回复 第5楼 的 Tracy X：Actually, I haven't got a general solution. What I did was to generate multivariate normal variables with a correlation matrix (to be determined later), transform each dimension to uniform, and to the specified beta. Then for each pair of the beta variables, find the joint distribution. Compute the correlation of the beta variable in terms of the MVN correlation. Solve this equation to get the specified correlation in betas.
The problem with is that I can get all positive correlations. But if the requested correlation (in beta) is close to -1, there does not exist such a MVN correlation as the root for the last equation.
