Direct and indirect effects in a logit model

Area of Science:

• Statistics
• Econometrics
• Biostatistics

Background:

• The decomposition of effects in logit models is crucial for understanding complex relationships.
• Erikson et al. (2005) proposed a method for effect decomposition, but it had limitations.

Purpose of the Study:

• To extend the Erikson et al. (2005) method for decomposing effects in logit models.
• To generalize the method to accommodate variables with any distribution.
• To introduce standard error estimation using bootstrapping.
• To enable the inclusion of control variables in the decomposition.

Main Methods:

• Generalization of the original logit effect decomposition method.
• Application of the bootstrap method for standard error estimation.
• Inclusion of control variables within the decomposition framework.
• Implementation of the extended method in the R package 'lddecomp'.

Main Results:

• The extended method successfully decomposes total effects into direct and indirect components.
• The generalized approach accommodates mediator variables with non-normal distributions.
• Bootstrap standard errors provide reliable estimates for the decomposed effects.
• The inclusion of control variables allows for more robust and nuanced analysis.

Conclusions:

• The enhanced method offers a more flexible and comprehensive approach to effect decomposition in logit models.
• The 'lddecomp' package provides a practical tool for researchers to implement these advanced techniques.
• This work contributes to a deeper understanding of causal pathways in statistical modeling.