[EMAIL PROTECTED] writes: >Yet r is estimated by averaging the estimates of r >across multiple imputations. In general, these estimates will not agree: >if r>0, then the estimate of R^2 will be less than the squared estimate of >r. If the estimator of r is unbiased, then the proposed estimate of R^2 must >be biased.
Let me say that the problem is similar to trying to estimate the unit variance of a variable from several studies. Should you average variances or standard deviations? In your case both r and R-square have the characteristic of being unaffected if every unit is repeated the same number of times (though the confidence intervals will be affected). Thus if you had ten identical data sets and concatenated them, your correlations would be the same as for any one of them. Now suppose the ten were not identical, but differed for the imputed values. Intuitively, I would think the r and R-square that would result from concatenating the ten data sets would be a consistent (unbiased ???) estimate. Let me say that I am not clear if you should convert both variables to z scores within each data set first. But the approach of combining data sets to obtain a single point estimate seems reasonable.