5-HT receptor subtypes have been shown to suppress impulsive action and cocaine cue-reactivity in preclinical studies (Anastasio et al. 2011; Anastasio et al. 2015; Cunningham et al. 2013; Fink et al. 2015; Fletcher et al. 2011; Fletcher et al. 2007; Sholler et al. 2019; Winstanley et al. 2004).

Recognizing the involvement of 5-HT in the neurobiology of CUD has prompted investigation of serotonergic agents as potential therapeutic modalities. Several selective 5-HT reuptake inhibitors (SSRIs) have been examined, including fluoxetine, sertraline and citalopram. Early fluoxetine trials failed to show reductions in cocaine use (Grabowski et al. 1995; Schmitz et al. 2001; Winstanley et al. 2011), possibly owing to a pharmacokinetic interaction between cocaine and fluoxetine that is accompanied by elevated brain levels of cocaine (Fletcher et al. 2004). Having little or no effect on the metabolism of cocaine, sertraline and citalopram trials have yielded predominantly positive findings. In two similarly designed clinical trials of sertraline in abstinent patients with CUD and depressive symptoms, rates of lapse and relapse were significantly lower in the active medication group (200 mg) relative to those who were treated with placebo (Mancino et al. 2014; Oliveto et al. 2012). We conducted a randomized clinical trial to assess the effect of citalopram, a selective 5-HT_2C agonist, in a sample of 76 treatment-seeking outpatients with CUD (Moeller et al. 2007). Relative to placebo, those receiving 20 mg citalopram daily for 12 weeks showed a significant reduction in cocaine-positive urines during treatment. There were no group differences in retention, medication adherence, or side effects, supporting further evaluation of the potential clinical utility of citalopram for the treatment for CUD.

To date, the highest dose of citalopram evaluated for CUD has been 20 mg; substantially lower than doses shown to be beneficial for the treatment of obsessive-compulsive disorder (60 mg) or depression (40 mg). Uncertainty regarding the adequate dose of citalopram for treatment of CUD prompted us to design the current double-blind trial comparing two active medication arms (20 mg; 40 mg) with placebo. The primary aim of the study was to select the most promising treatment arm to evaluate in a subsequent larger confirmatory trial, thus, we selected a drop-the-loser (DTL) Bayesian adaptive design.

The use of adaptive and Bayesian trial designs over conventional fixed parallel group designs in drug development has been encouraged with guidance from the FDA (e.g., US Food and Drug Administration 2010). One such design is the two-stage DTL trial in which multiple treatment arms versus placebo are assessed at a predefined interim analysis, with only the most promising treatment arm(s) carried forward to a second stage, thus reducing the sample size required and increasing the number of patients exposed to more efficacious doses. For the current trial, we carried out a simulation study to arrive at a set of decision rules for retaining the best-performing active condition, which we hypothesized to be citalopram at 40 mg (Rathnayaka 2017). Here we report interim and final results from this DTL trial.

2. MATERIALS AND METHODS

2.1. Participants

Participants were seeking treatment for CUD and were recruited by advertisements in the local media and through clinical referrals. A two-phase enrollment protocol was employed. Participants were first screened by phone, then provided informed written consent to participate in a 7–10 day general intake evaluation consisting of a psychiatric evaluation (SCID DSM-IV: First 1997), Addiction Severity Index (ASI: McLellan et al. 1992), physical examination and laboratory testing (chemistry screen, complete blood count, urinalysis, and a 12 lead EKG). For this study, major inclusion criteria were age (18 – 60 years old), meeting DSM-IV diagnostic criteria for current cocaine dependence, and providing at least one cocaine-positive urine drug screen (UDS) during intake. Having current dependence on any drug except cocaine, alcohol, nicotine, and cannabis was exclusionary. Physiological dependence on alcohol requiring medical detoxification was exclusionary. Individuals diagnosed with a non-substance induced Axis I psychotic, depressive, or anxiety disorder were excluded, as were those presenting with cardiovascular disease (CVD) or with symptoms determined by electrocardiogram suggestive of CVD problems not related to drug use, such as hypertension (treated or untreated), stroke, chest, pain. Individuals having other medical conditions or using medications that would adversely interact with citalopram were excluded.

The University of Texas Health Science Center at Houston (UTHealth) Center for Neurobehavioral Research on Addiction (CNRA) served as the primary study site, however a second site at Virginia Commonwealth University (VCU) enrolled a subgroup of participants (n=7) following the PI’s (FGM) relocation to this institution. The study was approved by the UTHealth Committee for the Protection of Human Subjects (local IRB) in accord with the Belmont Report and the Declaration of Helsinki, and registered at ClinicalTrials.gov [NCT01535573].

2.2. Design and Procedures

The trial used a double-blind, two-stage DTL design with initial randomization of eligible participants into one of three treatment arms comparing placebo (PLC) to citalopram 20 mg (C20) or 40 mg (C40). As described in the published protocol (Rathnayaka 2017), a planned interim analysis was performed at approximately 50% of data gathering to drop or “prune” the active medication group performing the least well. Following application of the interim decision rules, described below, additional participants were randomly allocated to the remaining treatment conditions. Data collected from participants in all three treatment arms was used in the final analysis.

The 10-week trial began with a 1-week dose escalation (10 mg days 1–3; 20 mg days 4–7, to reach dose levels of 20 mg or 40 mg by the start of week 2), followed by maintenance for 7 weeks, and a 1-week dose reduction at week 9. Riboflavin (50 mg) was added to all capsules and used as a marker to monitor compliance. Participants attended 3 clinic visits per week (MWF) to receive study medication and provide a urine specimen for drug screening and fluorescent detection of riboflavin. A side-effects checklist was completed each week, with moderate to severely rated items evaluated by the study nurse and reviewed by the study physician (FGM). Once weekly individual cognitive-behavioral therapy (CBT) was provided. A prize-bowl contingency management (CM) intervention was used to reinforce clinic visit attendance. We used the standard escalating schedule in which participants earned a draw each time they attended a scheduled clinic visit (Petry et al. 2012). The prize-bowl contained 500 slips, with cash values of $0 (50%), $1 (41.9%), $25 (8.0%), and $100 (.2%). Earnings were given in the form of a gift cards.

Self-report data on cocaine use was collected at each clinic visit using a timeline follow-back procedure. The primary outcome was longest duration of abstinence (LDA), based on the number of consecutive cocaine-negative (benzoylecgonine values < 300 ng/ml) UDS, as used in the initial citalopram trial (Moeller et al. 2007). LDA serves as a composite metric of both treatment retention and abstinence. The secondary outcome was proportion of cocaine-negative UDS collected 3 times weekly during treatment.

2.3. Data Analytic Strategy

Sample characteristics were evaluated via descriptive statistics, including measures of frequency and central tendency. All analyses were performed on a modified intent-to-treat sample of randomized participants who received at least one capsule (PLC n = 43; C20 n = 21; C40 n = 43). Negative binomial regression (i.e., generalized linear modeling; GLM) was used to model LDA as a function of treatment condition. Multilevel logistic regression (i.e., generalized linear mixed modeling; GLMM) was used to model cocaine-negative (vs. positive) UDS as a function of time, treatment, and the interaction between time and treatment. This longitudinal analysis was first performed with listwise deletion of missing outcome data and again with all missing imputed as cocaine-positive. These analyses did not provide different inferences; as such, the current manuscript focuses on the results from the first model featuring all available observed data. The baseline UDS result was included in the longitudinal analysis as part of the outcome. Cox proportional hazards regression was used to evaluate time to dropout across and between groups. Parameter estimates from negative binomial, logistic, and Cox regression were exponentiated to provide incidence rate ratios (IRR), odds ratios (OR), and hazard ratios (HR), respectively. All models statistically controlled for the stratification variables of cocaine and marijuana use severity (high vs. low based on use in past 30 days). As due diligence, treatment site (CNRA, VCU) was evaluated as a potential confounder of the treatment effect on each outcome. Including treatment site did not influence model inferences or parameter estimates for any other variables and was thus excluded from the final analyses. Primary outcomes were analyzed using the R Statistical Computing Environment (2019) using packages brms and rstan (Bürkner 2017). Time to dropout was analyzed using SAS (2006).

As specified by the protocol, Bayesian statistical inference was used to directly quantify the probability that model effects exist, given the present data and vague, neutral priors (μ=0; σ² = 1 × 10⁶). Assumptions of Bayesian statistical analyses were evaluated via effective sample size, scale convergence factors (“rhat”), and graphical posterior predictive checking (i.e., comparing the distribution of the observed data to simulated datasets from the posterior predictive distribution); these assumptions were satisfied for all models. The median of the posterior distribution was used to derive a point estimate for each model effect (IRR, OR, HR), with 95% credible intervals (CrI) to describe uncertainty. Multilevel models used leave-one-out information criteria (Vehtari et al. 2016) to determine the optimal random effect structure for each model, finding that the best-fitting model included level 2 intercepts and slopes. Model inferences rely on interpreting the chance that an effect exists by calculating the proportions of the posterior distribution that are greater and less than the null effect (e.g., IRR = 1), also called the posterior probability (PP). As an example, a posterior distribution that is 75% greater and 25% less than 1.0 could be written as PP(IRR > 1) = 75%. Posterior probabilities are also sometimes rescaled to Bayes factors (BF): the ratio of the PP in favor and against an effect (e.g., PP(IRR > 1) = 75% is akin to BF = 3 = 75%/25%).

The protocol for the current study provided a specific posterior probability threshold PP ≥ 95% that would provide a significant degree of evidence to conclude that at a difference exists between the retained active treatment condition and PLC with respect to the primary outcome (LDA). This threshold was chosen in part to (a) provide a generic conceptual bridge from traditional frequentist inference and (b) necessitate a substantial degree of evidence in favor of the active treatment. The protocol did not establish a threshold of evidence for the secondary outcome (cocaine-negative UDS). Given that the PP is not the complement of the p-value (i.e., 95% PP ≠ 0.05 p-value), and that Bayesian inference does not rely on a monolithic value for the PP akin to p < 0.05, a supplemental interpretation of results following broad heuristics for various thresholds described in the literature is provided for all analyses: PP 50–74% (BF = 1 to 3) as “anecdotal” evidence; PP 75–90% (BF = 3 to 10) as “moderate” evidence; PP 91–96% (BF = 10 to 30) as “strong” evidence; and PP ≥ 97% (BF > 30) as “very strong” to “extreme” evidence (Andraszewicz et al. 2015; Jeffreys 1961; Lee and Wagenmakers 2014).

2.4. Interim Analysis

Planned interim analyses, performed by a third-party statistician and reported back to the pharmacist to avoid disrupting the double-blind, were carried out to retain the best-performing active (non-placebo) condition. Timing of the interim analysis was planned when enrollment reached approximately half of the expected total sample size. Decision rules regarding treatment arm retention were elaborate and previously reported in the literature (Rathnayaka 2017). In brief, both active treatment conditions met the drop criteria and therefore only the condition with the largest effect size was retained. Post-trial unblinding determined that the dropped treatment arm was C20.

3. RESULTS

3.1. Sample Description and Treatment Retention

Figure 1 displays the progress of all participants through the trial before and after interim analysis. Before the interim analysis, N = 66 participants were randomized into one of the 3 conditions. After the interim analysis, N = 42 (35 at the CNRA; 7 at VCU) were randomized into one of the remaining 2 conditions. Of the total sample of participants who started treatment (N = 107), sixty (56%) completed all 9 weeks of treatment. Bayesian Cox regression indicated no difference in time to dropout for C20 versus PLC (HR = 1.01; 95% CrI = [0.47, 2.10], PP = 51.0%) or C40 versus PLC (HR = 1.05; 95% CrI = [0.58, 1.88], PP = 56.0%). Total earnings from the attendance-based prize-bowl CM were $237.43, $239.59, $236.47 and not different for C20, C40, and PLC, respectively (Kruskal-Wallis test p = 0.989). Characteristics of the randomized participants are shown in Table 1. Preliminary analyses (Fisher exact test; Kruskal-Wallis rank sum test) using frequentist inference (i.e., separate from the primary Bayesian analyses described below) determined that none of the characteristics were significantly different across treatment conditions.

Figure 1.

Flow chart of study participants.

Table 1.

Baseline Characteristics of Randomized Participants by Treatment Group

	Placebo	Citalopram 20 mg	Citalopram 40 mg
N	43	21	44
Male (n, %)	34 (79.1)	18 (85.7)	36 (81.8)
Race/ethnicity (n, %)
Hispanic	5 (11.6)	2 (9.5)	5 (11.4)
Black	31 (72.1)	18 (85.7)	32 (72.7)
White	7 (16.3)	1 (4.8)	7 (15.9)
Age, years (M, SD)	47.1 (8.83)	46.9 (7.76)	44.5 (9.60)
Years education (M, SD)	12.5 (1.94)	12.6 (2.27)	12.05 (1.82)
Cocaine use (M, SD)
Current (past 30 days)	14.7 (10.08)	12.2 (7.31)	16.6 (9.10)
Lifetime (years)	16.5 (8.20)	12.7 (8.03)	13.2 (9.84)
Alcohol use (M, SD)
Current (past 30 days)	6.9 (8.07)	6.3 (7.47)	7.2 (8.54)
Lifetime (years)	13.9 (12.56)	11.6 (10.49)	10.9 (11.99)
Cannabis use (M, SD)
Current (past 30 days)	5.8 (10.09)	4.2 (5.89)	8.8 (12.48)
Lifetime (years)	11.7 (10.92)	7.5 (8.99)	12.6 (12.62)

3.2. Longest Duration of Abstinence

Mean (SD) LDA by treatment condition was 2.5 (5.0), 2.6 (4.6), and 3.5 (6.0) for PLC, C20, and C40, respectively. Bayesian GLM modeling of LDA as a function of treatment condition determined that, relative to PLC, the expected number of days in the LDA was 22% higher for C20 (IRR = 1.22, 95% CrI = [0.46, 3.55]) and 44% higher for C40 (IRR = 1.44, 95% CrI = [0.62, 3.22]). The posterior probability that these incidence rates were greater than PLC were PP (IRR > 1) = 64.8% for C20 and PP (IRR > 1) = 82.1% for C40. Neither treatment exceeded the protocol-based 95% threshold; in the strictest sense, evidence was not attained to support a statistically reliable difference. Heuristic-based interpretation suggests these PP values provide an anecdotal and a moderate level of evidence in favor of C20 and C40, respectively.

Figure 2 provides graphical representations of the posterior distributions for each treatment condition (each relative to placebo), shaded in red (C20) and blue (C40). Succinctly, each posterior distribution describes the probability of different effect sizes. This figure illustrates the intuition behind the current analytic approach: as opposed to traditional frequentist inference, wherein a regression coefficient is considered to have one fixed value, Bayesian inference considers that coefficient to have a range of possible values. Figure 2 also displays the method for deriving the above-noted expected values for number of days in the LDA for each active treatment condition, relative to placebo: the median of the posterior distribution for each comparison provided the point estimate: (C20: IRR = 1.22; C40: IRR = 1.44), with the shaded region above the reference line to describe the posterior probability that each effect is greater than IRR = 1.0.

Figure 2. Posterior probability of longest duration of abstinence.

The posterior distributions for the incidence rate ratios for citalopram 20 mg (red) and 40 mg (blue), relative to placebo. The x-axis represents the magnitude of the effect (IRR) and the y-axis represents the density of the posterior distribution. The solid black line at IRR = 1.0 represents a null effect. The chance that the effect of each treatment confers benefit (relative to placebo) is represented by the proportion of each distribution that to the right of the solid black line (PP(IRR > 1); as described above these were 64.8% (C20) and 82.1% (C40). The median and 95% CrI of each distribution are represented by dashed and dotted lines, respectively. For each distribution, the median provides the best single point estimate of the effect and the 95% CrI provides the equal-tailed interval that has 2.5% of the distribution on either side of its limits.

3.3. Cocaine Use over Time

The proportion of cocaine-negative urines averaged across visits in treatment (i.e., before dropout) by condition was M (SD) = 0.18 (0.29), 0.18 (0.24), 0.25 (0.33) for PLC, C20, and C40, respectively. Recalculating these values with all missing imputed as cocaine-positive (including beyond dropout) provided lower average proportions of cocaine-negative urines across all 27 possible treatment visits (3 visits/week for 9 weeks): M (SD) = 0.15 (0.25), 0.15 (0.23), and 0.20 (0.31), for PLC, C20, and C40, respectively (see Supplement S1 for mean proportions of cocaine-negative UDS by treatment over time). Bayesian GLMM did not support an interaction of time and treatment for predicting cocaine-negative (versus positive) urines; that is, there was no evidence that the probability of cocaine-negative urines changed differently for the three treatments over time. Reducing to main effects, the model demonstrated a moderate PP(OR > 1) = 83.6% effect of time, such that the odds of cocaine-negative urines decreased by 2.8% for each additional study visit (OR = 0.97, 95% CrI = [0.91, 1.03]). Relative to PLC, the model demonstrated anecdotal and very strong evidence for main effects of C20 and C40: PP(OR > 1) = 63.1% and 97.4%, respectively.

Figure 3 provides a plot of the posterior distributions for the effects of each treatment (relative to placebo). Across the entire duration of the study, C20 was related to 29.2% higher odds of cocaine-negative urines (OR = 1.29, 95% CrI = [0.28, 6.23]), while C40 was related to 270.4% higher odds (OR = 3.70, 95% CrI = [1.00, 14.67]). Finally, Figure 4 describes the observed probability of cocaine-negative UDS for each treatment condition over time.

Figure 3. Posterior probability of cocaine-negative urine drug screens.

The posterior distributions for the odds ratios for citalopram 20 mg (red) and 40 mg (blue), relative to placebo. Interpretation follows from Figure 2.

Figure 4. Observed probability of cocaine-negative UDS over time.

The green, red, and blue shapes describe observed probability of cocaine-negative urine drug screens for each treatment condition (PLC, C20, and C40, respectively) over time.

3.4. Medication Tolerability and Compliance

The most frequently reported side effects with average ratings moderate-to-severe included drowsiness (PLC = 4; C20 = 2; C40 = 10), not sleeping well (PLC = 17, C20 = 7, C40 = 10), and headache (PLC = 5; C20 = 2; C40 = 6). Statistical analyses via chi-square test determined that the frequencies of each of these reported side effects were not significantly different between groups. Across all reported side effects, significantly different frequencies were found between groups for dizziness and shortness of breath; however, all reports of these side effects (four of each) occurred in the placebo condition. One participant was hospitalized for an IV drug infection (placebo) and another participant received emergency treatment for suicide ideation with a plan (40 mg). These serious adverse events were judged to be unrelated to the study medication.

Rates of medication compliance based on urinary riboflavin were 72.4% (PLC), 78.8% (C20), and 81.1% (C40), with slightly lower rates based on self-reported pill consumption with return of unused pills: 69.3% (PLC), 70.9% (C20), and 69.2% (C40).

4. DISCUSSION

The current study demonstrates the utility of applying adaptive decision rules to optimize the efficiency of a trial in allocating participants to the most efficacious dose of citalopram for the treatment of CUD. Pre-specified criteria were used to drop the low dose condition (20 mg) at the midpoint of the trial for inferior efficacy (relative to 40 mg). Final analysis of the primary outcome showed that, relative to placebo, both active doses of citalopram produced longer durations of abstinence, with the strongest evidence in favor of citalopram 40 mg, associated with a moderate effect (82.2% PP). Overall, the odds of having a cocaine-negative urine decreased during the course of treatment, but remained higher for participants receiving 40 mg citalopram, relative to placebo (97.5% PP).

The premise for the current trial was based on our initial randomized clinical trial of citalopram 20 mg per day (Moeller et al. 2007). In that trial, citalopram treated participants showed a significant reduction in cocaine-positive urines during 12 weeks of treatment, compared to placebo. The mean number of consecutive cocaine-negative urines was 5.06 and 2.13 for citalopram and placebo, respectively. These positive findings prompted the current investigation to identify optimal dosing, based on the hypothesis that a higher dose of citalopram would maximize therapeutic benefit. The current findings complement and extend our previous findings by using a Bayesian approach that more readily maps onto clinical decision making via directly interpretation of the probability of the alternative hypothesis (i.e., that an effect of citalopram exists). Decision makers may find interpreting posterior probabilities described here (e.g., an 82.1% chance that C40 results in greater LDA than PLC) to be more useful for assigning treatment to prospective patients than the circuitous interpretation afforded by traditional frequentist inference (e.g., a given treatment did not a meet the requirements of a monolithic statistical significance criterion, given that the null hypothesis is true). Moreover, future research may rely on the Bayesian approach to incrementally improve treatment of cocaine use disorder by incorporating the present results as informative priors that would inform posterior probabilities derived from novel data.

At interim analysis, citalopram 20 mg was dropped for insufficient efficacy over placebo, with final results indicating only a 64% probability of benefit (vs. placebo) for LDA. The smaller effects for citalopram 20 mg observed in the current trial may reflect differences in the behavioral therapy platform. Participants in the initial trial received once-weekly individual CBT sessions along with a voucher-based CM intervention targeting consecutive cocaine-negative urine samples. In contrast, the current trial provided CBT with a prize-bowl CM intervention targeting clinic visit attendance. We selected this lower cost and less intense CM platform to enhance the detection of medication effects on abstinence, independent of abstinence achieved via targeted CM incentives. That citalopram showed stronger effects when paired with abstinence-based CM may be viewed as consistent with trials of sertraline, another SSRI agent, in which medication effects have been observed in the context of recent abstinence from cocaine use (Bashiri et al., 2017; Oliveto et al., 2012). This combined medication-behavioral therapy approach is one important direction for future research studies of citalopram for CUD. Another direction would be to combine citalopram with other agents having different mechanisms of action with the goal of achieving a larger and more clinically useful effect. Such combination treatments for stimulant use disorders have shown promise (Levin et al. 2019; Trivedi et al. 2021).

Consistent with our initial study, the present findings suggest that citalopram is a safe and acceptable treatment. Following a 1-week titration, both maintenance doses of citalopram were well-tolerated. Adverse effects were generally infrequent and unrelated to study discontinuation. Medication compliance rates ranged from 69% to 81%, depending on measurement by urinary riboflavin or pill count. That only 56% of participants completed the 9-week trial is low, but not inconsistent with other outpatient medication trials enrolling active cocaine users. In a recently conducted comprehensive review and meta-analysis of dropout rates of in-person psychosocial treatments for substance use disorders (Lappan et al. 2020), the average rate across 151 studies was 30.4%, with higher rates found in studies targeting cocaine (46.8%) and methamphetamine (53.5%). Whereas telemedicine is being increasingly used to overcome the geographical and financial barriers of attending frequent clinic-based treatment visits, the use of technology for remote data collection in the context of a medication trial has received little study. Mobile applications that offer capabilities to track medication adherence, conduct remote substance testing, provide psychosocial therapy, and deliver CM financial rewards, show promise in moving the field toward more “virtual” clinical trials that will improve participant access and retention.

The present study defined PP ≥ 95% as the threshold of evidence for concluding that the retained condition has substantial benefit over placebo on LDA (longest number of consecutive cocaine-negative urines). Citalopram 40 mg failed to meet this threshold which, in hindsight, may have been too stringent. To date, few other candidate medications for CUD have shown this level of “very strong or extreme” evidence of efficacy when using continuous abstinence outcomes (e.g., 2+ consecutive weeks). Unlike classical or frequentist paradigms where a p-value below a specified threshold (α = 0.05) is interpreted as failing to reject the null hypothesis, the Bayesian paradigm shifts the focus on calculating the probability in favor of the alternative hypothesis (Jeffreys 1961; Kass and Raftery 1995). In the present case, the BF for LDA was 4.59 (PP(IRR > 1)/PP(IRR < 1) = 82.1%/17.9% = 4.59). As with the posterior probability thresholds outlined in the data analytic strategy, this has also been interpreted in the literature as a moderate strength of evidence (BF > 3) in favor of the alternative hypothesis (Jeffreys 1961; Johnson 2013). The strength of evidence should be interpreted in tandem with the range of credible effect sizes provided via the posterior distributions of C40 (relative to PLC) for each model, with median effect sizes IRR = 1.44 and OR = 3.70 for LDA and UDS, respectively. In other words, C40 was related to 44% higher rates of days in the longest duration of abstinence compared to PLC and 270% higher odds of having a cocaine-negative urine drug screen on a given day of the trial.

Although 40 mg citalopram was retained as the “winner” in this DTL adaptive trial, there was considerable uncertainty around the value of this estimated effect. As shown in the PP distribution (Figure 2), 82% probability associated with the parameter (IRR > 1) included a wide range of effect sizes (Credible Interval range 0.62 to 3.22). Despite allocating more participants to this condition, the small sample size coupled with a high drop-out rate may have limited the precision in estimating the treatment effect. With the Bayesian approach, findings from the present study can be employed as priors in future research; effectively combining the information gleaned from this trial with evidence from additional participants to update treatment estimates with greater statistical power.

The current trial was limited with respect to the relatively smaller sample size for C20, the dropped condition, which constrained estimation precision for that dose. The short duration of the trial may be viewed as another limitation. Whereas 9 weeks was set to optimize efficiency and speed in identifying the optimal dose of citalopram, most medication trials for treatment of CUD are longer (e.g., 12 weeks), including our initial citalopram trial (Moeller et al. 2009), making cross-trial comparisons difficult. The trial was also influenced by dropout. Only 56% of participants completed all 9 weeks of treatment, which is lower than other studies using CM approaches to improve retention (e.g., Levin et al., 2020; Kampman et al., 2015; Schierenberg et al., 2012).

In summary, this study presents some methodological innovation to the field of medication development for CUD. DTL and other flexible trial designs offer an efficient and informative alternative to traditional fixed designs when it comes to selecting the most promising treatment or dose to advance. Despite these advantages, adaptive designs have been underutilized in clinical trials of pharmacotherapies for drug addiction, particularly in comparison to other areas of medication development (Pallmann et al. 2018). Here we demonstrate the clinical utility of using a DTL design for testing optimal dose response of citalopram. It has been argued that most candidate medications for CUD have been prematurely dismissed without considering dosing and other procedural factors that may affect treatment outcome (Brandt et al., 2020). Citalopram 40 mg shows very strong evidence of efficacy in reducing cocaine use and should be further evaluated for the treatment of CUD.

HIGHLIGHTS.

• An adaptive Drop-The-Loser 3-arm trial compared two citalopram doses to placebo.

• A planned interim analysis at 50% of recruitment dropped the 20mg dose.

• The 40mg dose had an 82.1% chance of greater longest duration of abstinence.

• The 40mg dose had a 97.4% chance of higher odds of cocaine-negative drug screens.