Dear Imputers, thanks to Rod and Jae I got ahead in distinguishing between Bayesianly proper and "frequentist" proper MI.
The procedure I was asking about is the generalization of Rubin's proposal for statistical matching (in Europe we call it data fusion), see Rubin (1987), p. 187-188, "two variables never jointly observed". I programmed it among others in SPLUS and did some simulation studies. It`s frequentist properties are obviously nice so far. My point of interest is estimating the unknown correlation rho of the variables never jointly observed, which clearly depends on the prior used for rho. But is it a "frequentist" proper MI? Many thanks again! Susanne P.S.: If there is interest, I can attach a postscript file containing the algorithm. ------------------------------------------------------------------- Dr. Susanne R?ssler Institute of Statistics and Econometrics University of Erlangen-Nuernberg, Germany email: [email protected]
