IADR Abstract Archives

Parametric and Nonparametric Propensity Score Weighting in a Non-randomized Study

Objective: Lack of randomization in group assignment for intervention studies can lead to imbalances in pre-intervention covariates and biased effect estimates. Researchers are often skeptical of propensity score (PS) analyses, a common method for balancing groups, because of the potential for unobserved confounders. We examined an Early Head Start (EHS) dataset with rich pretreatment information, rendering the no unobserved confounders assumption plausible. However, simultaneously balancing dozens of pretreatment variables is challenging.  We compare PS results using standard logistic regression models (LRM) versus generalized boosted models (GBM).

Method: We estimated propensity scores using 45 socio-demographic characteristics and EHS enrollment criteria obtained by parent interviews from a state-wide sample of 637 EHS and 930 Medicaid-matched control children. LRM and GBM were used to estimate propensity scores for EHS enrollment. Because standard LRM drops observations with missing values, we imputed means to estimate propensity scores for all subjects. Performance of both approaches was evaluated by similarities in both (1) pre-treatment covariate distributions between treated and control subjects; and (2) propensity score weights measured by the effective sample size.

Results: Distributions of all variables were balanced for EHS versus non-EHS groups using propensity score weights calculated with LRM and GBM.  GBM resulted in the preferred lower maximum effect sizes and Kolmogorov-Smirnov distances, both of which are measures of similarity between treated and propensity score weighted control distributions. The effective sample size of the controls decreased from 930 subjects to 500 with GBM and to 300 with LRM.

Conclusion: Although propensity scores derived from GBM and LRM both effectively balanced pre-intervention covariates, GBM resulted in better balance. LRM cannot automatically handle missing values and resulted in a smaller effective sample size. PS adjustment using GBM is an effective statistical method to reduce confounding because of imbalanced distributions of pre-intervention covariates in this EHS intervention study.

Division: AADR/CADR Annual Meeting
Meeting: 2014 AADR/CADR Annual Meeting (Charlotte, North Carolina)
Location: Charlotte, North Carolina
Year: 2014
Final Presentation ID: 9
Abstract Category|Abstract Category(s): Behavioral, Epidemiologic, and Health Services Research
Authors
  • Burgette, Jacqueline  ( University of North Carolina at Chapel Hill, Chapel Hill, NC, USA )
  • Preisser, John  ( University of North Carolina, Chapel Hill, NC, USA )
  • Rozier, Richard Gary  ( University of North Carolina, Chapel Hill, NC, USA )
  • SESSION INFORMATION
    Oral Session
    Dental Public Health
    03/19/2014