Opioid Analgesic Recommendation Approaches

• Opioid Analgesic Recommendation is the process of tailoring opioid dosing, adjunct strategies, and refill protocols to provide effective pain relief while minimizing adverse outcomes.
• Studies employ methods like deep reinforcement learning, causal estimands, and mark-specific recurrent-event models to quantify opioid-sparing effects and optimize dosing in various clinical settings.
• Clinical implications include personalized pain management, reduced opioid exposure in trauma and chronic pain, and dosage-specific refill guidance based on patient demographics and surgical context.

Searching arXiv for the cited papers to ground the article in current records. Opioid analgesic recommendation denotes the formal and clinical problem of selecting opioid dosing, duration, co-medication, and refill strategies so that analgesia is preserved while unnecessary exposure and downstream harms are reduced. In the cited work, this problem is addressed through four distinct but related paradigms: deep reinforcement learning for morphine dosing in critical care, causal estimands for NSAID add-on policies that quantify opioid-sparing effects, modified treatment policy analyses that isolate associations between pain-management strategies and incident opioid use disorder (OUD), and mark-specific recurrent-event models that estimate dosage-specific refill hazards after surgery (Lopez-Martinez et al., 2019, Stoltenberg et al., 19 Aug 2025, Rudolph et al., 2024, Yang et al., 14 Sep 2025).

1. Conceptual scope of opioid analgesic recommendation

In critical care, the recommendation problem is framed as sequential decision making. Opioids are the preferred medications for the treatment of pain in the intensive care unit, but undertreatment leads to unrelieved pain and poor clinical outcomes, whereas excessive use puts patients at risk of experiencing multiple adverse effects. The morphine-focused reinforcement-learning work therefore treats recommendation as a real-time, patient-specific dosing problem conditioned on evolving pain and physiological state (Lopez-Martinez et al., 2019).

In trauma and post-discharge care, recommendation is framed differently. The central question is whether supplementing opioid treatments with other analgesics, such as nonsteroidal anti-inflammatory drugs (NSAIDs), can reduce opioid consumption, and the relevant policy object is not a fixed dosing rule but an add-on regime: a time-varying strategy that adds NSAIDs whenever opioids are administered while preserving the naturally intended NSAID level when opioids are not administered. This contrasts with static regimes, which assign NSAIDs at predefined time points regardless of clinical context, and with conventional dynamic regimes, which define treatment decisions at every time point during the treatment period (Stoltenberg et al., 19 Aug 2025).

In chronic pain management, recommendation is inseparable from misuse risk. The Medicaid analysis considered nine pain-management treatments—prescription opioid dose, duration, number of opioid prescribers, opioid co-prescription with benzodiazepines, muscle relaxants, and gabapentinoids, non-opioid pain prescription, physical therapy, and other pain treatment modality—and estimated their associations with incident OUD while holding other pain-management treatments at their observed levels (Rudolph et al., 2024).

In postoperative care, recommendation is dosage-specific rather than merely event-specific. The recurrent-event framework treats refill dosage as a continuously varying mark, allowing inference about whether particular patient factors are associated with low-dosage or high-dosage refill hazards rather than with refill occurrence alone. This makes the recommendation problem one of selecting both whether and how much opioid to prescribe at refill (Yang et al., 14 Sep 2025).

2. Sequential morphine dosing in critical care

The critical-care study is titled "Deep Reinforcement Learning for Optimal Critical Care Pain Management with Morphine using Dueling Double-Deep Q Networks" and presents a sequential decision making framework for opioid dosing based on deep reinforcement learning. It focuses on morphine, one of the most commonly prescribed opioids, and states that the framework provides real-time clinically interpretable dosing recommendations, personalized according to each patient’s evolving pain and physiological condition. The abstract further states that retrospective data from the publicly available MIMIC-3 database were used to train and evaluate the model, and that the results demonstrate that reinforcement learning may be used to aid decision making in the intensive care setting by providing personalized pain management interventions (Lopez-Martinez et al., 2019).

The named architecture is Dueling Double-Deep Q Networks (DDDQN). The supplied synthesis states that standard DDDQN equations are implied but not reported explicitly in the paper text. Those implied formulations are:

$$Q(s,a; \theta, \alpha, \beta) = V(s; \theta, \beta) + A(s,a; \theta, \alpha) - \frac{1}{|\mathcal{A}|}\sum_{a'} A(s,a'; \theta, \alpha)$$

$$y = r + \gamma \; Q\big(s', \arg\max_{a'} Q(s',a'; \theta); \theta^{-}\big)$$

with squared-error loss

$$L(\theta) = \mathbb{E}\big[(y - Q(s,a; \theta))^2\big].$$

The same synthesis also makes clear the limits of what can be said from the provided text. Cohort details, inclusion and exclusion criteria, ICU types, sample size, demographics, state variables, time discretization, morphine dosing schema, reward design, optimizer, learning rate, batch size, training epochs, baseline comparators, off-policy evaluation method, quantitative performance, and clinical integration mechanisms are all described as not reported. Consequently, the paper supports the existence of a morphine recommendation framework, but not any extractable bedside morphine dosing protocol or safety rule beyond that high-level claim (Lopez-Martinez et al., 2019).

A common misconception is that an algorithmic opioid recommendation is synonymous with an explicit titration schedule. In this instance, that would overstate what the available text supports. What is reported is a DDDQN-based sequential decision framework for personalized morphine dosing; the operationalization of “personalized,” the action constraints, and the evaluation details are not reported.

3. Add-on NSAID regimes and opioid-sparing recommendation

The add-on regime framework was developed to quantify opioid-sparing effects of NSAIDs in a policy-relevant way. Formally, the add-on rule is

$$A_t^{g_{\text{add},j}} = \begin{cases} j, & Y_t^g > 0,\ A_t^g, & Y_t^g = 0, \end{cases} \quad j \in \{0,1\}.$$

Here, $j=1$ means “add NSAIDs when opioids are given,” and $j=0$ means “withhold NSAIDs when opioids are given.” When opioids are not administered, the regime requires no intervention and preserves the natural level of NSAIDs. This differs from static regimes such as “always NSAIDs” versus “never NSAIDs,” and from conventional dynamic regimes that assign NSAIDs at every time point based on the current history (Stoltenberg et al., 19 Aug 2025).

The opioid-sparing estimand is defined in terms of cumulative opioid consumption, measured in Oral Morphine Equivalents (OMEQ), interchangeable with MME. If $Y^g := \sum_{k=1}^{K} Y_k^g$, then the difference contrast is

$$\Delta_{\text{diff}} = \mathbb{E}[Y^{g_{\text{add},1}}] - \mathbb{E}[Y^{g_{\text{add},0}}],$$

with negative values indicating opioid-sparing. Identification relies on consistency, positivity, and sequential exchangeability. The paper argues that these assumptions are easier to assess than those used in existing methods because the add-on regime intervenes only when opioids are given, so the relevant adjustment set is anchored to opioid decision points rather than to all treatment times (Stoltenberg et al., 19 Aug 2025).

Estimation is given through the parametric g-formula, inverse probability weighting (IPW), and broader doubly robust approaches. The implementation described in the application used parametric models and Monte Carlo simulation through the R package gfoRmula, while the broader identification result is stated to enable sequentially doubly robust estimators, for example via the R package lmtp (Stoltenberg et al., 19 Aug 2025).

The empirical application used NTRplus, which links the Norwegian National Trauma Registry to the Prescription Database, Cause of Death Registry, Norwegian Patient Registry, and Statistics Norway. The cohort comprised $N=8{,}718$ trauma patients from 2015–2018 with serious injuries; inclusion required discharge date, NSAID eligibility, at least one post-discharge opioid dispensation in month 0, and survival through month 0. Follow-up used months $k \in \{0,\dots,21\}$, with month 0 defined as the first month after initial opioid dispensation within the first post-discharge month, and examined both a short treatment period $$y = r + \gamma \; Q\big(s', \arg\max_{a'} Q(s',a'; \theta); \theta^{-}\big)$$0 and a long treatment period $$y = r + \gamma \; Q\big(s', \arg\max_{a'} Q(s',a'; \theta); \theta^{-}\big)$$1. No censoring occurred (Stoltenberg et al., 19 Aug 2025).

Contrast	Treatment period	Total opioid dose difference (95% CI)
Static “always NSAID” vs “never NSAID”	Short, $$y = r + \gamma \; Q\big(s', \arg\max_{a'} Q(s',a'; \theta); \theta^{-}\big)$$2	$$y = r + \gamma \; Q\big(s', \arg\max_{a'} Q(s',a'; \theta); \theta^{-}\big)$$3 OMEQ $$y = r + \gamma \; Q\big(s', \arg\max_{a'} Q(s',a'; \theta); \theta^{-}\big)$$4
Add-on-1 vs add-on-0	Short, $$y = r + \gamma \; Q\big(s', \arg\max_{a'} Q(s',a'; \theta); \theta^{-}\big)$$5	$$y = r + \gamma \; Q\big(s', \arg\max_{a'} Q(s',a'; \theta); \theta^{-}\big)$$6 OMEQ $$y = r + \gamma \; Q\big(s', \arg\max_{a'} Q(s',a'; \theta); \theta^{-}\big)$$7
Static “always NSAID” vs “never NSAID”	Long, $$y = r + \gamma \; Q\big(s', \arg\max_{a'} Q(s',a'; \theta); \theta^{-}\big)$$8	$$y = r + \gamma \; Q\big(s', \arg\max_{a'} Q(s',a'; \theta); \theta^{-}\big)$$9 OMEQ $$L(\theta) = \mathbb{E}\big[(y - Q(s,a; \theta))^2\big].$$0
Add-on-1 vs add-on-0	Long, $$L(\theta) = \mathbb{E}\big[(y - Q(s,a; \theta))^2\big].$$1	$$L(\theta) = \mathbb{E}\big[(y - Q(s,a; \theta))^2\big].$$2 OMEQ $$L(\theta) = \mathbb{E}\big[(y - Q(s,a; \theta))^2\big].$$3

Monthly patterns were negative early and attenuating toward zero over time for the short treatment period, whereas the long treatment period remained negative throughout follow-up. Static regimes showed larger negative effects than add-on regimes, but were described as less clinically relevant. The paper’s interpretation is that NSAID add-on is associated with clinically meaningful opioid-sparing in this trauma cohort, particularly when the add-on rule is applied over a longer treatment period (Stoltenberg et al., 19 Aug 2025).

The practical recommendations follow directly from this structure. NSAIDs are to be added when opioids are initiated or dispensed for acute post-trauma pain, provided there are no contraindications; NSAIDs are not to be forced when opioids are not used; and safety screening is to include renal impairment, active or high-risk gastrointestinal bleeding, peptic ulcer disease, uncontrolled heart failure, known NSAID hypersensitivity, late pregnancy, high bleeding risk, and intensive antiplatelet or anticoagulant therapy. The suggested implementation steps include confirming opioid indication, screening for contraindications and drug interactions, initiating NSAID concurrently with the first opioid dose if safe, reassessing pain control and adverse effects, titrating opioids downward as adequate analgesia is achieved, repeating trigger-based reassessment at subsequent opioid administrations, and documenting decisions and outcomes (Stoltenberg et al., 19 Aug 2025).

4. Chronic pain prescribing and incident opioid use disorder

The Medicaid study addressed opioid analgesic recommendation by estimating the extent to which pain-management strategies were associated with incident OUD among chronic pain patients, controlling for baseline demographic and clinical confounding variables and holding other pain-management treatments at their observed levels. It used T-MSIS Analytic Files from 25 Medicaid expansion states, years 2016–2019, and included non-pregnant, non-dual-eligible Medicaid adults aged 35–64 with 12 months of continuous enrollment. Two chronic pain subgroups were analyzed: chronic pain with co-occurring physical disability (CPPD; $$L(\theta) = \mathbb{E}\big[(y - Q(s,a; \theta))^2\big].$$4) and chronic pain without physical disability (CP; $$L(\theta) = \mathbb{E}\big[(y - Q(s,a; \theta))^2\big].$$5). Baseline covariates and subgroup status were defined in months 1–6, exposures in months 7–12, and incident OUD outcomes in months 13–24, with a shorter months 13–18 window used in sensitivity analysis (Rudolph et al., 2024).

The causal estimand was a modified treatment policy (MTP) effect of the form

$$L(\theta) = \mathbb{E}\big[(y - Q(s,a; \theta))^2\big].$$6

where binary treatments could be set to 1 for all and continuous treatments could be multiplied by $$L(\theta) = \mathbb{E}\big[(y - Q(s,a; \theta))^2\big].$$7. Estimation used a doubly robust, nonparametric targeted minimum loss-based estimator (TMLE) for longitudinal MTP effects, with cross-fitting and Super Learner ensembles comprising GLMs, gradient boosted machines, multivariate adaptive regression splines, neural networks, random forests, and intercept-only learners. The measure of association reported was the adjusted relative risk (RR) with 95% confidence intervals (Rudolph et al., 2024).

The main prescribing-relevant findings were concentrated in concurrency, dose, and duration. Concurrent opioids plus benzodiazepines were associated with RR $$L(\theta) = \mathbb{E}\big[(y - Q(s,a; \theta))^2\big].$$8 $$L(\theta) = \mathbb{E}\big[(y - Q(s,a; \theta))^2\big].$$9 in CP and RR $$A_t^{g_{\text{add},j}} = \begin{cases} j, & Y_t^g > 0,\ A_t^g, & Y_t^g = 0, \end{cases} \quad j \in \{0,1\}.$$0 $$A_t^{g_{\text{add},j}} = \begin{cases} j, & Y_t^g > 0,\ A_t^g, & Y_t^g = 0, \end{cases} \quad j \in \{0,1\}.$$1 in CPPD. Concurrent opioids plus gabapentinoids were associated with RR $$A_t^{g_{\text{add},j}} = \begin{cases} j, & Y_t^g > 0,\ A_t^g, & Y_t^g = 0, \end{cases} \quad j \in \{0,1\}.$$2 $$A_t^{g_{\text{add},j}} = \begin{cases} j, & Y_t^g > 0,\ A_t^g, & Y_t^g = 0, \end{cases} \quad j \in \{0,1\}.$$3 in CP and RR $$A_t^{g_{\text{add},j}} = \begin{cases} j, & Y_t^g > 0,\ A_t^g, & Y_t^g = 0, \end{cases} \quad j \in \{0,1\}.$$4 $$A_t^{g_{\text{add},j}} = \begin{cases} j, & Y_t^g > 0,\ A_t^g, & Y_t^g = 0, \end{cases} \quad j \in \{0,1\}.$$5 in CPPD. Maximum daily opioid dose in MME, modeled as a 20% increase in observed maximum daily MME, was associated with RR $$A_t^{g_{\text{add},j}} = \begin{cases} j, & Y_t^g > 0,\ A_t^g, & Y_t^g = 0, \end{cases} \quad j \in \{0,1\}.$$6 $$A_t^{g_{\text{add},j}} = \begin{cases} j, & Y_t^g > 0,\ A_t^g, & Y_t^g = 0, \end{cases} \quad j \in \{0,1\}.$$7 in CP. Proportion of days covered (PDC) by opioids, modeled as a 20% increase, was associated with RR $$A_t^{g_{\text{add},j}} = \begin{cases} j, & Y_t^g > 0,\ A_t^g, & Y_t^g = 0, \end{cases} \quad j \in \{0,1\}.$$8 $$A_t^{g_{\text{add},j}} = \begin{cases} j, & Y_t^g > 0,\ A_t^g, & Y_t^g = 0, \end{cases} \quad j \in \{0,1\}.$$9 in CP and a significant increase of approximately $j=1$0 per 20% higher PDC in CPPD. Physical therapy was associated with RR $j=1$1 $j=1$2 in CP. By contrast, muscle relaxant co-prescriptions, number of opioid prescribers, non-opioid pain prescriptions, and the composite of other pain modalities were not significantly associated with OUD after adjustment (Rudolph et al., 2024).

These estimates were translated into explicit guidance. The study recommends avoiding concurrent opioids plus benzodiazepines whenever possible and avoiding concurrent opioids plus gabapentinoids unless a strong indication exists and risks are mitigated. It recommends using the lowest effective daily MME, avoiding rapid escalation, minimizing total days covered, planning shorter courses and longer opioid-free intervals, centralizing opioid prescribing in one clinician or team and one pharmacy, and offering physical therapy as first-line or adjunctive treatment. The study also recommends prescription drug monitoring programs, overdose education, naloxone provision, and frequent reassessment, including within 1–2 weeks of initiation or dose change and monthly thereafter if opioids are continued (Rudolph et al., 2024).

A common misconception is that OUD risk can be managed primarily through daily dose ceilings alone. The Medicaid analysis shows that duration and continuity, represented by PDC, and concurrent benzodiazepine or gabapentinoid exposure were at least as salient as maximum daily MME.

5. Dosage-specific refill recommendation after surgery

The post-surgical refill paper proposes a mark-specific proportional hazards model for recurrent events on the gap-time scale:

$j=1$3

where $j=1$4 is a continuous dosage mark, for example oral morphine equivalents (OMEs) of the refill, and $j=1$5 varies smoothly with dosage. Estimation at a target dosage $j=1$6 uses two weights: proximity-to-mark weighting $j=1$7 and informative recurrence weighting $j=1$8, with combined weight $j=1$9. The estimator is obtained through a weighted pseudo-partial likelihood and Newton–Raphson, with inference based on a sandwich variance estimator. Under smoothness, bounded-moment, positivity, kernel, bandwidth, and process assumptions, the paper establishes uniform consistency and asymptotic normality (Yang et al., 14 Sep 2025).

In simulation, the proposed mark-specific model was near-unbiased across marks, with average bias typically between $j=0$0 and $j=0$1, and achieved coverage of approximately $j=0$2 at interior marks; the non-mark-specific comparator showed substantial bias and poor coverage at extremes. Average estimated standard errors closely matched empirical standard deviations, which the paper presents as support for the robust sandwich variance estimator (Yang et al., 14 Sep 2025).

The applied analysis linked 11,492 opioid-naïve surgical patients from the Michigan Surgical Quality Collaborative (2017–2019) to the Michigan Automated Prescription System. Time at risk began on the day of surgery. Median refills were 2 (IQR 1–5), median gap-time between refills was 64 days (IQR 13–434), and median refill dosage was 112.5 OMEs (IQR 60–300). Dosage per refill was transformed by square root and normalized to $j=0$3, and analyses were reported at 50, 100, 150, 300, and 450 OMEs (Yang et al., 14 Sep 2025).

The salient results were dosage-differential. Smoking was associated with elevated hazards across dosages, especially at higher dosages, including HR $j=0$4 $j=0$5 at 300 OMEs and HR $j=0$6 $j=0$7 at 450 OMEs. Very high BMI ($j=0$8) was associated with higher hazards of high-dosage refills versus BMI 25–30 and little or no elevation at lower dosages. Cancer was not significantly different at 50–150 OMEs but was associated with significantly higher hazards at 300–450 OMEs. Open surgery was associated with higher hazards at high dosages, including HR $j=0$9 $Y^g := \sum_{k=1}^{K} Y_k^g$0 at 300 OMEs, while low-dosage differences were not significant. Inpatient surgeries elevated refill hazards across all dosages. Black race versus White race was associated with higher hazards of low-dosage refills and lower hazards of high-dosage refills, with HR $Y^g := \sum_{k=1}^{K} Y_k^g$1 $Y^g := \sum_{k=1}^{K} Y_k^g$2 at 50 OMEs, HR $Y^g := \sum_{k=1}^{K} Y_k^g$3 $Y^g := \sum_{k=1}^{K} Y_k^g$4 at 100 OMEs, HR $Y^g := \sum_{k=1}^{K} Y_k^g$5 $Y^g := \sum_{k=1}^{K} Y_k^g$6 at 300 OMEs, and HR $Y^g := \sum_{k=1}^{K} Y_k^g$7 $Y^g := \sum_{k=1}^{K} Y_k^g$8 at 450 OMEs (Yang et al., 14 Sep 2025).

The paper translates these findings into dosage-specific guidance. For minor outpatient general surgery such as laparoscopic cholecystectomy, laparoscopic appendectomy, and inguinal hernia, it recommends starting at 50–100 OMEs total and targeting 15–30 MME/day for $Y^g := \sum_{k=1}^{K} Y_k^g$9 days, with reassessment if use exceeds 5 days. For intermediate procedures or higher expected pain burden, including open approach or inpatient stay, it recommends starting at 100–200 OMEs total for 3–5 days and never exceeding 300 OMEs without senior clinical review, with an MME/day ceiling of 30–40 for $$\Delta_{\text{diff}} = \mathbb{E}[Y^{g_{\text{add},1}}] - \mathbb{E}[Y^{g_{\text{add},0}}],$$0 days. In higher-risk groups—high BMI, smokers, cancer, open surgery, or inpatient surgery—the recommended adjustments are to use the lower end of initial ranges, prefer multimodal analgesia, implement an explicit taper reducing daily opioid dose by 20–25% every 1–2 days as pain improves, avoid automatic refills, require a structured pain-function check and PDMP review, keep warranted refills to $$\Delta_{\text{diff}} = \mathbb{E}[Y^{g_{\text{add},1}}] - \mathbb{E}[Y^{g_{\text{add},0}}],$$1 OMEs, and require in-person reassessment if $$\Delta_{\text{diff}} = \mathbb{E}[Y^{g_{\text{add},1}}] - \mathbb{E}[Y^{g_{\text{add},0}}],$$2 OMEs are requested (Yang et al., 14 Sep 2025).

The paper also gives an equity-oriented interpretation of the race-specific hazard pattern. Because Black patients showed higher hazards of low-dosage refills but lower hazards of high-dosage refills, it recommends standardized, procedure-specific dosing ranges applied uniformly across race, escalation and de-escalation based on pain and function rather than assumptions, and service-line audits of prescribing by race (Yang et al., 14 Sep 2025).

6. Limitations, controversies, and emerging synthesis

The most important limitation across these lines of work is heterogeneity in what is actually estimated. The critical-care reinforcement-learning paper establishes the existence of a DDDQN morphine recommendation framework, but the supplied text does not report cohort construction, state and action spaces, reward function, training details, safety rules, baseline comparators, or quantitative outcomes. By contrast, the trauma, Medicaid, and postoperative studies report formal estimands, identification assumptions, and effect estimates in much greater detail (Lopez-Martinez et al., 2019, Stoltenberg et al., 19 Aug 2025, Rudolph et al., 2024, Yang et al., 14 Sep 2025).

A second limitation is observational measurement. In the NSAID add-on study, dispensation records do not include over-the-counter and institutional dispensations, and dispensation is not the same as actual ingestion. In the Medicaid study, exposures and outcomes are claims-based, over-the-counter non-opioids are poorly captured, and residual confounding remains possible despite rich baseline and time-updated covariates. In the postoperative recurrent-event study, marks reflect OMEs per refill rather than within-refill dose variability or adherence, and boundary marks are less stable because kernel localization reduces effective sample size (Stoltenberg et al., 19 Aug 2025, Rudolph et al., 2024, Yang et al., 14 Sep 2025).

The literature also corrects several recurring misunderstandings. One is that clinically relevant opioid-sparing policies must be static schedules. The add-on regime work argues that static NSAID rules can be clinically infeasible or ethically inappropriate because they disregard context, while trigger-based add-on policies preserve natural NSAID use when opioids are not administered. Another is that refill risk is dosage-invariant; the mark-specific recurrent-event results show that smoking, BMI, cancer, surgical approach, and race may have different associations at 50 OMEs than at 300 or 450 OMEs. A third is that safer opioid recommendation can be reduced to lowering nominal dose alone; the Medicaid results indicate that continuity of exposure and co-prescribing patterns are central (Stoltenberg et al., 19 Aug 2025, Rudolph et al., 2024, Yang et al., 14 Sep 2025).

The reported research directions are correspondingly heterogeneous. The NSAID add-on paper calls for randomized target trials, heterogeneity analyses by injury type, age, comorbidity, and baseline opioid use, study of optimal NSAID agent, dose, and duration, investigation of broader multimodal bundles and combined add-on-and-opioid-down-titration regimes, enhanced data capture of actual administration and over-the-counter use, and methods for time-varying effect modification, mediation via competing events, and robustness to unmeasured confounding. The postoperative refill paper proposes embedding mark-specific estimators into electronic health record decision support so that intended dosage marks can be evaluated in real time, with hazard estimates, covariate-specific hazard ratios, and uncertainty conveyed through robust standard errors and confidence intervals (Stoltenberg et al., 19 Aug 2025, Yang et al., 14 Sep 2025).

Taken together, these studies define opioid analgesic recommendation not as a single dosing rule but as a family of decision problems. In one setting it is sequential morphine control in the ICU; in another it is trigger-based multimodal analgesia that preserves clinical context; in another it is minimizing OUD-associated exposures in chronic pain care; and in another it is dosage-specific refill governance after surgery. The shared technical theme is that recommendation improves when the target of inference is aligned with the actual clinical decision—dose, duration, concurrency, timing of adjuncts, or refill size—rather than collapsed into a single undifferentiated opioid-use measure.