Detecting Heterogeneity of Intervention Effects in Comparative Judgments

Area of Science:

• Statistics
• Psychology
• Public Health

Background:

• Comparative measures like paired comparisons and rankings are vital for assessing health states and quality of life.
• Existing models may not fully capture treatment effect heterogeneity in such data.

Purpose of the Study:

• Introduce log-linear Bradley-Terry (LLBT) models for intervention effectiveness evaluation with paired comparison or ranking data.
• Present a combination of LLBT and model-based recursive partitioning (MOB) to detect treatment effect heterogeneity.
• Enable identification of subgroups with differing preference orders and intervention effects on choice behavior.

Main Methods:

• Log-linear Bradley-Terry (LLBT) models were employed.
• Model-based recursive partitioning (MOB) was combined with LLBT to detect treatment effect heterogeneity.
• The MOB LLBT approach was demonstrated using artificial and real-world data examples.

Main Results:

• The MOB LLBT model successfully recovered the true model in the artificial data example.
• In real-world data, LLBT confirmed situational drug-harm trivialization among festival visitors when peer behavior was accessible.
• MOB LLBT revealed this trivialization effect is context-dependent, most pronounced in moderately intoxicated individuals who sought counseling.

Conclusions:

• MOB LLBT models provide more nuanced assessments of intervention effectiveness.
• The approach is applicable to paired comparisons, rankings, and rating data.
• R code examples are provided for implementing MOB LLBT models.