So I got ahold of the article. It is a little down the page here (Red Meat Consumption and Mortality): http://archinte.jamanetwork.com/issue.aspx?journalid=71&issueid=23009&direction=P

The reviewers and editor(s) are dipshits. Basically. The statistics they did are: 1) not explained well, and 2) very flawed.

On the top of the right column on page 2 of the article they discuss (I use that word lightly) how they separated out meat consumption from the other independent variables (BMI, smoking, etc… the dependent variable is mortality as determined by their ). They used a multivariate technique. They don’t describe the specifics of the technique they used (there is not just one way to do a MANOVA). The best I can tell they standardized the independent variables (other than meat consumption) to meat consumption. So, theoretically, they were calculating just the effect of meat consumption on mortality when they did their univariate analysis.

However, there is a BIG problem. They fail to account for the multicollinearity of the independent variables. If there is a high degree of multicollinearity even among two variables, then the whole thing falls apart unless you exclude those variables. Gee, could hypertension be highly correlated with high cholesterol? How about diabetes (yes/no) with activity level? How about activity with hypertension? I could go on, but you get the picture.

So, their conclusions are based off of INVALID statistics. If their research question was: “what factors/interactions of factors out of a,b,c,x,y,z are associated with mortality?” this study would be perfectly fine. But it was: “what is the quantitative relationship between red meat consumption and mortality?” It is impossible to get rid of the collinearity among the independent variables statistically. And even if you could deal, you still have other factors that relate to mortality that were not measured (mental health, for example). It was a very bad research question. This is what happens when you come up with a research question based on pre-existing data, instead of coming up the question first then designing the study.

They could have tried a factorial regression model that integrated all of the independent variables and the effect each one had individually, and all possible interactions of them interacting with each other, on mortality. But either they were too lazy, too dumb, or they knew if they did that, that one of the first IVs to drop out of the model (you eliminate variables/variable combinations from the model until only the ones that really effect the independent variable are left) would be red meat consumption. And there goes their headlines.

They did actually do a regression (bottom of page 2). But they did it among red meat and what they considered risk factors (cardiovascular mortality and cancer mortality) of death. In other words, meat consumption was the independent variable. This excludes every other factor in the universe that could have influenced cancer and cardiovascular mortality in the subjects.

At the end of the results they confess there were different numbers of people in each of the quintiles for total meat intake. With a non-even distribution of measurements along the x-axis (meat consumption) the data does not meet one of the requirements for doing a regression.

This was a very, very, very questionable study in terms of how the data was analyzed (I didn’t even get into the problems with how it was collected). I’m half tempted to write a letter to the editor. This is normally a good journal. Dumbass reviewers and editor(s). I wonder if they had a personal agenda they were trying to push and let bad science slide. Or maybe it wasn’t a blind review, and they let it go because the authors were from HHHaaaavvard.