Combined estimators for the correlation coefficient

From MaRDI portal





The common correlation coefficient \(\rho\) of k binormal distributions is to be estimated by means of k random samples taken independently. The author derives three differently combined estimators of \(\rho\), functions of the k sample correlation coefficients \(r_ i\) and sizes \(N_ i\). The approximate maximum likelihood estimator \({\hat \rho}\) is the unique solution in (-1,1) of a cubic equation, obtained with the help of series expansions from the exact maximum likelihood equation based on the distribution of \(r_ i\) in Hotelling's form, i.e. involving the hypergeometric function \(F(\frac{1}{2},\frac{1}{2},N_ i- \frac{1}{2},(1+\rho r_ i)/2).\) The second estimator \({\tilde \rho}=(N-k)^{-1}\sum^{k}_{i=1}(N_ i-1)r_ i[1+(1-r^ 2\!_ i)/2(N_ i-4)], N=\sum^{k}_{i=1}N_ i\), is a linear combination of k approximate minimum variance unbiased estimators of \(\rho\), given by \textit{I. Olkin} and \textit{J. W. Pratt} [Ann. Math. Stat. 29, 201-211 (1958; Zbl 0094.144)], whereas the third one is defined by \({\bar \rho}=\tan gh {\bar \xi}\), where \({\bar \xi}=(N- 3k)^{-1}\sum^{k}_{i=1}(N_ i-3)z_ i\) is a weighted function of the Fisher-transformed \(z_ i=\ln \sqrt{(1+r_ i)/(1-r_ i)},\) and therefore asymptotically normal with variance \((N-3k)^{-1}.\) Properties of the considered estimators are studied, and compared by analyzing computer simulated data \((k=3\), for each of 17 different values of \(\rho\) and of 8 different combinations \((N_ 1,N_ 2,N_ 3)\) by a Monte Carlo experiment with 1000 repetitions).











This page was built for publication: Combined estimators for the correlation coefficient

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3316399)