Estimation of conditional and marginal odds ratios using the prognostic score

Area of Science:

• Epidemiology
• Biostatistics
• Statistical modeling

Background:

• The prognostic score (PGS) is analogous to the propensity score (PS), primarily used for estimating marginal treatment effects.
• Previous evaluations of PGS methods focused on collapsible situations where conditional and marginal effects are equivalent.
• Applied researchers need clarity on the specific treatment effects estimated by different PGS methods for robust application.

Purpose of the Study:

• To evaluate existing and introduce novel PGS-based methods for estimating conditional treatment effects (CTE) and marginal average treatment effects (ATE/ATT).
• To compare the performance of PGS methods against PS-based methods and multivariate regression, particularly using the non-collapsible odds ratio.
• To provide guidance on selecting appropriate PGS methods based on the desired treatment effect and exposure prevalence.

Main Methods:

• Monte Carlo simulations were employed to assess the performance of various PGS-based methods.
• The study compared existing and newly proposed PGS methods with traditional PS-based techniques and multivariate regression.
• Performance was evaluated for estimating conditional treatment effect (CTE), average treatment effect (ATE), and average treatment effect on the treated (ATT).

Main Results:

• Existing PGS methods failed to estimate ATE and performed poorly with high exposure prevalence.
• New PGS methods demonstrated better performance for marginal effect estimation in low-prevalence scenarios, with coverages closer to nominal values.
• For CTE estimation, new PGS methods performed comparably to standard multivariate regression analysis.

Conclusions:

• The proposed PGS-based methods offer a viable alternative to PS methods for estimating specific treatment effects.
• New PGS methods show promise for estimating marginal effects, especially when exposure is infrequent.
• PGS methods, particularly the new ones, provide reliable estimation of conditional treatment effects, similar to traditional regression techniques.