There is an interesting article from Sunday’s N. Y. Times magazine (“…what makes us healthy?“) about the problems in epidemiology of using non-randomized data to draw conclusions. For instance, there is much uncertainty as to whether hormone-replacement-therapy increases, decreases or has no effect on the probability a woman will have heart disease.
The article talks about a few important biases which occur when researchers analyze observational data:
- Healthy user bias. People who faithfully take recommended pharmaceuticals, are more likely to take better care of themselves (i.e.: eat a healthier diet, exercise more, go to doctor for regular check-ups). Thus, researchers may erroneously conclude that a people who take a certain drug become healthier, when in fact it is really the case that healthy people are the ones who are using the drug.
- Compliance bias. People who comply with the doctors orders may also be different than those who do not comply. The author cite a Coronary Drug Project study which found that “…faithfully taking the placebo cuts the death rate by a factor of two.”
- Prescriber Effect. A physician may prescribe statins to a healthy patient in order to lower their cholesterol. On the other hand, “a physician is not going to take somebody either dying of metastatic cancer or in a persistent vegetative state or with end-stage neurologic disease and say, ‘Let’s get that cholesterol down, Mrs. Jones.’ The consequence of that, multiplied over tens of thousands of physicians, is that many people who end up on statins are a lot healthier than the people to whom these doctors do not give statins.”
There are also some interesting comments on Seth’s blog.