Improved estimation for a new class of parametric link functions in binary regression (Q2188756)

The authors discuss binary regression problems where usually the probability \(\mu=P(Y=1)\) is modelled by \(\mu=F(\eta)\) where \(F\) is a cumulative distribution function (CDF) and \(\eta\) follows a linear regression setup \(\eta=x^T\beta\). Usually \(F^{-1}\) is called link funktion. The authors use a new class of asymmetric link functions \(F=G^\lambda\) where \(G\) is a CDF of a symmetric distribution. The parameter \(\lambda\) characterizes the skewness of the link function. For estimation of \(\beta\) and \(\lambda\) the Maximum Likelihood (ML) method is used which typically leads to biased estimators for finite samples. The chief goal of the paper is to derive a closed-form expression for the bias of the ML estimators which is used then to define bias corrected ML estimators. Numerical results from Monte Carlo experiments show that the bias corrected ML estimators are nearly unbiased.