D. Freedman, Evaluation Review, December 2006:
"Experimental data are often analyzed by fitting regression models and their ilk. As shown in section 10, randomization does not justify such models. Experimental data should therefore be analyzed first by comparing rates or averages, according to the intention-to-treat principle. Only then should models be deployed. It must be emphasized that statistical adjustments to experimental data often depend substantially on assumptions, not just on randomization."