Prolonged Opioid Use and Pain Outcome and Associated Factors after Surgery under General Anesthesia: A Prospective Cohort Association Multicenter Study

This was a multicenter, prospective observational study. After approval in the Multicenter Perioperative Outcomes Group consortium peer review forum, all of the consortium’s active member institutions were invited to participate. Details about the study’s methods have been published previously.²²  The primary outcome variable of this study was opioid consumption at 3 months postoperatively. Study approval was obtained from the University of Utah (Salt Lake City, Utah) Institutional Review Board (IRB), which served as the Single IRB for three of the participating sites. The remaining nine participating institutions obtained approval from their local IRBs. The approved University of Utah Single IRB and individual institutions’ IRB protocols used a waiver of documentation of informed consent or full consent.

As described by Stuart et al.,²²  we did not formally estimate the sample size, since the effect sizes of perioperative data on the primary outcomes are not established. Similarly, we did not perform a power analysis. We determined an enrollment of 150 patients per week per institution during the 2 weeks of perioperative data collection to be feasible. Ten institutions would allow for a total size of 3,000 patients to be included.

Each institution, using convenience sampling, enrolled patients during a 2- to 4-week period between September and October 2017, if they were scheduled to undergo spine surgery, open thoracic surgery, knee or hip surgery, mastectomy, or abdominal surgery (including laparoscopic surgery) under general or regional anesthesia. These procedures were chosen because they are known to be associated with a higher incidence of persistent postoperative pain.^9,23–26  Patients were not eligible for inclusion if (1) their surgery was minor without the need for regional or general anesthesia, (2) they participated in a randomized clinical trial, which involved blinding interventions or administered medications that were relevant to this study, or which did not accept patients to be coenrolled in an observational study, (3) they were younger than 18 yr of age, (4) they did not speak English, or (5) they were cognitively impaired.

Using the surgical schedule, patients were pre-creened for eligibility using the above criteria and a list of 415 Current Procedural Terminology codes (see Table S1, Supplemental Digital Content 1, https://links.lww.com/ALN/D22, “Current Procedural Terminology Codes Included and Their Mapping to Surgery Types”), consented, and enrolled during the preoperative period in the preoperative waiting areas or admission suites.

Preoperatively, questionnaires were administered in person to the enrolled patients in order to collect information about their physical characteristics and demographics, and home opioid and nonopioid analgesic use. Patients were asked whether they were currently using pain medication. If they answered yes, follow-up questions elicited more detailed information regarding which particular medications, and for the opioids for how long they were taking the opioids, at which dose, and how often.

In addition, validated questionnaires were used to assess the patients’ pain intensity (Brief Pain Inventory severity questions for pain at the site of surgery and overall body pain²⁴ ); comorbid central nervous system symptoms (2011 Fibromyalgia Survey Criteria using the Symptom Severity Index²⁷ ); widespread pain (Michigan Body Map²⁸ ); physical function (Patient Reported Outcome Measurement System Physical Function short form 4a²⁹ ); anxiety (Patient Reported Outcome Measurement System Anxiety short form 4a²⁹ ); depression (Patient Reported Outcome Measurement System Depression short form 4a²⁹ ); catastrophizing, i.e., thoughts about symptoms (Patient Reported Outcome Measurement System Sleep Disturbance short form 5a²⁹ ); and expectations of surgery.

Additional information was collected via chart review, namely time until readiness for discharge from the postanesthesia care unit, postoperative intensive care unit length of stay, hospital length of stay, postoperative day 0 and postoperative day 1 pain scores, in-hospital opioid and nonopioid analgesic medications, reintubation, oxygen dependence, new noninvasive ventilation requirements, length of hospital stay, and postoperative myocardial injury.

Enrolled patients were contacted by phone at 3 months postoperatively. If patients could not be reached, three more contact attempts were made at different weekdays and times of day. Patients were asked whether they had taken any opioid or nonopioid analgesic medication since their surgery. If they answered yes, follow-up questions elicited more detailed information regarding which particular medications, and for the opioids for how long they were taking the opioids, at which dose, and how often. Using the same instruments that were used preoperatively, patients were also asked about pain at the site of surgery, overall body pain, symptom severity index, Michigan Body Map, Patient Reported Outcome Measurement System Physical function short form 4a, Patient Reported Outcome Measurement System Anxiety short form 4a, Patient Reported Outcome Measurement System Depression short form 4a, Patient Reported Outcome Measurement System Sleep Disturbance short form 5a, catastrophizing, satisfaction with surgery, and the occurrence of adverse events.

Case report forms to collect these data were fashioned after a study by Brummett et al.^26,30  (see Supplemental Digital Content 3, “Case Report Forms,” https://links.lww.com/ALN/D24).

Intraoperative data were collected from the Multicenter Perioperative Outcomes Group database.³¹  These included the American Society of Anesthesiologists Physical Status, admission diagnosis, comorbidities, type of anesthesia, intraoperative anesthetic technique including all drugs administered, and discharge International Classification of Diseases code. A detailed table with all variables collected, together with the timepoint and the manner in which they were collected, can be found in Supplemental Table S2 (Supplemental Digital Content 2, https://links.lww.com/ALN/D23, “Information Collected”).

In the selection of information to collect, we aimed to include variables that were likely relevant to the primary outcomes of prolonged postsurgical opioid use and pain. The selection was guided by experts we had among our authors and previous publications,^{6–8,26,32–38}  and aided by an internal peer review process within the Multicenter Perioperative Outcomes Group consortium, while—for the manually collected information—attempting to limit the number of variables to a number that would not overwhelm data collectors and participants.

The primary outcomes were opioid consumption and surgical site pain at 3 months as elicited from the telephone survey conducted 3 months postoperatively, described above and detailed in Supplemental Digital Content 3, “Case Report Forms” (https://links.lww.com/ALN/D24). The responses were collated into binary events (opioid use present or absent, surgical site pain present or absent at 3 months).

Patterns in the raw data including missingness were inspected by histograms, density plots, boxplots, x-y plots, and cross-tabulation. The distributions of covariate values for subjects with and without the primary outcome were compared by standardized mean differences. Standardized mean differences greater than 0.2 were interpreted as showing an imbalance of means.^39,40 

Logistic multivariate, multivariable models were estimated for the primary outcomes; the two outcomes were jointly and simultaneously modeled on the covariates. All covariates considered to putatively influence the outcome were included in the models; these were (1) age, (2) sex, (3) body mass index, (4) race, (5) occupation, (6) relationship status, (7) American Society of Anesthesiologists Physical Status, (8) surgery type, (9) preoperative anxiety score, (10) preoperative depression score, (11) preoperative physical function score, (12) pain at the surgery site preoperatively, (13) preoperative opioid use, (14) preoperative nonopioid analgesic use, (15) anesthesia duration, (16) intraoperative parenteral morphine equivalents, (17) intraoperative nonopioid analgesics, (18) general anesthesia, (19) neuraxial anesthesia, and (20) Multicenter Perioperative Outcomes Group institution (group or random effect). These covariates were chosen by expert opinion and from previous publications.^{6–8,26,32–38}  The preoperative psycho-social scores used the Patient Reported Outcome Measurement System scores (Physical Function short form 4a, Anxiety short form 4a, and Depression short form 4a).²⁹  Subjects with missing covariate data were included in outcome models using 20-fold multiple chain imputation techniques to replace missing values; imputation algorithms were predictive mean matching (numeric data), logistic regression (factors with two levels), and polytomous regression (factors with three or more levels). Missing outcomes (opioid consumption and surgical site pain at 3 months) were not imputed.

Model fit was by hierarchical Bayesian regression methods using Markov chain Monte Carlo algorithms, specifically Hamiltonian Monte Carlo with the No-U-Turn Sampler having more rapid convergence for high-dimensional models.^41,42  Bayesian analysis allows direct probability interpretation of intervals bounding the mean values and avoids the frequent misapplication of null hypothesis significance testing.⁴³  We used a set of increasingly informative prior distributions: noninformative (improper flat prior), weakly informative (Student’s t), and informative (horseshoe). The weakly informative and informative prior distributions enforced regularization of parameter estimates to prevent overfitting of model coefficients. The model was estimated 20 times, once for each imputed data set; four chains with 4,000 iterations with 50% warmup and a 1-to-1 thinning ratio were used. The posterior distribution was a pooling of the 20 models with a total size of 640,000 draws (2,000 draws times 20 models).

Convergence characteristics of the posterior distribution of parameters were assessed by the $R^$ statistic, effective sample size, chain mixing, and chain autocorrelation. Model covariates were checked for collinearity. The posterior predictive distribution was used to generate a predictive accuracy metric as measured by leave-one-out cross-validation.⁴⁴  A posterior projection of the model was performed and model predictive performance assessed by expected-log-predictive-density and root mean square error, for variable selection.^45,46  Covariate significance was assessed using the region of practical equivalence procedure⁴⁷  with a range of –0.1 to 0.1 as suggested by Kruschke,⁴⁸  by evaluating the probability of direction, and by inspecting the maximum a posteriori-based P value.⁴⁹ 

Institutions were included in the statistical model as group effect. Because of observed differences between institutions in the completion rate of the 3-month surveys, in enrolled patients, and in the mix of surgery types, a variance partition coefficient analysis was performed to assess the contribution of the institutions to the overall observed variance.

Model results are presented as means, medians, SDs, and 95% credible intervals. A 95% credible interval has a 95% probability of containing the true parameter value. By contrast, a 95% CI is interpreted under the assumption that if a large number of analyses are repeated, in 95% of these analyses, the 95% CI will contain the true parameter value. Model coefficients are also presented with density plots to show the probability of direction.⁵⁰  Inferences on model coefficients followed methods suggested by Kruschke⁴⁸  and Makowski et al.⁵⁰  Statistical modeling was done in the R software (available at https://www.r-project.org/, accessed February 13, 2023) using the mice, bayestestR, brms, loo, mcmcplot, posterior, tidybayes, and projpred packages. We followed guidelines for reporting Bayesian analysis.⁵¹ 

Twelve Multicenter Perioperative Outcomes Group member institutions participated in the study (Cleveland Clinic [Cleveland, Ohio], Columbia University [New York, New York], University of Michigan [Ann Arbor, Michigan], Oregon Health & Science University [Portland, Oregon], University of Utah [Salt Lake City, Utah], Utrecht University [Utrecht, the Netherlands], Vanderbilt University [Nashville, Tennessee], University of Vermont [Burlington, Vermont], University of Virginia [Charlottesville, Virginia], University of Washington [Seattle, Washington], Washington University [St. Louis, Missouri], and Yale University [New Haven, Connecticut]). Eleven of the participating institutions were academic hospitals in the United States; one participating institution (Utrecht University) was in the Netherlands.

Between September and October 2017, 3,505 surgical procedures met inclusion criteria. From these patients, a sample of 1,110 cases were enrolled (reasons for not enrolling included night or weekend cases, availability of study recruitment staff, some institutions focusing their enrollment on certain hospitals within their network). Seventeen cases were excluded (fig. 1), leaving 1,093 patients. One institution did not perform any follow-up on its 24 patients and was dropped from the study. Of the 1,069 remaining patients at 11 institutions, 389 were lost to follow-up at 3 months, leaving 680 (64%) for outcome analyses (fig. 1).

There was considerable imbalance between participating institutions in terms of number of enrolled patients per institution (14 to 190) and type of surgery (see Table S4, Supplemental Digital Content 4, https://links.lww.com/ALN/D25, “Case Numbers by Institution and Type of Surgery”). Additionally, completion rate of the 3-month follow-up varied between institutions: all but four institutions had completion rates of more than 66%, but the remaining four had rates of 34%, 45%, 53%, and 62%, respectively.

Of those variables that were identified as being strong predictors for the outcome variables, the standardized mean difference between participants for whom the 3-month survey was completed and those who were lost to follow-up, preoperative taking of opioids, and preoperative pain at the site of surgery were below the standardized mean difference threshold of 0.2, and surgery type was found to have a standardized mean difference of 0.21, i.e., just slightly above that threshold (table 1).

The vast majority of cases were abdominal surgeries, accounting for 59% of enrolled cases, with none of the other surgery types contributing more than 17% (see Table S4, Supplemental Digital Content 4, https://links.lww.com/ALN/D25, “Case Numbers by Institution and Type of Surgery”).

Using a threshold of 0.2 for the standardized mean difference,^39,40  age, institution, race, and surgery type were the modeled variables for which there was a larger than small to medium imbalance between patients who were lost to the 3-month follow-up compared to those who participated in the follow-up (table 1; Table S5, Supplemental Digital Content 5, https://links.lww.com/ALN/D26, “Patient and Case Characteristics and Data Availability”).

Three months after surgery, 96 (14%) patients of the 680 patients in the final dataset were taking opioids. Comparisons of patient and case characteristics between patients taking opioids at 3 months and those not taking opioids at 3 months are shown in table 2 (and in greater detail in Table S6, Supplemental Digital Content 6, https://links.lww.com/ALN/D27, “Patient and Case Characteristics and Outcome ‘Taking Opioids at Three Months’”; and Table S7, Supplemental Digital Content 7, https://links.lww.com/ALN/D28, “Patient and Case Characteristics and Outcome ‘Taking Opioids at Three Months’ (Univariable Testing)”). Patients’ preoperative use of opioids, their physical function score, and whether patients preoperatively had pain at their site of surgery were the three variables with the largest standardized mean difference on this primary outcome measure.

The vast majority of the 680 patients in the final analysis (545, 80%) did not take opioids preoperatively (table 3). Four percent (23) of these patients reported taking opioids at 3 months after their surgery. In contrast, more than half (73, 54%) of the 135 patients who did take opioids preoperatively were still taking opioids at 3 months.

Of the 656 patients for whom data were available for surgical site pain preoperatively and at 3 months, a little more than half (341, 52%) reported no pain at the site of their surgery preoperatively (table 4). Thirteen percent (45) of these patients reported surgical site pain 3 months after their surgery. More than half of patients (183, 58%) who did report surgical site pain preoperatively (315, 48%) reported no surgical site pain at 3 months postsurgery. Comparison of patient and case characteristics between patients having surgical site pain at 3 months and those who did not are shown in table 5 (and in greater detail in Table S8, Supplemental Digital Content 8, https://links.lww.com/ALN/D29, “Patient and Case Characteristics and Outcome ‘Surgical Site Pain at Three Months’”; and Table S9, Supplemental Digital Content 9, https://links.lww.com/ALN/D30, “Patient and Case Characteristics and Outcome ‘Surgical Site Pain at Three Months’ (Univariable Testing)”).

Of the 675 patients for whom pain data were available at 3 months postsurgery, 51 patients (8%) reported both that they took opioids and that they had pain at the site of surgery at 3 months postsurgery. This rate is about half of that observed preoperatively, when 110 patients (17%) of the 661 patients, for whom data on preoperative pain were available, reported both surgical site pain and use of opioids.

Overall, 1.5% of the covariates had to be imputed; in 130 patients, one or more of the covariates’ values were missing, while complete data sets were collected from 550 patients.

The Bayesian statistical model had good estimation properties with good chain mixing, absence of chain autocorrelation by lag 2, $R^$ close to 1.01, and more than adequate effective sample size (see Table S10, Supplemental Digital Content 10, https://links.lww.com/ALN/D31, “Summary of Posterior Distribution for Taking Opioid at Three Months,” and Table S11, Supplemental Digital Content 11, https://links.lww.com/ALN/D32, “Summary of Posterior Distribution for Surgical Site Pain at Three Months”). All Pareto k estimates were smaller than 0.7, indicating a good model fit. In addition, the covariates did not show collinearity. The posterior predictive distribution had a good fit with the observations.

The preoperative taking of opioids showed a very strong probability of direction (greater than 99.99%) and the largest mean parameter values for taking opioids at 3 months after surgery (fig. 2). This variable is associated with an adjusted odds ratio of 18.6 (credible interval, 10.3 to 34.5). This variable’s probability of significance is so large (greater than 99.9%) and the probability of it being inside the region of practical equivalence is so low (0%) that the hypothesis that its parameter density includes zero (i.e., that coefficient would have no effect on the model’s performance) must be rejected. At the same time the Bayesian maximum a priori–based P value for this variable is low—and the lowest among all variables (less than 0.001), and the probability of direction is high (100%), indicating that its observed positive direction of effect is not due to random sampling (see Table S10, Supplemental Digital Content S10, https://links.lww.com/ALN/D31, “Summary of Posterior Distribution”). Preoperative physical function is the only other variable that seems to have an effect, albeit a weak one (table 2). These results are consistent with the projected prediction analysis (Supplemental Digital Content S12, https://links.lww.com/ALN/D33, “Projected Prediction Analysis for Taking Opioid at Three Months”), which indicates that including preoperative taking of opioids into the submodel yields the largest model prediction performance improvement. Additionally including preoperative physical function into the model provides little incremental performance gain. However, including any additional predictor variables beyond those two changes model performance only very little if at all.

For surgical site pain at 3 months, the presence of preoperative surgical site pain, surgery type knee replacement, and surgery type spine surgery showed very strong probability of direction (greater than 99.99%) and low probability of being inside the region of practical equivalence (0%), such that the hypothesis that their parameter densities include zero (i.e., that coefficient would have no effect on the model’s performance) must be rejected. The adjusted odds ratios (credible interval) for these predictors were 2.58 (1.45 to 4.4), 4.1 (1.73 to 8.9), and 2.75 (1.39 to 5.0), respectively. Surgery type mastectomy also has a very strong probability of direction (97%), but in contrast to the other variables, its credible interval is not completely above zero, so the null hypothesis for it cannot be rejected. In contrast, the intraoperative parenteral morphine equivalent has greater than 99.99% probability of being inside of the region of practical equivalence, indicating the null hypothesis (no effect on model performance) should be accepted (see Table S11, Supplemental Digital Content S11, https://links.lww.com/ALN/D32, “Summary of Posterior Distribution, Outcome Surgical Site Pain”). These results are consistent with the projected prediction analysis (Supplemental Digital Content S13, https://links.lww.com/ALN/D34, “Projected Prediction Analysis for Surgical Site Pain at Three Months”), which indicates that including surgery type into the submodel yields the largest model prediction performance improvement. Additionally including preoperative surgical site pain into the model provides some incremental performance improvement. However, including any additional predictor variables beyond those two changes model performance only very little if at all.

No observable correlation was found between the two outcome variables, opioid taking at 3 months and surgical site pain at 3 months. The contribution of the institutions to the observed variance in the outcome variables was found to be less than 2% for either outcome variable.

In this multicenter, prospective observational study, we identified one variable as the strongest predictor of opioid use at 3 months postoperatively: preoperative use of opioids. Other factors showed statistical significance in univariable analysis: institution, not being employed, preoperative taking of nonopioid analgesics, anxiety, depression, physical function, preoperative pain at the surgical site, surgery type, and intraoperatively administered morphine equivalent. However, in multivariable statistical modeling analysis (table 2), preoperative use of opioids was the only predictor of opioid use at 3 months postoperatively. Its adjusted odds ratio (18.6; credible interval, 10.3 to 34.5) was large in comparison to other variables and in absolute terms.

In the univariate analysis, a number of predictors achieved statistical significance in their association with the presence of surgical site pain 3 months postsurgery: institution, female sex, not being employed, preoperative taking of opioids or nonopioid analgesics, anxiety, depression, physical function, preoperative surgical site pain, surgery type, general anesthesia, and neuraxial anesthesia. However, in the multivariate model, only three strong predictors were found: preoperative surgical site pain, knee replacement surgery, and spine surgery. Mastectomy showed an elevated signal but did not reach statistical significance.

Notably, no correlation was found between the two outcome variables. This study was not a controlled study, which might have contributed to no correlation being observed. However, this finding was similar to that of Goesling et al.,⁹  who did not find an association between persistent opioid use and a change in joint pain in knee and hip arthroplasty patients, and the suggestion by Brummett et al. from nationwide insurance claims data that prolonged opioid use after surgery may not be associated with pain.²⁶ 

The findings about long-term opioid use are similar to findings of mostly retrospective or database-based studies.^52–55  For example, in a retrospective analysis in 490 shoulder arthroplasty patients, patients who had used opioids preoperatively were seven times more likely to still use them 1 yr after discharge.⁵⁶  Gil et al., analyzing insurance claims data of 104,154 shoulder arthroscopy patients, found that filling an opioid prescription in the month before the procedure, was one of the factors associated with the highest odds ratios for prolonged opioid use after the procedure.⁵²  Others found that preoperative opioid use is associated with less improvement or more difficult surgical recovery.^8,53,57  For example, a retrospective analysis of claims data from more than 34,000 adult orthopedic surgery patients found that patients who used opioids preoperatively had worse outcomes in terms of length of stay and revision rates and a 64% smaller rate of opioid use discontinuation in the 18 months after their procedure.⁵³ 

In contrast to the current study, most reports of factors associated with prolonged use of opioids after surgeries were from retrospective studies.⁵⁸  Findings in those studies included an association of prolonged postoperative opioid use with preoperative depression, anxiety, or mood disorders,^6,26,32  age and lower household income,⁷  type of surgery,^6–8  preoperative pain,^23,32  preoperative use of opioids,^32,33,59  and discharge prescription of opioids.^59–61  No association was found between prolonged opioid use and nerve blocks^62,63  or minor versus major surgery.²⁶  Our prospective study included most of these factors, and did not confirm most findings of these retrospective studies, other than preoperative opioid use. However, our study findings were consistent with a prospective study of 574 patients undergoing total knee or total hip arthroplasty that found that patients who reported opioid use on the day of surgery had a much higher rate (41.7%) of still using opioids 6 months after surgery than patients who were opioid-naive on the day of surgery (9.8%).⁹ 

The incidence of prolonged postsurgical pain in our study is within the range found in previously reported studies.^12–14  While our study’s findings about the association of preoperative pain and surgical type are consistent with previous studies,^15,16  they did not confirm other risk factors found in previous studies, including female sex, surgery duration, preoperative use of opioids, and anxiety and depression.^{12,14,17–19} 

Our multicenter study may not have detected previously identified associations with prolonged postoperative opioid use or prolonged postsurgical pain due to larger variability of care in multicenter versus single-center studies. Thus, our study may have been underpowered to detect these previously identified associations,^32,64  which might also have played a role in the relatively large credible intervals of the odds ratios (e.g., table 2, table 5). The addressing of imperfect data by using Bayesian analysis techniques and imputation is a strength of our study. The probability of direction, an inherent output from Bayesian analysis, showed a very strong indication of the emerging predictor variables’ effect.

Our study had the following limitations. Outcomes, including the primary outcomes, were patient-reported, which is known to be associated with inherent shortcomings and might have introduced bias.⁶⁵  The 3-month completion rate of 64% potentially introduced nonresponse bias. Of those variables that were identified as being strong predictors, the standardized mean differences between participants for whom the 3-month survey was completed and those who were lost to follow-up were below the threshold, except for surgery type, which was slightly above the threshold—indicating that the impact of nonresponse bias on those variables was small. Institution was one of the variables that were not identified as strong predictors but showed an above-threshold standardized mean difference. The contribution of less than 2% by institutions to the observed variance on either outcome variable indicates a limited impact of completion rate differences between institutions. Age also had an above-threshold standardized mean difference: patients lost to follow-up were a few years younger than those who responded at 3 months. The difference in age, however, was not clinically significant. Finally, consolidated race showed a standardized mean difference larger than 0.2, with nonwhite respondents having a somewhat larger loss to follow-up than white respondents. Considering the overall small portion of nonwhite participants, the strength of the signals of the variables that were identified as impactful predictors, and the very small signals of those that were not, the bias introduced by this difference is limited.

This study’s completion rate was comparable to other postoperative follow-up studies, especially considering the duration of the follow-up period. In a single-site study of opioid use after cesarean delivery, Bateman et al.⁶⁶  lost 252 of 975 patients (25.9%) to follow-up by phone, but in a much shorter time after discharge than in this study. A study of 330 general surgery patients yielded a 38% response rate via phone after 12 months.⁶⁷  Given these comparisons and the lack of evidence for nonresponse bias in our enrolled sample, we suggest that nonresponse bias is not a major concern for interpretation of our study results.

There were limitations related to the information we were able to collect. For example, while the inclusion of variables in the questionnaires was guided by experts, literature, and a consortium-internal peer review process, we had to limit the number of variables so as not to overwhelm data collectors and participants, lest the data quality would suffer. As another example, while we can confidently detect opioid use, the data quality of patient-reported opioid dosage did not allow us to draw reliable quantitative conclusions about the opioid amounts taken. In addition, not all potential covariates could be included in the models, and we studied only a limited set of type of surgeries, selected for their propensity to cause postoperative pain and prolonged opioid use.

The numbers for the different race categories the participants identified as were imbalanced. Most patients identified as “White,” and for many patients (10%), “no response” was recorded in place of their race. The two largest self-reported racial groups were “White” and “Black or African American.” The number of participants who identified as “White” was 10 times larger than the number of participants who identified as “Black or African American,” the next largest category. None of the nonwhite categories rose above single-digit percentages.

The statistical modeling included all patients for whom data about their 3-month opioid use was available. Some of these patients had missing data in other variables, which we addressed by using imputations. Another limitation is that the statistical model did not consider interactions between variables because of estimation difficulties. Relying on a convenience sample (e.g., no night or weekend cases, availability of study recruitment staff, some institutions focusing their enrollment on certain hospitals within their network) might have introduced bias and did not ensure that recruitment was representative of the general population. There is a wide variation of postsurgical opioid prescribing patterns between individual prescribers⁶⁷ ; however, both single-institution reports (e.g., Nobel et al.⁶⁸ ) and reports about national trends⁶⁹  indicate that postsurgical opioid prescribing is changing over time. With that, another limitation of this report is the age of the data, which were collected toward the end of 2017. Finally, both preoperative and prolonged opioid use might be affected by factors that are challenging to determine in a patient phone survey, including cultural norms and expectations, physician prescribing behaviors, and risk factors for addiction. The use of opioids preoperatively was found to be a predictor of opioid use at 3 months after surgery with a very strong statistical indication of effect.

This research was supported in part through computational resources and services provided by Advanced Research Computing, a division of Information and Technology Services at the University of Michigan (Ann Arbor, Michigan). Additionally, the authors thank the following people whose hard work on this project made it possible: Marcia E. Birk, R.N., Anesthesiology, University of Virginia, Charlottesville, Virginia; Amber D. Bledsoe, M.D., Anesthesiology, University of Utah, Salt Lake City, Utah; Jacqueline F. van Dijk, Ph.D., Anesthesiology, University Medical Center Utrecht, Utrecht, Netherlands; Lisa Y. Flint, B.S., Anesthesiology and Pain Medicine, University of Washington, Seattle, Washington; Alexander F. Friend, M.S., Anesthesiology, University of Vermont, Burlington, Vermont; Scott R. Junkins, M.D., Anesthesiology, University of Utah, Salt Lake City, Utah; Yuri Kida, M.S.W., Anesthesiology, University of Utah, Salt Lake City, Utah; Sherry L. McKinnon, M.S., Anesthesiology, Washington University, St. Louis, Missouri; Jordan Oberhaus, B.S., Anesthesiology, Washington University, St. Louis, Missouri; Nicole A. Pescatore, M.P.H., Anesthesiology, University of Michigan, Ann Arbor, Michigan; Julia L. White, B.S., Anesthesiology, University of Utah, Salt Lake City, Utah; Troy S. Wildes, M.D., Anesthesiology, Washington University School of Medicine, St. Louis, Missouri; and Josh Zimmerman, M.D., Anesthesiology, University of Utah, Salt Lake City, Utah.

Funding was provided by departmental and institutional resources at each contributing site. In addition, partial funding to support underlying electronic health record data collection into the Multicenter Perioperative Outcomes Group registry was provided by Blue Cross Blue Shield of Michigan/Blue Care Network (Detroit, Michigan) as part of the Blue Cross Blue Shield of Michigan/Blue Care Network Value Partnerships program. Although Blue Cross Blue Shield of Michigan/Blue Care Network and Multicenter Perioperative Outcomes Group work collaboratively, the opinions, beliefs and viewpoints expressed by the authors do not necessarily reflect the opinions, beliefs, and viewpoints of Blue Cross Blue Shield of Michigan/Blue Care Network or any of its employees.

Dr. Brummett reports being a consultant for Heron Therapeutics (San Diego, California), Vertex Pharmaceuticals (Boston, Massachusetts), Benter Foundation (Pittsburgh, Pennsylvania), and Alosa Health (Boston, Massachusetts), and providing expert medical malpractice testimony. Dr. Domino declares no competing interests. She reports research funding to the institution provided by Mathematica (Princeton, New Jersey) and Edwards Lifesciences (Irvine, California), unrelated to this project. Dr. Johnson declares no competing interests. He reports research funding unrelated to this work from Medtronic (Dublin, Ireland) and the National Institute of Neurologic Disorders and Strokes (Bethesda, Maryland). He also reports being an equity partner in Applied Medical Visualizations, L.L.C. (Salt Lake City, Utah), also unrelated to this work. Dr. Kuck declares no competing interests. He reports research funding, unrelated to this work, from the National Science Foundation (Alexandria, Virginia), Foundation for Anesthesia Education and Research (Schaumburg, Illinois), National Institute of General Medical Sciences (Bethesda, Maryland), Dynasthetics L.L.C. (West Valley City, Utah), and Medtronic. He also reports equity interest, unrelated to this work, in KSCube, L.L.C. (Park City, Utah), and CKC Medical, L.L.C. (Park City, Utah), and financial interest, unrelated to this work, from patents assigned to Drägerwerk (Lübeck, Germany) and the University of Utah Research Foundation (Salt Lake City, Utah). Dr. Posner declares no competing interests. She reports research funding, unrelated to this work, to the institution provided by Mathematica, Microsoft Research (Seattle, Washington), American Society of Anesthesiologists (Schaumburg, Illinois), Society for Anesthesia and Sleep Medicine (Milwaukee, Wisconsin), Anesthesia Patient Safety Foundation (Schaumburg, Illinois), and Anesthesia Quality Institute (Schaumburg, Illinois); and travel and speaker funding, unrelated to this work, to the author from The Doctors Company (Napa, California). Dr. Saager declares no competing interests. He reports receiving funding, unrelated to this work, from The Surgical Company International (Amersfoort, Netherlands). Dr. Pace declares no competing interests. He reports receiving funding, unrelated to this work, from Elute Inc. (Salt Lake City, Utah). The other authors declare no competing interests.

Supplemental Digital Content 1. Table: Table S1, Current Procedural Terminology Codes, https://links.lww.com/ALN/D22

Supplemental Digital Content 2. Table: Table S2, Information Collected, https://links.lww.com/ALN/D23

Supplemental Digital Content 3. Forms: Case Report Forms, https://links.lww.com/ALN/D24

Supplemental Digital Content 4. Table: Table S4, Case Numbers by Institution, https://links.lww.com/ALN/D25

Supplemental Digital Content 5. Table: Table S5, Patient and Case Characteristics and Data Availability, https://links.lww.com/ALN/D26

Supplemental Digital Content 6. Table: Table S6, Patient and Case Characteristics, Outcome “Taking Opioids at Three Months,” https://links.lww.com/ALN/D27

Supplemental Digital Content 7. Table: Table S7, Patient and Case Characteristics, Outcome “Taking Opioids at Three Months” (Univariable Testing), https://links.lww.com/ALN/D28

Supplemental Digital Content 8. Table: Table S8, Patient and Case Characteristics, Outcome “Site Pain at Three Months,” https://links.lww.com/ALN/D29

Supplemental Digital Content 9. Table: Table S9, Patient and Case Characteristics, Outcome “Site Pain at Three Months” (Univariable Testing), https://links.lww.com/ALN/D30

Supplemental Digital Content 10. Table: Table S10, Summary of Posterior Distribution (Taking Opioids at Three Months), https://links.lww.com/ALN/D31

Supplemental Digital Content 11. Table: Table S11, Summary of Posterior Distribution (Site Pain at Three Months), https://links.lww.com/ALN/D32

Supplemental Digital Content 12. Table: Table S12, Projected Prediction Analysis for Taking Opioid at Three Months, https://links.lww.com/ALN/D33

Supplemental Digital Content 13. Table: Table S13, Projected Prediction Analysis for Surgical Site Pain at Three Months, https://links.lww.com/ALN/D34

Group Authors

Multicenter Perioperative Outcomes Group Enhanced Observation Study Investigator Group

Mitchell F. Berman, M.D., M.P.H., Anesthesiology, Columbia University, New York, New York. Professor of Anesthesiology at the Columbia University Medical Center. Coordinated the performance of the study at the local site, recruited study subjects at the local site, collected and entered data at the local site, created software and database infrastructure for the project, consulted/commented on the manuscript.

Beatriz Raposo Corradini, M.S., Herbert Irving Comprehensive Cancer Center, Columbia University Irving Medical Center, New York, New York. Clinical Research Manager. Supervised the performance of the study at the local site, coordinated the performance of the study at the local site, recruited study subjects at the local site, collected and entered data at the local site, reviewed, contributed to, approved of the final version of the manuscript.

Shawn L. Mincer, M.S.W., Anesthesiology & Pain Medicine, University of Washington, Seattle, Washington. Research Coordinator. Provided scientific advice and guidance of the overall project, supervised the performance of the study at the local site, coordinated the performance of the study at the local site, recruited study subjects at the local site, collected and entered data at the local site, reviewed, contributed to, approved of the final version of the manuscript.

Sydney E. Rose, M.D., Anesthesiology and Perioperative Medicine, Oregon Health & Science University, Portland, Oregon. Anesthesiologist. Coordinated the performance of the study at the local site. Recruited study subjects at the local site. Collected and entered data at the local site, reviewed, contributed to, approved of the final version of the manuscript.

Wilton A. van Klei, M.D., Ph.D., Anesthesiology, University Medical Center Utrecht, Utrecht, Netherlands. Anesthesiologist. Supervised the performance of the study at the local site, interpreted the results, reviewed, contributed to, approved of the final version of the manuscript.

Nonauthor Collaborators

Multicenter Perioperative Outcomes Group Enhanced Observation Study Collaborator Group

David A. Edwards, M.D., Ph.D., Anesthesiology, Vanderbilt University Medical Center, Nashville, Tennessee. Chief, Division of Pain Medicine. Supervised the performance of the study at the local site, consulted and commented on the manuscript.

Olivia O.A. Lamers, M.D., Parasitology, Leiden University Medical Center, Leiden, Netherlands. Ph.D. student. Recruited study subjects at the local site, collected and entered data at the local site, reviewed, contributed to, approved of the final version of the manuscript.

Michelle T. Vaughn, M.P.H., Anesthesiology, Michigan Medicine, Ann Arbor, Michigan. Lead Statistician. Planned and designed the overall study, supervised and directed the overall project, provided scientific advice and guidance of the overall project, supervised the performance of the study at the local site, coordinated the performance of the study at the local site, created software and database infrastructure for the project, interpreted the results, reviewed, contributed to, approved of the final version of the manuscript, consulted and commented on the manuscript.