|dc.description.abstract||Background: The propensity score is often used to obtain unbiased effect estimates in the presence of confounding. The prognostic score is a method similar to the propensity score that estimates the conditional probability of the outcome. While the prognostic score has been applied to address confounding, an application to selection bias has not been explored. The existing methods to correct for selection bias are limited when dealing with non-participation in forming the study population. The prognostic score could be potentially applied to obtain unbiased effect estimates in the presence of selection bias.
Objective: This thesis explores the use of the prognostic score as a method to address confounding and selection bias in epidemiologic research. The first investigation of the prognostic score will employ multiple forms of the prognostic score to estimate the unbiased effect in the presence of confounding. We will show that the propensity score and a modified version of the prognostic score yield equivalent effect estimates for a logistic model. The second investigation compares the use of the prognostic score in the presence of selection bias to existing methods including inverse probability of selection weights (IPSW) and direct adjustment with and without the presence of confounding.
Design: Based on several directed acyclic graphs, Monte Carlo simulations compared several approaches to isolating the effect estimate. In the presence of confounding, weighting using three variations of the prognostic score were compared to weighting using the propensity score. Approaches to combining the prognostic and propensity score were also investigated. In the presence of selection bias, weighting using the three prognostic score approaches, IPSW, and direct adjustment were compared.
Main Outcomes: Percent relative bias, robust variance estimates, Monte Carlo variance estimates, and mean squared error with respect to the marginal and conditional odds ratios were compared between all of the methods.
Results: In the presence of confounding, the stabilized modified prognostic score weights and stabilized inverse probability of exposure weights yielded the marginal odds ratio, while the combination prognostic and propensity score approaches, the unexposed prognostic score weights, and the full population prognostic score weights resulted in the conditional odds ratio. For the selection bias simulations, the unexposed and full population prognostic score weights estimated the conditional odds ratio and were comparable to direct adjustment methods. The modified prognostic score yielded a result that appeared to be a mix of the marginal and conditional odds ratio. In the presence of unmeasured selection variables, the prognostic score approaches and direct adjustment were biased.
Conclusions and Relevance: The prognostic score is an acceptable alternative method to adjust for confounding and for selection bias except for when the selection variable acts as a collider in the presence of unmeasured variables.||