JMP uses the multivariate radial strata method for each factor that uses the Normal Weighted distribution. This seems to work better than a number of Importance Sampling methods, as a multivariate Normal Integrator accurate in the extreme tails.
First, define strata and calculate corresponding probabilities and weights. For d random factors, the strata are radial intervals as follows.
NStrata – 1
|
1.
|
2.
|
Determine a random n-dimensional direction by scaling multivariate Normal (0,1) deviates to unit norm.
|
The resulting factor distributions are multivariate normal with the appropriate means and standard deviations when estimated with the right weights. Note that you cannot use the Distribution standard deviation with weights, because it does not estimate the desired value. However, multiplying the weight by a large value, like 1012, and using that as a Freq value results in the correct standard deviation.