Resveratrol does not reduce CVD risk, prolong life – A statistical point of view

I read the paper “Resveratrol Levels and All-Cause Mortality in Older Community-Dwelling Adults” by Semba et al. (JAMA Intern. Med. 2014; 174(7):1077e1084) [1], where the authors want to determine whether resveratrol levels achievedwith diet are associatedwith inflammation, cancer, cardiovascular disease, andmortality in humans. In this paper, the authors concluded as “In older community-dwelling adults, total urinary resveratrol metabolite concentration was not associated with inflammatory markers, cardiovascular disease, or cancer or predictive of all-cause mortality.” I also read a review “Resveratrol Does Not Reduce CVD Risk, Prolong Life”. It says “In terms of patient recommendations, Semba offered a simple one, that patients can continue to enjoy red wine in moderation, but it isn't going to help them live longer or protect them against heart disease and cancer.” This research is very interesting. However, from a statistical point of view, I do not think Dr. Semba et al. can have that strong statement “In older community-dwelling adults, total urinary resveratrol metabolite concentration was not associated with inflammatory markers, cardiovascular disease, or cancer or predictive of all-cause mortality”. Based on the analysis in the paper, it is statistically wrong to have this strong statement or give the recommendation “patients can continue to enjoy red wine in moderation, but it isn't going to help them live longer or protect them against heart disease and cancer”. When P-value is small (usually < 0.05), we can say we have the evidence to reject the null hypothesis and have the conclusion of association. But when the P-value is large, we can only say we do not have the evidence to reject the null hypothesis, but we cannot say we have proved the null hypothesis (no association). For example, when a study does not have enough sample size, we can have large P-value, evenwrong trend when the association is real. Unfortunately, there is no sample size or power calculation in this paper. However, in the discussion section, the authors mentioned “a much larger sample size would be needed to detect the association”. Overall, I think the authors have made a fundamental statistical error in selling their findings.