Skip to main content
Article
Thinking of having a higher predictive power for your first-stage model in propensity score analysis? Think again
Health Services & Outcomes Research Methodology (2008)
  • Liang Li
Abstract
The predictive power of the first-stage propensity score (PS) model is commonly reported in clinical publications via c-statistics for logistic regressions. A c-statistic greater than 0.80 was recommended in a recent publication. However, we argue that a cutoff like this may not be the best determinant of the first stage PS model, and it is a misconception that the higher predictive power always implies a better PS model. A better way to assess the PS model is to study the relationships between variables of observed confounders, treatment assignment, and outcomes, while the c-statistic can help check the model adequacy. We recommend researchers not blindly craving for high predictive powers with large c-statistics when building the first-stage PS models.
Publication Date
2008
Citation Information
Liang Li. "Thinking of having a higher predictive power for your first-stage model in propensity score analysis? Think again" Health Services & Outcomes Research Methodology (2008)
Available at: http://works.bepress.com/LiangLi-Biostatistician/23/