Effect of Multidisciplinary Transitional Pain Service on Health Care Use and Costs Following Orthopedic Surgery

The study patient cohort was stratified into either NOU or COU groups based on opioid use before surgery. Preoperative COU was defined as > 25% nonzero days (calculated using the cabinet supply method) in the 180 days before surgery admit date time (> 45 nonzero days) or ≥ 1 opioid prescription for ≥ 28-day supply released within 90 days before surgery date.^12,13 For NOU patients, the postoperative outcome of interest was new postoperative prolonged opioid use. This was defined by Page and colleagues as ≥ 1 opioid prescription released between surgery discharge date and postdischarge day 44, ≥ 1 opioid prescription released between postdischarge day 45 and 89, and ≥ 1 opioid prescription released between postdischarge day 90 and 180.¹⁴ For COU patients at the time of surgery, the postoperative outcome measure of interest was continued COU, defined as > 25% nonzero days between postdischarge day 90 and 240 (> 37 nonzero days) or ≥ 1 opioid prescription for ≥ 28-day supply released between postdischarge day 90 and 180.

Outcome Variables

Outcome variables included health care use and costs during 1-year pre- and postperiods from the date of surgery. VA health care costs for outpatient, inpatient, and pharmacy services for direct patient care were collected from the Managerial Cost Accounting System, an activity-based cost allocation system that generates estimates of the cost of individual VA hospital stays, health care encounters, and medications. Health care use was defined as the number of encounters for each visit type in the Managerial Cost Accounting System. All costs were adjusted to 2019 US dollars, using the Personal Consumption Expenditures price index for health care services.¹⁵

A set of sociodemographic variables including sex, age at surgery, race and ethnicity, rurality, military branch (Army, Air Force, Marine Corps, Navy, and other), and service connectivity were included as covariates in our regression models. In addition, ICD-9 and ICD-10 codes were used to define the indicators of whether veterans had a diagnosis of mental illness (including anxiety, bipolar disorder, depression, or trauma) or substance use (including alcohol, cannabis, opioids, or tobacco). Finally, pharmacy records were used to create indicators for prescribed opioid-relevant pharmacotherapy (including antidepressants, benzodiazepines [BZD], gabapentinoids, muscle relaxants, non-BZD sedative hypnotics) and active antidepressant drug use during the 1-year preindex period.

Statistical Analyses

Descriptive analyses were used to evaluate differences in baseline patient sociodemographic and clinical characteristics between pre- and postperiods for TPS intervention and control cohorts using 2-sample t tests for continuous variables and χ² tests for categorical variables. We summarized unadjusted health care use and costs for outpatient, inpatient, and pharmacy visits and compared the pre- and postintervention periods using the Mann-Whitney test. Both mean (SD) and median (IQR) were considered, reflecting the skewed distribution of the outcome variables.

We used a DID approach to assess the intervention effect while minimizing confounding from the nonrandom sample. The DID approach controls for unobserved differences between VAMCs that are related to both the intervention and outcomes while controlling for trends over time that could affect outcomes across clinics. To implement the DID approach, we included 3 key independent variables in our regression models: (1) an indicator for whether the observation occurred in the postintervention period; (2) an indicator for whether the patient was exposed to the TPS intervention; and (3) the interaction between these 2 variables.

For cost outcomes, we used multivariable generalized linear models with a log link and a Poisson or Υ family. We analyzed inpatient costs using a 2-part generalized linear model because only 17% to 20% of patients had ≥ 1 inpatient visit. We used multivariable negative binomial regression for health care use outcomes. Demographic and clinical covariates described earlier were included in the regression models to control for differences in the composition of patient groups and clinics that could lead to confounding bias.