4 ms·

The whole notion of 'controlled variables' is a ridiculous exercise. At some point we have to cop to the fact that biological interaction is too complex for our

The whole notion of 'controlled variables' is a ridiculous exercise. At some point we have to cop to the fact that biological interaction is too complex for our shitty parameterization. We're carving mountains with a tack hammer. Maybe there is a better math out there that can manage, but I haven't seen it yet.

jacques_chester 10y ago
Two things. First, "it's hard and we can't control everything" isn't the same as "it's impossible to control anything". Second, large sample sizes are preferred exactly to help even out the variables that are uncontrolled or, especially, unknown.
- astazangasta 10y ago
  I'm not saying "it's impossible to control anything", I am complaining about the idea that the world can be represented as a set of variables, that everything can be quantified and broken down into axes, and this parameterization is a useful or helpful way to model the vast interacting complexity of a biological system. If you have ever tried to do PCA on a gene expression dataset, you know my pain. The exercise is futile; you are attempting to produce a description of the world based on a presumption that is wholly inadequate.
  - jacques_chester 10y ago
    My point was that you're introducing the nirvana fallacy: rejecting all attempts because they cannot satisfy a perfect outcome. That it's impossible to perfectly model reality is not news in any scientific field. In any case, I was talking about grossly observable variables such as "does this person exercise at all?", or "do they smoke?" or "do they drink?". You'd be surprised how many studies don't bother to check simple factors like these.
throw_away_777 10y ago
Great points. Often times studies would be much better served collecting more data and looking at interesting correlations than trying and failing to "control" for all the hidden variables and winding up with biased sample sizes of 20 or less.
- stdbrouw 10y ago
  A confounder remains a confounder even when you've got a bazillion datapoints. Bigger sample sizes increase precision, they don't decrease bias.
  - throw_away_777 10y ago
    A confounder can remain a confounder even when you think you have controlled for all variables. Clearly social scientists are not properly analyzing their data, why should I trust them to be properly controlling for counfounding variables? I'd rather just see studies with enough data to measure the strength of a correlation than studies which are completely unreliable because of bad statistics or not enough data, as explained in the article. For example, you could do a large study where you measured the correlation of wine drinking with obesity. This study would at least tell me interesting information, even though I wouldn't know exactly if there is a causative relationship. But if the correlation was high and I wanted to lose weight I would stop drinking wine and see what happens. Maybe there is a confounding variable, maybe there isn't, I'd take that chance if I wanted to lose weight. This is much better than the current state of food and exercise studies where none of the information is reliable because of cherry picking stats and low sample sizes. As detailed in the article, we don't know anything about food and exercise from studies besides the basic physics that to lose weight you need to use more calories than you take in.
    - jschwartzi 10y ago
      The problem is that a study of the correlation of wine drinking with obesity would find an inverse correlation, as people who drink wine are more likely to be affluent and thus less likely to overeat. I think that when we get down to it the obesity epidemic in the US is a consequence of people not having enough time or money to buy unprocessed foods and cook them, and not a consequence of their lack of control.

jacques_chester 10y ago

Two things. First, "it's hard and we can't control everything" isn't the same as "it's impossible to control anything". Second, large sample sizes are preferred exactly to help even out the variables that are uncontrolled or, especially, unknown.

astazangasta 10y ago
I'm not saying "it's impossible to control anything", I am complaining about the idea that the world can be represented as a set of variables, that everything can be quantified and broken down into axes, and this parameterization is a useful or helpful way to model the vast interacting complexity of a biological system. If you have ever tried to do PCA on a gene expression dataset, you know my pain. The exercise is futile; you are attempting to produce a description of the world based on a presumption that is wholly inadequate.
- jacques_chester 10y ago
  My point was that you're introducing the nirvana fallacy: rejecting all attempts because they cannot satisfy a perfect outcome. That it's impossible to perfectly model reality is not news in any scientific field. In any case, I was talking about grossly observable variables such as "does this person exercise at all?", or "do they smoke?" or "do they drink?". You'd be surprised how many studies don't bother to check simple factors like these.

throw_away_777 10y ago

Great points. Often times studies would be much better served collecting more data and looking at interesting correlations than trying and failing to "control" for all the hidden variables and winding up with biased sample sizes of 20 or less.

stdbrouw 10y ago
A confounder remains a confounder even when you've got a bazillion datapoints. Bigger sample sizes increase precision, they don't decrease bias.
- throw_away_777 10y ago
  A confounder can remain a confounder even when you think you have controlled for all variables. Clearly social scientists are not properly analyzing their data, why should I trust them to be properly controlling for counfounding variables? I'd rather just see studies with enough data to measure the strength of a correlation than studies which are completely unreliable because of bad statistics or not enough data, as explained in the article. For example, you could do a large study where you measured the correlation of wine drinking with obesity. This study would at least tell me interesting information, even though I wouldn't know exactly if there is a causative relationship. But if the correlation was high and I wanted to lose weight I would stop drinking wine and see what happens. Maybe there is a confounding variable, maybe there isn't, I'd take that chance if I wanted to lose weight. This is much better than the current state of food and exercise studies where none of the information is reliable because of cherry picking stats and low sample sizes. As detailed in the article, we don't know anything about food and exercise from studies besides the basic physics that to lose weight you need to use more calories than you take in.
  - jschwartzi 10y ago
    The problem is that a study of the correlation of wine drinking with obesity would find an inverse correlation, as people who drink wine are more likely to be affluent and thus less likely to overeat. I think that when we get down to it the obesity epidemic in the US is a consequence of people not having enough time or money to buy unprocessed foods and cook them, and not a consequence of their lack of control.