I love it when my visitors challenge me! Most recently a physician researcher E-mailed me a question about composite endpoints. My exposure to this statistical method is somewhat limited, but I just needed an excuse to research the matter further. The results of my research can be found on the Statistical Functions Page, under the new section entitled "Composite Endpoints." For the several days I spent on this I have found that this is a recent phenomenon that was first mentioned in 1999. Of course, it could have been around longer, but the earliest Net mention was 1999.

Basically, to save on the cost of conducting clinical trials, researchers have resorted to the use of composite endpoints that are sort of summary indices for more than one endpoint, or a global assessment measure. It is unclear how these composite endpoints are determined, as I could not find any methodological treatise to review. Nevertheless, this is kind of like the newest reincarnation of the "p Value" - one number that can be used to measure the importance of a research finding. However, if you have been a consumer research maven, you will notice that p Values have been replaced with confidence intervals as more appropriate ways to gauge the importance of research findings. Feise's article provides a thoughtful look at the use of p Values with multiple outcome measures.

You will find composite endpoints reported in studies dealing with chronic diseases, as cardiovascular, renal, arthritis, etc. and disabling diseases, as stroke. That's because such diseases have multifactorial causes, multiple therapies and multiple outcomes. Current therapies are not necessarily "cures" in the sense that people get well. More than likely, they just get "better" and we all know that this is a qualitative opinion at best. Thus, to capture the outcomes of numerous therapies "all at once," researchers came up with the idea of looking at what would happen if a patient were to do this, this and that, compared to doing that, that and something else. This is the basic premise (but, you can correct me if I am wrong). Feise asserts using composite endpoints could replace the need for p-value adjustments. Then again, Shih does raise the important issue of missing data in clinical trials where participants drop out. He asserts keeping the dropouts in the denominator unless you can explain why the data are missing.

Nevertheless, there is much controversy around the use of composite endpoints, which I have noted with a . From the big picture standpoint, composite endpoints may have more to do with study design than the actual statistical manipulation involved. I foresee that its use would probably make meta analyses difficult to conduct, AND we do need meta analyses to make sense and consolidate what is actually relevant from the slew of research being conducted for any particular topic.

Excellent critiques have been published, most notably from the FDA, and I like the American College of Emergency Physicians' "GP IIb/IIIA Inhibitors: A Hill of Beans?". Users of composite endpoints will tell you that it will help physicians to explain the benefits of various treatments to patients with chronic diseases so they will become better informed to choose options wisely. From what I have reviewed, this will hardly be the case in most instances. People do not usually think statistically (if they did, there would be no casinos) when making decisions. Keep in mind that there was a time that hormone replacement therapy (HRT) was hailed as the cure-all for women of certain age, and they only had to choose between better heart outcomes or risk developing breast and/or endometrial cancer. Pragmatically speaking, these are hardly choices any woman would want to make. Most recent studies have pooh-poohed the use of HRT and there are now studies that have linked HRT to dementia . In retrospect, it would seem that those women who were reluctant to use HRT because it didn't seem natural are probably better off from thinking with their gut than to obsess over what significant findings were saying back then. Furthermore, scientific research always leaves room for reversals when most of what drive research to begin with are hypotheses (or guesses) that seek to explain what we see happening around us.