Cause and effect

For a good look at how hard it is to make sense when discussing cause and effect in complex systems with ill-understood dynamics, try this article and the associated discussion in the comments.  I'm not sure most of us ever get beyond the magic stage.

22 comments:

Grim said...

I liked this comment he made, when someone asked if he could name 10 things that he thinks should increase crime:

"Off the top of my head – rising inequality, decreasing % population rural, dysgenics, decline of manufacturing/rise of needing college degree, Bowling-Alone-style social atomization, kids being raised by daycare rather than their parents, rising divorce rate, weaker gun control, welfare reform creates needier poor people, epidemic of crystal meth and bath salts, rising incarceration rates create more hardened criminals."

But they don't, somehow. So what's really going on?

I think "weaker gun control" accounts for a measurable part of the decline, actually. Maybe we should ask criminals. "Hey, why don't kids rob liquor stores as much anymore?"

E Hines said...

Your responder is missing at least part of the point regarding modeling. It’s more likely to occur when you don’t have many real world data sets to assess. In the case of "crime rates in the U.S." you have, depending on how you want to look at it, only one data set.

We do this all the time in an organization that models stock behavior--to very good effect--and we have exactly one data set (or two): the historical behavior of a stock and the historical behavior of a selection of stocks. In such an event, you segregate the data set and use one part to construct your model and the other set to test its efficacy. You never use the whole data set to build your model and then go for it, hoping for the best, in the unrolling world. He's right about over fitting, though (it's called, in statistics, over specifying, and in our stock behavior modeling, curve fitting. I know you live and breathe for the trivia of jargon).

But it's also easy to see if you've curve fit, and there even are statistical tests that assess the degree of that. There also are statistical techniques for adding factors to the correlation until you get a "good" set of factors and for starting with the kitchen sink and taking out factors until you're left with a "good" set of factors.

Much of the problem with trying to find whys and what causeds in dynamic systems of any degree of complexity is impatience. Especially if you're dealing with time series data like crime (or stocks) is the problem that time is required to assess the real world fit of your model. That assessment can't be done right d*n now, like a simple rat-in-a-maze experiment can be. And over that time, the relevant factors may well evolve, too.

Eric Hines

Texan99 said...

Ten proposed causative factors, none of which is demonstrably the cause. So we should conclude that each of them contribute 10%? Then all we have to do is explain how 10 separate factors accidentally lined up to produce the same effect. Or they could be correlated, but we have no idea why.

This kind of thinking isn't cause-and-effect, it's curve-fitting.

E Hines said...

This kind of thinking isn't cause-and-effect, it's curve-fitting.

Yup. It's a textbook example of what an old stat prof advised: if your theory posits a linear relationship, collect two data points. If your theory needs a curve, collect three.

Eric Hines

Grim said...

I'm not sure we understand causation very well above the chemical level. We can explain with great fluency why salt dissolves in water, and why it stops dissolving in water past a certain point. What causes both of those things is the structures of the chemicals and the ways in which they 'fit together.'

I suspect people believe that something like that kind of explanation is possible at the level of politics or ethics, if only we had enough knowledge and processing power. It may really just not be possible. 'What caused the decline in crime rates?' suggests that you could repeat the effect if you replicated the cause in another society. But everyone in that society is a different human being from everyone in the first society. There may not be an explanation of behavior that fits the expectation of clarity we draw from chemistry and physics.

Texan99 said...

I saw a glossary yesterday somewhere: "In my experience" decodes to "once"; "In case after case" decodes to "twice." I forget what "three times" was, but something like "it's universally accepted."

Grim said...

It may be, to rephrase it in stronger terms, that there is no cause because the effect is illusory. There isn't really a 'declining crime rate.' There's just a bunch of discrete cases of individuals who didn't elect to commit crimes. Bunching them up this way or that way makes it look like they're exhibiting group behavior, but it could be that the individual choices are not explicable in terms of any factors that affect 'society as a whole.'

Texan99 said...

Do you think so? I'd have said that in our ordinary lives we have almost no trouble at all with a rough but useful rule of cause and effect. It's only when we're trying to make a contentious point with murky data that we allow ourselves to drift into the kind of absurdity outlined in the linked article.

The main problem usually is an unwillingness to admit that we don't have any firm understanding of what's going on, and therefore have nothing sensible to say about cause and effect. In a more straightforward context, we're quite capable of saying "the practice of walking out in front of a speeding car causes me to get killed," in the philosophically unsophisticated sense of choosing between a couple of alternatives on the basis of the likely result. The crime-reduction study was about practical policy alternatives, after all, not about the deep meaning of time or action at a distance.

Texan99 said...

"There isn't really a 'declining crime rate.' There's just a bunch of discrete cases of individuals who didn't elect to commit crimes."

I can't make a helpful distinction between those two statements. It's possible to conclude that an apparent decline in the crime rate is an inconsequential squiggle in a noisy signal, of course, but when a trend goes on long enough that's hard to do.

Grim said...

Well, the speeding car thing is physics: I think we do understand physics and chemistry in a way that allows us to give causal answers pretty well (and ever-better). I can tell you not just that but exactly how the speeding car can cause the death of an organism.

On the other hand, even in intimate relationships, human beings have trouble with cause and effect. Well, maybe especially with intimate relationships! But how many marriages dissolve over lack of communication, where spouses attribute 'causes' to their spouse that don't exist in any sort of reality?

I mean, I think we can say in a rough and ready way that ISIS/Daesch is murdering people because they think it will be effective in expanding their power and letting them impose their will on others. But how to stop them? Well, there's an explanation from physics that's pretty reliable, just like the speeding car thing. If we try to move them via appeals to morality, ethics, politics, or even deterrence, we may well find that we don't understand how their will gets formed well enough to manipulate it.

Stick to physics, in that case.

Assistant Village Idiot said...

That bunch of "discrete cases who didn't elect.." can be seen as random when there isn't much movement on the dial. But when there is a pattern, we search for explanations. Sometimes it helps to look at incentives and disincentives at the margin, as economists do. There is likely a core group of predators who are unlikely to be affected by much of anything police or legislatures do one way of another. Yet there may be a large group influenced by point-of-inflection questions of risk and reward.

E Hines said...

I can tell you not just that but exactly how the speeding car can cause the death of an organism.

There's a big difference between "causing the death of an organism" and "causing the death of the particular organism that Texan posited." You can't get to the particular from a general collection of "discrete cases" no matter how well understood is the overall pattern of, say, a Brownian motion.

It's hard, too, to posit causality in the face of the crap shoot (or coin toss) that is quantum physics while living in a universe that is that crap shoot scaled up. Too, the predictability of high correlation (even perfect correlation) does not necessarily flow from causal relations.

I think we can say in a rough and ready way that ISIS/Daesch is murdering people because they think it will be effective in expanding their power and letting them impose their will on others. But how to stop them?

We also can say in a rough and ready way that Daesh is murdering people because they exist. The question remains, and the answer to the question remains unchanged. Causality requires a degree of rationality (even though we may not understand it yet); correlation offers a degree of predictability but requires no rationality at all. When we're dealing with humans, we're often dealing in the irrational, which eliminates (the need for) causation but leaves correlation intact. Thus: roughly like matter/anti-matter, Daesh and non-Daesh don't mix and often (enough for predictability) result in the death of the non-Daesh. That's correlation enough to answer the question: kill Daesh on encounter. I also don't need much of an explanation of that pattern to reach that executable answer.

With less catastrophic mixes, an explanation of a pattern could be very useful; manipulation and incentives may be more optimal.

Eric Hines

Texan99 said...

If we decided there was no way to tell what public policies did or didn't improve the crime rate, we could save a lot of time and effort and start throwing darts at a dartboard instead. But I think that some complicated systems can yield useful rules of thumbs even though they can't be reduced to simple formulas. The trick is to know when you've got a data set that's good enough to draw some conclusions from, and when you don't. Bad data's not too dangerous as long as you don't fool yourself into thinking it's good.

Grim said...

So, the point about 'individuals who didn't elect' is just this:

Today, Bob might have robbed a liquor store, but he broke his ankle. Jane was going to shoplift, but she got religion. Joe was going to steal a car, but he stepped in front of a moving one (per Mr. Hines).

Alice was going to commit bank fraud, but she decided it was too confusing to bother with. Mike wanted to knock over a guy at the ATM, but they just installed CCTV. Melissa was going to drive so recklessly as to risk vehicular homicide, but her tires are worn out and she hasn't got the money to replace them, so she was more careful.

Ultimately, the whole society is made up of cases like this. Why didn't they commit crimes? None of them had the same reason. So what caused the decline in the crime rate in this population? Nothing.

Texan99 said...

Every molecule in a bar of iron might individually, spontaneously, and randomly line up its N-S pole with all of the others. In practice, though, we find this to be so unusual that we assume a strong field has been brought to bear on the whole group, and this assumption rarely steers us wrong, even though we find it difficult to avoid discussing it without recourse to some kind of terms involving cause and effect. It may be intellectually unsound, but it's awfully practical.

E Hines said...

[T]he whole society is made up of cases like this. Why didn't they commit crimes? None of them had the same reason.

Not every data set has a pattern that can be teased out. Some data sets do, though. It's how we know the wind blows in a more or less coherent pattern, but we can't predict the motion of the individual atoms and molecules that make up the air in the wind over even a very short distance. It's how we can predict a culture's overall behavior, even though we can't tell what any of those individual humans of "your cases like this" might do or get done to them.

We just need to be able to tell the difference between those two sets. And there might not even be a pattern to give us hints in that discernment--we just can play the hot hand until it's not hot (at some unpredictable and/or unrecognizable point) anymore.

Eric Hines

Grim said...

The analogy to magnets is another analog from physics, though. You're making my point, in a way: this is the kind of explanation we want, because we find our physical and chemical explanations very gratifying.

Should we expect one here? These are free actors making choices for reasons of their own. I gave an example of five or six people with different reasons, but we know that all ten or thirty million people will be more like our six people than like any atoms of iron. They're all doing things for reasons that may not be very similar.

A libertarian should love this idea: it's like Nozick or Margret Thatcher's insight that 'there is no such thing as society.' In just the same way, there may be no such thing as a 'crime rate' -- we can extract one from the data, but it's a false idea. It's an illusion that there's a thing called 'society' from which to extract a crime rate.

Texan99 said...

I suppose it's an illusion that there are "men" and "women"?

MikeD said...

I'd say that the causes in the cases you listed are indeed "random". But how many potential criminals are deterred when concealed carry rights are granted/expanded? Sure, it's an individual choice on the part of each of the affected potential criminals, "do I mug this guy, or do I think he might have a gun?" But ultimately, if 10% of potential muggers decide that it's not worth the risk as a means to make a fast buck, then perhaps we can conclude that at least some of the decline can be traced to the fact that some potential criminals fear being shot by a citizen practicing concealed carry? And thus, we can conclude that expanding/creating concealed carry laws "lower the crime rate"?

Grim said...

I suppose it's an illusion that there are "men" and "women"?

We can talk about that too, but it's not quite the same question. When Thatcher says there is no such thing as society, she means all that there is are individual men and women -- what we're calling society is just a relationship between them.

To say there aren't men and women, then, is to make a different sort of claim. You're not making an ontological claim: we agree that the individual men and women exist, whereas we might disagree about whether a "society" exists in the same independent way. What you're disputing is the definition of 'men' or 'women.'

It's possible to do that: strict Nominalism denies all universals. And of course there are ways in which every man or woman is unique. But it's not the same kind of claim as 'there is no such thing as society.'

Texan99 said...

Maybe there are several billions individuals who happen to act like men, and several billion who happen to act like women, and it has nothing to do with their having been caused to do so by anything innate. :-)

Grim said...

Feminism as a kind of highly implausible Nominalism? I'd rather that than the kind of Marxism it usually turns out to be. At least Nominalism is honorably Medieval. :)