From nwang@genie.tamu.edu Wed Aug 14 16:59:21 1996 Received: from genie.tamu.edu (GENIE.TAMU.EDU [165.91.112.236]) by helix.nih.gov (8.6.13/8.6.12) with ESMTP id QAA12115 for ; Wed, 14 Aug 1996 16:59:18 -0400 Received: (from nwang@localhost) by genie.tamu.edu (8.7.5/8.7.3) id PAA03376; Wed, 14 Aug 1996 15:59:06 -0500 (CDT) From: Naisyin Wang Message-Id: <199608142059.PAA03376@genie.tamu.edu> Subject: Re: framingham analysis To: xlin@sph.umich.edu (Xihong Lin) Date: Wed, 14 Aug 1996 15:59:06 -0500 (CDT) Cc: bobby@genie.tamu.edu (Roberto Gutierrez), rayc@helix.nih.gov, nwang@genie.tamu.edu (Naisyin Wang) In-Reply-To: from "Xihong Lin" at Aug 14, 96 04:11:50 pm X-Mailer: ELM [version 2.4 PL24] MIME-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Status: RO > > I have also thought about using MLE+SIMEX. However, this > would require a more programming burden for Bobby. Also, we may > need to rerun all simulations, since it is not consistent if > we use MLE in the framingham analysis and use CPQL in the simulation. > This is clearly a lot of work for Bobby. I am not sure whether we > want to do this. However, I can certainly help Bobby with programming > using MLE if everyone feels that this is the only way out. If we want to use MLE+SIMEX, then we have to be consistent throughout the paper. That is, we need to use MLE in the simulation as well. > > > > Xihong, is the large SE for the estimates of \theta due to the small > > number of exams (since the success probability has been increased)? > > > The larger SE is due to the smaller number of subjects in > the analysis, since previously we used 932, now I used 580 subjects > whose age<55. > > The reason I asked is that there seems to be something else besides the smaller sample sizes. It's a wild guess due to the "unproportional" increasing SE. > > > > (1) In the framingham data set, could we find another variable > > (binary) which indicates heart disease, for example, the CHD. > > Let a new variable HIST = 1, if CHD <=2 (previous history); > > =0, otherwise. This may help to reduce the variation among > > individuals. > > > > The variable CHD is whether a person had heart disease. > ECG basically is an surgorate marker for CHD. I hence do not feel > comfortable to use CHD as a covariate. If a person has CHD, it > will not go away. > > If the variable CHD is an indicator for family history > of CHD, it is appropriate to be used as a covariate. Do we want to give it a try? CHD indicates that 1st occurrence of heart disease, which is an indicator of disease history, since we are using exam 2 to 6. CHD = 1, or 2 (or just 1) would indicate an existing condition. I got an impression that ECG is more a short term symptom; there are patients who have ECG = 1 in the first several exams but change to have ECG = 0 in the last couple exams and vise versa. It seems to me that putting a history indicator would be able to reduce the variation across individuals since it divides the patients into 2 groups and allow a covariate to explain the difference between the two groups. However, since there are patient which have ECG change from 1 to zero, or zero to 1, the SBP could be used to explain the short term variation of ECG (the short term dependence of ECG on SBP). This set up would change the scientific question to "is there a short term dependence between ECG and SBP"? We could even narrow down the population to those who have at least one ECG > 0 and then see how large the \theta would be. This may reduce the # of subjects to 90 or so. What do you think? Would you mind narrowing down the population using the current data set and give it a try? Naisyin