Propensity score approaches in quantifying effects of treatment from observational data

ABSTRACT

Introduction@#Despite the growing popularity of utilizing observational studies for determining associations with public health implications, there is limited literature using them for examining and quantifying the effects of exposures or treatments The study compared traditional regression with scoring approaches in estimating treatment effects considering the noted limitations in the dataset. @*Methods@#We conducted a secondary analysis of previously collected retrospective cohort data derived from maternal-neonatal dyads delivered prematurely in a tertiary hospital. Propensity scores (PS) were estimated using logistic and boosting regression. These scores were implemented into matching, stratification, and weighting models. The estimated measures of effect from traditional regression and PS-adjusted models were compared using certain metrics (i.e., the width of CI, SE, AIC, BIC). Sensitivity analysis was also performed. @*Results@#We included data from 562 patients (123 untreated and 439 treated). Both the estimated scores demonstrated satisfactory fit and reduction in the standardized differences between the groups. However, the logit-estimated scores had better prediction (AUC 0.71 vs 0.66) and forecasting properties (Brier 0.15 vs 0.17) than the boosting-estimated scores. All generated statistical models demonstrated a reduction in the occurrence of respiratory morbidity among preterm neonates exposed to a single-dose antenatal corticosteroid (ACS) (ORs ranged from 0.37 to 0.59). The estimated average treatment effects (ATE) and effect among those treated (ATET) from various models suggested a small benefit attributed to the single-dose ACS (ATEs range from -0.09 to -0.41; ATETs range from -0.07 to -0.17). @*Conclusion@#PS estimated using logistic regression performed better than those estimated using machine learning strategies. The matching model using the said scores demonstrated better fit and parsimony over conventional and propensity-adjusted models. Future studies are recommended to improve the application of these analytic techniques in real-world data.