Tuesday, September 4, 2012

Why Observational Epidmeiology is frustrating

Andrew Gelman has a post up on the history of cigarette smoking research, based on a book he was reading a while back.  It's pretty interesting but what really caught my eye was this comment:

Vague statistical inference can not possibly establish such a causal link. Even valid associative inference should establish a 50-100% correlation between smoking and cancer, but it does not even come close. Most people who smoke don’t get lung cancer, and at least 10% of Americans who do get lung cancer- do not smoke. There are also huge international/ethnic variations among smokers and cancer rates. There is currently no proof whatsoever for the alleged smoking-cancer causal link. None. Smoking is a disgusting and silly habit. But all that one can now objectively say is that it is a risk factor for cancer and increases the incidence of lung cancer.
 
And this was not the only person in the comments who was casting doubt on this association.  As an epidemiologist, I want to scream.  If people will not believe this evidence then they really will not believe any level of evidence for observational epidemiology.  We have cohort studies going back 50 or more years (Richard Doll has one). Even better, the members of this cohort did not initially know that smoking was harmful (and I recall that the original hypothesis was automobile fumes and not smoking, although my memory may be failing me here).  So we don't even suspect a healthy abstainer effect. 

The requirement for a 50 to 100 correlation seems to ask for smoking to be directly causal of lung cancer instead of increasing the underlying risk of lung cancer.  Consider skiing and broken legs.  Not all broken legs are due to skiing and many people ski and do not break a leg.  But there is no question that skiing is a risk factor for broken legs.  Another good example is collapsed disks in the back.  If you are working with a veterans population, the first question you ask when you see a compression fracture in the spinal cord is "were you a paratrooper?".  Not all paratroopers have compression fractures and not all compression fractures are due to jumping out of airplanes, but it is a pretty direct link to increased risk. 

There is a libertarian line of defense here: people ski because they value the enjoyment of skiing more than the risk of a broken limb.  I am not always delighted by it, but it is at least an arguable position.  But directly denying the link between smoking and lung cancer seems to be setting a very aggressive standard of proof. 

8 comments:

  1. I have seen some evidence in pop culture that indicates that smoking had a rep as being slightly unhealthy (particularly for athletes -- it hurt their 'wind') in the first half of the 20th Century so there might have been a very small selection bias, though certainly not enough to explain the observed effects.

    ReplyDelete
  2. All observational studies have small degrees of bias. But a rate of lung cancer 16 times the expected value among smokers would be one heck of a selection effect.

    ReplyDelete
  3. @Mark, this isn't to knock your point so much as to respond to the commentators who think that the evidence of smoking -> lung cancer is non-existent. Is it possible, in the logically possible sense, that this inference could be overturned by additional data? yes.

    But boy, who would want to bet on it??

    ReplyDelete
    Replies
    1. Lots of interesting points here. One is the difference between what common sense can spot and what statistical tools can identify. People were able to see that smokers tended to have less endurance. That was an effect that appeared almost immediately and went away when you quit. The causal relationship between smoking and lung cancer was even stronger, but was not as intuitively obvious.

      Delete
    2. Rare events with a long lag time are precisely what human beings are very poor at predicting. We overstate the risks of nuclear stuff as much as we understate the risks of smoking.

      Studies can help but it's slow progress to do it well.

      Delete
  4. Another point is the commenter's apparent assumption that causality means "really high correlation."

    ReplyDelete
    Replies
    1. That one confused me too. I suspect that they are using cause in the very strong sense where it is "I drop the fall and it falls" as opposed to the "crash was caused by slippery roads and high speeds".

      Delete
  5. Another factor here is in sorting out specific types. For instance, Bronchioalveolar Carcinoma of the lungs is not associated with smoking. It helps to know that so it is not in the smoking studies and also so that you start to work on what is associated with it.

    ReplyDelete