PMC:3987104 / 8159-9246 JSONTXT

{"project":"2_test","denotations":[{"id":"24669828-18816511-8120650","span":{"begin":285,"end":287},"obj":"18816511"}],"text":"The GA-MM methodology makes use of the Simple Genetic Algorithm (Table 1), completely analogous to GA-OLS, producing a ranking of variables by their frequency in a set S of GA solutions. However, there is no single commonly used definition for the R2 statistic as is the case for OLS [16,17]. Several definitions have been suggested that all have different interpretations in the presence of correlated errors. Here, we used the marginal R2MM definition from [18], quantifying the variance explained by the fixed effects. As new data will originate from other subjects than those used for the training of the model, the random effects cannot be used for prediction. In [1] it has also been described that conditional R2 (variance explained by the entire model, including the random effects) should not be used for fixed-effect variable selection. For us, the main motivation for using R2MM was that the MM can be fitted using REML, resulting in better estimates for the variance components, needed in the estimation of the fixed effects, especially in models with many fixed effects [7]."}