Logistic Liu estimator under stochastic linear restrictions (Q2423199): Difference between revisions

The general logistic regression model is considered. The following stochastic prior information is given $h=H\beta+v$, $E(v)=0$. Here $h$ is a known vector, $H$ is a full rank known matrix, $\beta$ is a vector of regression coefficients to be estimated, and $v$ is a random vector of disturbances with known positive definite variance-covariance matrix. It is assumed that $v$ is stochastically independent of the noise in the main model. \par In the presence of exact linear restrictions (i.e., for the case $v=0$) \textit{D. E. Duffy} and \textit{T. J. Santner} [Commun. Stat., Theory Methods 18, No. 3, 959--980 (1989; Zbl 0696.62140)] proposed the restricted maximum likelihood estimator (RMLE). \par In the paper under review, a new estimator called stochastic restricted Liu maximum likelihood estimator (SRLMLE) is constructed by incorporating Liu estimator (see \textit{K. Liu} [Commun. Stat., Theory Methods 22, No. 2, 393--402 (1993; Zbl 0784.62065)]) to the logistic regression with the stochastic linear restrictions. The conditions for superiority of SRLMLE over MLE and restricted Liu logistic estimator (RLLE) are derived w.r.t. mean square error criterion. The simulation study shows that SRLMLE is superior over the two other estimates in all the cases except the case of very high degree of collinearity $\rho=0.99$ and small sample size $n=25, 50$.