To add to this -- whenever one does analysis based on very broad correlations, and where there is much potential for confounding variables that can never be per

To add to this -- whenever one does analysis based on very broad correlations, and where there is much potential for confounding variables that can never be perfectly controlled for, it is very easy to cut the data in a way that results in a 5-20% effect either way. That is why there are studies that show pretty much everything from broccoli to butter both causes and prevents cancer -- http://www.vox.com/2015/3/23/8264355/research-study-hype http://www.vox.com/2015/3/23/8264355/research-study-hype. So when looking at correlations, you need to ask, is this a "broccoli causes cancer result" (a small effect, easily caused by confounding variables and/or p-value hacking) or is this a "cigarettes cause lung cancer result" (a giant effect that cannot be explained any other way). The study in this article is a "broccoli prevents cancer" type of result. It is possible there is a small effect from the intervention, it is also possible there is no effect. The only information we can actually take away from the study is that there is not an obvious large effect.