Third attempt to send to nmusers. Ignore if you have already received this message.
Bob: First. My apologies in case you are getting repeated e-mails from me. I am finding I need to attempt several times for nmusers list-serv system to publish my e-mails. Something to do with >40000 characters. So I need to cut off some of the trailing e-mail portions that have already been published. In my original investigations, the normal and t-distribution were using Box-Mueller type random sample generations, which of course did not work well, so I switched to inverse CDF methods. For normal distribution (DF=0), the IMP uses the modified GASDEV, where you see that when ISELECT=5 (Sobol), DINVNORM is used, and the the Box-Mueller method is used when Sobol method is not used. Robert J. Bauer, Ph.D. Vice President, Pharmacometrics, R&D ICON Development Solutions 7740 Milestone Parkway Suite 150 Hanover, MD 21076 Tel: (215) 616-6428 Mob: (925) 286-0769 Email: [email protected]<mailto:[email protected]> Web: www.iconplc.com<http://www.iconplc.com/> From: Bob Leary [mailto:[email protected]] Sent: Monday, May 19, 2014 2:36 PM To: Bauer, Robert; [email protected]<mailto:[email protected]> Subject: RE: [NMusers] SAEM and IMP Thanks, Bob. This indeed has been an interesting and thought provoking discussion. I too took another look at this. I thought, without really doing any analysis, that the directionality bias using independent univariate t-components would get really bad as the dimensionality increased, because that’s the way it works with using fat-tailed independent double exponentials (samples there tend to fail disproportionately near the coordinate axes, and this gets arbitrarily bad as the dimensionality increases without bound). But when I actually did the analysis in the student t-case, it’s not bad at all and simply goes to a limit as the dimensionality increases. So in fact, there may be some real advantages, particularly in the Sobol quasi-random case, to doing it the way you do with independent univariate t’s rather than using the multivariate t. So my original message to Emmanuel to discourage him from using Sobol in combination with the t-distribution was based on the erroneous assumption that IMP was using the multivariate-t rather than Independent univariate t’s. So his suggestion is probably OK (except as you noted, that an even value of DF might be cleaner than DF=7 given the way tdev2 works). But there is still the question of why you were seeing a bias with your original method before going to tdev2 –was it simply a different method of generating the same distribution? Along somewhat similar lines, I did notice that your routine GASDEV for generating normal N(0,1) components (presumably used in the DF=0 case) actually offers two different ways of doing it – Box-Muller and inverse cdf of a uniform 0-1. It is not clear which one you get when you simply specify DF=0. The inverse cdf method of course is by definition OK for the quasi-random case – the cdf of the random vectors generated with the Sobol distribution using the inverse cdf method will simply be that Sobol distribution. But for Box Muller, this certainly will not be the case, and it is not clear what the discrepancy of the Box Muller generated CDF will be like. Certainly Box Muller is obviously OK in the pseudorandom case, and will have the same discrepancy as the inverse cdf method using pseudorandom uniform 0-1 inputs, but whether the CDF of a Box Muller – generated sequence of Normal random vectors using Sobol vector inputs is actually a low discrepancy sequence (or even a relatively low discrepancy sequence) is not at all clear. I know this has been looked at in 2-dimensions and there Box Muller empirically looks OK, but I don’t know if anything has actually been proved. The low discrepancy property tends to be fragile with respect to nonlinear transformations. So it is possible that tdev2 actually does preserve this property , or at least does relatively little damage to it compared to whatever you were using before. From: Bauer, Robert [mailto:[email protected]] Sent: Monday, May 19, 2014 11:08 AM To: Bob Leary; [email protected]<mailto:[email protected]> Subject: RE: [NMusers] SAEM and IMP Bob: Yes, these are 1 Dimensional t-distribution random deviates generated by TDEV2, followed by scaling p sets of them with an offset vector and cholesky matrix to provide covariance and mean. It provides the tails as you say (property a), and is conveniently used in the Sobol process. Your discussion on the properties of various t-distribution sample creation techniques and their possible impact in importance sampling is interesting. With that in mind, I just completed testing example6, which has an 8 dimensional eta space, and found no bias in the objective function evaluation, or in parameter assessment when setting DF=2 or 4 (I have not tried others), with or without using Sobol (RANMETHOD=3S2). Also, using the 1 dimensional t-distribution random deviates retained the Sobol method’s stochastic noise reduction ability in this example. When I used the multivariate t-distribution algorithm, Sobol’s stochastic noise reduction was not as good, confirming what you related earlier. As you point out as well, property b (radial symmetry) may or may not be needed depending on the posterior density. I am inclined to think that since the posterior density is not going to be perfectly t-distributed or normal distributed anyway, then the sampling density matching the posterior density regarding the radial symmetry property may be of less of relevance. The more important aspect may be to choose a sampling density that has long tails in cases where the posterior density also has long tails, to promote the general efficiency of the sampling density for that posterior density. It is possible that because the sampling density is also dynamically scaled to best fit the posterior density, this reduces any inefficiency in fitting the posterior density that might occur from the sampling density not having radial symmetry. Robert J. Bauer, Ph.D. Vice President, Pharmacometrics, R&D ICON Development Solutions 7740 Milestone Parkway Suite 150 Hanover, MD 21076 Tel: (215) 616-6428 Mob: (925) 286-0769 Email: [email protected]<mailto:[email protected]> Web: www.iconplc.com<http://www.iconplc.com/> <br /><br /> ICON plc made the following annotations. ------------------------------------------------------------------------------ This e-mail transmission may contain confidential or legally privileged information that is intended only for the individual or entity named in the e-mail address. If you are not the intended recipient, you are hereby notified that any disclosure, copying, distribution, or reliance upon the contents of this e-mail is strictly prohibited. If you have received this e-mail transmission in error, please reply to the sender, so that ICON plc can arrange for proper delivery, and then please delete the message. Thank You, ICON plc South County Business Park Leopardstown Dublin 18 Ireland Registered number: 145835
