Reducing Bias in Diagnostic Tests of Health Measures with the Propensity Score Method

Xian Liu, Uniformed Services University of the Health Sciences (USUHS)
Charles Engel, Jr., Uniformed Services University of the Health Sciences (USUHS)
Kristie Gore, Walter Reed Army Medical Center
Michael Freed, Walter Reed Army Medical Center

This study develops a propensity score method seeking to evaluate biases in performing diagnostic tests of health measures. The propensity score serves as a distributional balance of covariates and is predicted by the binomial logistic regression with disease diagnosis as the response variable and a number of covariates as predictor variables. We develop two multivariate regression models to validate health measures. For health measures that are dichotomized, we establish a binomial logistic regression model to estimate sensitivity, specificity, and the likelihood ratio. For health measures involving more than two levels, we use the multinomial logit regression to estimate the probability of a true positive result, the probability of a false negative result, and the likelihood ratio at each measurement level. Our empirical examples demonstrate that without considering an individual’s demographic, socioeconomic and other relevant characteristics, results from diagnostic tests of health measures can be biased and misleading.

  See paper

Presented in Poster Session 1