Abstract: | For linear models, randomization of the assignment to the levels of a variable A is a sufficient condition to obtain correct estimates of the effect of A on another variable B. For logit models, the omission of a relevant variable Z often results in invalid estimates of the effect of A and in a loss of power of the significance test. The values of this underestimation and of the loss of power are evaluated here asymptotically and from simulation studies. Consequences for practical use of logit models in randomized settings are underlined. |