The paper tackles a problem which arises during the analysis of binary models, and which is the heteroskedasticity of a random element manifested by the variable value of variance. In the paper, the following probability models, used in the analysis of a dichotomic variable, were considered: a logit model, probit model, and raybit model, which is a model proposed by the authors. The following measures of goodness of fit, present in the field literature, were considered: MSE, MAE, WMSE, and WMAE. A new measure of goodness of fit of a model was proposed, which limits the amplitude of varying values of variance.