Day-of-Surgery Gabapentinoids and Prolonged Opioid Use

A Retrospective Cohort Study of Medicare Patients Using Electronic Health Records

Jessica C. Young, PhD; Nabarun Dasgupta, PhD; Brooke A. Chidgey, MD; Til Stürmer, MD, PhD; Virginia Pate, MS; Michael Hudgens, PhD; Michele Jonsson Funk, PhD


Anesth Analg. 2021;133(5):1119-1128. 

In This Article


This study was approved under the University of North Carolina Institutional Review Board (IRB) 18–1248, and the requirement for written informed consent of these retrospective data was waived by the IRB. This article adheres to the applicable STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) guidelines.

Data Source

Electronic health records (EHRs) dating from April 4, 2014 to December 16, 2019 from a large integrated health care system in the United States were used. The EHR contains detailed clinical and administrative data for patient care provided across 11 hospitals and over 700 clinics. These data provide an in-depth view of medical encounters, including longitudinal data on diagnosis and procedure codes from any encounter. Relevant to the current study, these data include date and timestamp for start of surgery, surgical procedure code, preoperative pain scores, outpatient medication orders, and inpatient medication administrations. The EHRS also include demographic data including height, weight, race, and ethnicity, in addition to self-reported alcohol and tobacco use.

Study Population

Patients undergoing major therapeutic surgical procedures (nonocular) between January 1, 2016 and September 16, 2019 within the 2 main surgical facilities in the integrated health care system were identified.[15] Both inpatient and outpatient surgeries were included. For patients undergoing multiple surgeries, only the index surgery was examined. Inpatient surgeries were limited to those with a total length of stay of ≤4 nights, with patients discharged home for self-care.

Patients with any outpatient medication orders for gabapentinoids or diagnoses of epilepsy or postherpetic neuralgia before surgery were excluded. Patients with a documented history of opioid abuse, addiction, or dependence, or who had evidence of prolonged opioid use (opioid orders in 3 consecutive months) at any time in the 12 months before surgery were excluded. To assess prolonged opioid use, patients were required to have at least 90 days of follow-up after discharge from surgery. Patients who underwent additional surgical procedures, died, or disenrolled from the database during the 90-day follow-up were excluded. The impact of this exclusion on each cohort is reported.


The primary exposure was preoperative gabapentinoids, defined using inpatient medication administration records. Administration records for oral gabapentinoids must have been on the day of surgery with an administration start time stamp before the start of the surgery, with a description of "GIVEN." We identified whether pregabalin or gabapentin was administered and the dosage in milligrams.


We examined the proportion of patients with prolonged opioid use following surgery. Prolonged opioid use was defined as at least 1 outpatient opioid order in each of 3 consecutive 30-day windows immediately following surgical discharge.[16]

Potential Confounding Variables

We reported and adjusted for demographic factors that have been found to be associated with health care delivery and opioid-prescribing practices, including patient gender, age, and patient-reported race (Black, White, and other). Because use of preoperative gabapentinoids and rates of opioid prescribing have changed over calendar time and by institution, we controlled for calendar time using 6-month increments and for medical facility of surgery. To address baseline health imbalances, we adjusted for maximum recorded preoperative pain (0, 1–3, 4–6, and 7+), number of outpatient prescriptions in the previous 6 months (0, 1–6, and 7+), patient-reported smoking history (current smoker, former smoker, never smoker, and other), patient-reported alcohol use (yes versus no), and body mass index (BMI) categorized according to the US Centers for Disease Control and Prevention. We also adjusted for pain-related medications and diagnoses using binary variables indicating the presence of prescriptions for opioids, benzodiazepines, and pain-related diagnoses (arthritis, cancer, depression, chronic back pain, fibromyalgia, neuralgia, headache/migraine, and abdominal pain) at baseline.

Our main analyses controlled for calendar time in 6-month intervals. Because calendar time may be an important confounding variable, we conducted 2 additional analyses using different specifications to account for calendar time. The first modeled surgery date as a continuous variable (number of days from the start of the study period) using a quadratic term, and the second used a cubic spline with 3 knots at the tenth, 50th, and 90th percentiles.[17] Because the health care system in this study first implemented Enhanced Recovery After Surgery (ERAS) protocols including preoperative gabapentinoid recommendations on March 1, 2018, we also conducted analyses examining whether the association between preoperative gabapentinoids and prolonged opioid use differed in the period before any ERAS protocols (surgery between January 1, 2016 and February 28, 2018) and after the implementation of ERAS protocols (surgery between March 1, 2018 and September 16, 2019). We assessed the interaction between the exposure and period and conducted stratified analyses by period.

Statistical Analyses

Because of the rapidly evolving landscape surrounding opioid prescribing and pain management, we report the percentage of patients receiving preoperative gabapentinoids and having prolonged opioid use following surgery by 6-month intervals based on the date of surgery.

Crude and adjusted risk and adjusted risk ratios (adjRR) with 95% confidence intervals (CIs) of prolonged opioid use in the exposed (received gabapentinoids on the day of surgery) and unexposed (no gabapentinoids on day of surgery) were calculated using log-binomial regression.

Logistic regression adjusting for the potential confounders detailed above was used to calculate propensity scores predicting administration of preoperative gabapentinoids. Adjusted estimates were calculated using stabilized inverse probability of treatment weights (IPTWs) to reduce bias due to measured confounders.[18] Exposed subjects received a weight of (1/ps) ×p, where ps represents the predicted probability of exposure to gabapentinoids and p represents the proportion of patients observed as treated with gabapentinoids. Unexposed patients received a weight of (1 − p)/(1 − ps).[19] Asymmetric trimming at the first and 99th percentiles of the propensity score was used to define a study population with greater treatment equipoise resulting in more clinically relevant estimates.[20,21] IPTWs were recalculated among the trimmed population, and balance between the weighted groups was assessed using absolute standardized mean differences (ASMDs), with ASMD <0.1 indicating balance.[22] Separate propensity score models were fit for the main, secondary, and sensitivity analyses described below, and asymmetric trimming with stabilized IPTW was repeated within each analysis. Due to extreme weights, 1.5% asymmetric trimming was used in the calendar time–stratified analyses.

Main Analysis. The main analysis was conducted in the total population of surgeries meeting the inclusion criteria.

Secondary Analysis. A secondary analysis was conducted on a subset of 4 surgical procedures (colorectal resection, hip arthroplasty, knee arthroplasty, and hysterectomy), for which at least 30% of all patients received a preoperative gabapentinoid, in an effort to focus on a clinical population with higher equipoise for whom preoperative gabapentinoids appeared to be a more common part of care. Results in the population undergoing surgical procedures in which <30% of patients received preoperative gabapentinoids are also presented in the Supplemental Digital Content, Supplemental Materials, Due to small sample size, the secondary analysis controlled for calendar time in 1-year increments instead of 6-month increments.

Sensitivity Analysis. Because we observed only the health care and medications received within the health care system from which the EHRs were extracted, we conducted a sensitivity analysis restricting the population to patients with at least 1 outpatient visit and 1 outpatient medication order in the health care system in the 182 days before surgery. This subset of patients represents a group that has a more regular history of interaction with this health care system, for whom we have higher confidence that baseline and follow-up care will be captured in the data. Results in the population who did not meet these criteria are also presented in the Supplemental Digital Content, Supplemental Materials, Due to small sample size, the sensitivity analysis controlled for calendar time in 1-year increments instead of 6-month increments.

Quantitative Bias Analyses. Quantitative bias analyses estimate the impact systematic error (such as outcome misclassification) may have on effect estimates. To examine the potential bias due to imperfect capture of opioid prescriptions during follow-up, we linked a subset of the patients to insurance claims data and conducted 2 sets of quantitative bias analyses. The first analysis addressed the potential of underestimating prolonged opioid use and assumed that any patients with prolonged opioid use in either EHR or Medicare claims data were correctly classified as having prolonged opioid use ("gold standard" was the combined EHR and claims data). The second analysis addressed the potential of overestimating prolonged opioid use in the EHR and treated the Medicare claims data as the "gold standard".[23,24]

For both bias analyses, we estimated the positive predictive value (PPV) and negative predictive value (NPV) of prolonged opioid use measured by the EHR data. Using these estimates and corresponding estimated standard errors (SEs), we conducted a probabilistic bias analysis that reclassifies the data to present bias-adjusted risk ratios (adjRRs) incorporating uncertainty in the measurement of the outcome as well as random error. We resampled the population with replacement to create 1000 pseudopopulations. For each iteration, the NPV and PPV were randomly drawn from normal distributions with mean and standard deviation equal to the estimated NPV and PPV and the corresponding SEs. The mean risk ratio over the 1000 iterations is reported as the bias-adjRR, and the 2.5th and 97.5th percentiles of the 1000 iterations were reported as the 95% CI.[23]

All analyses were conducted using SAS 9.4.