Dr. Len’s Cancer Blog, written by Dr. Len Lichtenfeld, Deputy Chief Medical Officer of the American Cancer Society, offers a terrific example of how to scrutinize confidence intervals in a study.
He commented on a study that got a lot of news coverage – suggesting that women with breast cancer who took tamoxifen had 440% greater chance of developing a more aggressive, hormone insensitive tumor compared to women who did not take the drug. He then jumped in with a lesson:
In statistics, there is something we call the “confidence interval.” In simple terms, it means if you kept repeating the same experiment in different populations or with more women in the same population, what is the chance you would come up with the same result? What are the possible other numbers that may show up?
In this study, the confidence intervals for that 440% number vary from 1.03 (a 3% increased risk) up to 19 (a 1900% risk). In our world, that is what we call “not very tight.” That 1.03 number–in reality–just gets you over the line of what we call statistical significance.
He had other problems with the way the news of the study was disseminated. Read his entire thoughtful blog posting.
(Thanks to Ivan Oransky for pointing out Dr. Len’s posting.)