Re: Digital Visual Fortran vers 5.0 vs 6.6

From: Nick Holford Subject: Re: [NMusers] Digital Visual Fortran vers 5.0 vs 6.6 Date: Wed, 20 Mar 2002 13:04:21 +1200 Paul Hutson wrote: My answer is in two parts. First of all I found a bug in NM-TRAN. The bug produces a TIME item in FDATA with inconsistent format and causes data truncation. One kind of data set structure produces FDATA that is E7.0 for TIME whereas changes in data items produces a format of E8.0 for TIME. e.g. FDATA 96 112 1173.05164.271 986.08 14.08 64 2 89 1 1 (E1.0,E3.0,E4.0,E7.0,E1.0,E6.0,E1.0,E6.0,2E2.0,3E1.0,2E7.0,5E3.0) TIME is 1173.0 in data item 4. Changing the structure of the data set (but not the values used by the model) e.g. by dropping one or more items so that there are fewer data items preserves an additional digit of precision for TIME: FDATA 96 112 1173.08 5 164.27 1 986.08 14.08 64 2 89 1 (2(9E8.0/),1E8.0) TIME is 1173.08 in data item 4. I would like to submit details of the data to Globomax so that they can investigate and fix the NM-TRAN bug. I am waiting for a reply from the owner of the data so I can send it to Globomax. Hint to Dr XX at the YYY company to please reply! Now for the second part which relies on these small differences being produced by the NM-TRAN bug. Here are results showing OBJ and SigDig with 3 DF compiler versions using the truncated and full precision values for TIME created by the NM-TRAN bug: Data Version OBJ SigDig Trunc 6.6 20067.293 2.7 Full 6.6 20089.472 2.5 Trunc 6.5 20100.58 . Full 6.5 20096.119 2.6 Trunc 6.0A 20091.387 3.4 Full 6.0A 20091.676 2.9 NONMEM was compiled for all 3 versions with these options /fltconsistency /optimize:4 /fast. DF 6.0A gets essentially the same OBJ. DF 6.5 has higher OBJ for both data sets but can tell them apart. DF 6.6 has 22 unit lower OBJ with the trunc vs full data sets and both OBJ values are lower than with the earlier versions. So I would conclude that DF 6.6 is more sensitive to small differences in the data and can get to a lower objective function. These differences suggest that this is because of better numerical precision using DF 6.6. Does anyone have a plausible alternative explanation? Nick -- Nick Holford, Divn Pharmacology & Clinical Pharmacology University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand email:n.holford@auckland.ac.nz tel:+64(9)373-7599x6730 fax:373-7556 http://www.health.auckland.ac.nz/pharmacology/staff/nholford/