1. Introduction

The staggering impact of opioid use disorder (OUD) is compounded by polysubstance use (Cicero et al., 2020; Compton et al., 2021). While estimates on the percentage of people with OUD who use multiple substances vary, the most recent estimates suggest that polysubstance use is the norm in people with OUD (Cicero et al., 2020; Makarenko et al., 2018; Soyka et al., 2017). Rates of polysubstance use among people in treatment for OUD range between 65% (Jarlenski et al., 2017) to 85% (Raffa et al., 2007). While some overdoses occur in the context of unintentional ingestion of multiple substances, such as carfentanil mixed with heroin, other people actively seek out multiple drugs (Ataiants et al., 2020). However, little is known about the patterns in which various types of drugs are used. Some combinations, such as opioids with benzodiazepines, are frequently used together and likely contribute to overdose events (Hernandez et al., 2018; Seth et al., 2018). Less is known about other combinations, such as opioids with marijuana or opioids with alcohol, and their impact on treatment response is unclear (Hassan and Le Foll, 2019; Wagner et al., 2018). There is a need to better characterize which substances are used together and whether the patterns of use change prior to engagement in treatment.

Polysubstance use researchers focusing on drugs other than opioids, such as cannabis (Connor et al., 2013, 2014) and tranquilizers (Votaw et al., 2020), have used a classic statistical method, Latent Class Analysis (LCA) (Collins and Lanza, 2009), to identify drug use patterns. An extension of LCA, repeated measures latent class analysis (RMLCA), considers repeated drug exposure patterns and how they cluster through time (Collins and Lanza, 2009). Knowing which drugs are used together and how this does or does not change over time allows for identifying patterns in multiple comorbid conditions, such as depression and anxiety frequently co-occurring in “wide-range substance users.” Understanding these patterns may, in turn, lead to more targeted treatments (Carlsen et al., 2020).

While methods like LCA are useful for finding such patterns, particularly when dealing with a small number of grouping features (e.g., 6 to 12 drugs or classes of drugs), modern machine learning methods such as market basket analyses (MBA) are adept at identifying patterns when there are many features (Lantz, 2019). The method has been applied to identify co-occurring patterns in a host of domains, such as detecting patterns of genes associated with disease, co-occurring food allergies, and even detecting tactical patterns in elite beach volleyball (Aguinis et al., 2013; Wenninger et al., 2019). This ability to detect patterns, even in the presence of rarely used drugs out of a pool of many possible substances, makes MBA a useful tool for studying polysubstance use. MBA, as opposed to RMLCA, focuses on polysubstance use over short periods (e.g., in the same day).

Although studies have documented the rate of polysubstance use and/or reported on drugs frequently used together with opioids (Hser et al., 2017; Northrup et al., 2015; Ruglass et al., 2019), few have looked at the fine details of drug use patterns and how substance use changes through time in people with OUD before entering treatment. Here we explore the daily patterns of substance use self-reported in the “Timeline Followback” in the 28 days before the onset of Medication for Opioid Use Disorder (MOUD) in three National Institute on Drug Abuse (NIDA) Clinical Trials Network (CTN) studies. Data from these three CTN studies were harmonized as part of a fourth CTN study (CTN-0094) to allow more extensive secondary analyses of OUD treatment. Different from the previous studies that model the trajectories of outcomes (Hser et al., 2017; Northrup et al., 2015; Ruglass et al., 2019), we focused on modeling trajectories of substance use prior to treatment initiation as predictors for relapse. We report patterns of drug use using MBA and RMLCA. MBA is useful to identify patterns of all individual drugs used on any given day, while RMLCA allows exploring patterns over longer periods of time. We examine daily reports of use across 12 broad classes of substances over the 28 days before starting MOUD. Associations between selected RMLCA classes with relapse outcomes were also examined. We hypothesized that subgroups of polysubstance use, with different prevalence, can be identified by MBA and RMLCA, and that some of these patterns are more likely than others to be associated with relapse to regular opioid use during treatment.

2. Material and methods

2.1. Study population

Full protocols for the individual studies were previously published (Lee et al., 2016; Saxon et al., 2013; Weiss et al., 2011). Briefly, all three studies enrolled individuals who met the criteria of DSM-IV-TR for opioid dependence (CTN-0027/CTN-0030) or DSM-5 diagnosis of OUD (CTN-0051). The trials enrolled adult participants over 18, with very broad pragmatic inclusion and few exclusions except for major medical and unstable psychiatric comorbidities. CTN-0027 was the most inclusive, whereas CTN-0030 excluded individuals with OUD who only used heroin, and CTN-0051 excluded individuals currently receiving methadone treatment. CTN-0027 randomized individuals to buprenorphine and methadone for outpatient treatment for 24 weeks. CTN-0030 randomized individuals to two different types of psychotherapy programs, and all participants received identical medication treatment in two phases: in the initial phase, participants received a buprenorphine taper, and those who relapsed (about 90% of all participants), were treated with buprenorphine maintenance over 12 weeks in the second phase. CTN-0051 randomized patients from inpatient treatment facilities and they either received buprenorphine or extended-release naltrexone after discharge and were followed as outpatients for 24 weeks.

2.2. Substance use measures

We focused our analysis on polysubstance use prior to treatment initiation as a baseline predictor of treatment outcomes. For all participants who enrolled in the above three studies, substance use history in the 28 days prior to treatment initiation was surveyed with the Timeline Followback method (Sobell and Sobell, 1992). The free text drug names were harmonized to account for spelling errors and differences in names (generic, brand versus street) and to remove superfluous information (e.g., mg/capsule/oz), resulting in a set of 44 distinct substances. These data included both high-frequency drugs of use which were also identified in structured questions and low-frequency substances of use (e.g., cathinones N = 5 events), drugs of concern (e.g., gabapentin N = 21), as well as other drugs such as antiemetics (N = 18), non-opioid pain killers (e.g., acetaminophen N = 5) and antipsychotics (N = 3). These values were used for MBA. To allow LCA to operate on a more tractable set of options, these harmonized drug names were grouped into twelve “drug categories.” Table 1 lists the drugs that were surveyed as part of structured questions in the three studies and how they were grouped for LCA. Other repeatedly occurring drugs were: fentanyl and opium (grouped-with/classified-as Heroin for LCA); GHB (Gamma hydroxybutyrate) (Depressants); K2 (Synthetic cannabinoids); merperidine, tramadol, oxymorphone (Opioids); barbiturate, sedative-hypnotic (Depressants); MDMA (3,4-Methyl​enedioxy​methamphetamine), hallucinogen unspecified (Hallucinogens); muscle relaxant unspecified, soma, methocarbamol, flexeril, baclofen, carisoprodol (Relaxant).

Table 1. Structured drug use variables from three harmonized studies and how they were categorized for repeated measure latent class analyses.

_{⁎Can be calculated.}

_{⁎⁎Only a count of days.}

_{⁎⁎⁎If yes to opiates, then follow up for which drug.}

2.3. Relapse

Relapse was defined by four consecutive opioid use weeks between 21 days after treatment initiation and the end of the 12-week treatment period. An opioid use week was defined as having either a urine drug screen (UDS) positive for any non-prescribed opioid or a missing/refused UDS in that week. We defined relapse as a categorical variable representing three possible outcomes: 0 indicated that the definition of relapse was not met (50% of the sample), 1 indicated at least one opioid-positive UDS during four consecutive weeks with positive or missing UDSs (indicating that the participant, albeit positive for non-prescribed opioids, showed up to the clinic at least once; 29% of the sample), and 2 indicated missing UDS data for four consecutive weeks (21% of the sample). This definition closely followed a more complex definition from CTN-0051, but only required UDS results and not the self-reported drug use information gathered on the Timeline Followback.

2.4. Statistical analysis

MBA was originally designed to guide business decisions using massive datasets, such as all the transactions in a supermarket for a month. In these scenarios, traditional hypothesis testing with p-values becomes meaningless because anything is statistically significant with a big enough sample size. Instead of focusing on p-values, MBA methods rely on three statistics to assess the importance of an association: lift, support, and confidence. These measures assess the chances of seeing drug A and drug B (or a set of drugs which we will label A and a different set that we will label as B) appearing on a drug screening on the same day. Lift, which assesses if an association exists, is calculated as 𝑃⁡(𝐴∩𝐵)𝑃⁡(𝐴)*𝑃⁡(𝐵). It is the ratio of the actual probability of drug A and drug B occurring on a given day divided by the probability of seeing drug A on any day times the probability of seeing drug B on any day. In other words, it is the increased or decreased probability of seeing both drugs relative to their overall chance of either one appearing (assuming their co-occurrence is random). Support is synonymous with probability of seeing the various sets of drugs being examined. MBA practitioners focus on the support for A and B, 𝑃⁢(𝐴∩𝐵), that is, the probability of seeing both items on the same day. Support is of decreasing utility as the number of possible drugs in the set increases. Confidence, which describes the probability that a set of items appears given that one has already occurred, is calculated as 𝑃⁡(𝐴∩𝐵)𝑃⁡(𝐴). In other words, it is the conditional probability of seeing drug B given that you see A, 𝑃⁡(𝐵|𝐴).

Using these concepts, it is possible to think of predictive association rules where A leads to B (which is written as A -> B). It is useful to think of the relationship between lift and confidence: 𝑙⁢𝑖⁢𝑓⁡𝑡⁡(𝐴→𝐵)=𝑃⁡(𝐵|𝐴)𝑃⁡(𝐵). Importantly, A and B are not interchangeable in the confidence equations. That is, the chances of seeing chips after seeing guacamole (guacamole -> chips) is not the same as the chances of seeing guacamole after seeing chips (chips -> guacamole). This “directional” property allows MBA to make predictions that can describe the greatly increased risk of drug B if the person has used drug A, but seeing drug B may not increase the chances of seeing drug A.

For MBA, the a priori algorithm was used after excluding drugs with a support of less than 0.001. That is, drugs that occurred with a frequency of less than twice a month per transaction (28*2/56,000) were excluded. The algorithm was set to extract all rules with a confidence of at least 0.5. We then selected the top ten rules based on the lift (i.e., ratio of the support of the antecedent drugs co-occurring with the consequent drugs, divided by the probability that the antecedent and consequent drugs co-occur if the two are independent) and created a binary class to indicate the particular basket of polydrug use. Participants who reported using all substances in the basket rule during the study period were identified and categorized into different polydrug baskets.

RMLCA for the drugs used in the 28 days leading up to MOUD initiation was conducted using the drug groups described above. First, class enumeration was done without covariates by estimating models with increasing numbers of classes until the sample size in each latent class was considered to be too small for practical interpretation (less than 5% of the total sample size) and/or information criteria showed worse model fit. To ensure the models converged to the global maximum, 1000 random starts and 100 replicated likelihoods were used for each model. Model fit statistics for each of the models were used to determine the model that best fit the data, including entropy (Celeux and Soromenho, 1996) and penalized information criteria (Bayesian Information Criteria (BIC), Akaike's Information Criteria (AIC) (Vrieze, 2012), Vuong-Lo-Mendell-Rubin, Lo-Mendell-Rubin adjusted LRT tests (Nylund et al., 2007), and bootstrapped likelihood ratio tests (BLRT) (Asparouhov and Muthén, 2014; Feng and McCulloch, 1996). Second, once the best-fitting class structures were determined, a 3-step approach using the R3STEP (Asparouhov and Muthén, 2014) procedure was used to examine between-class differences in the relapse covariate using multinomial logistic regression. The 3-step procedure ensures that the inclusion of the covariates does not change the class structure.

Logistic regressions were used to test for the effects of RMLCA classes and MBA top lift baskets on relapse controlling for which of the three trials, and treatment arms within the trials. Odds ratios (OR) and 95% confidence intervals (CIs) were reported. P-values < 0.05 for two-sided tests were considered statistically significant. Demographic and substance use data from CTN-0027, 0030, and 0051 were harmonized using SAS 9.4. Exploratory data analyses and MBA were conducted with R 3.6.2 with packages including haven (Wickham and Miller, 2021) (version 2.2), tidyverse (Wickham et al., 2019) (version 1.3), arules (Hahsler et al., 2022, 2011; Hornik et al., 2005) (version 1.6–5), and arulesViz (Hahsler, 2017; Hahsler and Chelluboina, 2021) (version 1.3–3). Haven was used to export the data from SAS, and tidyverse was used for data cleaning. Arules and arulesViz were used for MBA. RMLCA analyses were conducted in Mplus (version 8.3).

3. Results

3.1. Parent study and participants

All people who provided drug use Timeline Followback from CTN-0027 (N = 1300), CTN-0030 (N = 661) and CTN-0051 (N = 676) were included in this study. Participants typically provided full drug use information for the 28 days prior to initiation of MOUD, and 97.9% provided more than three weeks of history. Despite different participant selection criteria and differences in available data across the three trials, Table 2 shows similar patterns in terms of baseline demographic and psychiatric co-morbidities.

Table 2. Demographic details on 2637 participants from three clinical trials, who self-reported drug use in the 28 days before initiation of medication for opioid use disorder and randomization.

3.2. Market basket

The (N = 2637) participants reported using at least one drug on 66,098 (89%) days in the 28 days prior to MOUD initiation and randomization. Fig. 1 shows the top 10 frequently used substances. The most commonly occurring substances were heroin (reported on N = 40,740 days), oxycodone (N = 12,289 days), cannabis (N = 9183 days), hydrocodone (N = 7336), and methadone (N = 3663 days). Almost two-thirds (64%) of drug use days involved using a single substance (mean drugs used per day = 1.45). Participants used two substances on 29%, three substances on 6%, and four substances on 6% of drug use days. One person reported using 13 substances on a single day.

Fig. 1.

_{Substance use frequency (Top 10).}

The a priori MBA algorithm identified 34 association rules. As can be seen in Table 3, these rules included extremely strong predictions for rare events, for example, a 16-fold increase (with 71% confidence) in the chance of seeing heroin with the use of opium (lift = 16, count = 69). Predictions were also strong for relatively more common events, such as using heroin with crack (lift = 1.46, count = 2884) or cocaine (lift = 1.35, count = 2598). The algorithm noted several large effect-size combinations, such as the 13-fold increase in the chances of reported cocaine use when a combination of heavy drinking, opioid, and cannabis was observed, and a 6.13-fold increase in the chances of cannabis in a combination of cocaine, heavy drinking, and opioid. While the majority of the rules (76%) predicted heroin use after consuming other drugs, 18% of the rules were associated with cannabis use.

Table 3. Market basket analysis result, sorted by lift, shows which drug is predicted (C is consequent) given the presence of other drugs (+ shows antecedents) which were used on the same day.

_{* +: Antecedent; C: Consequent.}

3.3. RMLCA

RMLCA with three to eight classes was fit to all the harmonized data. Fit indices for each LCA model are presented in Appendix Table A1. The Vuong-Lo-Mendell-Rubin test and Lo-Mendell-Rubin adjusted LRT test, both with p-values of 0.811, suggested that a six-class solution was an adequate fit and that seven classes were not needed. However, the bootstrapped parametric likelihood ratio test, with a p-value < 0.0001, suggested seven classes. Given the large improvement in adjusted BIC between the five and the six-class solutions and the fact that the seven-class solution has some classes with smaller size groups, we selected the six-class solution (Table A1).

The daily probability and prevalence of drug use by latent class for the six-class solution are shown in Fig. 2 (by days) and Fig. 3 (by drug use). More comprehensive figures by days, and by drug use, stratified by the three studies are provided as in the appendix (Figs. A1–A6). The largest class was C1-“All time low” (n = 829, 31.4%), which describes a group of patients who used few substances, including any opioids, in the 28 days prior to treatment initiation. This group primarily included participants from CTN-0030 (n = 466), CTN-0051 (n = 208) and CTN-0027 (n = 155). The second largest class was C2-High opioid decreasing (n = 824, 31.3%). This class was comprised of individuals who had a self-initiated decrease of prescription opioids in the 28 days prior to the date of assessment on treatment entry. This group likely represents individuals who voluntarily decreased opioid intake in preparation for either inpatient (CTN-0051, n = 13) or outpatient treatment (CTN-0027 n = 811); no CTN-0030 participants were represented in this class. As can be seen in the corresponding figures, across all classes except the low substance use group, there was a precipitous drop-off in substance use in the two weeks leading up to treatment initiation, particularly heroin, methadone, opioid, and cannabis use. However, the degree of reduction differed by drug.

Fig. 2.

_{Six class solution by RMLCA class (Overall).}

Fig. 3.

_{Six class solution by drug use (Overall).}

3.4. Association between relapse and top 10 lift baskets

We selected the top 10 lift rules based on the MBA (Table 4). Individuals co-using cocaine, heroin, opioids, and cannabis had a higher risk of having a relapse event defined by the presence of at least one opioid-positive UDS (OR = 2.82, 95% CI = 1.13, 7.03) or exclusively missing UDS for four consecutive weeks (OR = 5.13, 95% CI = 1.44, 18.22) compared to individuals who did not report use of this particular drug combination, controlling for other polydrug baskets and trial. Individuals who reported polydrug use of cocaine, heavy drinking, cannabis, and opioids had a lower risk of relapse defined by four consecutive missing UDS (OR = 0.44, 95% CI = 0.25, 0.78). Similarly, we identified decreased odds of relapse for individuals who reported polydrug use of benzodiazepine, heroin, opioid, and cannabis (OR = 0.58, 95% CI = 0. 39, 0.86). The number of participants with polydrug patterns represented by the top 10 lift baskets is listed in Appendix Table 2.

Table 4. Association between relapse and Market Basket top 10 lift baskets.

4. Discussion

This study provides novel evidence on patterns of polysubstance use in the immediate period prior to treatment initiation among individuals with OUD. Polysubstance use was common, and both RMLCA and MBA offer valuable insights into the specific patterns. RMLCA is a repeated measures extension of LCA and person-centered data analytic technique. Therefore, this approach helps identify latent patterns of responses to categorical items with varying probabilities of endorsement. Compared with other data segmentation methods, such as hierarchical clustering, RMLCA derives clusters using a formal probabilistic approach and can be used in conjunction with multivariate methods to estimate parameters. It helps determine how many patterns of responses/behavior are present in the data, how prevalent each pattern is, and how likely item endorsement is in each latent class. The optimal number of classes minimizes the degree of relationship among cases belonging to different classes. To select the optimal number, methods such as the Bayesian Information Criterion are used, which capitalize on the value of the negative log-likelihood function, a well-established measure of the goodness of fit of a statistical model. MBA was used to reveal the most common combinations of substances used together on the same day and it affords additional insights that are otherwise masked by broad drug classes needed for traditional LCA.

Our results indicate, for example, that opium users were also using heroin, alcohol, and cannabis, and there were several unidirectional relationships where the use of specific drugs (e.g., benzodiazepine, cocaine, and cannabis) was likely to lead to heroin use. RMLCA identified six broad classes of polysubstance use patterns prior to treatment initiation and revealed that a large number of patients either had shown low use of all substances, including opioids or substantially reduced use prior to treatment initiation. Further, both methods used in our analyses revealed subgroups of polysubstance use patterns associated with relapse outcomes.

We found some unexpected associations between specific MBA baskets and relapse by week 12: the observance of a combination of cocaine, cannabis, and heavy drinking with prescription opioids did not negatively impact the odds of eventual relapse but instead was correlated with decreased odds of relapse. In contrast, the appearance of the same combination with heroin predicted a marked increase in the odds of relapse. This suggests that a polysubstance “binge” is not automatically associated with relapse to regular opioid use or non-attendance at scheduled clinic visits (i.e., missed UDS). In fact, the occurrence of a polysubstance heroin binge portends a worse prognosis, and a polysubstance binge with prescription opioids much less so.

Most importantly, the different RMLCA classes and MBA polydrug use groups had different relationships with relapse. For example, RMLCA revealed that people in both the “high opioid decreasing” and “high heroin decreasing” classes were more likely to relapse to regular opioid use. The implication is that even if patients decrease their use of opioids prior to treatment engagement, their risk of relapse is still substantially higher compared to those who start with lower levels of use or who have more sporadic usage patterns. MBA also showed that there were subgroups among opioid users and one of the subgroups who also consumed heroin, cocaine, and cannabis, was more likely to have a potentially worse outcome—consecutive non-attendance at scheduled clinic visits (i.e., missed UDSs) for four weeks between 21 days after randomization and the end of the 12-week treatment period. These results demonstrate that there are quantifiably different subgroups among polydrug users associated with clinically meaningful differences in treatment outcomes.

These results should be considered in light of several limitations. Measures of substance use are self-reported and may be subject to error because of social desirability and/or recall bias. Comparing the two models, MBA is a marginal model across time and therefore does not allow conclusions about the timing of the use of different drug combinations. In contrast, RMLCA specifically clusters the data on time patterns. This may be useful in evaluating predictors of treatment success due to the importance of a period of abstinence prior to induction on extended-release naltrexone and to a lesser extent, buprenorphine. Strengths of our study include the use of geographically diverse sites and large samples from three randomized clinical trials. In addition, we were able to establish a temporal relationship between identified RMLCA classes and MBA groups before treatment initiation and relapse outcomes after treatment initiation. Observing the pattern of polysubstance use prior to initiation of treatment may facilitate early identification of patients with unique needs and different probabilities of responding to treatment. Our ongoing work, using both traditional and modern machine learning methods to predict response to MOUD, will use these features to help identify who is best suited for different kinds of treatment.

Our data harmonization efforts highlight the need to standardize Timeline Followback questions. The fact that CTN-0030 did not gather data on unspecified, non-opioid substances was a missed opportunity to assess details of how other unexpected drugs may impact OUD. Differences in the breadth of drug categories, grouping hallucinogens with MDMA versus listing individual drugs as structured questions versus free text, may also limit the interpretability and clinical usefulness of our results. Furthermore, inconsistencies in the way in which substances are described or entered as free text (leading to more than a dozen spellings of “street buprenorphine” and half a dozen spellings of Adderall), and instances where respondents reported a set of drugs as a single free text “event” suggests a need for additional tools to support uniform Timeline Followback data collection. Future directions include tools that would automatically map clinical datasets, such as those obtained from medical records, to structured datasets such as Timeline Followback, providing a venue to replicate our findings on larger observational datasets.

In summary, this study suggests the presence of subgroups with distinct patterns of polysubstance use among individuals with OUD prior to MOUD treatment initiation. We reported the most prevalent polysubstance use combinations and identified six subgroups with different substance use behaviors during the 28 days leading up to MOUD initiation as well as associations between the effects of RMLCA classes and MBA top lift baskets to predict relapse in treatment. Our results suggest that patterns of substance use prior to treatment initiation may be useful in tailoring OUD interventions. Future studies should replicate these findings in larger datasets and different treatment contexts.

1. Introduction

Opioid Use Disorder (OUD) is a major public health issue, and its challenges are made worse by the use of multiple substances, known as polysubstance use. Most individuals with OUD use more than one substance. Studies show that between 65% and 85% of people in OUD treatment use multiple drugs. While some overdoses happen when substances are accidentally mixed, others result from intentionally using several drugs. However, the specific ways different drugs are used together are not well understood. For instance, opioids and benzodiazepines are often used together, which can lead to overdoses. Less is known about other combinations, such as opioids with marijuana or alcohol, and how these affect treatment success. More research is needed to understand which substances are used together and if these patterns change before a person enters treatment.

Researchers studying polysubstance use, including drugs like cannabis and tranquilizers, have used a method called Latent Class Analysis (LCA) to find patterns of drug use. A related method, Repeated Measures Latent Class Analysis (RMLCA), looks at how drug use patterns change over time. Understanding which drugs are used together and how these patterns evolve can help identify links with other health issues, such as depression and anxiety, which often occur in people who use many substances. This knowledge can then lead to more specific and effective treatments.

While methods like LCA are good for finding patterns when there are a limited number of drug types, modern machine learning tools like Market Basket Analysis (MBA) are better suited for identifying patterns when many different substances are involved. MBA has been used in various fields, such as finding gene patterns linked to diseases or identifying common food allergies. Its strength lies in finding patterns even when some drugs are used rarely among many possibilities, making it valuable for studying polysubstance use. Unlike RMLCA, MBA focuses on drug use patterns over shorter periods, such as within a single day.

Previous studies have shown how often polysubstance use occurs and which drugs are often used with opioids. However, few have examined the detailed patterns of drug use and how they change over time in individuals with OUD before they start treatment. This study explores daily substance use patterns, reported through the “Timeline Followback” method, during the 28 days before individuals began Medication for Opioid Use Disorder (MOUD) in three National Institute on Drug Abuse (NIDA) Clinical Trials Network (CTN) studies. Data from these studies were combined to allow for broader analysis of OUD treatment. Unlike earlier studies that focused on treatment outcomes, this research examines substance use patterns before treatment as potential predictors of relapse. The study uses both MBA and RMLCA: MBA helps identify combinations of individual drugs used on a single day, while RMLCA explores patterns over longer periods. Daily use reports across 12 broad substance categories over 28 days before MOUD were analyzed. The links between specific RMLCA patterns and relapse outcomes were also studied. The researchers predicted that MBA and RMLCA would identify different subgroups of polysubstance use, and that some of these patterns would be more linked to relapse to regular opioid use during treatment.

2. Material and methods

2.1. Study population

Details of the individual studies have been published before. In brief, all three studies recruited adults aged 18 or older who met the criteria for opioid dependence or Opioid Use Disorder (OUD) according to standard diagnostic manuals (DSM-IV-TR or DSM-5). The studies had broad inclusion criteria and few exclusions, mostly for serious medical or unstable mental health conditions. CTN-0027 was the most open, while CTN-0030 did not include individuals with OUD who only used heroin, and CTN-0051 excluded those already on methadone. CTN-0027 assigned participants to buprenorphine and methadone for 24 weeks of outpatient treatment. CTN-0030 assigned participants to different psychotherapy programs, with everyone receiving the same medication treatment in two parts: an initial buprenorphine taper, followed by buprenorphine maintenance for 12 weeks for those who relapsed (about 90%). CTN-0051 assigned patients from inpatient settings to buprenorphine or extended-release naltrexone after discharge, with outpatient follow-up for 24 weeks.

2.2. Substance use measures

The analysis focused on polysubstance use before treatment as a way to predict treatment success. For all study participants, substance use history for the 28 days before treatment began was gathered using the Timeline Followback method. Drug names entered as free text were standardized to correct spelling and name differences (e.g., generic, brand, or street names), and to remove extra details like dosage. This process resulted in a list of 44 unique substances. This data included frequently used drugs, as well as less common substances like cathinones, gabapentin, antiemetics, non-opioid pain killers, and antipsychotics. These specific substance names were used for Market Basket Analysis (MBA). For Latent Class Analysis (LCA), which works best with fewer categories, these harmonized drug names were grouped into twelve broader "drug categories." Table 1 shows how the drugs from the structured questions in the three studies were categorized for LCA. Other frequently reported drugs, such as fentanyl and opium, were grouped with heroin, and so on for other categories like depressants, synthetic cannabinoids, opioids, hallucinogens, and relaxants.

2.3. Relapse

Relapse was defined as four consecutive weeks of opioid use between 21 days after treatment started and the end of the 12-week treatment. An opioid use week meant either a urine drug screen (UDS) positive for any non-prescribed opioid, or a UDS that was missing or refused during that week. Relapse was categorized into three outcomes: 0 meant no relapse (50% of participants), 1 meant at least one opioid-positive UDS during four consecutive weeks of positive or missing UDSs (meaning the participant was positive for non-prescribed opioids but attended the clinic at least once; 29% of participants), and 2 meant missing UDS data for four consecutive weeks (21% of participants). This definition simplified an earlier, more complex one by relying only on UDS results, not self-reported drug use.

2.4. Statistical analysis

Market Basket Analysis (MBA) was initially developed to help businesses make decisions by analyzing large datasets, such as all transactions in a supermarket over a month. In such cases, traditional statistical tests with p-values are not very useful because, with enough data, almost anything can appear statistically significant. Instead, MBA uses three measures to determine the importance of a relationship: lift, support, and confidence. These measures assess the likelihood of two drugs (or groups of drugs) appearing together on a drug screening on the same day. Lift shows if an association exists; it is calculated by dividing the actual probability of two drugs appearing together by the probability of each drug appearing independently. This indicates whether the co-occurrence of the drugs is more or less likely than random chance. Support refers to the probability of seeing a specific combination of drugs together on the same day. As the number of possible drugs increases, support becomes less useful. Confidence describes the probability of seeing a set of drugs appear, given that another drug has already appeared. This is the conditional probability of seeing drug B if drug A is present.

These concepts allow for creating predictive rules, where the presence of drug A predicts the presence of drug B (written as A -> B). The relationship between lift and confidence is important, as the confidence measure is directional; for example, the likelihood of seeing drug B after drug A is observed is not necessarily the same as seeing drug A after drug B. This directional property allows MBA to predict a significantly increased risk of drug B if drug A has been used, even if drug B's use does not increase the chance of drug A appearing. For MBA, drugs that appeared very infrequently (less than twice a month) were excluded. The analysis then identified all rules with a confidence of at least 0.5. The top ten rules, based on their 'lift' value (indicating how much more likely drugs are to appear together than by chance), were selected. A binary category was then created to identify participants who used all substances in a specific drug combination. Repeated Measures Latent Class Analysis (RMLCA) was performed on the drug groups described earlier, covering the 28 days before MOUD initiation. First, the number of distinct patterns (classes) was determined by testing models with increasing numbers of classes. The process stopped when classes became too small for meaningful interpretation or when the model fit worsened. To ensure reliable results, the models were run multiple times with different starting points. Statistical measures were used to identify the best-fitting model. Once the optimal class structures were found, a three-step method was used to examine how these classes differed in relation to relapse, ensuring that adding these factors did not alter the core class patterns. Logistic regressions were used to assess the impact of RMLCA classes and MBA’s top combinations on relapse, while accounting for the specific study trial and treatment group. The results were reported as odds ratios and 95% confidence intervals. A p-value less than 0.05 was considered statistically significant. Data from the three clinical trials were standardized using SAS software. Exploratory data analysis and MBA were performed using R software with specific packages. RMLCA analyses were carried out using Mplus software.

3. Results

3.1. Parent study and participants

This study included all individuals who provided drug use information via the Timeline Followback method from CTN-0027 (N=1300), CTN-0030 (N=661), and CTN-0051 (N=676). Most participants provided complete drug use history for the 28 days before starting MOUD, with 97.9% providing over three weeks of data. Even though the three trials had different selection criteria and data available, Table 2 indicates similar patterns in participants' demographic and mental health characteristics at the start of the study.

3.2. Market basket

Out of 2637 participants, at least one drug was reported on 66,098 days (89%) during the 28 days before starting MOUD. Figure 1 illustrates the ten most frequently used substances. Heroin was the most common (reported on 40,740 days), followed by oxycodone (12,289 days), cannabis (9,183 days), hydrocodone (7,336 days), and methadone (3,663 days). On nearly two-thirds (64%) of the drug use days, only one substance was used, with an average of 1.45 drugs used per day. Participants used two substances on 29% of days, three substances on 6% of days, and four substances on 6% of days. One individual reported using 13 substances on a single day. The Market Basket Analysis (MBA) algorithm identified 34 association rules. As shown in Table 3, these rules provided very strong predictions, even for rare events; for instance, opium use was associated with a 16-fold increased chance of also seeing heroin use. Strong predictions also appeared for more common events, such as heroin use with crack or cocaine. The algorithm also highlighted combinations with significant effects, like a 13-fold increase in cocaine use when heavy drinking, opioid, and cannabis were present together. Most of the rules (76%) predicted heroin use following other drugs, while 18% were linked to cannabis use.

3.3. RMLCA

Repeated Measures Latent Class Analysis (RMLCA) models, ranging from three to eight classes, were applied to the combined data. Fit indices for each model are presented in Appendix Table A1. Statistical tests suggested that a six-class solution was a good fit, indicating that seven classes were not necessary. However, another test suggested seven classes. Considering the significant improvement in model fit from five to six classes, and that the seven-class solution included some very small groups, the six-class solution was chosen (Table A1). Figures 2 and 3 show the daily probability and overall prevalence of drug use for the six-class solution. More detailed figures are available in the appendix (Figures A1-A6), broken down by study. The largest group, Class 1 ("All time low," n=829, 31.4%), consisted of patients who used few substances, including opioids, in the 28 days before starting treatment. This group mainly included participants from CTN-0030 (n=466), CTN-0051 (n=208), and CTN-0027 (n=155). The second largest group, Class 2 ("High opioid decreasing," n=824, 31.3%), included individuals who had reduced their prescription opioid use in the 28 days before entering treatment. This group likely represents those who reduced opioid use on their own in preparation for inpatient or outpatient treatment. As shown in the figures, across all groups except the low substance use group, there was a sharp decrease in substance use in the two weeks before treatment initiation, especially for heroin, methadone, opioids, and cannabis. However, the extent of reduction varied depending on the drug.

3.4. Association between relapse and top 10 lift baskets

The ten most impactful rules from the Market Basket Analysis (MBA) were selected (Table 4). Individuals using a combination of cocaine, heroin, other opioids, and cannabis had a higher risk of relapse. This relapse was defined either by at least one opioid-positive urine drug screen (UDS) or by consistently missing UDS tests for four weeks. This risk was higher compared to individuals not reporting this specific drug combination, even after accounting for other polysubstance use patterns and the study trial. In contrast, individuals reporting cocaine, heavy drinking, cannabis, and other opioids had a lower risk of relapse defined by four consecutive missing UDS tests. Similarly, a decreased chance of relapse was observed for individuals who reported using benzodiazepines, heroin, other opioids, and cannabis together. Appendix Table 2 lists the number of participants exhibiting the polysubstance use patterns identified by the top ten MBA rules.

4. Discussion

This study offers new insights into polysubstance use patterns among individuals with Opioid Use Disorder (OUD) just before they begin treatment. Polysubstance use was found to be common, and both Repeated Measures Latent Class Analysis (RMLCA) and Market Basket Analysis (MBA) provided valuable understanding of these patterns. RMLCA, a person-centered method, helps identify hidden patterns in how people respond to different substance use categories over time, determining the number and prevalence of these patterns. This method is more formal than other data grouping techniques, allowing for a probabilistic approach to clustering data and estimating parameters. MBA, on the other hand, effectively reveals the most frequent combinations of substances used on the same day, offering detailed insights that might be missed by broader drug classifications used in other analyses.

The findings suggest that individuals who used opium also frequently used heroin, alcohol, and cannabis. Additionally, certain drug uses, such as benzodiazepines, cocaine, and cannabis, often preceded heroin use. RMLCA identified six main categories of polysubstance use patterns before treatment. A significant portion of patients either used very few substances, including opioids, or had greatly reduced their substance use before starting treatment. Both RMLCA and MBA methods highlighted specific subgroups of polysubstance use patterns linked to different relapse outcomes.

Unexpectedly, certain Market Basket Analysis combinations showed varied links to relapse. For example, using cocaine, cannabis, and heavy alcohol with prescription opioids was linked to a decreased likelihood of relapse, rather than an increase. However, the same combination, when used with heroin, strongly predicted an increased risk of relapse. This indicates that not all polysubstance "binges" automatically lead to relapse or missed clinic visits. A polysubstance binge involving heroin suggests a poorer outlook, while one with prescription opioids appears less severe. Importantly, the various RMLCA classes and MBA polysubstance groups had distinct relationships with relapse. RMLCA showed that individuals in the "high opioid decreasing" and "high heroin decreasing" groups were more prone to relapse to regular opioid use. This suggests that even a reduction in opioid use before treatment does not eliminate the higher relapse risk compared to those with consistently low or sporadic use. MBA also identified subgroups among opioid users, with one group that also used heroin, cocaine, and cannabis being more likely to experience poorer outcomes, specifically consecutive missed clinic visits. These findings demonstrate that different patterns of polysubstance use among individuals are associated with meaningful differences in treatment outcomes.

This study has several limitations. Substance use data were self-reported, which means they might be affected by social desirability or memory bias. Also, MBA, unlike RMLCA, does not provide information about the timing of drug combinations. However, strengths of the study include its use of large, diverse samples from three clinical trials, and its ability to establish a timeline between pre-treatment polysubstance patterns and later relapse outcomes. Understanding pre-treatment polysubstance patterns could help identify patients with specific needs and predict their response to treatment. Efforts to standardize data collection, especially for the Timeline Followback method, are crucial for future research, as inconsistencies in reporting drug names or categories can limit the usefulness of results. In conclusion, this study identifies distinct subgroups of polysubstance use among individuals with OUD before they begin MOUD treatment. The research highlights common polysubstance combinations and six behavioral subgroups leading up to treatment, demonstrating how these patterns are linked to relapse. These findings suggest that knowing a person's substance use patterns before treatment could help tailor more effective OUD interventions. Future research should aim to confirm these findings using larger datasets and in different treatment settings.

1. Introduction

The significant challenges of opioid use disorder (OUD) are often made worse by the use of multiple substances. While exact figures vary, recent studies suggest that using more than one substance is typical for individuals with OUD. Rates of polysubstance use among those in treatment for OUD often range from 65% to 85%. Some overdoses happen when people unknowingly consume multiple substances, like fentanyl mixed with heroin. However, other individuals intentionally seek out and use various drugs together. Despite this, there is limited understanding of the specific patterns in which different types of drugs are used.

Combinations such as opioids and benzodiazepines are frequently used together and are known to contribute to overdose incidents. Less is known about other combinations, such as opioids with marijuana or alcohol, and how they affect treatment success. There is a clear need to better understand which substances are used together and whether these patterns change before someone enters treatment.

Researchers studying polysubstance use, including substances like cannabis and tranquilizers, have previously used a statistical method called Latent Class Analysis (LCA) to identify drug use patterns. An advanced version, repeated measures latent class analysis (RMLCA), examines drug exposure patterns over time and how they group together. Knowing which drugs are used together and how these patterns evolve can help identify concurrent conditions, such as depression and anxiety often seen in individuals who use a wide range of substances. This understanding could, in turn, lead to more targeted and effective treatments.

While methods like LCA are useful for finding patterns, especially with a limited number of drug types, modern machine learning approaches such as market basket analysis (MBA) are very good at identifying patterns when many different substances are involved. This method has been applied in various fields, including detecting gene patterns related to diseases, identifying co-occurring food allergies, and even analyzing tactical plays in sports. MBA’s ability to find patterns, even when some drugs are used rarely from a large pool of possibilities, makes it valuable for studying polysubstance use. Unlike RMLCA, MBA focuses on polysubstance use over short periods, such as within the same day.

Although studies have documented the overall rates of polysubstance use and commonly co-used drugs with opioids, few have deeply explored the detailed patterns of drug use and how substance use changes over time in individuals with OUD before they start treatment. This study examines the daily substance use patterns, as reported over 28 days before starting Medication for Opioid Use Disorder (MOUD), from three large clinical trials. The data from these trials were combined to allow for more extensive secondary analyses of OUD treatment. The study used both MBA and RMLCA to understand drug use patterns, with MBA identifying combinations of individual drugs used on a given day and RMLCA exploring patterns over longer periods. The study looked at daily reports across 12 broad categories of substances over the 28 days before MOUD began. It also investigated how selected RMLCA patterns were linked to relapse outcomes. Researchers hypothesized that MBA and RMLCA could identify distinct subgroups of polysubstance use with different frequencies, and that some of these patterns would be more strongly associated with relapse to regular opioid use during treatment.

2. Material and methods

This study included individuals who met the criteria for opioid dependence or OUD from three different National Institute on Drug Abuse (NIDA) Clinical Trials Network (CTN) studies (CTN-0027, CTN-0030, CTN-0051). All participants were adults over 18, with broad inclusion criteria and few exclusions, mostly for severe medical or unstable mental health conditions. Each trial had slightly different specifics; for example, CTN-0027 enrolled a wide range of participants, CTN-0030 excluded those who only used heroin, and CTN-0051 excluded individuals already receiving methadone. The trials involved different treatment approaches, such as buprenorphine and methadone maintenance, psychotherapy with buprenorphine taper, or buprenorphine/extended-release naltrexone after inpatient treatment.

The analysis focused on polysubstance use patterns in the 28 days before participants began treatment, aiming to use this information as a predictor for treatment outcomes. Substance use history was collected using the Timeline Followback method. The reported drug names, initially provided as free text, were standardized to correct spelling errors and remove extra details, resulting in 44 distinct substances. These detailed substance reports were used for the Market Basket Analysis (MBA). For Repeated Measures Latent Class Analysis (RMLCA), these substances were grouped into 12 broader categories to make the analysis more manageable.

Relapse was defined as four consecutive weeks of opioid use between 21 days after treatment began and the end of the 12-week treatment period. An opioid use week was marked by either a positive urine drug screen (UDS) for a non-prescribed opioid or a missing/refused UDS. Relapse was categorized into three outcomes: not meeting the definition of relapse, having at least one opioid-positive UDS during four consecutive weeks, or having missing UDS data for four consecutive weeks.

For the statistical analysis, Market Basket Analysis (MBA) was employed. This method was originally developed for business decisions using large datasets and assesses the importance of associations between items using three statistics: lift, support, and confidence. Lift indicates if an association exists and its strength (how much more or less likely drugs are to appear together than expected by chance). Support represents the overall probability of seeing drugs occur together. Confidence describes the conditional probability of one drug appearing given that another has already occurred. This directional property allows MBA to identify predictive associations. For MBA, drugs occurring less than 0.001 (rarely) were excluded, and the algorithm focused on rules with at least 0.5 confidence, selecting the top ten rules based on lift.

Repeated Measures Latent Class Analysis (RMLCA) was used to analyze drug use patterns over the 28 days leading up to MOUD initiation, based on the grouped drug categories. The process involved identifying the optimal number of classes that best fit the data, using various statistical criteria to ensure reliable class structures. Once the best-fitting classes were determined, a three-step procedure was used to examine differences in relapse outcomes between these classes, ensuring that the inclusion of outcome variables did not alter the class structure itself. Logistic regressions were then used to test the effects of RMLCA classes and MBA top lift baskets on relapse, controlling for the specific clinical trial and treatment arm.

3. Results

The study included 2,637 participants who provided substance use information through the Timeline Followback method. Participants generally provided complete drug use details for the 28 days before starting Medication for Opioid Use Disorder (MOUD). Despite some differences in participant selection and data availability across the three trials, baseline demographics and psychiatric conditions were largely similar among participants.

Participants reported using at least one substance on 89% of the days in the 28-day period leading up to MOUD initiation. The most frequently used substances were heroin, oxycodone, cannabis, hydrocodone, and methadone. On most drug use days (64%), only a single substance was used. However, participants used two substances on 29% of days, three substances on 6% of days, and four substances on another 6% of days, with one individual reporting the use of 13 substances on a single day.

The Market Basket Analysis (MBA) identified 34 strong association rules. These rules showed very strong predictions for less common events; for example, opium use was strongly associated with heroin use. Predictions were also significant for more common combinations, such as heroin used with crack or cocaine. The algorithm also highlighted combinations with large effects, like a significant increase in the likelihood of cocaine use when heavy drinking, opioid, and cannabis use were reported together, or an increased likelihood of cannabis use when cocaine, heavy drinking, and an opioid were combined. Most of the rules (76%) predicted heroin use following the consumption of other drugs, while 18% of the rules were linked to cannabis use.

Repeated Measures Latent Class Analysis (RMLCA) identified six distinct patterns of polysubstance use across the harmonized data. The largest pattern, labeled "All time low," included 31.4% of patients who reported using few substances, including opioids, in the 28 days before treatment. The second largest pattern, "High opioid decreasing," accounted for 31.3% of participants. This group consisted of individuals who had reduced their prescription opioid use in the month before treatment. Across most patterns, excluding the "low substance use" group, there was a sharp decline in substance use, especially for heroin, methadone, prescription opioids, and cannabis, in the two weeks leading up to treatment initiation. However, the extent of this reduction varied by drug type.

The study examined the association between the top 10 Market Basket (MBA) patterns and relapse outcomes. Individuals who used a combination of cocaine, heroin, opioids, and cannabis had a higher risk of relapse, defined either by a positive opioid urine drug screen or by consistently missing drug screens. In contrast, individuals who reported using cocaine, heavy drinking, cannabis, and opioids showed a lower risk of relapse defined by missing urine drug screens. Similarly, a decreased risk of relapse was observed for individuals who reported using benzodiazepines, heroin, opioids, and cannabis together.

4. Discussion

This study offers new insights into the patterns of polysubstance use immediately before treatment initiation in individuals with opioid use disorder (OUD). Polysubstance use was common, and both Repeated Measures Latent Class Analysis (RMLCA) and Market Basket Analysis (MBA) provided valuable understanding of these specific patterns. RMLCA, a person-centered statistical technique, helps identify hidden patterns in categorical responses over time, revealing how many behavior patterns exist, how common they are, and the likelihood of specific behaviors within each pattern. MBA, on the other hand, excels at revealing the most frequent combinations of substances used on the same day, offering detailed insights that broader drug categories might obscure. The findings indicate that, for example, individuals using opium were also typically using heroin, alcohol, and cannabis. The analysis also showed several one-way relationships where the use of certain drugs, such as benzodiazepines, cocaine, and cannabis, often preceded heroin use.

An unexpected finding was the association between certain MBA patterns and relapse by week 12. For instance, the combination of cocaine, cannabis, and heavy drinking with prescription opioids did not negatively impact the odds of eventual relapse; instead, it was correlated with a decreased risk of relapse. In contrast, the same combination, but with heroin, predicted a notable increase in relapse risk. This suggests that a polysubstance "binge" is not automatically linked to relapse to regular opioid use or missing clinic appointments. In fact, a polysubstance heroin binge may indicate a worse prognosis, whereas a similar binge involving prescription opioids may suggest a less severe outcome.

Most importantly, the various RMLCA classes and MBA polysubstance use groups showed different relationships with relapse. For example, RMLCA revealed that individuals in both the "high opioid decreasing" and "high heroin decreasing" classes were more likely to relapse to regular opioid use. This implies that even if patients reduce their opioid use before treatment, their relapse risk remains significantly higher compared to those who start with lower or more sporadic usage patterns. MBA also identified subgroups among opioid users, with one subgroup, which also consumed heroin, cocaine, and cannabis, having a greater likelihood of adverse outcomes, specifically consecutive non-attendance at scheduled clinic visits. These results demonstrate the existence of distinct subgroups among polysubstance users, associated with clinically significant differences in treatment outcomes.

These findings should be considered alongside several limitations. Substance use data are self-reported, which may introduce errors due to social desirability or recall bias. Comparing the two models, MBA focuses on patterns at a single point in time and does not provide insights into the timing of different drug combinations. RMLCA, however, specifically clusters data based on temporal patterns, which can be valuable for predicting treatment success, particularly given the importance of a period of abstinence before certain treatments. Strengths of this study include the use of data from diverse geographic locations and large samples from three randomized clinical trials. Furthermore, the study established a temporal link between identified RMLCA classes and MBA groups before treatment initiation and relapse outcomes after treatment. Understanding polysubstance use patterns before treatment could help identify patients with unique needs and different probabilities of responding to treatment earlier. Ongoing research using traditional and machine learning methods aims to use these patterns to determine which treatments are best suited for different individuals.

The efforts to combine data from different studies highlight the need for standardized questions in methods like the Timeline Followback. Inconsistencies in how substances were described or entered as free text, and instances where multiple drugs were reported as a single event, suggest a need for better tools to support uniform data collection. Future work should focus on tools that can automatically map clinical data from medical records to structured datasets, allowing for replication of these findings in larger observational studies. Overall, this study suggests the presence of distinct subgroups with different patterns of polysubstance use among individuals with OUD prior to Medication for Opioid Use Disorder (MOUD) treatment initiation. The research identified prevalent polysubstance use combinations and revealed subgroups with distinct substance use behaviors and associations with relapse during treatment. The results indicate that patterns of substance use before treatment may be useful in customizing OUD interventions. Future studies are encouraged to replicate these findings in larger datasets and in various treatment settings.

Introduction

The significant impact of opioid use disorder (OUD) becomes even more complex when individuals also use multiple other substances. While the exact percentage of people with OUD who use more than one substance varies, recent studies suggest that using multiple substances is common among those with OUD. Rates of polysubstance use in people undergoing OUD treatment range from 65% to 85%.

Some overdoses happen when people accidentally take a mix of substances, such as carfentanil with heroin. However, other individuals intentionally seek out and use multiple drugs. Despite this, little is known about the specific ways different types of drugs are used together. Certain combinations, like opioids and benzodiazepines, are frequently used and likely contribute to overdose deaths. Less is known about other combinations, such as opioids with marijuana or alcohol, and their effects on treatment success are unclear. There is a need to better describe which substances are used together and whether these patterns change before a person enters treatment.

Researchers studying polysubstance use, especially with drugs like cannabis and tranquilizers, have used a common statistical method called Latent Class Analysis (LCA) to find patterns in drug use. A more advanced version, repeated measures latent class analysis (RMLCA), looks at repeated drug use patterns and how they group together over time. Understanding which drugs are used together and how these patterns change allows for the identification of patterns in other co-occurring health issues, such as depression and anxiety, which are often seen in people who use a wide range of substances. Understanding these patterns may lead to more focused treatments.

While methods like LCA are good for finding such patterns, especially when dealing with a small number of drug types (e.g., 6 to 12 drugs), modern machine learning methods, such as market basket analysis (MBA), are very good at identifying patterns when many different substances are involved. This method has been used to find patterns in many areas, such as identifying genes related to diseases, co-occurring food allergies, and even tactical patterns in professional beach volleyball. This ability to find patterns, even for drugs that are rarely used among many possibilities, makes MBA a useful tool for studying polysubstance use. Unlike RMLCA, MBA focuses on polysubstance use over short periods, such as on the same day.

Although studies have reported on the rate of polysubstance use and drugs frequently used with opioids, few have looked closely at the specific details of drug use patterns and how substance use changes over time in people with OUD before they start treatment. This study examines daily patterns of self-reported substance use over 28 days before individuals began Medication for Opioid Use Disorder (MOUD) in three large research studies. Data from these studies were combined to allow for more in-depth analysis of OUD treatment. Unlike previous studies that looked at how treatment outcomes change over time, this study focused on modeling substance use patterns before treatment as factors that might predict relapse. The study reports drug use patterns using both MBA and RMLCA. MBA is useful for finding patterns of all individual drugs used on any given day, while RMLCA allows for exploring patterns over longer periods. Daily reports of use across 12 broad categories of substances were examined over the 28 days before MOUD initiation. The study also looked at connections between selected RMLCA classes and relapse outcomes. It was hypothesized that subgroups of polysubstance use, with different levels of occurrence, could be identified by MBA and RMLCA, and that some of these patterns would be more likely than others to be linked to relapse to regular opioid use during treatment.

Material and Methods

Study Population

Complete details for the individual studies have been previously published. In short, all three studies included individuals who met the diagnostic criteria for opioid dependence or opioid use disorder. The trials enrolled adult participants over 18 years old, with broad inclusion criteria and few exclusions, except for major medical conditions and unstable mental health issues. One study was very inclusive, while another excluded individuals with OUD who only used heroin. The third study excluded individuals currently receiving methadone treatment. Participants in one study were randomly assigned to buprenorphine and methadone for outpatient treatment for 24 weeks. Another study randomly assigned individuals to two different types of therapy programs, and all participants received the same medication treatment in two phases: an initial buprenorphine taper, followed by buprenorphine maintenance for 12 weeks for those who relapsed (about 90% of participants). A third study randomly assigned patients from inpatient treatment facilities to receive either buprenorphine or extended-release naltrexone after discharge and were followed as outpatients for 24 weeks.

Substance Use Measures

The analysis focused on polysubstance use before treatment began as a baseline predictor of treatment outcomes. For all participants in the three studies mentioned above, their substance use history in the 28 days before starting treatment was surveyed using the Timeline Followback method. The self-reported drug names were standardized to correct spelling errors and differences in names (e.g., generic, brand, or street names) and to remove unnecessary information (e.g., mg/capsule/oz). This resulted in 44 distinct substances. This data included frequently used drugs, which were also identified in structured questions, as well as less frequently used substances (e.g., cathinones, gabapentin), and other drugs like antiemetics, non-opioid pain killers, and antipsychotics. These values were used for MBA. To make the LCA analysis more manageable, these standardized drug names were grouped into twelve "drug categories." Table 1 lists the drugs surveyed as structured questions in the three studies and how they were grouped for LCA. Other drugs that appeared repeatedly included fentanyl and opium (grouped with/classified as Heroin for LCA); GHB (grouped as Depressants); K2 (Synthetic cannabinoids); meperidine, tramadol, oxymorphone (grouped as Opioids); barbiturate, sedative-hypnotic (grouped as Depressants); MDMA, unspecified hallucinogens (grouped as Hallucinogens); and various muscle relaxants (grouped as Relaxant).

Table 1. Structured drug use variables from three harmonized studies and how they were categorized for repeated measure latent class analyses.

Relapse

Relapse was defined as four consecutive weeks of opioid use between 21 days after treatment began and the end of the 12-week treatment period. An opioid use week meant having a urine drug screen (UDS) positive for any non-prescribed opioid, or a missing/refused UDS during that week. Relapse was defined as a categorical variable with three possible outcomes: 0 indicated that the relapse definition was not met (50% of participants), 1 indicated at least one opioid-positive UDS during four consecutive weeks with positive or missing UDSs (meaning the participant, despite testing positive for non-prescribed opioids, showed up to the clinic at least once; 29% of participants), and 2 indicated missing UDS data for four consecutive weeks (21% of participants). This definition was similar to a more complex one from a previous study but only required UDS results, not the self-reported drug use information from the Timeline Followback.

Statistical Analysis

Market Basket Analysis (MBA) was originally designed to help businesses make decisions using huge amounts of data, such as all transactions in a supermarket for a month. In these situations, traditional hypothesis testing with p-values becomes unhelpful because anything can be statistically significant with a large enough sample size. Instead of p-values, MBA methods use three measures to assess the importance of a connection: lift, support, and confidence. These measures evaluate the likelihood of seeing drug A and drug B (or a set of drugs labeled A and a different set labeled B) appearing on a drug screen on the same day.

Lift assesses if an association exists. It is the ratio of the actual probability of drug A and drug B occurring on a given day, divided by the probability of seeing drug A on any day multiplied by the probability of seeing drug B on any day. In simpler terms, it shows the increased or decreased chance of seeing both drugs together compared to their overall likelihood of appearing separately (assuming their co-occurrence is random). Support refers to the probability of seeing the various sets of drugs being examined. MBA users focus on the support for A and B, which is the probability of seeing both items on the same day. Support becomes less useful as the number of possible drugs in the set increases. Confidence describes the probability that a set of items appears given that one has already occurred. It is the conditional probability of seeing drug B given that drug A is present.

Using these ideas, it is possible to think of predictive association rules where A leads to B (written as A -> B). It is helpful to consider the relationship between lift and confidence: lift(A→B) = P(B|A) / P(B). Importantly, A and B are not interchangeable in the confidence calculations. For example, the chance of seeing chips after seeing guacamole (guacamole -> chips) is not the same as the chance of seeing guacamole after seeing chips (chips -> guacamole). This "directional" property allows MBA to make predictions that can describe a greatly increased risk of drug B if a person has used drug A, but seeing drug B may not increase the chances of seeing drug A.

For MBA, the a priori algorithm was used after excluding drugs that appeared very rarely (less than 0.001 support). This means drugs that were used less than twice a month per transaction were excluded. The algorithm was set to find all rules with a confidence of at least 0.5. The top ten rules were then selected based on their "lift" (the ratio of the likelihood of the first drugs appearing with the second drugs, divided by the likelihood if they occurred independently). A binary category was created to show a specific "basket" of polydrug use. Participants who reported using all substances in a particular basket rule during the study period were identified and grouped into different polydrug baskets.

RMLCA for the drugs used in the 28 days leading up to MOUD initiation was performed using the drug groups described earlier. First, the number of classes was determined without considering other factors by estimating models with an increasing number of classes until the sample size in each class was too small for practical interpretation (less than 5% of the total sample size) or until information criteria indicated a worse model fit. To ensure the models found the best possible solution, 1000 random starts and 100 replicated likelihoods were used for each model. Model fit statistics for each model were used to determine the model that best fit the data, including entropy and penalized information criteria. Second, once the best-fitting class structures were found, a three-step approach was used to examine differences in relapse rates between classes using multinomial logistic regression. This three-step process ensures that adding other factors does not change the basic class structure. Logistic regressions were used to test the effects of RMLCA classes and MBA top lift baskets on relapse, while accounting for which of the three trials and which treatment arm within the trials. Odds ratios (OR) and 95% confidence intervals (CIs) were reported. P-values less than 0.05 for two-sided tests were considered statistically significant. Demographic and substance use data from the three studies were standardized and analyzed. Exploratory data analyses and MBA were conducted with R statistical software. RMLCA analyses were conducted in Mplus software.

Results

Parent Study and Participants

All individuals who provided drug use information from the three studies (a total of 2637 participants) were included in this study. Participants typically provided complete drug use information for the 28 days before starting MOUD, with 97.9% providing more than three weeks of history. Despite different participant selection criteria and data availability across the three trials, Table 2 shows similar patterns in terms of baseline demographics and mental health conditions.

Table 2. Demographic details on 2637 participants from three clinical trials, who self-reported drug use in the 28 days before initiation of medication for opioid use disorder and randomization.

Market Basket

The 2637 participants reported using at least one drug on 66,098 days (89% of days) in the 28 days before MOUD initiation and randomization. Fig. 1 shows the top 10 most frequently used substances. The most common substances were heroin (reported on 40,740 days), oxycodone (12,289 days), cannabis (9,183 days), hydrocodone (7,336 days), and methadone (3,663 days). Almost two-thirds (64%) of drug use days involved using a single substance (the average number of drugs used per day was 1.45). Participants used two substances on 29% of days, three substances on 6% of days, and four substances on 6% of drug use days. One person reported using 13 substances on a single day.

Fig. 1.

Substance use frequency (Top 10).

The a priori MBA algorithm identified 34 association rules. As shown in Table 3, these rules included very strong predictions for rare events; for example, a 16-fold increase (with 71% confidence) in the chance of seeing heroin with the use of opium (lift = 16, count = 69). Predictions were also strong for more common events, such as using heroin with crack (lift = 1.46, count = 2884) or cocaine (lift = 1.35, count = 2598). The algorithm noted several combinations with large effects, such as a 13-fold increase in the chances of reported cocaine use when a combination of heavy drinking, opioid, and cannabis was observed, and a 6.13-fold increase in the chances of cannabis in a combination of cocaine, heavy drinking, and opioid. While most of the rules (76%) predicted heroin use after consuming other drugs, 18% of the rules were associated with cannabis use.

Table 3. Market basket analysis result, sorted by lift, shows which drug is predicted (C is consequent) given the presence of other drugs (+ shows antecedents) which were used on the same day.

RMLCA

RMLCA with three to eight classes was applied to all the combined data. Fit indices for each LCA model are presented in Appendix Table A1. Statistical tests suggested that a six-class solution was a good fit, and that seven classes were not needed. However, another statistical test suggested seven classes. Given the significant improvement in model fit between the five- and six-class solutions, and the fact that the seven-class solution included some smaller groups, the six-class solution was chosen (Table A1). The daily probability and prevalence of drug use by latent class for the six-class solution are shown in Fig. 2 (by days) and Fig. 3 (by drug use). More detailed figures are provided in the appendix. The largest group was C1-"All time low" (829 participants, 31.4%), which describes a group of patients who used few substances, including opioids, in the 28 days before starting treatment. The second largest group was C2-"High opioid decreasing" (824 participants, 31.3%). This group included individuals who had voluntarily reduced their prescription opioid use in the 28 days before assessment for treatment entry. This group likely includes individuals who decreased opioid intake to prepare for either inpatient or outpatient treatment. As seen in the figures, across all groups except the low substance use group, there was a sharp decline in substance use in the two weeks leading up to treatment initiation, especially for heroin, methadone, other opioids, and cannabis use. However, the extent of this reduction varied by drug.

Fig. 2.

Six class solution by RMLCA class (Overall).

Fig. 3.

Six class solution by drug use (Overall).

Association Between Relapse and Top 10 Lift Baskets

The top 10 lift rules from the MBA were selected (Table 4). Individuals who used cocaine, heroin, opioids, and cannabis together had a higher risk of relapse, defined by at least one opioid-positive urine drug screen (Odds Ratio = 2.82) or by exclusively missing urine drug screens for four consecutive weeks (Odds Ratio = 5.13), compared to individuals who did not report using this specific drug combination, even when controlling for other polydrug combinations and the study type. Individuals who reported using a combination of cocaine, heavy drinking, cannabis, and opioids had a lower risk of relapse defined by four consecutive missing urine drug screens (Odds Ratio = 0.44). Similarly, a decreased chance of relapse was found for individuals who reported using benzodiazepines, heroin, opioids, and cannabis together (Odds Ratio = 0.58). The number of participants with polydrug patterns represented by the top 10 lift baskets is listed in Appendix Table 2.

Table 4. Association between relapse and Market Basket top 10 lift baskets.

Discussion

This study provides new information about patterns of polysubstance use in the period immediately before individuals with OUD began treatment. Using multiple substances was common, and both RMLCA and MBA offered valuable insights into specific patterns. RMLCA is a statistical technique that helps identify hidden patterns in how people respond to categorical items, with varying probabilities. Compared to other methods, RMLCA uses a formal statistical approach to create groups and can be used with other methods to estimate factors. It helps determine how many response or behavior patterns are present in the data, how common each pattern is, and how likely an item is to be endorsed within each hidden group. The best number of groups minimizes the connections between cases belonging to different groups. Methods like the Bayesian Information Criterion, which use the negative log-likelihood function (a measure of how well a statistical model fits the data), are used to select the best number of groups. MBA was used to reveal the most common combinations of substances used together on the same day, and it provides additional insights that are hidden when using broader drug categories needed for traditional LCA.

The results show, for example, that individuals who used opium also used heroin, alcohol, and cannabis. There were also several one-way relationships where the use of specific drugs (like benzodiazepines, cocaine, and cannabis) was likely to lead to heroin use. RMLCA identified six broad classes of polysubstance use patterns before treatment began, revealing that a large number of patients either used few substances overall, including opioids, or significantly reduced their use before starting treatment. Furthermore, both methods used in this analysis identified subgroups of polysubstance use patterns that were linked to relapse outcomes.

Some unexpected connections between specific MBA drug combinations and relapse by week 12 were found: the observation of cocaine, cannabis, and heavy drinking combined with prescription opioids did not negatively impact the odds of eventual relapse. Instead, this combination was linked to decreased odds of relapse. In contrast, the same combination with heroin predicted a clear increase in the odds of relapse. This suggests that a polysubstance "binge" is not automatically linked to relapse to regular opioid use or missing scheduled clinic visits. In fact, a polysubstance heroin binge suggests a worse outlook, while a polysubstance binge with prescription opioids suggests a much less severe outcome.

Most importantly, the different RMLCA classes and MBA polydrug use groups had different connections to relapse. For instance, RMLCA showed that people in both the "high opioid decreasing" and "high heroin decreasing" classes were more likely to relapse to regular opioid use. This suggests that even if patients reduce their opioid use before starting treatment, their risk of relapse remains significantly higher compared to those who start with lower levels of use or have more inconsistent usage patterns. MBA also showed that there were subgroups among opioid users, and one subgroup who also used heroin, cocaine, and cannabis was more likely to experience a potentially worse outcome—missing scheduled clinic visits for four consecutive weeks. These results show that there are measurable differences among polysubstance users that are associated with clinically meaningful differences in treatment outcomes.

These results should be considered in light of several limitations. Substance use measures are self-reported and may be inaccurate due to a desire to be seen favorably or problems remembering past events. When comparing the two models, MBA is a general model that doesn't consider timing, so it cannot draw conclusions about when different drug combinations were used. In contrast, RMLCA specifically groups data based on time patterns. This can be useful for predicting treatment success, especially due to the importance of a period of abstinence before starting certain medications like extended-release naltrexone and, to a lesser extent, buprenorphine. Strengths of this study include using sites from diverse geographic areas and large samples from three randomized clinical trials. In addition, a temporal relationship could be established between the RMLCA classes and MBA groups identified before treatment began and relapse outcomes after treatment started. Observing polysubstance use patterns before treatment initiation may help identify patients with unique needs and different likelihoods of responding to treatment early on. Ongoing work, using both traditional and modern machine learning methods to predict response to MOUD, will use these features to help determine who is best suited for different kinds of treatment.

Efforts to combine data in this study highlight the need to standardize questions asked during the Timeline Followback. The fact that one study did not collect data on unspecified, non-opioid substances was a missed opportunity to assess how other unexpected drugs might affect OUD. Differences in the breadth of drug categories, such as grouping hallucinogens with MDMA versus listing individual drugs as structured questions versus free text, may also limit how our results can be interpreted and their clinical usefulness. Furthermore, inconsistencies in how substances are described or entered as free text (leading to many spellings of "street buprenorphine" and Adderall), and instances where respondents reported a group of drugs as a single free text "event," suggest a need for better tools to support consistent Timeline Followback data collection. Future directions include developing tools that would automatically link clinical datasets, such as those from medical records, to structured datasets like Timeline Followback, providing a way to confirm these findings using larger observational datasets.

In summary, this study suggests the existence of distinct subgroups with different patterns of polysubstance use among individuals with OUD before starting MOUD treatment. The most common polysubstance use combinations were reported, and six subgroups were identified with different substance use behaviors during the 28 days leading up to MOUD initiation. The study also found connections between the effects of RMLCA classes and MBA top lift baskets in predicting relapse during treatment. The results suggest that patterns of substance use before treatment begins may be useful in customizing OUD interventions. Future studies should repeat these findings in larger datasets and in different treatment settings.

Introduction

When someone has opioid use disorder (OUD), it often gets harder because they also use many other drugs. This is called polysubstance use. Most people with OUD use more than one drug. Sometimes, people take drugs that are mixed without them knowing, like heroin with a strong pain medicine called carfentanil. Other times, people choose to use many drugs together. It is not always clear which drugs are often used at the same time. For example, opioids and medicines like benzodiazepines are often used together, and this can lead to overdoses. Not as much is known about using opioids with marijuana or alcohol, and how these mixes affect treatment. More information is needed about which drugs are used together and if how people use them changes before they start treatment.

Experts who study drug use patterns have a special way to find them. This method is called Latent Class Analysis, or LCA. There is also a version of LCA that looks at drug use patterns over time, called RMLCA. Learning these patterns helps show if other health problems, like sadness or worry, often happen with certain drug use. Understanding these patterns can then help doctors find better ways to treat people.

While LCA is helpful for looking at a few drugs, a newer method called Market Basket Analysis (MBA) works well when there are many drugs. MBA is used in many fields to find patterns, like which foods cause allergies or even how sports teams play. This method can find drug use patterns even if some drugs are not used very often. Unlike RMLCA, MBA focuses on drug use patterns over short times, like on the same day.

Many studies have shown how often people use more than one drug, but few have looked closely at the daily ways drugs are used and how this changes before treatment for OUD. This study looks at what drugs people reported using each day for 28 days before they started medicine for OUD (MOUD). Information from three different studies was put together for this work. Unlike other studies that looked at treatment results, this study looked at drug use patterns before treatment to see if they could show who might have problems later. We used MBA and RMLCA to find these patterns. MBA helps find patterns of single drugs used on the same day. RMLCA helps explore patterns over longer times. The study looked at 12 main groups of drugs over 28 days. It also checked if certain RMLCA patterns were linked to future relapse. The researchers believed they would find different groups of people based on their polysubstance use, and that some patterns would make relapse more likely during treatment.

Material and methods

This study included people from three past studies who had opioid use disorder. All participants were adults, over 18 years old. Most people could join these studies, unless they had serious health problems or very unstable mental health issues. One study (CTN-0027) gave people two different medicines, buprenorphine or methadone, for 24 weeks. Another study (CTN-0030) gave people talking therapy; people first received a medicine taper, and if they started using opioids again, they then received buprenorphine medicine for 12 weeks. The third study (CTN-0051) included people who had been in a hospital for treatment, and after leaving, they received either buprenorphine or naltrexone for 24 weeks.

For this study, the focus was on how people used many drugs together before they started treatment. Each person was asked to remember and report all the drugs they used each day for 28 days before starting treatment. This was done using a method called "Timeline Followback." The names of drugs were made uniform, so there were 44 different drug names. These 44 names were used for the MBA part of the study. For the RMLCA part, these drugs were put into 12 larger groups, like "opioids" or "cannabis."

Relapse meant that a person used opioids for four weeks in a row after starting treatment. An "opioid use week" meant that a urine test showed they used an opioid not given by a doctor, or they missed or refused to give a urine test that week. There were three types of outcomes for relapse: the person did not relapse (half of the people), the person had at least one positive urine test for opioids during four weeks in a row (about 29% of people), or the person missed all their urine tests for four weeks in a row (about 21% of people).

Market Basket Analysis (MBA) was first used in business to understand what items people buy together, like at a grocery store. Instead of standard statistical tests, MBA uses three main measures: "lift," "support," and "confidence." "Lift" shows how much more likely it is to see two drugs together than if they just happened to appear by chance. "Support" is how often two drugs are seen together. "Confidence" tells how likely it is to see one drug if another drug has already been used. MBA can show if using one drug makes it more likely to use another. For example, using "Drug A" might make it very likely to use "Drug B." For the MBA part of the study, very rare drug uses were not included. The study looked for drug patterns where there was at least a 50% chance of one drug appearing with another. The top ten patterns with the highest "lift" were chosen.

For the RMLCA part, the study looked at drug use patterns over the 28 days before MOUD started. The researchers tested different numbers of patterns, and they chose six main patterns that fit the information best. They then used another step to see how these patterns were linked to relapse. Finally, different math models were used to see how both the RMLCA patterns and the MBA patterns were linked to relapse. This also took into account which of the three original studies the person was in and what treatment they received. The results show how much more or less likely a person was to relapse.

Results

This study included all 2,637 people from the three original studies who reported their drug use. Most of them (about 98%) provided a full 28-day history of their drug use before starting treatment. Even though the original studies were a little different, the people who took part had similar backgrounds and mental health issues.

The 2,637 people in the study reported using at least one drug on most days (89% of the time) in the 28 days before starting MOUD. Heroin, oxycodone, cannabis, hydrocodone, and methadone were the drugs used most often. On most days (64%), people used only one drug. On average, people used about one and a half drugs each day. Some used two drugs, some three, and some four drugs on a single day. One person even used 13 drugs on one day.

The MBA method found 34 patterns of drug use. For example, it showed that when people used opium, they were 16 times more likely to also use heroin on the same day. Using crack or cocaine also made it more likely to use heroin. There were also strong patterns showing that using cocaine was 13 times more likely when someone used heavy alcohol, another opioid, and cannabis together. Most of the patterns (76%) showed heroin use after other drugs, and 18% showed cannabis use.

The RMLCA method showed that there were six main patterns of drug use. The biggest pattern, called "All time low" (31.4% of people), included people who used very few drugs, including opioids, in the 28 days before treatment. Another large pattern, called "High opioid decreasing" (31.3% of people), included people who used a lot of opioids but started using less in the month before treatment. This group likely decreased their opioid use to get ready for treatment. Across most of the patterns, people used much less of certain drugs like heroin, methadone, other opioids, and cannabis in the two weeks right before starting treatment.

The study also looked at how the top 10 drug patterns found by MBA were linked to relapse. People who used cocaine, heroin, other opioids, and cannabis together had a higher chance of relapse. Their chance of having a positive drug test was almost 3 times higher, and their chance of missing tests was over 5 times higher. However, some patterns were linked to a lower chance of relapse. For example, people who used cocaine, heavy alcohol, cannabis, and other opioids together had a lower chance of missing drug tests. Also, using benzodiazepines, heroin, other opioids, and cannabis together was linked to a lower chance of relapse. This was a surprise, as it suggests that using many drugs at once ("polysubstance binge") does not always mean a higher chance of relapse. For example, using many drugs with heroin seemed to lead to a worse outcome, but using many drugs with prescription opioids did not seem as bad. This shows that different drug patterns were linked to different chances of relapse.

The RMLCA also showed that people in the groups who were decreasing their opioid use or heroin use were still more likely to relapse. This means that even if people cut down on their drug use before treatment, they might still have a higher risk of relapse compared to those who started with very little or irregular drug use. The MBA findings also showed that among opioid users, some groups who also used heroin, cocaine, and cannabis had a worse outcome, often missing their clinic visits. These results prove that there are clear groups among people who use many drugs, and these groups have different chances of success in treatment.

Discussion

This study provides new information about how people use many drugs together right before they start treatment for opioid use disorder. It found that using many drugs is common. Both the RMLCA and MBA methods were very helpful in understanding these specific patterns. RMLCA looks at groups of people and finds hidden patterns in how they answer questions about drug use. This method helps show how many patterns exist, how common each pattern is, and how likely it is that people in a pattern use certain drugs. MBA, on the other hand, helps find the most common combinations of drugs used on the same day. This gives details that might be missed when using broader drug groups.

For example, the study found that people who used opium often also used heroin, alcohol, and cannabis. It also found that using certain drugs, like benzodiazepines, cocaine, or cannabis, often happened before heroin use on the same day. RMLCA showed six main patterns of polysubstance use before treatment. It found that many people either used very few drugs, including opioids, or had greatly reduced their drug use before starting treatment. Importantly, both methods showed that different groups of drug users had different links to relapse.

Some surprising links were found between the MBA drug patterns and relapse. For instance, using cocaine, cannabis, heavy alcohol, and prescription opioids together did not lead to more relapse; instead, it was linked to a lower chance of relapse. But using the same combination of drugs with heroin instead of prescription opioids led to a big increase in relapse. This means that using many drugs at once is not always linked to relapse. A "binge" with heroin seemed to lead to a worse outcome, but a "binge" with prescription opioids did not seem as bad. This shows that different drug patterns were linked to different chances of relapse.

RMLCA also showed that people in the groups who were decreasing their opioid use or heroin use were still more likely to relapse. This means that even if people cut down on their drug use before treatment, they might still have a higher risk of relapse compared to those who started with very little or irregular drug use. The MBA findings also showed that among opioid users, some groups who also used heroin, cocaine, and cannabis had a worse outcome, often missing their clinic visits. These results prove that there are clear groups among people who use many drugs, and these groups have different chances of success in treatment.

This study has some limits. People reported their own drug use, which might not always be perfectly accurate. Also, MBA does not show when drugs were used together, but RMLCA does look at patterns over time, which can be helpful. However, the study has strong points too. It used information from many different places and a large number of people. It also showed a clear link between drug use patterns before treatment and how well treatment worked after starting. Knowing these patterns before treatment can help doctors find people who need special support and decide which treatment might work best for them. Future work will continue to use these patterns to help predict who will do best with different types of MOUD treatment.