Application of mediation modeling techniques in social science research

Abstract

Mediation analysis examines whether and how predictor X affects outcome Y via mediator M. Mediation is widely applied in social science research and is useful in evaluating conceptual models and underlying mechanisms of various interventions. Traditional mediation analysis such as the Baron and Kenny approach, Sobel test, and moderated causal steps suffers from methodological issues (biased estimates, lower power, and potential misinterpretations). Given the complexity of constructs in social science, sound research methodology is essential in shaping better evidence-based research. Methodological literature has contributed to significant advance in mediation analysis in the past decades. However, these modern techniques are not well-known among applied researchers and applications in substantive research lag behind. Such applications could allow more systematic examinations and accurate inference of substantive research. The present thesis aimed to bridge the gap between modern methods in the statistical literature and those used in the applied research. Methodological pitfalls of existing mediating methods were briefly reviewed, followed by delineation of contemporary methods on topics of longitudinal mediation, moderated mediation, causal mediation, and multilevel mediation. The utility of these modern methods was demonstrated via four independent studies in behavioral science. Study 1 examined longitudinal mediation among neurological soft signs, psychiatric symptoms, and functional outcomes in 151 Chinese schizophrenia patients. Using path analysis with bootstrapping, neurological soft signs in motor coordination showed significant and negative indirect effects on functional outcomes via psychiatric symptoms. Study 2 conducted moderated mediation analysis on the role of subjective perceived stress in the physiological effect of Dance Movement Therapy in 121 Chinese breast cancer patients. Despite the lack of significant overall treatment effect and indirect effect, perceived stress significantly moderated the direct treatment effect. The findings implied treatment heterogeneity in terms of improved endocrinological functioning for patients with heightened stress. Behavioral science research often classifies continuous data into binary variables based on cutoff values for diagnosis. The binary mediators/outcomes require causal effect formula to derive the causal direct and indirect effects based on potential outcomes. Using the Add Health sample of 4545 American adolescents, Study 3 evaluated the indirect effect from adiposity to glucose dysregulation via inflammation from a counterfactual perspective. Causal moderated mediation analysis showed significant and positive indirect effects from BMI growth to glycated hemoglobin via C-reactive protein across gender. The significant exposure-mediator interaction effect implied stronger indirect effects via CRP for adolescents with faster BMI growth. Conventional multilevel analysis suffers from Lüdtke’s bias and observed centering issues. Based on 6217 Chinese adults recruited from 218 communities in China Health and Nutrition Survey, Study 4 conducted multilevel mediation analysis with latent centering on the effects of urbanization growth on metabolic syndrome via lifestyle factors across gender. Urbanization showed significant and positive indirect effects on metabolic syndrome via changes in physical activity and fast food intake in males and females, respectively. This suggests gender-specific mediating role for lifestyle factors in the urbanization effect on metabolic syndrome. Finally, study limitations, empirical contributions, and future directions were discussed with reference to the substantive research fields in the social sciences.