Confounding adjustment performance of ordinal analysis methods in stroke studies

Area of Science:

• Biostatistics
• Clinical Trials
• Neurology

Background:

• Ordinal logistic regression (OLR) is commonly used for analyzing modified Rankin Scale (mRS) outcomes in stroke research.
• The non-parametric Mann-Whitney measure of superiority (MWS) is an alternative, but comparative performance with confounding adjustment is unclear.

Purpose of the Study:

• To quantitatively assess the performance of OLR and MWS under various confounding variable scenarios.
• To compare adjusted OLR (aOLR) with stratified MWS (sMWS) and propensity score-stratified MWS (psMWS).

Main Methods:

• A simulation study was conducted with three scenarios: dichotomous, continuous, and stroke-mimicking confounding variables.
• Compared aOLR and sMWS, and also utilized propensity scores for MWS stratification (psMWS).
• Evaluated bias, point estimate variation, and coverage probability, transforming OLR estimates to MWS for comparability.

Main Results:

• In scenario 1 (dichotomous confounders), both methods showed no bias and similar performance.
• Scenario 2 (continuous confounders) revealed more bias and variation with sMWS compared to aOLR.
• Scenario 3 (stroke-mimicking) showed minimal bias for both, but aOLR had lower point estimate variation; psMWS demonstrated excellent performance with no bias and low variation.

Conclusions:

• Bias was similar between methods in the stroke simulation scenario.
• Propensity score stratification significantly improved MWS performance, reducing point estimate variation.
• Stratified MWS is a viable alternative to aOLR, especially when propensity score stratification is employed, broadening its use.