Estimating the Association Between Initial Postpartum Discharge Opioid Prescription Dose and Subsequent Opioid Use: The Role of Different Modeling Strategies

Background: Prior work suggests postpartum opioid use is associated with an increased risk of persistent opioid use. Whether low-dose opioid use is ‘safe' after delivery is of great clinical interest, yet the use of simple regression approaches (linear/categorical) to study this question can mischaracterize effects when there is a high proportion of non-users or low-dose data are sparse.

Objectives: To compare modeling strategies for estimating the association between low-dose initial postpartum opioid use after vaginal delivery and total opioid dose used in the subsequent postpartum period.

Methods: We assembled a retrospective cohort of opioid-naïve women with vaginal delivery enrolled in Tennessee Medicaid (2007-2014). Prescription fill data were used to characterize total opioid dose (MME) filled in the initial postpartum period (exposure: day 3 pre-delivery to postpartum day 5) and in the subsequent postpartum period (outcome: postpartum day 6 to 42). Covariates were measured from day 180 pre-delivery to postpartum day 5. We used linear regression, adjusting for covariates, to compare strategies for modeling initial opioid dose, such as: (1) categorical terms, (2) a linear term, (3) natural cubic splines, and (4) fractional polynomials. For continuous models (2-4), we examined the use of a "spike" indicator term to distinguish initial users from non-users. With initial non-users as reference, we examined the dose response curves at low doses ( < 75 MME) across modeling strategies.

Results: Among 147,414 vaginal births, 53.3% of women filled an initial postpartum opioid; few filled doses < 75 MME (2.5%). In models without a "spike" term, the dose-response curves were forced through a y-intercept of 0 and had narrow confidence intervals (CI) (despite sparse data in the low-dose range) that indicated a significant positive association with subsequent opioid dose as initial dose increased from 0 to 75 MME. In contrast, the categorical and spline models with a "spike" term had dose-response curves with wide CI that were not significantly different from the null value at low doses and showed a negative association as initial dose increased from 0 to 75 MME. Importantly, all models demonstrated a significant positive association between an initial opioid dose >150 MME and subsequent postpartum opioid dose.

Conclusions: The association between low-dose initial postpartum use and subsequent opioid use is sensitive to the modeling strategy in the presence of non-linear relationships and infrequent low-dose prescriptions. Studies of opioid doses should use flexible modeling approaches that do not presume a "through-the-origin" dose response curve to prevent a false illusion of safe vs. non-safe doses where they do not exist.