|
|
|
|
 |
Daily Speculations The Web Site of Victor Niederhoffer & Laurel Kenner Dedicated to the scientific method, free markets, deflating ballyhoo, creating value, and laughter; a forum for us to use our meager abilities to make the world of specinvestments a better place. |
Write to us at:
(address is not clickable)
2/9/2005
Correlation is Not Causality, by Tom Ryan
As someone who has a science background, and who on occasion is called
upon to provide expert witness testimony, and one who participates
periodically on a MSHA board investigating fatal accidents in mining,
the first rule that one follows in the case of causal inference, which
dates back to John Stuart Mill, is that correlation is not causality.
This is particularly relevant when one is trying to go from a part to
the whole, that is, when ones sample used for developing a prediction is
small compared to the total predictive effect, say a pie baked with macintosh apples (my personal fav)to a pie baked with one macintosh and
eleven other different apples. In particular you have many problems with
using the results of one time period, and applying it to the cumulative
result of the next 11 time periods:
- Temporality can easily be confused with causality
- You may be ignoring multiple causes/effects
- In a time series there may be an underlying cause affecting both your
sample and your prediction (post hoc fallacy)
- One is ignoring the probability of random intervening factors (11
times more probable for the prediction period because of its length
compared to the length of the sample in this case)
- what about counter examples?
In engineering for example we deal with this a lot in terms of scale
effects and extrapolations, that is what works at one scale does not
necessarily work at another scale. This is discussed at length in Petroski "Case Histories of Error and Judgment in Engineering", with
the classic case of the dangers of extrapolation being the design and
collapse of the Dee Bridge in England.