Скачать 1.21 Mb.
Health Study Examiner's Handbook. |
DR. GOUGH: May I ask, did we start with is this yours?
DR. HARRISON: No.
DR. GOUGH: Who wrote the one little page that said: Open Issues Not Addressed in this Document?
I think those are good questions. I haven't any idea about 2 through 6 because I'm not a statistician, but they seem like really good questions, Joe.
DR. MICHALEK: There is another question I forgot to
DR. HARRISON: Oh, no. Let me --
MR. COENE: Maybe we should clarify, what are you
DR. GOUGH: Well, on the back of the cover page, there were open issues, and the first number is, Increase the number of invited Comparisons.
DR. HARRISON: Oh, I see. That's this red .
DR. CAMACHO: Are we on the draft now?
DR. HARRISON: Yes, we're going I'm not sure what we're doing.
We're supposed to be discussing the statistical models right now, and the questions --
DR. STOTO: Those issues are statistically --
DR. HARRISON: Yes, the questions that are here are fair enough to be a part of that.
DR. GOUGH: I have a question.
DR. HARRISON: Go ahead.
DR. GOUGH: What has been the policy about replacing comparisons?
DR. HARRISON: If you read this document, it basically says all the comparisons that you can find are going to be invited.
DR. MICHALEK: That's true. The overriding rule is, anyone who has ever been invited to the study will be re-invited.
Secondly, any comparison who was invited last time and refuses to come this time is to be replaced. And there are rules for replacement. There are rules associated with replacement, too, completely elicited; and they are all spelled out in the Statement of Work.
DR. HARRISON: But in your Statement of Work, you make the statement that all Ranch Handers and all comparisons in the first six categories are invited. The seventh category, for the comparisons, are comparisons that you can't find.
DR. MICHALEK: What paragraph is that?
DR. HARRISON: Page I think it should be page 8.
I believe it's invitation to the exam.
DR. STOTO: 188.8.131.52.
MR. COENE: 3137.
DR. STOTO: It starts on page 6; 3131.
DR. MICHALEK: All right, would you like a summary? Would you like to talk or would you like me to talk?
Now you guys need to correct me if I make a mistake. Bill Grubbs, there you are.
Anyone who has ever been invited to the study in past years will be invited again. The only people we intentionally do not call are individuals who have distinguished themselves as being hostile: "Do not call me again or I will sue you." "You call me again, I'm going to go out to Scripps with a rifle." Those individuals are called hostile.
They're about, how many of them? Maybe a hundred, total.
DR. GOUGH: I hope they're evenly divided between Comparisons and Ranch Handers.
DR. MICHALEK: Well, I don't know; we haven't done a careful analysis of that.
But those are specially listed and are not to ever be called again. We know who they are and we do not call them. This rule was specified early on in the study; anyone who was ever invited will be invited again.
In addition to that there is a very detailed replacement strategy that was spelled out in the protocol, and has been implemented since the beginning. And that's what all this text is about in these paragraphs.
The fundamental idea of replacement is that an individual in the comparison group who fails to be compliant; says "I don't want to come, period" is questioned about his health. "What's your status of health? Excellent, good, fair or poor?" And he's asked that during the telephone interview, telephone contact that was there to invite him.
DR. HARRISON: Joel, all that I wanted to say was that in the first sentence here, it simply states that all subjects in the three Ranch Hand categories and the first six comparison categories listed above shall be invited to the examination even though this may mean at times that two or more comparisons assigned to a single Ranch Hand are invited.
DR. MICHALEK: Exactly right.
DR. HARRISON: That seemed to me that that said that everybody got invited.
DR. MICHALEK: Exactly right. And it has happened; there are more than one comparison matched to a Ranch Hander that has come to the physical.
DR. GOUGH: But there are comparisons that have not been called, aren't there?
DR. MICHALEK: There are comparisons that have not been called because we haven't had to call them --
DR. HARRISON: That's not what this says. That's not what that sentence says.
DR. MINER: Yes, it does.
DR. MICHALEK: Please clarify.
DR. MINER: Who were never contacted to be invited. Perhaps if that were added.
DR. HARRISON: Well, perhaps if that were added it might say that. I'm saying right now right now it just says the first six --
DR. MINER: On Category 7.
DR. HARRISON: The first six categories, on category 7, is replacement comparisons who were never
DR. MICHALEK: Yes. Who were never contacted means there are up to 10 comparisons matched to Ranch Hander based on date of birth, military rank and job in Vietnam, and race. They are matched perfectly almost on those factors. They are matched to within one year on date of birth, most of them. They are matched perfectly on all the other factors.
The individuals in that matched set consist of one Ranch Hander and up to ten controls. The ten controls are randomized, according to their position in the matched set. At the very beginning of the study, the first person in that randomized comparison array was invited. He was called a so-called "original comparison."
Then, as the study proceeds --
DR. HARRISON: I understand now, the only difference is never contacted.
DR. MICHALEK: Never contacted. Because --
DR. HARRISON: That answered my question.
DR. MICHALEK: what can happen is that the first comparison contacted, complies, "Sure, I'll come" and he continues to come, and so we never need to call any of the other comparisons in that matched set, and we never do. And so the contractor is not to call those individuals because they had never been contacted and didn't have to be contacted. But anyone who was called would be called again.
DR. GOUGH: But didn't you have a system that, if not enough of them are compliant, then you add that's what this question is about; increase the number of invited comparisons?
DR. MICHALEK: That was Bruce Burnham's suggestion.
Bruce, would you like to amplify that?
LTC BURNHAM: Well, it was just the idea of increasing the power of our study, because one of the limitations is the lack of power; and so I thought really, you could add up to five controls per study subject. Technically, not financially.
DR. STOTO: But you wouldn't add much power.
DR. MICHALEK: No, a small amount of power. Very small.
DR. GOUGH: Because you can't increase the number of Ranch Handers, right?
DR. MICHALEK: Cannot increase the number of Ranch Handers, no.
At the last physical, we had about 900 Ranch Handers and 1,300 controls came.
DR. HARRISON: So what is matched analysis?
DR. MICHALEK: That's another way to analyze data without resorting to multivariate modeling, which is one of the methods we've used, is multivariate logistic regression, for example. Well, in the area of diabetes, we've been using that model a lot.
The advantage of the model is that it's well known and its properties are well understood and we've been using the model for years and years, and it's in textbooks and it's a standard technique of analysis.
However, it's not easy to visualize matching, the logistic model. Another approach is a 1:1 or 1:many matched analysis where individuals are literally selected and matched on factors relevant to diabetes, such as family history and father, mother, brother or sister.
Body fat in Vietnam, date of birth, and change in body fat from Vietnam to the present.
Very careful 1:1 matching of individuals based on those criteria can yield a defensible and a very compelling matched analysis of the data; and this approach is advocated in some textbooks; for example Rothman's text on Epidemiology.
And it can be used as a supplement or as a complement to the standard multivariate analysis that we've seen in multivariate analysis that we've seen in all of our reports. And we've been doing that recently; we've been doing matched, 1:1, very tight matching to within 1 percent of body fat in Vietnam, matched perfectly on all the risk factors that we know about for diabetes.
And sometimes when we do that, the results become very clear, and they're very easy to defend and present to an audience, because they don't rely on any multivariate statistical model. What you end up with at the end is a 1:1 matched dataset where the computations become very trivial. In the area of a continuous analysis such as insulin, you're simply doing a paired T test, that's all there is to it.
DR. STOTO: So that the pairs include one from the Ranch Hand --
DR. MICHALEK: We've been doing one to one.
DR. STOTO: and one from the control.
DR. MICHALEK: Yes, one to one.
DR. STOTO: But they always include one from the Ranch Hand and --
DR. MICHALEK: Right.
DR. STOTO: at least one from the control.
DR. MICHALEK: Right. Always one Ranch Hander and one control, that's true. And they have computations on relative risk and confidence interval that says, you know, from the Rothman's text and others; and the Mantels papers become very clear and very compelling type of analysis.
That's why the matched analysis is a favorite in some studies.
DR. STOTO: This makes use of the design of the study --
DR. MICHALEK: Exactly.
DR. STOTO: which Mike should like.
DR. GOUGH: Oh, I like matching with comparisons.
DR. MICHALEK: And it eliminates the need--
DR. GOUGH: I like data, actually.
DR. MICHALEK: -- to present and summarize the sometimes very complicated multivariate model, which to a statistician may be clear, but to non-statisticians may not be clear. So that's why the matched analysis is such a favorite.
DR. HARRISON: Are you required to show that if you match comparisons that you would not see a difference?
DR. MICHALEK: Matched comparisons 1 to 1? is that what you're saying?
DR. HARRISON: Yes.
DR. MICHALEK: You could ask that well, what's the purpose of your question? To ask whether nonexposure is related to nonexposure?
See, the purpose of the hypothesis in the matched analysis was exposed versus non-exposed. Now you're matching unexposed versus unexposed.
DR. HARRISON: Yes.
DR. MICHALEK: Mike.
DR. STOTO: Are you saying this to see whether or not --
DR. HARRISON: I mean, it sounds like a stupid question, but I agree. But you know, one way if I matched like-to-like and I saw no difference, and I matched unlike-to-unlike and saw a difference then it would seem to me that it would be even more compelling than if I just assumed that like-to-like would not show a difference.
DR. STOTO: I think that the logic behind that is that you're saying, just to be sure that we didn't get a fluke by here, let's do a matching where we don't expect to find a result, and make sure we really don't.
DR. HARRISON: Yes.
DR. MICHALEK: The answer is yes, that could be done. We could randomize the comparison group into two sub-cohorts and match them 1:1 on the same factors. And we expect to see nothing.
DR. STOTO: But there are statistical tests that tell you that, answer that question in a more formal way, and that use all the data; and that's presumably what you would do in the matched analysis.
DR. HARRISON: What is the question about the matched analysis?
DR. MICHALEK: The issue is, is exposure related to diabetes after controlling for the risk factors that we already know about that are related to diabetes?
DR. HARRISON: Is the question, should a matched analysis be added? To the Statement of Work.
DR. MICHALEK: There's a question about whether we should add a matched analysis as a component of the Statement of Work.
DR. GOUGH: Would you do that; do you do the matched analysis or does SAIC do the matched analysis?
DR. MINER: Depends on whether it's in the Statement of Work or not.
DR. GOUGH: Oh, I see. All right.
DR. MICHALEK: First of all, I've given you all the shining attributes of the match analysis, so let me tell you all the downside of the matched analysis.
DR. GOUGH: Sort of data-intensive?
DR. MICHALEK: Yes. If you were to require the contractor, namely SAIC, to do the matched analysis in several clinical areas, what you would have would be a different sample size, for one thing. You'd have different cohorts being analyzed in different chapters.
For example, when you match on body fat, family history of diabetes, age and other factors, you're going to get a different matched cohort than you would if you were matching for factors related to heart disease or cancer. So you're going to have different matched sets in every chapter, which would maybe confuse people, but you're not talking about the same people when you write the diabetes chapter, as when you're writing the cardio chapter or the cancer chapter.
[Speakerphone connection interrupted.]
DR. MICHALEK: So that's one complication. Another is that you have to write special software to do the matching. We have Fortran code that we wrote that does a 'nearest neighbor' match on these factors. And we've written it over a period of several months and we use it a lot.
If we required Bill Grubbs and SAIC to do that, they would either have to write their own code, or we'd have to give to them and they would have to test it out. I'm not sure they have Fortran on their machine. They'd have to try to translate that into SAS.
DR. HARRISON: I'm not sure that anybody has Fortran on their machine.
DR. GOUGH: Except for the U.S. Air Force.
DR. MICHALEK: Right. We have got Fortran, we also wrote the program in SAS, too.
In other words, there's a task there of writing the software to create the matched datasets, and that's not trivial.
DR. HARRISON: Okay, hold on here. Isn't it the committee's the committee's scope of work is to say whether or not this is desirable. It's up to the Air Force to figure out whether or not they have the money to fit it in.
DR. STOTO: Well, I think it's a slightly different issue. I think first of all we're talking about Section 3.6.1 now, right? Statistical Models.
DR. MICHALEK: Yes, that's a good place to talk about.
DR. STOTO: Yes. And it's been the tradition over the last couple of things that they have these four models that they're running, and that they do separate and sometimes more sophisticated analyses in their own papers that are not in these blue books.
DR. MICHALEK: That's right.
DR. STOTO: And in fact, you are doing some of this on your own right now.
|Advisory committee on immunization practices||Medical Devices Advisory Committee|
|Veterinary medicine advisory committee||National Vaccine Advisory Committee (nvac)|
|External Advisory Committee on Cities and Communities||Wildlife Diversity Policy Advisory Committee|
|Peer reviewed by the Arizona Department of Commerce Economic Research Advisory Committee||Food and drug administration national institutes of health advisory Committee on: transmissible spongiform|
|Advisory Committee, Cuyahoga Valley School-to-Career Consortium, Broadview Heights, Ohio 1996-2002||Jane D. Siegel, md; Emily Rhinehart, rn mph cic; Marguerite Jackson, PhD; Linda Chiarello, rn ms; the Healthcare Infection Control Practices Advisory Committee|