Secondary Logo

Journal Logo

Contents: Original Research

Development and Validation of a Model to Predict Postdischarge Opioid Use After Cesarean Birth

Osmundson, Sarah S. MD, MS; Halvorson, Alese MS; Graves, Kristin N. BA; Wang, Clara BA; Bruehl, Stephen PhD; Grijalva, Carlos G. MD, MPH; France, Dan PhD, MPH; Hartmann, Katherine MD, PhD; Mokshagundam, Shilpa MD; Harrell, Frank E. Jr PhD

Author Information
doi: 10.1097/AOG.0000000000004759

More than half of opioids prescribed after surgery go unused.1 Excess prescribing generates opioids available for nonmedical use with important public health implications. For years, the opioid epidemic in the United States has paralleled trends in legal opioid prescribing.2 Most individuals engaging in nonmedical use report using opioids legitimately prescribed in the past for their own care or prescribed to family members or friends.3 Additionally, larger opioid prescriptions are associated with persistent opioid use.4–7 Yet, opioid prescribing recommendations after surgery remain vague, but urge clinicians to prescribe the lowest dose and shortest duration possible.8

Cesarean birth is the most common major surgical procedure in the United States,9 with 1.2 million cesarean births performed annually.10 Opioid prescribing after cesarean birth is problematic as up to 75% of prescribed opioids go unused,11 and opioids prescribed postoperatively are linked to serious opioid-related adverse events.12,13 Yet, a consistent minority of patients report using all prescribed opioids and having unmet pain needs.11,14 Uniform reductions in opioid prescribing risk poor pain management in vulnerable patients with implications for postpartum depression and other adverse outcomes.15 Although the American College of Obstetricians and Gynecologists recommends that, “Therapy should be individualized based on the patient's condition…,”16 most prescribing after cesarean birth reflects clinician prescribing patterns rather than patient characteristics.17 Clinicians lack tools and concrete recommendations for efficiently individualizing opioid prescribing.

To address these needs, we collected detailed patient characteristics and opioid use data to develop, validate, and calibrate a prediction model for postdischarge opioid use in patients undergoing cesarean birth.


We conducted a prospective cohort study of patients older than age 18 years who were undergoing cesarean delivery between November 15, 2019, and January 15, 2021, at a single academic medical center. Patients who did not speak English or Spanish, had diagnoses of substance use disorder or chronic pain, experienced major birth-related complications, or tested positive for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection were excluded. All procedures were approved by the Vanderbilt Institutional Review Board, and informed written consent was obtained before collecting any survey or patient data. The study followed TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) guidelines for the transparent reporting of a multivariable prediction models for individual prognosis or diagnosis.18

Patients were enrolled postoperatively and administered three questionnaires during inpatient stay: a questionnaire assessing inpatient pain satisfaction and medication use, the ORT (Opioid Risk Tool), and the ERS (Euphoric Response Subscale). The ORT is a five-question screening tool to assess risk for opioid abuse among individuals using opioids.19 The ERS retrospectively assesses subjective responses to first ever use of opioids. Higher ORT and ERS scores are associated with lower opioid effect, less sedation, and more euphoria with opioid use.20 Demographic variables not reliably available from the electronic health record (EHR) were also queried, including highest education level attained and detailed tobacco use history. We asked patients about a personal history of depression or anxiety and antidepressant use, and assessed the availability of these variables in the EHR. Race and ethnicity were self-identified for enrolled participants, and these data were collected given prior research highlighting potential disparities in inpatient opioid dispensing.21 A broad range of data were collected to identify known and novel predictors of discharge opioid use.

Standardized prescriptions at discharge included 30 tablets of ibuprofen 600 mg and 30 tablets of hydrocodone 5mg–acetaminophen 325 mg (150 morphine milligram equivalents). The hydrocodone prescription was selected to comply with state laws regulating opioid prescribing and provide enough opioid to observe a range of use frequencies.11 The pharmacy delivered all study medications directly to participants' rooms. Starting 14 days after cesarean birth, participants were contacted to complete additional measures. They were asked to locate their opioid container, count the number of leftover tablets, and record how many were used. If participants stated they were still taking opioids, they were recontacted every 7 days until use stopped. For participants who could not be contacted, there were at least three contact attempts performed before they were deemed lost to follow-up. We accessed the Tennessee Controlled Substance Medication Database for all participants to assess whether additional opioid prescriptions were filled during the postpartum period and to confirm dispending of opioids prescribed at hospital discharge. Our prior research demonstrated high correlation between patient-reported opioid use and real-time electronic medication caps in this population.22

The primary outcome in our prognostic model was the amount of the opioid prescribed at discharge used until 6 weeks postpartum expressed as morphine milligram equivalents. For our primary model we examined participants who were confirmed via CMSD to have received only the prescribed study opioid (30 tablets of hydrocodone 5mg–acetaminophen 325 mg or 150 morphine milligram equivalents). Participants receiving nonstudy opioids from their clinicians based on review of the Controlled Substance Medication Database were examined in a planned separate secondary analysis.

We identified predictor variables a priori informed by prior research, clinical expertise, and literature review.11,23 All predictors were available before hospital discharge either as data collected from the EHR or patient-reported data from inpatient questionnaires. This resulted in a total pool of 34 predictors, which were ranked by hypothesized strength of association with the primary outcome. Ranking was informed by prior research and literature review.

To avoid model overfitting, hierarchical cluster analysis using Spearman's rho as a similarity measure helped identify correlated predictors. Principal component analysis was used to compute principal components for each variable cluster. These techniques condensed 27 of 34 predictors into seven groups while preserving overall information. The top seven predictors hypothesized to be highly associated with the outcome were not included in the principal component analysis (Table 1).

Table 1.:
Individual Predictors in Each Model

Because previous work suggested that inpatient opioid use was highly associated with postdischarge opioid use, particular attention was given to this predictor. Specifically, we considered two methods to characterize inpatient opioids administration: 1) total morphine milligram equivalents used per hour of inpatient stay, and 2) total morphine milligram equivalents used in the 24 hours preceding discharge. Additionally, given concerns about underreporting depression or anxiety in the EHR, we asked women about personal history of anxiety or depression or antidepressant use to compare information obtained from the EHR only (depression or anxiety, EHR only) compared with information from all sources including patient-reported outcomes (depression or anxiety, all sources). We also explored the effect of the participants’ bedside nurses on opioid use by asking participants whether, in general, they received opioids when offered by their nurse, when they initiated a medication request, or both. Using opioids primarily because they were offered was included as a binary predictor in the model (bedside nurse).

We constructed our primary model using all predictors, with seven principal component groups and seven individual predictors (full model). We then evaluated which measures of inpatient opioid use and depression improved model performance the most, ultimately selecting prior 24-hour opioid use and depression or anxiety from all sources for subsequent models (Appendix 1, available online at The second model (EHR model) used 17 predictors, three of which were included as principal components, that are routinely available within the EHR (Table 1). For our third model (reduced model), we used forward selection to add predictors in prespecified ranked order until the model's performance reached 90% of the Likelihood Ratio χ2 achieved for the full model.

Our prognostic model for outpatient opioid use was created using a semiparametric proportional odds ordinal logistic model, which allows for arbitrary outcome distributions.24 Restricted cubic splines were used for continuous predictors to allow for nonlinear associations. Partial effects plots and nomograms were drawn. We calculated and graphically displayed the fraction of explainable outcome contributed by each predictor based on their partial χ2 values. The sample size was selected to accommodate up to 18 degrees of freedom (df) based on the formula 15×df, or 270 participants. Therefore, we were adequately powered using only the cohort of women who received standard prescriptions. Predictors and estimated degrees of freedom were selected a priori.

We compared models using assessments of information (likelihood ratio χ2 statistic), complexity-penalized information (Akaike information criterion), and predictive accuracy (discrimination and calibration). Specifically, discrimination (ability to differentiate which participants needed which total opioid dose) was assessed using Somers' Dxy rank correlation (Dxy). Somers' Dxy measures the strength of association between ordinal dependent and independent variables and is related to the popular concordance index (c-index) by the equation C=Dxy+12. When Dxy or C=1, the predictions are perfectly discriminating. Both C and Dxy are expected to be lower for ordinal continuous outcomes than for binary outcomes. Finally, we used the van Houwelingen and le Cessie heuristic shrinkage factor to quantify the amount of overfitting present and estimate the likelihood that the model will reliably predict new observations in another population. A shrinkage factor greater than 0.9 is considered adequate. We then examined the potential performance of our models in future populations by using internal bootstrap validation with 500 replications with replacement. Analyses were conducted by using Stata 15.1 and R 3.6.2.


Four hundred fifty-nine of 552 (83.2%) eligible patients enrolled in the study, of whom 398 (86.7%) completed the primary outcome, self-reported opioid use (in morphine milligram equivalents). Age, insurance status, and EHR-identified race or ethnicity were not different for patients who enrolled in the study and did not enroll in the study. Eleven patients were excluded owing to complications they experienced after enrollment. After accessing the Controlled Substance Medication Database for the remaining 387 participants, 108 (28.0%) participants were noted to have filled a discharge prescription other than the exact study prescription. This occurred primarily because their clinicians did not realize patients were participating in a study and wrote separate prescriptions. The 279 participants who received and filled the standard study prescription were used for the primary model and analysis (Appendix 2, available online at On average, participants who received nonstudy prescriptions (either different number of tablets or oxycodone, or both) received less morphine milligram equivalents (90 [60–90] morphine milligram equivalents) than the study prescription of 150 morphine milligram equivalents. These patients had shorter lengths of stay, which may have contributed to not receiving the study prescription. They also reported higher inpatient pain scores but used less opioid per hour of inpatient admission. Other similar demographic and clinical characteristics were similar (Table 2).

Table 2.:
Characteristics of the Study Population by Prescription Type

Among the entire population (N=387), participants spent a median 2.8 days in the hospital and used the median equivalent of seven tablets of hydrocodone 5 mg–acetaminophen 325 mg during their hospital stay (35 morphine milligram equivalents). After discharge, they used a median of eight tablets or 40 morphine milligram equivalents (interquartile range 1–18 tablets) for a median of 8 days (interquartile range 3–11 days). A majority (53%, 207/387) used some opioid, 23.5% (91/387) used no opioid, and 23.0% (89/387) used all opioids. Using all opioids was more common among participants who received nonstudy prescriptions with a lower median total morphine milligram equivalents than the study prescription (49.2% vs 13.5%).

The EHR (problem list, history, or medication list) identified 17.1% (66/387) of participants as having depression or anxiety; however, patient-reported information (“have you ever been diagnosed with or taken medication for anxiety or depression”) almost doubled the frequency of a depression or anxiety diagnosis in our population (133/387, 34.4%).

Adjusted odds ratios for each predictor are presented in Table 1 and the proportion of explainable outcome variability is presented in Figure 1. Partial effects plots are presented in Appendix 3, available online at Adjusted odds ratios could not be estimated individually for predictors that were components of principal components but instead are presented for the principal component group. Of 34 predictors, four were associated with discharge opioid use: inpatient opioid morphine milligram equivalents used, depression or anxiety, tobacco use, and the bedside nurse. We then explored different measures of inpatient opioid use and depression or anxiety by comparing four full models with the combinations of inpatient morphine milligram equivalents per hour compared with inpatient opioid morphine milligram equivalents used in the 24 hours before discharge, and depression or anxiety from the EHR only compared with depression or anxiety from all sources. We selected inpatient opioid morphine milligram equivalents used in the 24 hours before discharge and depression or anxiety from all sources owing to better model performance (Appendix 1, Tobacco use and the bedside nurse were significant predictors of outpatient opioid use regardless of which inpatient opioid use and depression or anxiety predictor was selected. Combined, these four predictors accounted for 67.2% of the explainable outcome variation in the full model, and inpatient opioid use alone accounted for 56.4% of the explainable variation. The EHR model identified 18 predictors obtained only from the EHR. The reduced model required only three predictors—inpatient opioid use, depression or anxiety, tobacco use—to reach the prespecified model performance (90% of the full model).

Fig. 1.:
Contribution of individual predictors.

Each of the models demonstrated good predictive accuracy to estimate postdischarge opioid use (Somers' Dxy 0.46–0.52 and C-index 0.74–0.76 for all models, Table 3). The internal bootstrap validation results were comparable with the primary models (Bootstrap Dxy 0.47–0.48). We selected the reduced model as our final model given its similar model performance with the fewest number of predictors all obtained from the EHR. Application of the van Houwelingen and le Cessie heuristic shrinkage factor suggested that this model would perform 3% worse when applied to a new population (estimate 0.97). A nomogram based on the final reduced model estimates with these three predictors was constructed and is presented in Figure 2 and a web-based calculator that estimates the morphine milligram equivalents used after hospital discharge can be found at In a secondary analysis, we tested model performance among 108 participants who filled nonstudy prescriptions. Model performance was consistent despite variable opioid prescribing (Table 3).

Table 3.:
Model Performance
Fig. 2.:
Nomogram for estimating opioid use after cesarean delivery. To use the nomogram, assign points for each of the three variables by drawing a vertical line from the response to the first row labeled Points. Add up the total points and note this on the line labeled Total Points. Draw a vertical line from this position to the outcome measures below. For example, a patient who uses 10 morphine milligram equivalents (MME) in the prior 24 hours (30 points) has depression (10 points) and is not a tobacco user (0 points) would receive a total of 40 points. Based on a total of 40 points, the patient would be estimated to use a mean of 52 MME and a median of 37 MME after discharge. Their 90th percentile of use would be 140 MME. The probability of using more than 5 MME is more than 80%, using more than 40 MME is 52%, and using more than 95 MME is 20%.


Using prospective data collected from patients undergoing cesarean birth, we developed and validated three novel prognostic models to estimate outpatient opioid use postdischarge. Of these models, a simple model with three predictors that can be obtained from the EHR—inpatient opioid use, tobacco use, and depression or anxiety—performed as well as more complex models. Although predictive models should be replicated in other samples, our model has potential to provide clinicians individualized guidance on discharge opioid prescribing and ultimately reduce excess opioids available for misuse and diversion.

Predictive modeling has been used widely to predict a variety of obstetric outcomes (ie, postpartum hemorrhage, vaginal birth after cesarean)25,26 and to predict development of opioid use disorder in adults prescribed opioids.27,28 To the best of our knowledge, this is the first study to systematically build and validate a model that predicts the quantity of outpatient opioids used after discharge by patients who underwent cesarean birth (PubMed search query: ["Forecasting"{Mesh} OR forecasting{tiab} OR predictive model*{tiab} OR prediction{tiab}] AND ["Analgesics, Opioid"{Mesh} OR opioids{tiab} OR opioid use{tiab} OR opioid usage{tiab}[ AND ["Patient Discharge"{Mesh} OR patient discharge{tiab} OR hospital discharge{tiab}])

Multiple studies have examined associations between patient or surgical characteristics and postdischarge opioid use.29–33 Our study is consistent with previous observed associations. Studies from the surgical literature consistently find associations between inpatient and outpatient opioid use, but findings related to tobacco use and depression or anxiety are mixed and largely depend on whether these data were collected.29 One limitation of examining opioid use in the surgical population is anticipated variation in postoperative pain directly related to differences in types of surgery (ie, orthopedic vs urologic) and approach (open vs laparoscopic). Studying cesarean birth has the advantage of being common, with minimal variation in surgical technique.

In the population who underwent cesarean birth, Bateman et al found no association between tobacco history or antidepressant use and outpatient opioid use, although variability in how tobacco and depression variables were defined could explain these differences. Specifically, current tobacco use in pregnancy may not have been included in their definition, and they captured only one parameter of depression, current antidepressant use.14 In our study, 15% of participants taking antidepressant medications did not report having depression or anxiety suggesting underreporting of depression. Notably, our model confirms prior studies in the population who underwent cesarean birth that generally have not demonstrated associations between outpatient opioid use and surgical characteristics (duration of surgery, addition of tubal ligation, number of prior cesarean births), obstetric events (labor preceding cesarean birth, indication for cesarean birth), or anesthetic type (regional vs general).11,14,34

Our study provides a means for individualized opioid prescribing after cesarean birth, the most common major surgery in the United States. The finding that inpatient opioid use is the major predictor of outpatient opioid use aligns with existing literature and clinical experience that inpatient use patterns can be a template for outpatient behaviors. Although inpatient opioid use was the dominant predictor, still, 44% of the variability in the model was explained by other factors. The question of whether inpatient opioid use alone can be used to estimate the quantity of opioid prescribed at discharge should be tested prospectively incorporating rigorous assessment of patient pain and satisfaction and is beyond the scope of this analysis. Additionally, implementation of individualized opioid prescribing requires addressing barriers to clinical use, especially the burden of assembling and interpreting inpatient opioid use data. In this context, decision support tools in the EHR could facilitate prescribing but should be designed using methods that are sensitive to the needs of end-users.

The finding that primarily patient, not surgical, characteristics predict outpatient opioid use deserves further attention especially given prior work demonstrating associations between persistent postsurgical opioid use and tobacco use and depression.5,35 Both nicotine and depression may modulate the endogenous opioid system in ways that decrease pain tolerance and increase need for analgesic medication.36,37 Consideration of potential mechanisms underlying predictive effects may be valuable in future work.

Another consideration is the known association between increased opioid prescribing and increased opioid use.12,14,38,39 In our study, 28% of participants received nonstudy prescriptions and were excluded from the original model owing to concerns for this effect. Although our model appears to perform well when applied to the population who received nonstudy prescriptions (lower total morphine milligram equivalents), it is possible that our point estimates could change if we had selected a different quantity of opioid to dispense for the study. This could represent an avenue for further research.

Major strengths of our study include systematic collection of predictor and outcome variables and a robust approach to predictive modeling. We ranked this comprehensive list of predictors a priori by hypothesized strength of association with the primary outcome. This ranking permitted us to consider different methods to capture data about top predictors and identify optimal characterization for model performance. This selection process combined with modern data-reduction techniques decreases the likelihood of overfitting the model and underperforming in a new population. Our model selection deserves highlighting as well. Although opioid use is a continuous outcome, its distribution is not normal. Prior studies have used linear regression to model opioid use29,30,40; however, this model violates the assumption that residuals are normally distributed. In contrast, we employed the proportional odds model, which provides greater flexibility for nonlinear relationships and better predictive capacity under these conditions.24,41 Finally, despite excluding 28% of participants who received nonstudy prescription from our initial model, our sample size remained adequately powered to construct the model presented.

There are several important considerations with regards to how our three main predictors were collected. We assumed that patients used opioids during hospitalization as needed; however, this does not account for potential contributions of bedside nurses to inpatient opioid use.42 In our study, almost 40% of women reported using opioids because they were offered by the nurse. Given the association between inpatient and outpatient opioid use, we acknowledge that variation in inpatient nursing practice could ultimately affect our estimates for outpatient opioid use. We also acknowledge that patient-reported opioid use could be inaccurate; however, our prior study in this population found high correlation between patient-reported opioid use and data gathered from real-time electronic medication caps that record when a pill bottle is accessed.21 Finally, including patient-reported information about depression or anxiety increased ascertainment of this diagnosis compared with EHR documentation alone and improved model performance slightly. Future implementation of our model into clinical practice should emphasize robust ascertainment of depression or anxiety history at patient intake and during pregnancy.

We developed three prognostic models that accurately predict outpatient opioid use after discharge in patients who underwent cesarean birth. A simple model with inputs derived exclusively from the EHR performed similarly to more complex models. Our study suggests that opioid use can be predicted at hospital discharge, paving the way forward for individualized opioid prescribing and a reduction in unused opioids.


1. Schirle L, Stone AL, Morris MC, Osmundson SS, Walker PD. Leftover opioids following adult surgical procedures: a systematic review and meta-analysis. Syst Rev 2020;9:139. doi: 10.1186/s13643-020-01393-8
2. Guy GP, Zhang K, Bohm MK, Losby J, Lewis B, Young R, et al. Vital signs: changes in opioid prescribing in the United States, 2006-2015. MMWR Morb Mortal Wkly Rep 2017;66:697–704. doi: 10.15585/mmwr.mm6626a4
3. Substance Abuse and Mental Health Services Administration, Results from the 2013 National Survey on Drug use and health: summary of national findings, NSDUH Series H-48, HHS Publication No. (SMA) 14-4863. Accessed February 10, 2022.
4. Soneji N, Clarke HA, Ko DT, Wijeysundera DN. Risks of developing persistent opioid use after major surgery. JAMA Surg 2016;151:1083–4. doi: 10.1001/jamasurg.2016.1681
5. Sun EC, Darnall BD, Baker LC, Mackey S. Incidence of and risk factors for chronic opioid use among opioid-naive patients in the postoperative period. JAMA Intern Med 2016;176:1286–93. doi: 10.1001/jamainternmed.2016.3298
6. Osmundson SS, Wiese AD, Min JY, Hawley RE, Griffin MR, et al. Delivery type, opioid prescribing, and the risk of persistent opioid use after delivery. Am J Obstet Gyncol 2019; 220:405–7. doi: 10.1016/j.ajog.2018.10.026
7. Shah A, Hayes CJ, Martin BC. Characteristics of initial prescription episodes and likelihood of long-term opioid use - United States, 2006-2015. MMWR Morb Mortal Wkly Rep 2017;66:265–9. doi: 10.15585/mmwr.mm6610a1
8. Dowell D, Haegerich TM, Chou R. CDC guideline for prescribing opioids for chronic pain—United States, 2016. MMWR Recomm Rep 2016;65:1–49. doi: 10.15585/mmwr.rr6501e1
9. Healthcare Cost and Utilization Project. Hospital inpatient national statistics. Accessed September 10, 2021.
10. Martin JA, Hamilton BE, Osterman MJK, Driscoll AK. Births: final data for 2018. Natl Vital Stat Rep 2019;68:1–47
11. Osmundson SS, Schornack LA, Grasch JL, Zuckerwise LC, Young JL, Richardson MG. Postdischarge opioid use after cesarean delivery. Obstet Gynecol 2017;130:36–41. doi: 10.1097/aog.0000000000002095
12. Osmundson SS, Min JY, Wiese AD, Hawley RE, Mitchel E, Patrick SW, et al. Opioid prescribing after childbirth and risk for serious opioid-related events: a cohort study. Ann Intern Med 2020;173:412–14. doi: 10.7326/m19-3805
13. Bateman BT, Franklin JM, Bykov K, Avorn J, William S. Persistent opioid use following cesarean delivery: patterns and predictors among opioid-naïve women. Am J Obstet Gynecol 2016:215;353.e1–18. doi: 10.1016/j.ajog.2016.03.016
14. Bateman BT, Cole NM, Maeda A, Burns SM, Houle TT. Patterns of opioid prescription and use after cesarean delivery. Obstet Gynecol 2017;130:29–35. doi: 10.1097/aog.0000000000002093
15. Eisenach JC, Pan PH, Smiley R, Lavand’homme P, Landau R, Houle TT. Severity of acute pain after childbirth, but not type of delivery, predicts persistent pain and postpartum depression. Pain 2008;140:87–94. doi: 10.1016/j.pain.2008.07.011
16. Pharmacologic stepwise multimodal approach for postpartum pain management. ACOG Clinical Consensus No. 1. American College of Obstetricians and Gynecologists. Obstet Gynecol 2021;138:507–17. doi: 10.1097/AOG.0000000000004517
17. Badreldin N, Grobman WA, Chang KT, Yee LM. Opioid prescribing patterns among postpartum women. Am J Obstet Gynecol 2018;219:103.e1–8. doi: 10.1016/j.ajog.2018.04.003
18. Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. Ann Intern Med 2015;162:55–63. doi: 10.7326/m14-0697
19. Webster LR, Webster RM. Predicting aberrant behaviors in opioid-treated patients: preliminary validation of the Opioid Risk Tool. Pain Med 2005;6:432–42. doi: 10.1111/j.1526-4637.2005.00072.x
20. Bruehl S, Stone AL, Palmer C, Edwards DA, Buvanendran A. Self-reported cumulative medical opioid exposure and subjective responses on first use of opioids predict analgesic and subjective responses to placebo-controlled opioid administration. Reg Anesth Pain Med 2019;44:92–9. doi: 10.1136/rapm-2018-000008
21. Badreldin N, Grobman WA, change KT, Yee LM. Racial disparities in postpartum pain management. Obstet Gynecol 2019;134:1147–53. doi: 10.1097/aog.0000000000003561
22. Schirle LM, Dietrich MS, Lam L, Stone Al, Bruehl S, Osmundson SS. Accuracy of patient-reported versus real-time electronic postoperative opioid use outcomes condensation. Am J Obstet Gynecol MFM 2021;3:100347. doi: 10.1016/j.ajogmf.2021.100347
23. Osmundson SS, Raymond BL, Kook BT, Lam L, Thompson EB, Schornack LA, et al. Individualized compared with standard postdischarge oxycodone prescribing after cesarean birth. Obstet Gynecol 2018;132:624–30. doi: 10.1097/aog.0000000000002782
24. Harrell FE. Regression modeling strategies. 2nd ed. Springer International Publishing; 2015.
25. Venkatesh KK, Strauss RA, Grotegut CA, Heine RP, Chescheir NC. Machine learning and statistical models to predict postpartum hemorrhage. Obstet Gynecol 2020;135:935–44. doi: 10.1097/aog.0000000000003759
26. Grobman WA, Sandoval G, Rice MM, Bailit JL, Chauhan SP, Constantine MM. Prediction of vaginal birth after cesarean in term gestations: a calculator without race and ethnicity. Am J Obstet Gynecol 2021;225:664.e1–7. doi: 10.1016/j.ajog.2021.05.021
27. Glanz JM, Narwaney KJ, Mueller SR, Gardner EM, Calcaterra SL, Xu S, et al. Prediction model for two-year risk of opioid overdose among patients prescribed chronic opioid therapy. J Gen Intern Med 2018;33:1646–53. doi: 10.1007/s11606-017-4288-3
28. Reps JM, Cepeda MS, Ryan PB. Wisdom of the CROUD: development and validation of a patient-level prediction model for opioid use disorder using population-level claims data. PLoS One 2020;15:e0228632. doi: 10.1371/journal.pone.0228632
29. Howard R, Fry B, Gunaseelan V, Lee J, Waljee J, Brummett C. Association of opioid prescribing with opioid consumption after surgery in Michigan [published erratum appears in JAMA Surg 2019;154:368]. JAMA Surg 2019;154:e184234. doi: 10.1001/jamasurg.2018.4234
30. Saini S, McDonald EL, Shakked R, Nicholson K, Rogero R, Chapter M, et al. Prospective evaluation of utilization patterns and prescribing guidelines of opioid consumption following orthopedic foot and ankle surgery. Foot Ankle Int 2018;39:1257–65. doi: 10.1177/1071100718790243
31. Vattigunta S, Weiner S, Nayar S, Jenkins S, Srikumaran U. Opioid consumption following orthopedic shoulder surgery: a retrospective analysis. J Shoulder Elb Surg 2021;30:S153–8. doi: 10.1016/j.jse.2021.04.005
32. Swenson CW, Kelley AS, Fenner DE, Berger MB. Outpatient narcotic use after minimally invasive urogynecologic surgery. Female Pelvic Med Reconstr Surg 2016;22:377–81. doi: 10.1097/spv.0000000000000299
33. Holst KA, Thiels CA, Ubl DS, Blackmon SH, Cassivi SD, Nichols FC, et al. Postoperative opioid consumption in thoracic surgery patients: how much is actually used? Ann Thorac Surg 2020;109:1033–9. doi: 10.1016/j.athoracsur.2019.08.115
34. Emerson JB, Danilack VA, Kulkarni A, Kesselring C, Brousseau EC, Matteson KA. Outpatient opioid use after cesarean delivery. R Med J 2013;103:68–74. doi: 10.15585/mmwr.rr6501e1
35. Rajabi A, Dehghani M, Shojaei A, Farjam M, Motevalian SA. Tobacco smoking and opioid use disorders: a meta-analysis. Addict Behav 2018;92:225–35
36. Burns JW, Bruehl S, France CR, Schuster E, Orlowska D, Buvanendran A, et al. Psychosocial factors predict opioid analgesia through endogenous opioid function. Pain 2017;158:391–9. doi: 10.1097/j.pain.0000000000000768
37. Xue Y, Domino EF. Tobacco/nicotine and endogenous brain opioids. Prog Neuro-psychopharmacol Biol Psychiatry 2008;32:1131–8. doi: 10.1016/j.pnpbp.2007.12.012
38. Bartels K, Mayes LM, Dingmann C, Bullard KJ, Hopfer CJ, Binswanger IA. Opioid use and storage patterns by patients after hospital discharge following surgery. PLoS One 2016;11:e0147972. doi: 10.1371/journal.pone.0147972
39. Patel HD, Srivastava A, Patel ND, Faisal FA, Ludwig W, Joice GA, et al. A prospective cohort study of postdischarge opioid practices after radical prostatectomy: the ORIOLES Initiative. Eur Urol 2019;72:215–18. doi: 10.1016/j.eururo.2018.10.013
40. As-Sanie S, Till SR, Mowers EL, Lim CS, Skinner BD, Fritsch L, et al. Opioid prescribing patterns, patient use, and postoperative pain after hysterectomy for benign indications. Obstet Gynecol 2017;130:1261–8. doi: 10.1097/aog.0000000000002344
41. Sainani KL. Multinomial and ordinal logistic regression. PM R 2021;13:1050–5. doi: 10.1002/pmrj.12622
42. Leziak K, Yee LM, Grobman WA, Badreldin N. Patient experience with postpartum pain management in the face of the opioid crisis. J Midwifery Women’s Heal 2021;66:203–10. doi: 10.1111/jmwh.13212

Supplemental Digital Content

© 2022 The Author(s). Published by Wolters Kluwer Health, Inc.