A new approach to adjust for multivariate confounders in small randomized studies applied to dendritic cell...

In small studies, randomization alone is unlikely to eliminate confounding. In linear models, confounding can be addressed by including additional dependent variables, adding covariates, or stratifying the data for the analysis. When the functional relation between observed variables and underlying (latent) factors is unknown, however, methods based on ranks may be more appropriate than ANOVA. It is demonstrated, how the marginal likelihood principle can be used to provide objective and intrinsically valid procedures to adjust for (multiple) confounders when the assumptions of the linear model cannot be justified.