Quote:
Originally Posted by K Walt
BoBoGuy: Here's the problem:
"after controlling for nutrient intake, consumption of medical goods and services, income distribution, weather, and literacy, "
What does that mean, exactly?
How do you 'control' for variables like that? Do you ignore them? You play around with numbers until things look right? It's voodoo and hand-waving and Excel spreadsheet manipulations.
|
When they talk about controlling for variables such as these, it means that they were entered as covariates in the statistical analysis to try to isolate their effect on the outcome. Usually these factors are called confounders, since they confound the relationship between the variable of interest and the outcome. This can be a useful statistical tool, but it is definitely not realistic in some cases: for instance, I once saw a dataset where they had controlled for the effect of a child's age on height. Now, I don't know about you, but most children do not grow taller while their age remains the same--physics probably wouldn't allow it. I wouldn't call this voodoo, but it can sometimes leave something to be desired. I would agree that we need better statistical methods that would allow us to get at the relationship of interest without having to worry so much about confounding.
One other problem is that so-called 'residual confounding' may remain. It's possible that your measurement of the confounding factor is not precise enough to allow its effect to be removed from the analysis. Therefore, a residual effect may remain that would still be confounding the relationship of interest.
There is no use of Excel in this world--Excel does not have sophisticated statistical capabilities. SAS is commonly used, as are Stata, R, S-Plus, and SPSS. As for playing around with the numbers, any researcher worth their salt will have laid out an analysis plan for their study
before they begin the analytical process. This prevents any urge to play around with the numbers until they give you the answer you were hoping for. I am afraid that a large number of the low-fat hypothesis researchers that Taubes describes were probably guilty of this after-the-fact manipulation but were never called out on it by anyone with the authority to enforce statistical rigor.
Quote:
Originally Posted by K Walt
You have noted at least SIX things that vary widely between those populations. There may actually be 28 more. Or 9 more differences. There is no way to know. You are comparing native populations in Tanganyika with populations in Finland or Burma -- and are assuming that EVERYTHING is exactly the same between them, they are identical in every way, except for how much chicken or chick peas they eat.
|
Yes, this is a big issue in statistics. This is know as the unmeasured confounders. This is something people worry about quite a bit, as there's no way to truly account for every possible confounding factor. That is part of the reason that ecological studies are so unreliable. Those are the studies where they compare the rates in different countries--like Ancel Keys's countries study. Since you are looking at population level factors, you
cannot conclude anything at the level of the individual--you can only make generalizations at the population level, which are not terribly useful in a field like this. Ecological studies are really only good for hypothesis generation.
Quote:
Originally Posted by K Walt
Besides, in epidemiology, no matter how many cabillions of people you include, no matter how many variables you 'control for (which essentially means IGNORE), you can never, ever, ever, draw any causal conclusion. You can never, ever, measure with any certainty how much poultry or wheaties or whatever politically correct thing you want to measure.
http://www.nytimes.com/2007/09/16/m...emiology-t.html
Niney-eight times out of a hundred, when you take a 'hypothesis' generated from Excel spreadsheet manipulation, and subject it to a real-world, real-person clinical trial the results just don't hold up. We've seen that for fiber and colon cancer. Fat and heart disease. Fat and breast cancer. Hormones and post-menopausal heart disease. The literature is FULL of 'associations' that looked great in statistics but didn't pan out in actual tests.
Epidemiology can come up with hypotheses, with guesses. But it doesn't prove or confirm diddly. Never has, never will. And even epidemiologists will tell you that.
|
As an epidemiologist, I can say with confidence that
observational studies can never prove causation. Randomized trials are a subset of epidemiologic studies, but the randomization, blinding, etc, if done properly, can begin to answer the causation question with more certainty. Now clinical trials can also be fraught with statistical issues if the studies were not designed or conducted properly, as we saw with the ACCORD study. People act as if all clinical trials are perfect, but this is far from the truth. There have been clinical trials whose conclusions were later proven wrong by cleaner trials that did not have as many issues with their methods.
However, observational studies are far less expensive and allow us to accumulate evidence for a particular theory that can then be tested rigorously in a randomized trial. Therefore, observational epidemiologic studies do have value. What often frustrates me is when the media, the public, heck even sometimes the researchers themselves, make conclusions about an observational study that are just not supported by the data. One statistical association does not a causation make.
Ok, getting off my soapbox now. Thanks for reading this far.
--Melissa