with a model containing all the independent (predictor) variables of interest. Then, at each step the variable with the smallest F-statistic is deleted (if the F is not higher than the chosen cutoff level).