Opioid therapy trajectories of patients with chronic non-cancer pain over 1 year of follow-up after initiation of short-acting opioid formulations

Abstract

Objective

This study compared opioid utilization trajectories of persons initiating tramadol, short-acting hydrocodone, or short-acting oxycodone, and it characterized opioid dose trajectories and type of opioid in persistent opioid therapy subsamples.

Methods

A retrospective cohort study of adults with chronic non-cancer pain who were initiating opioid therapy was conducted with the IQVIA PharMetrics^® Plus for Academics data (2008–2018). Continuous enrollment was required for 6 months before (“baseline”) and 12 months after (“follow-up”) the first opioid prescription (“index date”). Opioid therapy measures were assessed every 7 days over follow-up. Group-based trajectory modeling (GBTM) was used to identify trajectories for any opioid and total morphine milligram equivalent measures, and longitudinal latent class analysis was used for opioid therapy type.

Results

A total of 40 276 tramadol, 141 023 hydrocodone, and 45 221 oxycodone initiators were included. GBTM on any opioid therapy identified 3 latent trajectories: early discontinuers (tramadol 39.0%, hydrocodone 54.1%, oxycodone 61.4%), late discontinuers (tramadol 37.9%, hydrocodone 39.4%, oxycodone 33.3%), and persistent therapy (tramadol 6.7%, hydrocodone 6.5%, oxycodone 5.3%). An additional fourth trajectory, intermittent therapy (tramadol 16.4%), was identified for tramadol initiators. Of those on persistent therapy, 2687 individuals were on persistent therapy with tramadol, 9169 with hydrocodone, and 2377 with oxycodone. GBTM on opioid dose resulted in 6 similar trajectory groups in each persistent therapy group. Longitudinal latent class analysis on opioid therapy type identified 6 latent classes for tramadol and oxycodone and 7 classes for hydrocodone.

Conclusion

Opioid therapy patterns meaningfully differed by the initial opioid prescribed, notably the presence of intermittent therapy among tramadol initiators and higher morphine milligram equivalents and prescribing of long-acting opioids among oxycodone initiators.

Introduction

Short-acting hydrocodone, short-acting oxycodone, and tramadol are the most commonly prescribed opioids.^1–3 Despite the near ubiquity of prescribing tramadol, short-acting hydrocodone, and short-acting oxycodone in acute and chronic pain settings, little is known about whether selecting one of these opioids as an initial analgesic treatment leads to different opioid therapy patterns, such as transitioning to persistent use. Among postoperative surgical patients, tramadol therapy was associated with a 41% increase in risk of long-term opioid therapy in comparison with other short-acting opioids.³ Likewise, initiation of tramadol was associated with an 11% reduction in the likelihood of opioid discontinuation at 1-year follow-up.¹ Furthermore, it is unknown whether persons initially prescribed one of these opioids are more likely to escalate doses, transition to other opioids, or add other opioids to the initial short-acting opioids. For instance, if individuals who initiated opioid therapy on tramadol were prescribed another opioid in addition to tramadol, it might suggest that the tramadol was not adequately managing pain.⁴

Statistical techniques for describing trajectories, such as group-based trajectory modeling (GBTM), growth mixture modeling, and latent class analysis are increasingly being used in pharmacoepidemiological studies.^5,6 Additionally, latent class analysis, a special type of finite mixture modeling that is used for identifying latent groups via categorical observed variables, has been used to identify subgroups in clinical trials or risk groups for substance use.⁷ The latent groups constructed through the use of GBTM, growth mixture modeling, or latent class analysis could be used along with subject matter knowledge to identify individuals likely to benefit the least from the treatment and also to design interventions that could change the risky therapy trajectories. Trajectory modeling approaches have been used for studying trajectories of days’ supply and dose after the first opioid prescription or trajectories of days’ supply after the initiation of long-term therapy.^8–10 These studies have not considered whether the trajectories might differ by the type of initial prescribed opioid.

Comparative study of opioid trajectories for initiators of tramadol, hydrocodone, and oxycodone could help us understand whether one drug is more likely to lead to riskier trajectories. As these are the 3 most prescribed opioid analgesics, identifying patients who follow potentially riskier trajectories could inform the selection of an initial opioid analgesic. The objectives of this study were (1) to identify the latent trajectories for weekly measures of opioid use, cumulative dose, and type of opioid (eg, short-acting, long-acting, or combination opioids) among initiators of tramadol, short-acting hydrocodone products, and short-acting oxycodone products; and (2) to describe the clinical and demographic characteristics associated with trajectory group membership. This study focused on adults with chronic non-cancer pain conditions (ie, back pain, neck pain, or osteoarthritis) who were prescribed opioid analgesics.

Methods

Data

A 10% random sample of the IQVIA PharMetrics^® Plus for Academics data (January 2008 to June 2018) was used. These data are from an administrative health insurance claims database of individuals enrolled in commercial health plans, Medicare Advantage, and Medicaid managed care plans.¹¹ The database has more than 10 million individuals and includes information on demographics (year of birth, sex, geographic region), inpatient stays, outpatient visits, emergency department visits, and prescription medications. Because the database did not contain patient identifiers, the study was deemed not to be human subjects research by the Institutional Review Board (protocol #261659).

Study design and cohort selection

A retrospective cohort study of ambulatory persons with chronic non-cancer pain who initiated treatment with tramadol or short-acting formulations containing hydrocodone or oxycodone was conducted. Individuals with an initial study opioid prescription between July 1, 2008, and June 30, 2017, were identified. Opioid prescriptions were identified from the pharmacy claims via Generic Product Identifier codes,¹² and a short-acting formulation was identified via the dosage form variable (eg, not long-acting, sustained release, or extended release). The date of the initial opioid prescription was designated as the index date. Subjects were required to be continuously enrolled with both medical and pharmacy benefits for 6 months before (baseline period) and 12 months after (follow-up period) the index date. To identify persons with chronic pain, subjects were required to have at least 1 diagnosis for back pain, neck pain, or osteoarthritis in the 90 days before the index date and 1 additional diagnosis for the same condition 30 days or more apart but within 180 days in either baseline or follow-up period including the index date (Figure 1). These conditions were identified from inpatient or outpatient claims through the use of International Classification of Diseases (ICD)-9-CM or ICD-10-CM codes (Table S1). To ensure that only new episodes of opioid treatment were identified, subjects with any opioid prescription in the 6-month baseline period were excluded. To ensure that the analysis focused on ambulatory adults with reliable demographic information who were free of cancer, substance use disorders, and more complicated pain diagnoses, subjects were excluded if they had any of the following in the baseline period: 1 or more diagnoses of cancer, rheumatoid arthritis, spinal injuries (quadriplegia and paraplegia), opioid use disorder or other substance use disorder, pregnancy, or organ transplantation; 1 or more claims for hospice or long-term care; or naloxone prescriptions. Individuals less than 18 years of age at the index date and those with missing sex were also excluded. Furthermore, individuals with potentially erroneous opioid analgesic prescriptions (prescriptions with negative days’ supply, more than 180 days of days’ supply, negative quantities, or more than 1000 units) at the index date or in the follow-up period were excluded. To ensure that subjects could be classified into one of the study groups, individuals with 2 or more of the study opioid analgesics of interest dispensed on the index date were excluded (Figure 2).

Figure 1.

Study diagram. Dx= diagnosis.

Figure 2.

Inclusion and exclusion criteria and final sample selection.

Study measures

Exposure

Three groups were created on the basis of the initial opioid analgesic prescription: tramadol initiators, short-acting hydrocodone initiators, and short-acting oxycodone initiators.

Opioid therapy measures in the follow-up period

Three separate trajectories based on weekly measures of any opioid therapy, morphine milligram equivalent (MME) dose, and type of opioid were developed for each of the 3 study groups over the 52-week follow-up period. A weekly time window was chosen, as it better captures the dynamics of opioid therapy (eg, gap, switching), unlike previous studies that used a monthly time window.¹³ Any opioid analgesic therapy was considered over the 52-week follow-up period, including opioids (ie, morphine, codeine) not considered in the formation of the trajectory groups. Because opioid dose and type of opioid are conditionally dependent on having any opioid therapy, these 2 measures were analyzed in subsamples of those following persistent opioid therapy trajectories.

After the classification of individuals into the 3 opioid groups, all opioid analgesic prescriptions in the 1-year follow-up were recorded. Start dates of prescriptions were adjusted to account for early refills in the following manner: If the number of overlapping days between 2 prescriptions was less than 30% of the days’ supply of the first prescription, then the start date of the second prescription was moved to the next day after the end date of the first prescription, on the premise that the prescription had been an early refill to be used after exhaustion of the initial days supplied. If the overlap was 30% or more, no such adjustment was made, on the premise that the initial opioid prescription had been exhausted or dual opioid therapy had been initiated.

Any opioid therapy was defined as having at least 1 day’s supply of any opioid analgesic in each of the 52 weeks. In other words, all opioid analgesics were considered, irrespective of the study opioid group into which subjects initially had been classified. Fifty-two dummy variables were created, with a value of 1 if the individual had at least 1 day of opioid therapy in the respective week and a value of zero otherwise.

The weekly total MME dose in each of the 52 weeks was calculated as a continuous measure among samples identified as having persistent opioid therapy on the basis of the GBTM of any opioid therapy measure. Each opioid ingredient dose was converted to an MME dose through the use of the Centers for Disease Control and Prevention MME conversion factor and a table from the North Carolina Association of Pharmacists specifically for parenteral formulations.^14,15 The weekly total MME dose was calculated from the sum of the total MME observed in each week. Because opioid dose is right-skewed, the weekly total MME dose was natural log-transformed.

The opioid type measure was constructed as multinomial variables with 5 categories for each of 52 weeks. For tramadol initiators, the categories were (1) tramadol only, (2) non-tramadol short-acting opioid only, (3) concomitant tramadol with another short-acting opioid, (4) long-acting opioid with or without short-acting opioids, and (5) no opioid. Similar categories were created for short-acting hydrocodone and short-acting oxycodone initiators.

Covariates

Covariate information was collected from the 6-month baseline period. Age, gender, region of residence, and payer type were obtained from the enrollment file. Index year was also included. The presence of other pain conditions (neuropathic pain, migraine, abdominal pain, and other pain conditions) was documented with ICD-9-CM and ICD-10-CM codes. The following mental health conditions were recorded: major depressive disorder, anxiety disorder, personality disorders, post-traumatic stress disorder, schizophrenia, and nicotine dependence. Prescriptions for benzodiazepines, hypnotics, skeletal muscle relaxants, antidepressants, and gabapentinoids were identified from pharmacy claims via Generic Product Identifier codes. Any surgery performed in the baseline period was identified via Current Procedural Terminology (CPT-4) codes. The Charlson Comorbidity Index was calculated and reported.¹⁶ All diagnosis, procedure, and drug codes are provided in Table S1.

Statistical analyses

GBTM was used to identify “any opioid therapy” trajectories separately for the 3 initiator groups. The binary logit distribution was used to model any opioid therapy. Models with 1–5 trajectories were fit by using up to the fourth-order polynomial of time (in weeks).⁶ The model output included the estimated percentage of all latent groups, the probability of belonging to each latent group for all individuals, and regression coefficients for time variables. The final models were selected on the basis of the combination of the following criteria: (1) Bayesian information criterion, with a larger value indicating better fit¹⁷; (2) at least 5% of sample estimated to belong in each of the latent groups; (3) an average posterior probability of 0.7 or larger in each of the latent groups; (4) closeness of the estimated latent group percentage and the percentage of individuals assigned to each group based on maximum probability; and (5) domain knowledge, with an emphasis on the clinical meaningfulness of the estimated additional trajectory.¹⁷

Patients following persistent opioid therapy trajectories, identified via GBTM of any opioid therapy in the 3 study groups (tramadol, short-acting hydrocodone, and short-acting oxycodone initiators), were included in the GBTM of MME dose. A censored normal distribution was used to model the log-transformed MME variables. The minimum and maximum values of the variables were selected empirically for the 3 groups. Models were fit by using up to fourth-order polynomial of time, similar to a previous analysis.⁶ Model fitting and final model selection were performed on the basis of the same criteria mentioned previously. The Proc traj procedure in SAS 9.4 was used to estimate the GBTM models.

In the subset of individuals with persistent therapy, longitudinal latent class analysis (LLCA), an application of latent class technique to longitudinal data, was used for analyzing the type of opioid separately for the 3 opioid groups. Multinomial logit distribution was used to model the variables. The final models were selected on the basis of the combination of the following criteria: (1) consistent Akaike information criterion, where lower value indicates better fit, and (2) domain knowledge, with a focus on the clinical meaningfulness of the estimated additional latent group. Proc lca in SAS 9.4 was used for this analysis.

Separate multinomial logistic regressions were used to explore factors associated with trajectory group membership of any opioid therapy for each of the 3 opioid initiator groups. Demographics, mental health conditions, pain conditions, and medication use variables were used as independent variables and the trajectory group as the dependent variable. Odds ratios and 95% confidence intervals were reported. Proc logistic in SAS 9.4 was used.

Sensitivity analyses

To test the robustness of the findings, we conducted sensitivity analyses. First, we changed the early refill criterion to a standard 3 days instead of a maximum overlap of 30% of days supplied: The prescription dates were adjusted if 2 opioid prescriptions had fewer than 3 days of overlap. Second, we adjusted for selected covariates (sex, index year, chronic back pain diagnosis, chronic neck pain diagnosis, chronic osteoarthritis, baseline surgery, and skeletal muscle relaxant prescriptions) when estimating the trajectory model for any opioid therapy. The probability of trajectory membership was estimated as the function of time (weeks) and the selected covariates mentioned previously through the use of a multinomial logistic model within the group-based trajectory framework. The covariates were selected on the basis of domain knowledge and on the differences observed across the 3 initiator groups. Third, we compared the baseline characteristics of patients who were excluded because they did not have continuous 12 months of post-enrollment data with the characteristics of our final sample to assess potential selection bias.

Results

A total of 1 999 409 individuals with an initial opioid prescription for tramadol, short-acting hydrocodone, or short-acting oxycodone were identified. After application of all the study inclusion and exclusion criteria, 226 520 individuals were retained in the final sample: 40 276 tramadol initiators (17.78%), 141 023 hydrocodone initiators (62.26%), and 45 221 oxycodone initiators (19.96%). Figure 2 describes the changes in the study sample after the application of each inclusion/exclusion criterion.

Approximately half of the individuals initiating one of the study drugs were 55 years of age or older, and this was similar across all the study drugs (Table 1). Fifty-six percent of hydrocodone and oxycodone initiators were female, whereas 65% of tramadol initiators were female. More than 70% of initiators had osteoarthritis, and back pain was present in at least 35% of initiators. Back pain was more prevalent in tramadol initiators than in others. The prevalences of depression, anxiety, mood disorders, personality disorders, and post-traumatic stress disorder were similarly low across the 3 groups, with the most frequently diagnosed mental health disorder being anxiety, which was observed in about 7% of initiators. A total of 35.91% of tramadol initiators had undergone surgery in the baseline period, whereas the corresponding prevalences for hydrocodone and oxycodone initiators were 48.93% and 65.81%, respectively. About 0.19% of tramadol initiators had another opioid drug on the index date, while the corresponding percentages for hydrocodone and oxycodone initiators were 0.27% and 2.04%, respectively. Among those with concurrent opioid therapy on the index date, the following opioids were commonly prescribed to tramadol initiators: propoxyphene (40%), codeine (32%), hydromorphone (21%), and meperidine (4%); the following were commonly prescribed to hydrocodone initiators: morphine (25%), propoxyphene (23%), meperidine (17%), and codeine (10%); and the following were commonly prescribed to oxycodone initiators: morphine (45%), propoxyphene (17%), hydromorphone (11%), and codeine (7%). The average days’ supply of initial opioid prescription among tramadol, hydrocodone, and oxycodone initiators were 13.56, 6.98, and 6.78, respectively (data not shown). The characteristics of the tramadol, short-acting hydrocodone, and short-acting oxycodone formulations are described in Table S2.

Table 1.

Baseline characteristics of tramadol initiators, short-acting hydrocodone initiators, and short-acting oxycodone initiators.

	Tramadol (n = 40 276)		Hydrocodone (n = 141 023)		Oxycodone (n = 45 221)
Characteristics	n	Percent	n	Percent	n	Percent
Age, years
Mean ± SD	51.51 ± 14.59		49.78 ± 14.49		50.25 ± 14.25
18–30 years	3809	9.46	15 960	11.32	4800	10.61
31–44 years	8370	20.78	31 888	22.61	9767	21.60
45–54 years	10 272	25.50	36 864	26.14	11 748	25.98
55–64 years	10 918	27.11	36 492	25.88	12 533	27.71
65 years and above	6907	17.15	19 819	14.05	6373	14.09
Sex
Female	26 099	64.80	79 325	56.25	25 137	55.59
Male	14 177	35.20	61 698	43.75	20 084	44.41
Region of residence
East	7799	19.36	23 414	16.60	13 278	29.36
Midwest	10 828	26.88	41 528	29.45	11 918	26.36
South	16 187	40.19	53 919	38.23	12 941	28.62
West	5462	13.56	22 162	15.72	7084	15.67
Index year
2008	4421	10.98	19 362	13.73	6178	13.66
2009	7016	17.42	32 651	23.15	8642	19.11
2010	5957	14.79	25 277	17.92	7054	15.60
2011	6056	15.04	20 781	14.74	6020	13.31
2012	3792	9.42	11 357	8.05	4319	9.55
2013	3096	7.69	8934	6.34	3594	7.95
2014	3240	8.04	8487	6.02	3060	6.77
2015	4546	11.29	10 165	7.21	4127	9.13
2016	1859	4.62	3538	2.51	1850	4.09
2017	293	0.73	471	0.33	377	0.83
Insurance type
Commercial only	28 447	70.63	108 848	77.18	33 677	74.47
Medicaid only	3988	9.90	8558	6.07	2350	5.20
Medicare only	2512	6.24	7167	5.08	2342	5.18
Others/missing	801	1.99	3113	2.21	1113	2.46
Self-insured only	4528	11.24	1337	9.46	5739	12.69
Pain type
Chronic back pain	18 790	46.65	58 490	41.48	16 500	36.49
Chronic neck pain	6649	16.51	22 050	15.64	6544	14.47
Chronic osteoarthritis	28 303	70.27	104 078	73.80	35 625	78.78
Neuropathic pain	2281	5.66	6526	4.63	2396	5.30
Migraine	1594	3.96	4528	3.21	1596	3.53
Abdominal pain	4417	10.97	15 594	11.06	5929	13.11
Chest pain	2947	7.32	8757	6.21	2939	6.50
Other pain conditions	6968	17.30	29 423	20.86	10 828	23.94
Mental health disorders
Major depressive disorder	1176	2.92	3742	2.65	1451	3.21
Anxiety disorders	3033	7.53	9949	7.05	3319	7.34
Mood disorders	518	1.29	1648	1.17	550	1.22
Personality disorders	66	0.16	224	0.16	89	0.20
Post-traumatic stress disorder	164	0.41	546	0.39	219	0.48
Nicotine dependence	311	0.77	1189	0.84	704	1.56
Any surgery	14 462	35.91	69 005	48.93	29 761	65.81
Charlson Comorbidity Index
0	29 520	73.29	109 517	77.66	34 457	76.20
1	7512	18.65	22 851	16.20	7539	16.67
2	1579	3.92	4179	2.96	1504	3.33
3 or more	1665	4.13	4476	3.17	1721	3.81
Baseline drug use
Benzodiazepines	4568	11.34	16 933	12.01	5869	12.98
Non-benzodiazepine hypnotics	340	0.84	1186	0.84	374	0.83
Gabapentinoids	2463	6.12	5343	3.79	1935	4.28
Skeletal muscle relaxants	11 417	28.35	32 017	22.70	7618	16.85
Antidepressants	8679	21.55	29 177	20.69	8799	19.46

When GBTM was used on the weekly measure of any opioid therapy, 3 trajectories were selected as the best-fitting models for hydrocodone and oxycodone, while 4 trajectories were selected for tramadol initiators (Figure 3). The model fit indices are reported in Table S3. The 4 trajectories among tramadol initiators were labeled as early discontinuers (39.00%), late discontinuers (37.90%), intermittent therapy (16.40%), and persistent therapy (6.70%). The average days’ supply of the initial opioid prescription among these 4 trajectories was 7.89, 15.13, 19.37, and 24.51, respectively. Three trajectories were identified among hydrocodone and oxycodone initiators: early discontinuers (hydrocodone: 54.10%; oxycodone: 61.40%), late discontinuers (hydrocodone: 39.40%; oxycodone: 33.30%), and persistent therapy (hydrocodone: 6.50%; oxycodone: 5.30%). Average initial days’ supply in the 3 hydrocodone trajectories was 5.07, 8.05, and 16.69, respectively. Likewise, average initial days’ supply in the respective 3 oxycodone trajectories was 5.23, 7.90, and 18.29.

Figure 3.

Latent trajectories based on group-based trajectory modeling of weekly opioid therapy of tramadol, short-acting hydrocodone, and short-acting oxycodone initiators.

There were 2687, 9169, and 2377 individuals with persistent therapy identified by the GBTM of any opioid therapy for tramadol, hydrocodone, and oxycodone, respectively. GBTM performed on opioid dose (log weekly MME) in each of the persistent therapy groups resulted in 6 similar trajectory groups in each of the 3 opioid initiator groups (Figure 4; Table S4). Three of the trajectories represented consistent therapy, differing on the level of dose, comprising 62%–69% of the samples. One trajectory represented an initial high-MME therapy followed by monotonic decline (8%–12%). The 2 remaining trajectories represented an initial decline followed by a steady increase (10%–13%). Interestingly, no trajectories were found that showed a consistent increase in dose over time.

Figure 4.

Latent trajectories identified from group-based trajectory modeling of opioid dose (morphine milligrams equivalent [MME]) in subsamples of persistent therapy among tramadol (n = 2687), short-acting hydrocodone (n = 9169), and short-acting oxycodone initiators (n = 2377).

In the LLCA of the persistent therapy subset, models with 6 latent classes of opioid therapy type were selected for tramadol and oxycodone initiators, whereas a 7-class model was selected for hydrocodone initiators (Figures 5–7; Table S5). Among tramadol initiators with persistent opioid therapy, the classes were (1) intermittent tramadol-only therapy (29.49%), (2) persistent tramadol-only therapy (19.64%), (3) persistent therapy with switching to other short-acting opioids (19.39%), (4) tramadol-only therapy with persistent initial therapy and intermittent late therapy (13.73%), (5) persistent therapy with addition of other short-acting opioids (11.57%), and (6) persistent therapy with switching to or addition of long-acting opioids (6.17%). Among hydrocodone initiators, the classes were (1) intermittent short-acting hydrocodone-only therapy (29.97%), (2) persistent short-acting hydrocodone-only therapy (24.46%), (3) intermittent therapy with switching to other short-acting opioids (17.47%), (4) short-acting hydrocodone-only therapy with late discontinuation (14.90%), (5) persistent therapy with mixed short-acting opioids (6.27%), (6) persistent therapy with switching to or addition of long-acting opioids (3.10%), and (7) an additional latent class of intermittent therapy with late switching to or addition of long-acting opioids (3.83%). For oxycodone initiators, the latent classes were (1) intermittent short-acting oxycodone-only therapy (28.43%), (2) persistent short-acting oxycodone-only therapy (26.36%), (3) intermittent therapy with switching to other short-acting opioids (23.30%), (4) persistent therapy with switching to or addition of long-acting opioids (11.49%), (5) persistent therapy with mixed short-acting opioids (5.92%), and (6) intermittent therapy with late switching to or addition of long-acting opioids (4.50%).

Figure 5.

Latent trajectories of opioid type identified from longitudinal latent class analysis in individuals with persistent therapy who initiated opioid use on tramadol (n = 2687). sa= short-acting.

Figure 7.

Latent trajectories of opioid type identified from longitudinal latent class analysis in individuals with persistent therapy who initiated opioid use on oxycodone (n = 2377). sa= short-acting.

Figure 6.

Latent trajectories of opioid type identified from longitudinal latent class analysis in individuals with persistent therapy who initiated opioid use on hydrocodone (n = 9169). sa= short-acting.

In the exploratory analyses of characteristics associated with trajectory group membership for any opioid therapy, older age, Medicaid coverage, chronic back pain, neuropathic pain, use of gabapentinoids, use of antidepressants, use of benzodiazepines, and higher levels of comorbidity were consistently associated with more persistent therapy among all 3 study drug initiator groups (Table 2; Tables S6–S8). Each of these factors increased the odds of persistent therapy with any opioid (vs being an early discontinuer) by at least 30% across all the study groups, and Medicaid coverage, compared with commercial insurance, increased the odds of persistent therapy by 300% or more (vs being an early discontinuer). Most of these factors increased the odds in a consistent fashion, where the odds of persistent therapy were higher than the odds of being a late discontinuer, both compared with being an early discontinuer. There were some differences across the 3 initiator groups with regard to the associations with the more persistent trajectories. Any surgery in the baseline period had no association with persistent therapy among tramadol initiators (odds ratio: 1.00 [0.92–1.10]), whereas it was associated with a lower risk of being a persistent user among hydrocodone and oxycodone initiators (hydrocodone odds ratio: 0.63 [0.60–0.66]; oxycodone odds ratio: 0.47 [0.42–0.51]) (Table 2; Tables S6–S8). On the contrary, among tramadol initiators, those with baseline skeletal muscle relaxant therapy had a lower risk of persistent therapy, whereas in the hydrocodone and oxycodone groups, baseline skeletal muscle relaxant therapy was associated with higher risks of persistent therapy (Table 2; Tables S6–S8).

Table 2.

Odds ratios of baseline clinical and demographic variables of being in a non–early-discontinuer trajectory among tramadol initiators (n = 40 276).

	Tramadol initiators
	Reference: Early discontinuers; odds ratio (95% CI)
Characteristics	Late discontinuers	Intermittent therapy	Persistent therapy
Age (ref = 18–30 years)
31–44 years	1.30 (1.20–1.42)	1.46 (1.29–1.67)	1.85 (1.53–2.24)
45–54 years	1.47 (1.35–1.60)	1.96 (1.73–2.22)	2.31 (1.92–2.79)
55–64 years	1.58 (1.45–1.72)	2.43 (2.14–2.75)	2.64 (2.19–3.19)
65 years and above	1.51 (1.36–1.68)	2.33 (2.01–2.69)	2.68 (2.16–3.32)
Sex (ref = male)
Females	0.93 (0.89–0.98)	0.84 (0.78–0.89)	0.78 (0.71–0.85)
Region of residence (ref = East)
Midwest	0.94 (0.88–1.00)	1.10 (1.01–1.20)	1.39 (1.22–1.58)
South	1.01 (0.94–1.07)	1.09 (1.00–1.18)	1.20 (1.06–1.37)
West	1.05 (0.96–1.14)	1.21 (1.09–1.35)	1.27 (1.09–1.48)
Index year (ref = 2008)
2009	1.04 (0.95–1.13)	0.94 (0.84–1.04)	0.93 (0.79–1.10)
2010	0.93 (0.85–1.02)	0.84 (0.75–0.94)	0.89 (0.76–1.06)
2011	0.92 (0.84–1.01)	0.77 (0.69–0.87)	0.86 (0.73–1.01)
2012	0.82 (0.74–0.91)	0.70 (0.62–0.80)	0.71 (0.59–0.86)
2013	0.81 (0.73–0.90)	0.60 (0.52–0.69)	0.61 (0.50–0.75)
2014	0.72 (0.65–0.80)	0.55 (0.48–0.63)	0.50 (0.40–0.61)
2015	0.74 (0.67–0.82)	0.60 (0.53–0.68)	0.78 (0.65–0.93)
2016	0.72 (0.63–0.82)	0.53 (0.45–0.63)	0.48 (0.37–0.61)
2017	0.62 (0.47–0.81)	0.49 (0.34–0.71)	0.44 (0.24–0.79)
Insurance type (ref = commercial)
Medicaid only	1.59 (1.45–1.74)	2.21 (1.97–2.48)	3.83 (3.33–4.42)
Medicare only	1.11 (0.99–1.24)	1.23 (1.07–1.41)	1.30 (1.07–1.58)
Others/missing	1.24 (1.05–1.46)	1.13 (0.91–1.41)	1.63 (1.23–2.17)
Self-insured only	1.16 (1.08–1.25)	1.02 (0.93–1.13)	0.99 (0.85–1.15)
Chronic back pain	1.05 (1.00–1.11)	1.14 (1.06–1.22)	1.24 (1.12–1.36)
Chronic neck pain	1.03 (0.97–1.10)	0.94 (0.86–1.03)	0.89 (0.78–1.01)
Chronic osteoarthritis	1.13 (1.06–1.19)	1.25 (1.16–1.35)	1.12 (1.01–1.25)
Neuropathic pain	1.21 (1.09–1.35)	1.30 (1.13–1.48)	1.34 (1.12–1.61)
Migraine	1.25 (1.11–1.41)	1.21 (1.04–1.42)	1.16 (0.94–1.43)
Abdominal pain	1.07 (1.00–1.16)	0.93 (0.84–1.02)	0.76 (0.66–0.88)
Chest pain	0.92 (0.85–1.01)	0.98 (0.88–1.09)	0.73 (0.61–0.86)
Other pain conditions	1.00 (0.94–1.06)	0.95 (0.88–1.03)	0.88 (0.78–0.99)
Major depressive disorder	0.95 (0.82–1.11)	1.03 (0.86–1.23)	1.02 (0.81–1.27)
Anxiety disorders	1.02 (0.92–1.12)	1.10 (0.97–1.23)	1.36 (1.17–1.58)
Mood disorders	1.22 (0.98–1.51)	1.12 (0.85–1.46)	1.34 (0.98–1.83)
Personality disorders	0.70 (0.37–1.31)	0.70 (0.31–1.58)	2.54 (1.30–4.99)
Post-traumatic stress disorder	1.12 (0.76–1.63)	0.73 (0.44–1.21)	1.50 (0.90–2.48)
Nicotine dependence	0.98 (0.75–1.28)	1.10 (0.80–1.53)	1.11 (0.72–1.71)
Any surgery	1.05 (1.00–1.11)	1.07 (1.00–1.14)	1.00 (0.92–1.10)
Charlson Comorbidity Index (ref = 0)
1	1.14 (1.08–1.22)	1.33 (1.23–1.43)	1.16 (1.04–1.30)
2	1.10 (0.97–1.24)	1.45 (1.25–1.67)	1.49 (1.23–1.82)
3 or more	1.35 (1.19–1.53)	1.47 (1.27–1.71)	1.73 (1.43–2.10)
Benzodiazepines	1.14 (1.05–1.23)	1.30 (1.18–1.43)	1.33 (1.16–1.52)
Non-benzodiazepine hypnotics	1.33 (1.00–1.77)	1.61 (1.17–2.21)	1.80 (1.21–2.69)
Gabapentinoids	1.53 (1.37–1.71)	2.18 (1.93–2.46)	3.03 (2.62–3.51)
Skeletal muscle relaxants	0.96 (0.91–1.01)	0.87 (0.81–0.93)	0.75 (0.67–0.83)
Antidepressants	1.21 (1.14–1.29)	1.56 (1.45–1.69)	1.76 (1.59–1.95)

Sensitivity analyses results

Nearly identical trajectories were observed when the 3-day criterion was used for early refill adjustment (Figure S1). The results for any opioid therapy were also similar to the main analysis when the trajectories were estimated with adjustment for the selected covariates (Figure S2). Finally, we identified 108 564 patients who were excluded for not fulfilling the continuous 12 months of enrollment in the follow-up but would not have been excluded on the basis of other exclusion criteria (Table S9). Those who were excluded were more likely to have started the opioid therapy in the latter years of the index window (2016 and 2017) than were the main sample; however, the demographic and clinical characteristics were similar.

Post-hoc analysis

To better understand the role initial days supplied has on trajectory membership and in particular its role on intermittent and persistent opioid therapy, we performed a post-hoc analysis by restricting the sample to those with 7 or fewer days’ supply on initial prescription. A total of 16 407 (40.74%), 107 005 (75.88%), and 34 568 (76.21%) tramadol, hydrocodone, and oxycodone initiators remained in the sample. Among tramadol initiators, 9.9% and 2.6% had intermittent and persistent therapy (Figure S3), respectively, which are both lower than the proportions from the original sample (16.4% intermittent, 6.7% persistent). More subtle reductions in persistent use were observed for the hydrocodone and oxycodone groups.

Discussion

This study sought to inform decisions about selection of an initial opioid by applying GBTM and LLCA to individuals with chronic non-cancer pain who initiated opioid therapy on tramadol, short-acting hydrocodone, or short-acting oxycodone. The trajectories developed for any opioid therapy showed that there are only subtle differences between oxycodone and hydrocodone initiators, and the proportions of individuals falling into each of the trajectories are comparable between initiators of these drugs. Tramadol initiators, on the other hand, had 4 distinct trajectories and were much less likely to discontinue opioids rapidly. Only 39% of tramadol initiators followed a rapid discontinuation trajectory, compared with 54% of hydrocodone initiators and 61% of oxycodone initiators. Although the proportion of tramadol initiators (6.7%) who were on persistent therapy was higher than that for hydrocodone (6.5%) or oxycodone initiators (5.3%), the differences between the groups were subtle. These persistent therapy findings are consistent with a previous study that used GBTM and monthly intervals and found that 6% of any opioid initiators had persistent therapy.¹³ Tramadol initiators were more likely to maintain intermittent therapy after opioid initiation than were initiators of other short-acting opioids. An intermittent therapy trajectory was not observed for oxycodone or hydrocodone initiators, but 16% of tramadol initiators had intermittent opioid therapy. Studies have shown that individuals on long-term opioid therapy who switch to intermittent therapy have pain intensity that is either similar to or lower than the pain intensity of those who maintain the long-term therapy.^18,19 However, this does not apply to intermittent opioid therapy immediately after opioid initiation, and studies conducted on Washington State workers’ compensation data reported that the majority of opioid overdoses occurred in intermittent or low-dose users.^20,21 This suggests that the risk–benefit profile of intermittent opioid therapy needs to be carefully evaluated.

Tramadol prescribing has increased over the years,²² and tramadol has been perceived to be safer than opioids such as hydrocodone and oxycodone, likely because of its schedule IV status.²³ This is reflected, in part, in the finding that tramadol initiators had greater days’ supply for the initial opioid prescription than did hydrocodone and oxycodone initiators, which might also account for the decreased likelihood of rapid discontinuation of tramadol. This is evidenced by the observation that the initial days’ supply was substantially different across the trajectory groups, with those prescribed a greater initial days’ supply being much more likely to have persistent therapy. In an attempt to disentangle the effect of the initial days’ supply from the perceived safer pharmacological profile of tramadol, the sensitivity analysis, which restricted the sample to those who had 7 or fewer initial days’ supply, still found that nearly 10% had intermittent therapy. This was lower than the value in the original analysis, but it makes clear that the initial days’ supply alone does not drive tramadol’s intermittent therapy trajectory.

Among individuals on persistent opioid therapy, MME trajectories were similar across the subsamples of tramadol, hydrocodone, and oxycodone initiators; however, the level of MME was higher among the oxycodone persistent therapy subsample. Individuals with persistently high MME among oxycodone initiators had average total weekly MMEs of more than 400, whereas the corresponding average weekly MMEs among the tramadol and hydrocodone subsamples were around 150. Prior studies have shown that higher MME use increases the risks of opioid use disorder and overdoses.^24,25 A study by Dunn et al. reported that individuals receiving 50–100 MME/day had a 273% increase in the rate of overdose as compared with individuals receiving 20–49 MME/day.²⁴ Likewise, another study reported doubling of the hazard of an opioid use disorder diagnosis with a dose of 50 MME/day or more on the initial script.²⁵ This suggests that the risk of adverse events could be higher in oxycodone initiators with persistent therapy, and the risk should be evaluated regularly against the benefits of pain control and better quality of life.

Dose escalation has been consistently associated with increased risk of overdose, opioid use disorder, accidents, and lower quality of life^26–28; however, we did not identify a trajectory in which consistent dose escalation was observed in any of the 3 subgroups of persistent users. A study documented that the release of a Centers for Disease Control and Prevention guideline on opioid prescribing reduced the population based rates of high-dose opioid prescriptions by 8 prescriptions per 100 000 persons per month.²⁹ Dose-tapering initiatives have also been adopted by many health systems, which could partly explain why continuous dose escalation was not observed, at least for the later years of the data.^30,31 We did, however, identify 2 trajectories in each of the 3 subsamples that represented dose escalation after an initial reduction, which comprised 22%–26% of the subsamples of those on persistent therapy. This finding should be interpreted with caution, however. Despite focusing the analysis on dose among those identified as having persistent therapy, the LLCA results reveal that among these, the probability of opioid therapy varies over time, and only 2–3 of the 6–7 LLCA trajectories show that opioid therapy is consistently at or near 100%. Several of the LLCA trajectories reveal that the probability of opioid receipt might decline and then rebound in later periods, which could explain, in part, the dose trajectories that were characterized by declining doses followed by increases in dose.

In the LLCA, 3 of the latent classes for the tramadol subsample included tramadol-only therapy with different levels of therapy (persistent, intermittent, and initial persistent with late intermittent), comprising 63% of the sample. Similarly, 69% and 54% of hydrocodone and oxycodone subsamples included short-acting hydrocodone-only and short-acting oxycodone-only therapy latent classes, indicating that the majority of individuals remained on the opioid initially prescribed. This indicates that the choice of first opioid script could have long-term implications. Six percent of the tramadol subsample switched to or added long-acting opioids, while 15% of the oxycodone subsample had an immediate (11%) or a delayed (4%) switch to or addition of long-acting opioids. The Centers for Disease Control and Prevention guideline on opioid prescribing for chronic non-cancer pain recommends against starting opioid treatment with long-acting formulations and recommends them only if adequate pain control is not achieved from short-acting opioids, non-opioid medications, and nonpharmacological interventions.³² It has also been recommended that long-acting formulations be tried only after a maximum of 8 doses of short-acting opioids per day does not provide meaningful pain control.³³ The late switching to or addition of long-acting opioids in our study appears to be in keeping with the clinical recommendations; however, the immediate use of long-acting opioids in all 3 subsamples could potentially incur more risk than benefits. Existing evidence does not indicate better efficacy of long-acting opioids, although they are better at maintaining stable analgesia in patients with persistent pain.^34,35 Furthermore, the use of long-acting opioids has been associated with increased risk of overdose and opioid use disorder.^36–38 Our study provides evidence that a fraction of individuals with persistent therapy in all 3 opioid initiator groups immediately start long-acting opioids and more ultimately add or switch to long-acting opioids. Long-acting opioid therapy was more likely to be observed among individuals initiating short-acting oxycodone. Given the similar efficacy and higher risk of long-acting formulations, clinicians should consider this differential likelihood for long-acting opioids before opioid initiation.

The exploratory multinomial logistic regressions identified older age, Medicaid insurance, higher comorbidities, benzodiazepine use, gabapentinoid use, and antidepressant use, among others, as correlates of persistent opioid therapy in all 3 initiator groups. In a study conducted on workers’ compensation data, incident opioid users with age 60 years or more had 92% higher odds of persistent opioid use than did 18- to 29-year-old individuals.³⁹ The incidence of long-term opioid use among incident opioid users in non-Medicaid plans has been estimated to be between 3% and 8%,^40,41 whereas it was around 14% for Medicaid enrollees.⁴² A systematic review reported that with an increase in the number of physical comorbidities, there was a 16% increase in the odds of persistent therapy.⁴³ The use of benzodiazepines and antidepressants has been associated with long-term opioid use.⁴⁴ In a study conducted in veterans, gabapentinoid use was associated with a 158% increase in the risk of persistent opioid use at 1 year of follow-up.⁴⁴ The risks associated with co-prescribing these drugs with opioids are well established; however, better treatment strategies are required for patients with complex health needs. Baseline surgery was associated with a lower risk of persistent therapy in hydrocodone and oxycodone initiators, whereas there was no association in tramadol initiators. This could be because hydrocodone and oxycodone were likely prescribed for postoperative pain in subjects who had undergone surgery and were discontinued after the surgical pain subsided.

The study should be interpreted in light of several limitations. First, the group-based trajectory method estimates discrete trajectories that encompass the underlying data distribution and does not quantify individual-level heterogeneity. Although we describe the heterogeneity at the trajectory level (Table S10), the overall individual-level heterogeneity was not assessed. Second, the relationships studied between trajectory group and demographic and clinical characteristics should be considered as exploratory and not causal, as the analysis was not conducted with a primary exposure of interest and adequate control for confounding. Third, the study database does not have information on variables such as pain intensity that could explain persistent use and high dose trajectories across the 3 drug groups. Fourth, information is lacking on opioid prescriptions paid out of pocket, which could translate to a few individuals being misclassified or trajectory groups being mis-estimated. Fifth, we did not account for opioid analgesics used in an inpatient/outpatient facility or office setting. A study reported that opioid use in hospitals after surgery and before discharge predicted opioid use at home.⁴⁵ Relatedly, another study reported that 8% of patients with inpatient opioid use within 12 hours before discharge were still using opioids 90 days after discharge, as compared with 4% of those without opioid use before discharge.⁴⁶ Not accounting for such opioid therapy could have resulted in misclassification of individuals into trajectories. Sixth, we did not study the effect of clinical guidelines or state-level opioid policies on the use trajectories. Seventh, as we required continuous enrollment in 1-year follow-up, which created a selected sample of those who maintained coverage for at least 1 year, selection bias would be likely if patients disenrolled for opioid- or pain-related reasons.

In summary, opioid therapy patterns differed meaningfully by the initial opioid. Approximately 15% of tramadol initiators appear to have intermittent opioid therapy, whereas intermittent therapy was not detected among short-acting hydrocodone and oxycodone initiators. This difference is likely due to a combination of greater days’ supply for tramadol and factors such as the perceived safety of tramadol. Furthermore, persistent therapy was similarly low (5%–7%) among hydrocodone, tramadol, and oxycodone initiators, suggesting that tramadol might not be as safe as its drug schedule leads us to believe. Initiators of short-acting oxycodone who had persistent therapy had higher MME doses than those in the tramadol and hydrocodone groups. Oxycodone initiators were also more likely to have long-acting opioid prescriptions over the next year than were tramadol and hydrocodone initiators, suggesting that long-term therapy among oxycodone initiators could be associated with higher risk of adverse outcomes. As a result, clinicians who are planning for long-term therapy for patients with chronic non-cancer pain might consider tramadol or short-acting hydrocodone as the initial opioid in place of short-acting oxycodone. Additionally, individualized treatment goals and regular assessment of risk–benefit profile, as has been recommended, could prevent patients from having potentially risky opioid use trajectories.

Supplementary material

Supplementary material is available at Pain Medicine online.

Funding

No funding was received for this study. The data used for this study were supported by the UAMS Translational Research Institute (TRI), NIH grant UL1TR000039.

Conflicts of interest: B.C.M. has received royalties from TrestleTree, LLC, for the commercialization of an opioid risk prediction tool that is unrelated to this investigation. Other authors have no conflicts to disclose.