This study has dual purposes: (1) survey of the specification and estimation of (i) the linear probability model, (ii) probit model, (iii) logit model, and (iv) Gompit model in which qualitative dependent variables are (i) dichotomous and (ii) polytomous, and (2) empirical estimations and comparison of predictability of these models in which holdings of television sets are presented as binary dependent variables. It is obtained that the predictability defined as revised MSE is greatest in cases of probit and logit models and somewhat less with a linear model. However, the differences are not very remarkable.