State sequence analysis (SSA) of daily methadone dispensing trajectories among individuals at U.S. opioid treatment programs (OTPs) before and following COVID-19 onset

Abstract

Background and aims:

U.S. regulatory changes allowed for additional methadone take-home doses following COVID-19 onset. How dispensing practices changed and which factors drove variation remains unexplored. We examined daily methadone dispensing trajectories over six months before and after regulatory changes due to COVID-19 using state sequence analysis and explored correlates.

Design:

Retrospective chart review of electronic health records.

Settings:

9 opioid treatment programs (OTPs) across 9 U.S. states.

Participants:

Adults initiating treatment in 2019 (n=328) vs. initiating one month after the COVID-19 regulatory changes of March 2020 (n=376).

Measurements:

Type of daily methadone medication encounter (in-clinic, weekend/holiday take-home, take-home, missed dose, discontinued) based on OTP clinic; cohort (pre vs. post-COVID-19); and patient substance use, clinical, sociodemographic characteristics.

Results:

Following COVID-19 regulatory changes, allotted methadone take-home doses increased from 3.5% to 13.8% of total person-days in treatment within the first six months in care. Clinic site accounted for the greatest variation in methadone dispensing (6.2% and 9.5% of the variation of discrepancy between sequences pre- and post-COVID-19 respectively). People who co-use methamphetamine had a greater increase in take-homes than people who did not use (from 3.7% pre-pandemic to 21.2% post-pandemic vs. 3.5% to 12.5%) and higher discontinuation (average 3.6 vs. 4.7 months among people who did not use methamphetamine pre-COVID-19; average 3.3 vs. 4.6 months post-COVID-19). In the post-COVID-19 cohort, females had a higher proportion of missed doses (17.2% vs. 11.9%) than males. People experiencing houselessness had a higher proportion of missed doses (19% vs. 12.3%) and shorter stays (average 3.5 vs. 4.5 months) when compared to those with stable housing.

Conclusion:

Daily methadone dispensing trajectories may depend more on treatment site practices than individual characteristics. Further research into take-home practices across clinics should be explored as potential targets for improving equitable access to more flexible treatment regimens across OTPs.

INTRODUCTION

The U.S. is currently experiencing the worst drug overdose epidemic in its history, with around 112,000 U.S. residents dying of drug overdose between August 2022 and August 2023, most of them involving opioids (1). Opioid use disorder (OUD) can be a relapsing and chronic condition (2), hence, the consequences of the epidemic will likely weigh heavily for decades. Methadone maintenance treatment (MMT) is a highly effective intervention for people with OUD (3), which has been shown to improve several outcomes, including substantially reducing overdose risk (4–10).

Despite ample evidence supporting the effectiveness of MMT, access has remained limited in the U.S. In 2019, only 408,550 people received MMT in OTPs (SAMHSA, CBHSQ, 2021), and 581,218 received buprenorphine, a small fraction of the estimated 7.6 million people with OUD (12,13). In the U.S. methadone can only be accessed through Opioid Treatment Programs (OTPs), which are subject to strict regulations by the US Drug Enforcement Administration (DEA), Substance Abuse and Mental Health Services Administration (SAMHSA), and State Opioid Treatment Authorities (SOTA) (14). State and local laws often place zoning restrictions on where OTPs can be located (15). Consequently, nearly 80% of U.S. counties (including the entire state of Wyoming) do not have OTPs (16). Historically, the U.S. OTP system has been particularly burdensome due to highly restrictive dispensing policies (17). Ultimately, this system poses a heavy burden among individuals in their treatment journey, and advocacy groups of patients in MMT have labeled the current delivery system as a form of “liquid handcuffs” (18,19).

COVID-19 stressors and mitigation measures, such as social distancing and stay-at-home policies, raised challenges for individuals receiving MMT (20–22). Before the onset of the pandemic, newly admitted individuals had to go in person at least 6 days a week to receive their medication during the first 90 days of treatment. Take-home doses were available when clinics were closed (e.g. Sundays or holidays) (23). This translated into long travel and waiting times to receive methadone, as well as crowding in clinics, putting already highly comorbid individuals (24–26) at risk during the pandemic. In response, regulatory measures were taken to increase flexibility in take-home dispensing practices in March 2020 to allow for up to 14 days of take-home regimens for “somewhat stable” and 28-day regiments for “stable” patients. In the case of newly admitted patients, 7 out of the first 14 days of treatment could be take-home doses (27). These regulatory measures were recently permanently adopted at a federal level by a new rule released by SAMHSA in January 2024 (“42 CFR Part 8”) (28). Although these changing policies have translated into more flexible medication schedules for some, early research from the pandemic showed that individual OTP providers often applied idiosyncratic discretion when translating these take-home guidelines into clinical practice (29,30), which ultimately resulted in great variability in the implementation of these new flexibilities (19,31–35). However, to our knowledge, there have been no longitudinal assessments of what factors contributed to take-home allocation at the individual-level, and what characteristics were associated with dispensing trajectories both before and following these regulatory changes.

State Sequence Analysis (SSA) is a fully non-parametric longitudinal technique that is used for analyzing and classifying sequences of categorical states. Originally developed for the analysis of DNA (36,37), SSA has found increasing use in social sciences (38) and has recently been applied to study care trajectories (39–43), and pharmacoepidemiology (44,45). In this study, we used this novel approach to 1) examine daily methadone take-home dispensing trajectories of individuals initiating MMT across 9 U.S. OTPs over six months before and following COVID-19 reforms, and 2) explore individual- and provider-level observed factors driving variation in take-home dispensing trajectories.

METHODOLOGY

Study design, settings, and participants

We used data from the Optimal Policies to Improve Methadone Maintenance Adherence Long-term (OPTIMMAL) study, a retrospective chart review study, whose primary aim was to evaluate how changes in regulations due to COVID-19 impacted 6-month retention among individuals in MMT (46). OTPs were eligible if: (i) they were licensed OTPs that possessed integrated electronic health record (EHR) systems; (ii) they could track medication dosing on a per-visit basis; (iii) could archive urine drug testing results; (iv) had admitted more than 50 new participants during the follow-up period; and (v) they were in continuous operation since at least January 1st, 2018. Site recruitment was conducted through the Clinical Trial Network (CTN). The purposive sampling of sites sought to consider the variability of expected changes in dispensing practices when implementing the new regulations (46). Priority was given to OTPs with prior experience successfully participating in research studies. OPTs were located in Connecticut, Florida, Massachusetts, Minnesota, New Mexico, New York, North Carolina, Ohio, Oregon, and Pennsylvania (46). For the current analysis, we collected data on new individuals entering 9 OTPs across the U.S. either before or after the COVID-19 reforms and followed them for up to 6 months (180 days) while in care. 6 months was decided according to a measure on continuity of pharmacotherapy for OUD developed by the National Committee on Quality Assurance (NCQA) and RAND Corporation, which was later supported by the National Quality Forum to the Centers on Medicare and Medicaid Services (47,48).

The total number of participants was 828. We excluded those with missing values on our explanatory covariates (N = 124, 15%). Participants with missing values in covariates were more likely to have enrolled in clinics 1 and 9 (p<0.001, data not shown), have no insurance (p=0.007), be employed (p=0.021), and declare no methamphetamine use at baseline (p=0.002). 704 adult participants were included in the complete case analysis. The post-COVID-19 cohort initiated methadone treatment between April 15 and October 14, 2020, accounting for a one-month delay after the initiation of the new SAMHSA regulations (n=376). The pre-COVID-19 cohort initiated treatment a year prior, in 2019 (n=328). There was no overlap between the two cohorts (see Williams et al. (46) for more details on the study methodology). The NYU Langone Health Institutional Review Board approved this study. The analysis for this study was not pre-registered, therefore the results should be considered exploratory. The STROBE statement is presented in Supplementary Material (Table S1).

Measurements.

Daily methadone medication encounter

Our main measure to create trajectories of methadone dispensing was day-level medication encounter type, which was defined as either (all mutually exclusive): (i) missed dose (e.g. no-show to the clinic but not discharged from MMT), (ii) in-clinic dose; (iii) weekend/holiday dose (i.e. available to all individuals, due to clinic closures); (iv) take-home dose (i.e. individuals were eligible to take additional take home from clinic); and (v) censored due to discontinuation of treatment. From these encounter types, we formed a 6-month (180 days) trajectory to characterize treatment episodes.

Treatment and medication dispensing pattern covariates

We included months in treatment, the number and the proportion of days in each type of medication encounter while engaged in treatment, and the maximum consecutive take-home days during follow-up (no take-home, 1 to 6 days, 7 to 13 days, 14 or more days).

Baseline covariates

We included the following information as potential correlates of MMT dispensing trajectories: (i) OTP clinic (1 to 9); (ii) sex (female, male); (iii) age group (18 to 29, 30 to 39, 40 to 49, 50 or more); (iv) race/ethnicity (white non-hispanic, african american, hispanic, other race/ethnicity [which includes people identifying asian, pacific islanders, native americans, multiracial, among others]); (v) type of insurance (none [self-pay], public [Medicaid or Medicare], other [commercial insurance, grant funded or other]); (vi) marital status (never married/single [unknown if ever married], married, formerly married); (vii) education (less than high school, high school or GED, more than high school); (viii) employment (employed, unemployed [looking], out of labor force [e.g., retired, disabled]); (ix) housing situation (houseless, unstable housing [e.g., transiently housed by family or friends or living in a motel], secure housing [including recovery houses and inpatient treatment]); (x) use of alcohol, cocaine, cannabis, benzodiazepines, and methamphetamine at intake (dichotomous one for each); (xi) main route of administration of opioids (oral, smoking, intranasal, injection); and (xii) any psychiatric comorbidities (no, yes) including anxiety, post-traumatic stress, depressive, schizophrenia/other psychotic, bipolar, and attention deficit/other disruptive behavior disorders.

Statistical methods

Bivariate analysis

We conducted a bivariate analysis comparing the sociodemographic and clinical characteristics of the two cohorts (pre- and post-COVID-19) to assess their comparability. We used t-tests for continuous variables and chi-square tests for categorical variables.

State Sequence Analysis (SSA)

SSA can be used for analyzing and classifying sequences of categorical states, often involving the estimation of a “distance” (or dissimilarity) between each pair of sequences over discrete time units (e.g., days or weeks), based on several algorithms (36,49). SSA offers the advantage of using the whole sequence as a statistical object rather than just one aspect of the process (e.g., retention, take-home allowances), without making assumptions about the data-generating process, making it suitable for exploring intensive longitudinal data such as ours (43). SSA is particularly helpful in analyzing trajectories of multinominal categorical outcomes like ours, as other methods, including latent class growth analysis, have been mostly implemented for binary, continuous, and count data (50). We used an extension of SSA, multifactor discrepancy analysis, that allows exploring how trajectories vary by observed groups (e.g., sex or OTP clinic) based on the dissimilarity measure (51–53). For a more extensive explanation of our approach please refer to the supplementary material. The analysis consisted of three main steps:

Step 1: Estimate the dissimilarity matrix.

Because sequences are composed of categorical states (e.g. in-clinic dose, take-home dose) there is no natural “metric” – a mean sequence cannot be observed, but differences between pairs can be constructed. Thus, the discrepancy between sequences is defined from their pairwise dissimilarities (the dissimilarity matrix), which can be built based on different algorithms or “distances” (52,54). Algorithms may focus on the sequencing (e.g., ordering), duration (e.g., time spent), and/or timing (e.g., when transitions occur) of states experienced over time (54). Building on previous studies of care trajectories (43,44,55) we used the longest common subsequence (LCS) distance algorithm to estimate the dissimilarity matrix. In LCS two sequences are considered similar if they present long subsequences in common while allowing gaps between different states (56). We decided to use the LCS because it is not a metric based on the exact position of states (e.g., two individuals experiencing a take-home on the same date), but rather accounts for non-aligned matches (e.g., the person first experienced in-clinic doses followed by weekend/holiday dose) (43). It is appropriate when sequencing and duration are more important than the precise timing of a transition between states. Moreover, it is more data-driven than other algorithms, such as those in which researchers decide the costs of deleting, inserting, or substituting states when estimating the distance between two sequences in the context of optimal matching (43,55).

As an example, Figure 1 depicts the hypothetical sequences of 10 individuals (P1–10) during their first 5 days receiving MMT. Sequences from individuals 1, 2, and 5 are likely to be considered more similar within, as they discontinued treatment early. The LCS for these individuals is composed of 1 in-clinic dose, 1 missed dose, 1 in-clinic dose, and 1 censored day. The LCS only includes 1 missed dose as individual 1 missed just one day, thus allowing for gaps between states that constitute the LCS (i.e., the second missed dose of individuals 2 and 5 is not accounted for, thus prioritizing the sequencing of states over time).

Figure 1:

Hypothetical daily methadone dispensing sequences for 10 individuals (P1–10) during their first 5 days in MMT

Note: On the upper panel, the unordered sequences of 10 individuals (P1 to P10, y-axis) during their 5 first days in MMT (days in the x-axis). On the lower panel, groups of sequences with less dissimilarity due to sharing a longest common subsequence (LCS). P1, P2, and P5 compose one group, with missed doses (grey) and discontinuation (white). P4, P7, and P10 compose another group, with 2 take-home doses (pink). P3, P6, P8, and P9 compose the last group, which includes in-clinic doses (light blue) and a weekend-holiday dose (blue).

Step 2: Multifactor discrepancy analysis.

The dissimilarity matrix is translated into a measure of discrepancy across observed groups (e.g. males, individuals within the same OTP clinic). The discrepancy measures the variability as the average distance to the sequence (named ‘medoid’ – think of as a ‘centroid’ for categorical data) that minimized the sum of distances to all sequences of the observed group (36,52). Thus, the discrepancy score will be low if the sequences are similar within the group (e.g. are composed of similar states over time) (52). This analysis assesses the contribution of each covariate (e.g. sex, OTP clinic) to the total discrepancy by decomposing it into explained between-group and residual within-groups variation, thus allowing for testing for differences between groups using permutation tests (52). It is analogous to an ANOVA for which the outcome is a latent position; the medoids are the group means, and the dissimilarities are the residuals within or between groups. As proposed by Studer et al. (52) we used 5,000 permutations to assess a significance threshold of 1 percent. Overall, this procedure provides information about which covariates influence the dispensing sequences, but it does not tell how they change on a given covariate value (52).

Step 3: Bivariate analysis and visualization.

We later described differences in our treatment and medication dispensing pattern covariates among the correlates that presented p<0.05 within each cohort in the previous step using bivariate analysis. Thus, we first identify measurements that are associated with dispensing trajectories within each cohort, generally, and then we look for differences within those identified factors to unveil differences in treatment. We used t-tests or ANOVA for continuous variables and chi-square tests for categorical variables.

Sensitivity analysis.

We repeated the main analysis using all available cases, including those with missing data in our explanatory correlates (with an indicator for missingness, n=828) to test for stability in our solution. The code for the analysis is available at https://github.com/idborquez/methadone_sequence_analysis.

RESULTS

Table 1 shows participant characteristics by cohort (pre- and post-COVID-19): 40.1% were females, 74.4% non-Hispanic white. Commonly used substances at intake (besides opioids) were cocaine (29.7%), cannabis (26.1%), and methamphetamine (17.6%). There were no meaningful differences between cohorts.

Table 1:

Participant characteristics according to cohort pre- (04/2019–10/2019) and post- (04/2020–10/2020) COVID-19

	Cohort		Total (N=704)	p-value
	Pre (N=328)	Post (N=376)
	N (%) or mean (s.d)
OTP clinic				0.075
1	11 (3.4%)	33 (8.8%)	44 (6.2%)
2	47 (14.3%)	45 (12.0%)	92 (13.1%)
3	42 (12.8%)	49 (13.0%)	91 (12.9%)
4	29 (8.8%)	46 (12.2%)	75 (10.7%)
5	43 (13.1%)	38 (10.1%)	81 (11.5%)
6	47 (14.3%)	41 (10.9%)	88 (12.5%)
7	41 (12.5%)	43 (11.4%)	84 (11.9%)
8	34 (10.4%)	43 (11.4%)	77 (10.9%)
9	34 (10.4%)	38 (10.1%)	72 (10.2%)
Sex				0.255
Male	204 (62.2%)	218 (58.0%)	422 (59.9%)
Female	124 (37.8%)	158 (42.0%)	282 (40.1%)
Age groups				0.241
18 to 29	70 (21.3%)	70 (18.6%)	140 (19.9%)
30 to 39	114 (34.8%)	159 (42.3%)	273 (38.8%)
40 to 49	68 (20.7%)	69 (18.4%)	137 (19.5%)
50 or more	76 (23.2%)	78 (20.7%)	154 (21.9%)
Race/Ethnicity				0.200
White Non-Hispanic	238 (72.6%)	286 (76.1%)	524 (74.4%)
African American	42 (12.8%)	30 (8.0%)	72 (10.2%)
Hispanic	31 (9.5%)	41 (10.9%)	72 (10.2%)
Other	17 (5.2%)	19 (5.1%)	36 (5.1%)
Type of Insurance				0.176
None (self-pay)	48 (14.6%)	47 (12.5%)	95 (13.5%)
Public	237 (72.3%)	292 (77.7%)	529 (75.1%)
Commercial Insurance	8 (2.4%)	12 (3.2%)	20 (2.8%)
Grant-funded or other	35 (10.7%)	25 (6.6%)	60 (8.5%)
Marital status				0.718
Never married/Single	195 (59.5%)	233 (62.0%)	428 (60.8%)
Married	77 (23.5%)	79 (21.0%)	156 (22.2%)
Formerly married	56 (17.1%)	64 (17.0%)	120 (17.0%)
Education				0.672
Less than high school	99 (30.2%)	110 (29.3%)	209 (29.7%)
High school graduate/GED	123 (37.5%)	153 (40.7%)	276 (39.2%)
More than High school	106 (32.3%)	113 (30.1%)	219 (31.1%)
Employment				0.909
Employed	90 (27.4%)	98 (26.1%)	188 (26.7%)
Unemployed (looking)	124 (37.8%)	143 (38.0%)	267 (37.9%)
Out of the labor force	114 (34.8%)	135 (35.9%)	249 (35.4%)
Housing situation				0.922
Houseless	53 (16.2%)	57 (15.2%)	110 (15.6%)
Unstable Housing	75 (22.9%)	85 (22.6%)	160 (22.7%)
Secure housing	200 (61.0%)	234 (62.2%)	434 (61.6%)
Substance use at intake (% Yes)
Alcohol	40 (12.2%)	42 (11.2%)	82 (11.6%)	0.672
Cocaine	99 (30.2%)	110 (29.3%)	209 (29.7%)	0.788
Cannabis	83 (25.3%)	101 (26.9%)	184 (26.1%)	0.639
Methamphetamine	51 (15.5%)	73 (19.4%)	124 (17.6%)	0.179
Benzodiazepines	28 (8.5%)	33 (8.8%)	61 (8.7%)	0.910
Route of administration of opioids				0.217
Oral	26 (7.9%)	18 (4.8%)	44 (6.2%)
Smoking	7 (2.1%)	11 (2.9%)	18 (2.6%)
Intranasal	108 (32.9%)	113 (30.1%)	221 (31.4%)
Injection	187 (57.0%)	234 (62.2%)	421 (59.8%)
At least one psychiatric comorbidity (% Yes)	197 (60.1%)	232 (61.7%)	429 (60.9%)	0.656

Note: p values were obtained using t-test for continuous variables and chi-square test for categorical variables.

Figure 2 illustrates daily medication dispensing trajectories for both cohorts, pre (left panel) and post (right panel) COVID-19 onset. The state distribution plots (a) indicate the proportion of states as a stacked bar chart at each time point (180 days since initiating treatment), showing the aggregate change in the distribution of states over time by disaggregating individual sequences. Take-home doses immediately increased post-COVID-19 onset with a slight decrease in weekend in-person dosing. Censoring occurs when a subject leaves treatment, and these rates and their timing appear similar across cohorts. The index frequency plots (b), sort individuals by their last observed state and length in that state, thus they provide information about the individual sequence of each participant which is represented by a line in the plot. When comparing dispensing practices across cohorts (Table 2), there was an increase in take-home doses after COVID-19 (from 3.5% to 13.8% of the total number of person-days, p<0.001), however, in-clinic doses remained the most common. Before COVID-19 two-thirds of individuals (65.5%) never received a take-home dose in the first six months of care, decreasing to 43.1% after COVID-19 onset. In both cohorts, long periods with continuous take-home doses were rare: only eight individuals across all clinics received more than 7 consecutive take-homes (2.4%) before COVID-19, increasing only to seventeen (4.6%) in the post-COVID-19 cohort. Although dispensing practices changed pre and post-COVID-19, the percentage of days with missing doses (i.e., no-show) remained similar (12.7% pre vs. 14.2% post-COVID-19, p=0.243), as did the average number of months retained in treatment (4.5 pre vs. 4.4 post-COVID-19, p=0.341).

Figure 2:

Daily methadone dispensing of newly admitted individuals pre- (n = 328, left panel) and post- (n = 376, right panel) COVID-19 onset cohorts over 6 months presented as (a) state distribution plots and (b) index frequency plots.

Note: (a) Upper panel depicts state distribution plots, which show the proportion of states in a stacked bar chart at each time point, allowing us to see the aggregate change in the distribution of states over time by disregarding individual sequences. In the y-axis the proportion of individuals. In the x-axis each day (1 to 180) since admission. (b) Lower panel depicts index frequency plots, which order subjects by the last observed state and length in that state, thus each participant’s sequence is a horizontal line in the plot. In the y-axis the frequency of individuals. In the x-axis each day (1 to 180) since admission. The left panel shows the pre COVID-19 cohort. The right panel shows the post COVID-19 cohort.

Table 2:

Treatment and medication dispensing pattern covariates according to cohort pre- (04/2019–10/2019) and post- (04/2020–10/2020) COVID-19 over the 6-month observation period (180 days).

	Cohort (person level)			p-value
	Pre (N=328)	Post (N=376)
	N (%) or mean (s.d)
Months observed	4.5 (2.1)		4.4 (2.2)	0.341
Maximum consecutive take-home days per participant				< 0.001
No take-home in the whole period	215 (65.5%)		162 (43.1%)
1 to 6 consecutive days	105 (32.0%)		197 (52.4%)
7 to 13 consecutive days	6 (1.8%)		10 (2.7%)
14 or more consecutive days	2 (0.6%)		7 (1.9%)
No weekend/holiday encounter	128 (39.0%)		152 (40.4%)	0.705
No missed dose encounter	66 (20.1%)		63 (16.8%)	0.249
	Cohort (encounter level)			p-value
	Pre (N=44,693)	Post (N=49,467)
	N (%) or mean (s.d)
Number of encounters (days)1
Total days	136.3 (63.5)	131.6 (66.6)		0.341
In-clinic days	106.3 (57.7)	88.1 (56.4)		< 0.001
Weekend/holiday days	17.5 (10.6)	11.5 (10.4)		< 0.001
Take-homes days	14.2 (21.0)	35.9 (35.6)		< 0.001
Missed dose days	18.0 (21.1)	19.4 (21.8)		0.449
Proportion of encounters (days)
In-clinic	0.761 (0.18)	0.672 (0.23)		< 0.001
Weekends/holidays	0.077 (0.07)	0.048 (0.06)		< 0.001
Take-homes	0.035 (0.11)	0.138 (0.19)		< 0.001
Missed dose	0.127 (0.16)	0.142 (0.17)		0.243

Note: p values were obtained using t-test for continuous variables and chi-square test for categorical variables.

¹The maximum number of encounters could be 180 per person.

Multifactor discrepancy analysis

Table 3 presents the results of the multifactor discrepancy analysis. All covariates together respectively accounted for 21.9% and 25.1% of the discrepancies of the daily methadone dispensing trajectories pre- and post-COVID-19 (pseudo-R² for Total in Table 3). The covariate that explained the most variation was the OTP clinic site in both cohorts. In the pre-COVID-19 cohort, if the clinic was removed from the model the pseudo-R² would decrease by 0.062 (p<0.001), the only variable with a significant pseudo-R² (p<0.05). Hence, this covariate accounted for 6.2% of the variation in dispensing trajectories.

Table 3:

Multi-factor discrepancy analysis for daily methadone dispensing sequences according to cohort pre- (04/2019–10/2019) and post- (04/2020–10/2020) COVID-19

Covariate	Pseudo F	Pseudo R2	p-value
Pre COVID-19 (N = 328)
OTP clinic	2.920	0.062	< 0.001
Race/ethnicity	1.813	0.015	0.058
Main route of administration: opioids	1.731	0.014	0.065
Type of Insurance	1.490	0.012	0.130
Age groups	1.229	0.010	0.240
Employment	1.305	0.007	0.209
Cannabis use at intake	1.727	0.005	0.123
Cocaine use at intake	1.491	0.004	0.163
Housing situation	0.713	0.004	0.648
Marital status	0.621	0.003	0.771
Methamphetamine use at intake	1.083	0.003	0.294
Education	0.526	0.003	0.885
Sex	0.963	0.003	0.354
Psychiatric comorbidity	0.912	0.002	0.378
Alcohol use at intake	0.764	0.002	0.487
Benzodiazepine use at intake	0.369	0.001	0.962
Total	2.339	0.219	< 0.001
Post COVID-19 (N=376)
OTP clinic	5.342	0.095	< 0.001
Age groups	1.877	0.012	0.0324
Housing	2.350	0.010	0.0194
Race/ethnicity	1.527	0.010	0.0906
Methamphetamine use at intake	4.161	0.009	0.0028
Type of Insurance	1.364	0.009	0.1454
Main route of administration: opioids	1.331	0.009	0.164
Sex	3.468	0.008	0.0104
Marital status	1.636	0.007	0.0904
Employment	1.239	0.005	0.23
Psychiatric comorbidity	1.827	0.004	0.0986
Education	0.658	0.003	0.7802
Cannabis use at intake	1.097	0.002	0.2808
Benzodiazepine use at intake	0.695	0.002	0.6196
Alcohol use at intake	0.575	0.001	0.7684
Cannabis use at intake	0.541	0.001	0.824
Total	3.132	0.251	< 0.001

Note: Estimation based on longest common subsequence (LCS) distance algorithm. Ordered by Pseudo R². The Pseudo R² represents how much variation of the discrepancy measure is explained by the predictor.

While we must cautiously interpret differences in discrepancies across cohorts, in the post-COVID-19 cohort, the OTP clinic accounted for a greater proportion of the variation – explaining 9.5% of the discrepancies (pseudo-R²=0.095, p<0.001). Other covariates explaining variation in dispensing trajectories in the post-COVID-19 cohort were: (i) age groups (pseudo-R²=0.012, p=0.032); (ii) housing situation (pseudo-R²=0.010, p=0.019); (iii) methamphetamine use at intake (pseudo-R²=0.009, p=0.003); and (iv) sex (pseudo-R²=0.008, p=0.010). The state distribution and index frequency plots and treatment and medication dispensing pattern covariates stratified by each of these correlates pre and post-COVID-19 can be found in Supplementary material, Figures S1–8 and Tables S2–11. The patterns reveal which groups utilize the take-home dispensing and at what points in their treatment.

Given the large role of the clinic in determining dispensing trajectories, we graphed the individual dispensing trajectories by OTP over 180 days, which showed great heterogeneity in practices across clinics before as well as after COVID-19 (Figure 3–4 and Tables S2–3). For instance, clinic 7 provided no take-homes to their newly admitted individuals in both cohorts, whereas clinic 3 greatly increased the proportion of days with take-home doses from 2.5% to 33.5%.

Figure 3:

Index frequency plot of the daily methadone dispensing trajectories among newly admitted individuals pre- (n = 328, left panel) and post- (n = 376, right panel) COVID-19 cohorts over 6 months by OTP clinic.

Note: Index frequency plots order subjects by the last observed state and length in that state, thus each sequence is a line in the plot. In the y-axis the frequency of individuals by OTP clinic. In the x-axis each day (1 to 180) since admission.

Figure 4:

State distribution plot of the daily methadone dispensing trajectories among newly admitted individuals pre- (n = 328, left panel) and post- (n = 376, right panel) COVID-19 cohorts over 6 months by OTP clinic.

Note: State distribution plots show the proportion of states in a stacked bar chart at each time point, allowing us to see the aggregate change in the distribution of states over time by disregarding individual sequences. In the y-axis the proportion of individuals by OTP clinic. In the x-axis each day (1 to 180) since admission.

When we assessed the role of patient-level covariates, we found that before the pandemic, individuals in the 40–49 age group had the least time in care (4.0 months, compared to 4.5 on average, p=0.041), while after the pandemic, individuals between 30 and 39 years of age spent the least time in care (4.0 months vs. 4.4 months on average, p=0.015) (Figure S1–2 and Table S4–5). There were no differences in the proportion of missed doses and time in treatment before COVID-19 among people of different housing situations (p=0.137 and p=0.424, respectively). However, individuals experiencing houselessness presented a higher proportion of missed doses (19% vs. 14.2% on average, p=0.014) and less time in treatment (3.5 vs. 4.4 months on average, p=0.007) after the pandemic (Figure S3–4 and Table S6–7). People who use methamphetamine showed a larger increase in take-home doses (from 3.7% to 20.2% vs. 3.5% to 12.5% among people who did not use methamphetamine) (Figure S5–6 and Table S8–9), more days with missing doses (19.4% vs. 11.4% among people who did not use methamphetamine pre COVID-19, p=0.001; 21.2% vs. 12.5% respectively post COVID-19, p<0.001) and less time engaged in treatment (3.6 vs. 4.7 months on average among people who did not use methamphetamine pre COVID-19, p<0.001; 3.3 vs. 4.6 months on average respectively post-COVID-19, p<0.001). Lastly, while there were no differences in the proportion of missed doses and time in treatment among females and males pre-COVID-19 (p=0.185 and p=0.218 respectively), in the post-COVID-19 cohort females showed more missed doses (17.2% vs. 11.9% among males, p=0.003) and less time in treatment (4.1 vs. 4.6 months on average, p=0.037) (Figure S7–8 and Table S10–11).

Sensitivity analysis

The multifactor discrepancy analysis using the whole sample with a missing indicator in the explanatory covariates yielded similar results, reaffirming our findings (Table S12). Some differences included that in the pre-COVID-19 cohort the type of insurance appeared as a potential predictor of discrepancies across trajectories (pseudo R²=0.017, p=0.035). In the post-COVID-19 cohort opioid main route of administration (pseudo R²=0.014, p=0.016), marital status (pseudo R²=0.011, p=0.033), and psychiatric comorbidities (pseudo R²=0.005, p=0.043) emerged as potential explanatory factors. Housing status nearly surpassed the p<0.05 threshold (pseudo R²=0.012, p=0.054). Although these characteristics emerge as potential correlates it is unclear if this is due to selection into missingness or differences between groups that were not captured in the main analysis.

DISCUSSION

In this study, we used a novel method that to our knowledge has never been applied to substance use research – state sequence analysis followed by discrepancy analysis – to characterize daily methadone dispensing trajectories among patients newly entering care across 9 U.S. OTPs before and following COVID-19’s change in take-home regulations. Our findings are consistent with previous research showing a modest increase in take-home doses after the regulatory changes (from 3.5% to 13.8% of the person-days engaged in treatment) (31,32,57–60), but add new insight as to how individual and clinic-level characteristics drive variation in dispensing trajectories. Although more newly admitted individuals received take-home doses after regulations were changed, our study highlights that OTPs did not substantially increase flexibility in terms of extended consecutive days with take-home doses (58). Importantly, we also found no increase in the proportion of days without a dose (i.e., missed) or treatment discontinuation (61).

Our results highlight the important role that providers (e.g. individual OTP clinics) play in determining the dispensing trajectories of individuals initiating MMT, as it was the covariate explaining more variation across sequences both before and following the pandemic changes. A prior synthesis of evidence (59) on the effect of COVID-19 in MMT practices emphasized that the proportion of individuals who received take-home doses varied greatly by study and clinic. We saw a similar pattern in our results, where there was great variability in the implementation of new flexibilities across sites, consistent with prior research (19,31–35). Additionally, our findings point to the already existing marked differences in daily dispensing trajectories across clinics even before COVID-19 and the new regulations. Furthermore, when analyzing both cohorts together clinic remained the most important covariate explaining variability in dispensing with a pseudo R² of 0.06 (vs. pseudo R² = 0.013 for the cohort variable, data not shown). Many have argued that the delivery system of methadone in the U.S. should move towards a more person-centered approach, given that one of the core principles of treatment for SUDs is individualized care (62). Although providers need to implement dispensing strategies that best suit their local environments (34), people’s preferences and circumstances should be considered when designing interventions for individuals with SUDs (62). Some individuals might prefer daily dispensing in a clinic, while others prefer a more flexible approach (19,34,63). Our findings question the extent to which OTP clinics are engaging in such person-level decision-making, given the culture and norms of the clinic itself seemed to be the greatest determinant of receiving take-home doses, despite each clinic having a wide mix of patients. In February of 2024, SAMHSA formalized more flexibility in take-home allowance on a federal level long-term (28), but our findings suggest that enforcing and supporting OTPs to seriously assess their clinic policies and practices around take-home decisions is critical. SOTAs play a major role in shaping the dispensing practices of clinics within their state (64,65) and are the authority that provides guidance, monitors, and oversees OTPs operations. Hence, they could promote the design of accountability mechanisms that include the perspectives of individuals in MMT to monitor the dispensing strategies of OTPs, make sure that clinics are adopting these policy changes, and promote more equitable practices across OTPs (17,64). Additionally, OTPs will need continuous technical assistance and incentives to adopt more flexible practices, as in-clinic visits will still be incentivized due to how methadone is billed through healthcare in some states (64).

Our study also points to some important characteristics associated with treatment retention early in treatment. We found that methamphetamine use at intake was associated with increased treatment discontinuation (66–68) and missed doses, which is consistent with previous findings using Medicare data during COVID-19 (69). Interestingly, people who use methamphetamine experienced a relatively greater increase in the proportion of take-homes – possibly due to more flexible guidelines on take-homes not being tied as heavily to abstinence (31,34), but their retention did not get worse or improve. With the current increase in overdoses involving stimulant and opioid use in the U.S. (70), OTP practices need to more seriously integrate efforts to address stimulant and polysubstance use, without making abstinence a requirement for engaging in treatment, which could risk further alienating patients (71). Houselessness was also associated with greater discontinuation during COVID-19 (19,72,73). People experiencing houselessness are of special interest as they face substantial barriers in accessing medication for OUDs, including stigmatizing attitudes from providers, and may particularly struggle with high threshold barriers to care and attendance requirements engrained in the OTP system (74,75). Approaches such as mobile medication units can improve access to these unserved populations by bringing treatment to them (76,77), however, under current regulations, only OTPs can operate them in the U.S. (78), limiting their reach. The Modernizing Opioid Treatment Access Act (MOTAA) bill, which is still waiting to be discussed amid 107,543 lives lost in 2023 (79), would allow other actors, such as physicians board-certified in Addiction Medicine or Addiction Psychiatry, to prescribe methadone outside of OTP settings, and community pharmacies to dispense methadone for OUD (64), potentially expanding access to MMT beyond OTPs.

Finally, our study emphasizes that SSA is a promising tool for analyzing the complex real-world care trajectories of patients with SUDs using data generated routinely in healthcare (80), which are increasingly available for research purposes. This would allow including the variety of services that individuals with SUDs encounter (e.g., syringe exchange, emergency room, behavioral treatment), and examine their care and medication trajectories longitudinally (43,44). Within the SSA framework, health services researchers can apply variable-centered approaches, like the analysis presented in this article, and/or person-centered approaches that aim to cluster and define subgroups of individuals that share similar trajectories (81). Likewise, researchers can analyze multiple domains (e.g., daily medication encounters and dosage in the case of MMT) simultaneously using its multichannel approach (82,83).

This article has several limitations. First, this study is observational and the findings depicted should be considered cautious regarding causality. Results derived from multifactor discrepancy analysis are not adjusted for other covariates. Second, we analyzed daily methadone dispensing trajectories among 9 OTPs, thus limiting the generalizability of our results as there are more than 1,800 functioning OTPs as of 2023 in the U.S. (84). Moreover, OTPs were purposively sampled to ensure expected variability in implementation of the new regulations and were largely affiliated with academic medical centers and not-for-profit. Although this might introduce selection bias, all OTPs were in states where increasing flexibility of take-home doses was feasible. Third, we only had information regarding one episode of care over a 6-month follow-up period. Thus, our censored participants might have started other types of treatment or MMT in other clinics, which would affect their trajectories. Longer periods of follow-up and information about multiple episodes of care across different types of services are needed to truly understand the care trajectories of individuals with OUD, which are highly heterogeneous and complex (85,86). Fourthly, the decision to use LCS to estimate the dissimilarity matrix is subjective and results might change if another algorithm is used. We based our decision on previous literature (43,44,55) and our interest in ordering and sequencing of states. Fifth, missingness in our variables of interest is likely not at random. Nonetheless, imputation techniques for multifactor discrepancy analysis have not been implemented to our knowledge, and most variables remained below p<0.05 after including the missingness indicator. Lastly, we lacked detailed clinic-level information (e.g. that could better explain why dispensing trajectories vary so greatly across sites).

CONCLUSION

In this study, we found that methadone dispensing trajectories both before and following COVID-19 regulatory changes appeared to depend more on a given treatment site’s practices than individual patient characteristics or response to treatment. Further research into take-home allowances should explore variability across clinics as potential targets for reducing treatment burden and improving equitable practices across OTPs. SSA offers an opportunity to answer important questions to better understand trajectories of healthcare utilization and pharmacoepidemiology of individuals with SUDs, especially with the increasing availability of real-world data like EHR, administrative, and claims data.