Originally Posted by
Jonathon Sullivan
Yeah...I wasn't talking about accupuncture. I'm talking about meta-analysis. I think that pooling data sets obtained from studies designed to address hypothesis x and using the combined data to test hypothesis y is problematic. Extremely. This approach might have some minimal value from an observational perspective, and for hypothesis generation, but I would regard any conclusions proceeding from such an enterprise with a very jaundiced eye.
Here is a crude analogy I picked up somewhere (can't remember where). I think of any particular data set arising out of an experiment as a sort of n-dimensional mathematical object. The study was hopefully designed to generate this data set to be analyzed in a specific manner to address a specific question. The data set-object is designed to be "sliced" or analyzed in that pre-determined, correct and statistically appropriate manner, established during study design and before data collection. When we "slice" our data set-object along that predetermined "plane," the resulting "shape" of the data, if it has the power, either supports or fails to support the tested hypothesis.
All well and good. The problem is that, if we want to inappropriately exploit the data, be sloppy, or just plain lie, we can retrospectively slice the data set in any way we choose, along any number of other "planes," hacking at it with all kinds of post-hoc analyses, statistical ledgerdemain, manipulation of groups, "data-mining," you name it, and get a different family of distorted "shapes" for the data--some of which, not incidentally, might serve our purposes (publication of a paper or renewal of a grant, say) better than the orthogonal projection of the data anticipated by the original study design. Again, if conducted transparently and with acknowledgement of the limitations, this might have some minimal value for hypothesis generation, but I believe it will actually mislead us most of the time.
And doing that to a bunch of different data set-objects from a bunch of different studies and then mixing them together to address a question none of them were specifically designed to answer? I think that's a recipe for shit on a shingle.