The traditional question is "How do I know if something is caused?" The extra-short answer provided by Hume is very simple: You don't. Full stop. Do not pass go, do not collect $200.
Fortunately, we gave up on this simple formulation a long time ago (back in Part I). Rather, our question is...what is our most successful method for predicting the future.
Fortunately...that problem has a known easy solution.
- Develop a mechanistic approach for making predictions
- Measure whether your predictions are correct
- When your approach fails to predict correctly, change the approach, or create boundary conditions.
- Part 3 is almost universally failed...by everyone all the time. It's why CAS is the uber-think.
How do you find a mechanistic approach? Turns out that's also a relatively solved problem:
- For simple problems (Accelerations under reasonable speeds, Gas volume under reasonable pressures, etc.), look for an equation that describes the situation. Use the equation as predictor until you find circumstances under which it doesn't work....then don't use that equation in the new circumstances.
- For less simple problems (anything involving people), look for correlations. With enough correlations (women like Cocky, Funny guys ~80% of the time, girls find competent guys pratfalls attractive) and enough boundary constraints (Cocky Funny works poorly when a girl's in a PMS week) , go Bayesian..
- Aside: There appears to be a school of thought that suggests that less simple problems can be manages with models that ignore variables. I find that this so universally fails step 2 above (and that the response in 3 is to tinker, rather than scrap) that the approach itself is faulty. See Econ-modeling, Climate-modeling.
You might note...the notion of Causality is entirely missing from this model. Why? Because it's entirely unnecessary. Give me correlations: F*** causality. The purpose of causality is to allow human beings stories to tell to one another because our natural language isn't math. Causality catches our memory better...but that's a defect in human brains, not a feature of the universe.
Posts
6 comments:
Aretae,
Causality is only important if your objective is to change what you're observing. Correlation is perfectly adequate for simple predictive purposes, sometimes even when the objects under prediction are sentient. For instance, ice cream sales correlate pretty well with crime, but they're clearly not causal. We couldn't reasonably expect to reduce the crime rate by banning Breyers.
But Jehu,
1. You don't have anything more than I've got...you've just added a fancy word.
2. I don't think you've EVER got anything better than Correlation when the objects under prediction are sentient. (And under statistical quantities, correlation IS prediction).
3. I'm perfectly happy to talk correlation with time-sequence. If A (t1) then usually B (t2). Except...most of the time I'm wrong...and my attachment to stories screws with my ability to see how wrong I am and how often.
Causality, correlation -- don't they elide a bit? Some say there is no correlation between money and happiness. Others say there is a big correlation, and they go on to advocate massive transfer payments in a scheme to cause more happiness to be created.
RSF,
The problem is harder than I'm making it sound for sure.
Issue: We can measure correlation. We can't measure causation... because causation is a story we tell ourselves, not a measureable event.
Even: We can measure time-sequence correlations... "If person A gets richer, (s)he gets happier, p>.95."
And on Happiness...the research is rather unambiguous: Richer = Happier.
The economics is also rather unambiguous: More money lets you buy more of what you actually want...even if what you want is more time with family, or more time to pray.
Wealth equaling happiness is not unambiguous. But I used that as an example of things not always being so clear, and correlation can be as much a story we tell ourselves as is causation.
Post a Comment