Table 4 Main characteristics of each approach used to model SNP-SNP interactions
Approach | Type of two-locus model detected | Pattern of complex interactions | Potential advantages | Potential limitations | Possible improvements |
|---|---|---|---|---|---|
LRM | Logical AND models – multiplicative models | Can not be investigated | Easy to fit | Curse of dimensionality | Logic regression MARS* |
CART | Conditional recessive or dominant models | Driven by SNP main effects and binary splits | Deals with sparse data Useful for risk characterization and prediction | Influence of main effects Redundancy | Random forest Boosting |
MDR | All types | Diverse | Deals with sparse data Useful for risk characterization and prediction | Over-fitting Difficult to find best models Inefficient with large number of SNPs | Limit plausible genetic models Use test statistic |