Thinking About Causality
July 31, 2016 3:49 AM Subscribe
Causation can be separated from antecedent conditions only if we can control them. This is a thought provoking essay by Joe Boswell on the notion of searching for the cause of cancer or the reason a marriage failed. It reaches back to Bertrand Russell's idea that science is the search for antecedents not causal relationships.
His examples are awfully slanted. In order to offer counter examples you have to offer arguments to say that women are to blame if they are raped, and that Islam is inherently associated with terrorism. I don't think you have to espouse those views to have reservations about Joe Boswell's analysis of causation.
posted by Segundus at 7:03 AM on July 31, 2016
posted by Segundus at 7:03 AM on July 31, 2016
I don't think he's even remotely offering you the choice of "women are to blame" or "Islam is inherently associated with terrorism" to refute his argument. The argument itself is "you can influence <some problem> by pushing or pulling on a 'handle' that (you claim) has a causal relationship to the problem." And if you want others to join you in pulling on that handle (rather than on a different handle someone else says everyone should pull), then you have to demonstrate why your handle is larger, fairer, or offers better leverage than the other options.
For example, rather than arguing that women are to blame for rape, or that modest clothes would solve rape, or education of boys, one could advocate for eliminating rape by killing all men. (Perhaps some already do!) And while that would certainly stop male-on-female rape -- just as not throwing a ball stops the ball from breaking a window -- and stop it far more reliably than the "clothes" or "education" approaches, it's challenging to make a case that it's fairer or better. So that handle has great leverage, but poor performance on other measures.
He's proposing a way of thinking about policies that lets you compare them more accurately, and forces proponents (of all policies) to do the harder work of saying why we should do <some thing>, what kind of results we can expect, and how that compares with the alternatives.
posted by spacewrench at 7:45 AM on July 31, 2016 [1 favorite]
For example, rather than arguing that women are to blame for rape, or that modest clothes would solve rape, or education of boys, one could advocate for eliminating rape by killing all men. (Perhaps some already do!) And while that would certainly stop male-on-female rape -- just as not throwing a ball stops the ball from breaking a window -- and stop it far more reliably than the "clothes" or "education" approaches, it's challenging to make a case that it's fairer or better. So that handle has great leverage, but poor performance on other measures.
He's proposing a way of thinking about policies that lets you compare them more accurately, and forces proponents (of all policies) to do the harder work of saying why we should do <some thing>, what kind of results we can expect, and how that compares with the alternatives.
posted by spacewrench at 7:45 AM on July 31, 2016 [1 favorite]
Segundus, that's one of the points of strong examples. Expressing doubt about a thesis is (way too) easy. Constructing an argument that defeats examples such as those is much harder (because constructing good arguments is inherently difficult). So, you should be more cautious about thinking your mere doubt is a good reason to discount a particular view.
As spacewrench and Boswell point out this is a pragmatic heuristic for trying to suss out what we can and should do about a (usually very complex) problem. It is not an analysis of causation.
posted by oddman at 7:57 AM on July 31, 2016 [2 favorites]
As spacewrench and Boswell point out this is a pragmatic heuristic for trying to suss out what we can and should do about a (usually very complex) problem. It is not an analysis of causation.
posted by oddman at 7:57 AM on July 31, 2016 [2 favorites]
Hmm, the two arguments he made in support of his idea don't seem to be as clearly separate as he's suggesting. I mean just as one can rule out removing religion from humanity and ending religious terrorism, people have claimed rape has evolutionary biological advantage, so with that perspective (which isn't at all my own point of view by the way), one could argue the male urge to rape is innate so seeking to change male behavior is as much an impossibility as removing religion by his example. They then could claim clothing is an area where society could gain some traction, and thus keep the blame on women.
He also speaks of using fairness as an argument, but fairness isn't a neutral position. Fair in what way and to whom has the requirement of asserting sets of beliefs that may or may not be equally held by all participants in the argument. Israel and Palestine, for example, have a little bit of space between their ideas over what a fair assessment of their situation might look like and what causal handles should be deemed meaningful.
I very well could be missing something, but I'm just not seeing much practical advantage to adopting Boswell's argument since it doesn't really seem to change anything much at all.
posted by gusottertrout at 8:29 AM on July 31, 2016
He also speaks of using fairness as an argument, but fairness isn't a neutral position. Fair in what way and to whom has the requirement of asserting sets of beliefs that may or may not be equally held by all participants in the argument. Israel and Palestine, for example, have a little bit of space between their ideas over what a fair assessment of their situation might look like and what causal handles should be deemed meaningful.
I very well could be missing something, but I'm just not seeing much practical advantage to adopting Boswell's argument since it doesn't really seem to change anything much at all.
posted by gusottertrout at 8:29 AM on July 31, 2016
I very well could be missing something, but I'm just not seeing much practical advantage to adopting Boswell's argument since it doesn't really seem to change anything much at all.
The author seems to be more or less advocating the "manipulability" theory of causation, which is also my pet theory of causation. The best advocate for it by far is (the philosopher) Nancy Cartwright, who argues that there aren't exactly laws of nature so much as "nomological machines", or systems in which we can say "when I do x, y happens". (What follows isn't really a good reading of her work but my own foggy recollection of it ). In this model it's considered kind of useless to speak of non-actionable regularities, things we can't manipulate, as causes. In a card game analogy, they're more like the cards we're dealt, or the rules of the game, than the reasons we win or lose. What matters is what we can manipulate to produce certain outcomes (the notion is very indebted to pragmatism). If I push on the gas pedal, the car accelerates, if I push on the brake, it stops. You can't "push" (or anyway it's non-trivial to do so) on the genetic inheritance that leads to male sexual aggression. But you can push on things like education and enculturation, and on criminal punishment and other incentives. In this view, genetics isn't a cause, so much as just the facts on the ground. The causes are the things we can change. In this view, one can argue: We can change the way we raise men and the attitudes they have towards women, so the cause of rape-culture is our failure to do so.
posted by dis_integration at 9:18 AM on July 31, 2016 [7 favorites]
The author seems to be more or less advocating the "manipulability" theory of causation, which is also my pet theory of causation. The best advocate for it by far is (the philosopher) Nancy Cartwright, who argues that there aren't exactly laws of nature so much as "nomological machines", or systems in which we can say "when I do x, y happens". (What follows isn't really a good reading of her work but my own foggy recollection of it ). In this model it's considered kind of useless to speak of non-actionable regularities, things we can't manipulate, as causes. In a card game analogy, they're more like the cards we're dealt, or the rules of the game, than the reasons we win or lose. What matters is what we can manipulate to produce certain outcomes (the notion is very indebted to pragmatism). If I push on the gas pedal, the car accelerates, if I push on the brake, it stops. You can't "push" (or anyway it's non-trivial to do so) on the genetic inheritance that leads to male sexual aggression. But you can push on things like education and enculturation, and on criminal punishment and other incentives. In this view, genetics isn't a cause, so much as just the facts on the ground. The causes are the things we can change. In this view, one can argue: We can change the way we raise men and the attitudes they have towards women, so the cause of rape-culture is our failure to do so.
posted by dis_integration at 9:18 AM on July 31, 2016 [7 favorites]
gusottertrout, those are good points. (And I don't want to die on a hill defending Boswell, but . . . ) to your first point, one way to respond to that is that the person asserting the evolutionary advantage "handle" would need to make the case and because that is an empirical claim (as opposed to either an analytic or conceptual claim) there are clear standards of evidence (i.e. the scientific method) that they would need to meet in order to present their claim as plausible. (After previewing they have to make the additional argument that clothing-control is a better solution has more leverage) than education, as dis_integration points out.)
Now, maybe they could, maybe they couldn't (I suspect the latter), but the very fact that there is a meaningful way to have this discussion and to (try to) suss out the good handles from the bad ones would, I suspect, be touted by Boswell as a step forward in and of itself.
posted by oddman at 9:22 AM on July 31, 2016
Now, maybe they could, maybe they couldn't (I suspect the latter), but the very fact that there is a meaningful way to have this discussion and to (try to) suss out the good handles from the bad ones would, I suspect, be touted by Boswell as a step forward in and of itself.
posted by oddman at 9:22 AM on July 31, 2016
Thanks, dis_integration. Now I have a new author to look up. (Although it turns out that the first Nancy Cartwright Google knows about is the voice of Bart Simpson... Eat my shorts, man!)
This Nancy Cartwright is the one we're talking about here.
posted by spacewrench at 9:51 AM on July 31, 2016
This Nancy Cartwright is the one we're talking about here.
posted by spacewrench at 9:51 AM on July 31, 2016
Yeah. Cartwright is great, but be warned her work (like a lot of recent philosophy of science) is super technical. She gives a lot of credit for her ideas to Elizabeth Anscombe's 1971 inaugural lecture "Causality and Determination" which is also quite difficult but at least spares you the calculus and symbolic logic. Causality is a huge area of research in philosophy of science these days, and fascinating stuff.
posted by dis_integration at 10:00 AM on July 31, 2016
posted by dis_integration at 10:00 AM on July 31, 2016
He calls politicians "single-factor fundamentalists," but that's not really fair. The interesting question would be why politicians, who often have very complex, nuanced ideas about why things are the way they are, have to dumb their rhetoric down to the point of absurd simplicity. I can think of a couple of causes.
posted by klanawa at 10:26 AM on July 31, 2016
posted by klanawa at 10:26 AM on July 31, 2016
Oh, sure, I can accept that as a reasonable position for any number of quandaries over cause and action. If, for example, I was worried about sunburn, I know I certainly can't do anything about the sun or the atmosphere, so I have to decide either to limit my exposure to the sun, and if that is an impractical handle, then wear sunscreen and appropriate clothing. No problem with those sorts of scenarios.
I was just thinking that with more complex issues, solutions can sometimes rely more on changing one's perception of the base premises of the problem.
With his talk of rape, for example, any argument about what men should do, respect women, or what women should do, dress less provocatively, isn't useful when a root premise of men having more power than women is left in place. It's only when that underlying premise is shown as a false root that the connected issue of the treatment of women can be better addressed. Seeking an answer to reducing rape without accepting women have the same societal power as men places the entire argument on an unstable framework as a starting place.
But then again I've always been more of an art history kind of guy where narrative, frames of perception, and non-discursive logic are as important than arguments of pure rationality.
posted by gusottertrout at 10:53 AM on July 31, 2016
I was just thinking that with more complex issues, solutions can sometimes rely more on changing one's perception of the base premises of the problem.
With his talk of rape, for example, any argument about what men should do, respect women, or what women should do, dress less provocatively, isn't useful when a root premise of men having more power than women is left in place. It's only when that underlying premise is shown as a false root that the connected issue of the treatment of women can be better addressed. Seeking an answer to reducing rape without accepting women have the same societal power as men places the entire argument on an unstable framework as a starting place.
But then again I've always been more of an art history kind of guy where narrative, frames of perception, and non-discursive logic are as important than arguments of pure rationality.
posted by gusottertrout at 10:53 AM on July 31, 2016
Ugh. I am sympathetic to interventionist theories of causation, but this is a super-naive variety of manipulability theory that really doesn't hold up to scrutiny. In many ways, it's similar to the Rubin-Holland view in statistics, for which see Holland's (1986) paper (pdf), especially Section 7. And it has all the same vices. For example, on such a naive manipulability theory, it doesn't make any sense to say that a person's race or gender caused him or her to lose a job or suffer ill treatment, since race and gender are not manipulable. It doesn't make sense to say that the moon's gravity causes the tides, since the moon's gravity is not manipulable. And so on. It's complete nonsense. Clark and Madelyn Glymour have an excellent short takedown of this view here.
It's also not at all true that we have to be able to control things in order to discover causal relations, as Clark Glymour, Judea Pearl, and their colleagues and students have been demonstrating for the last quarter century. The two big books in this area are Spirtes, Glymour, and Scheines' Causation, Prediction, and Search and Pearl's Causality.
Incidentally, Cartwright is not a naive manipulability theorist. She is closer to being a primitivist about causation. Her excellent 1979 paper Causal Laws and Effective Strategies (pdf) sets out one way probabilities and causal claims are (or might be) mutually constraining. She points out that one important reason for talking about causation is that causal laws are guides for action. This looks similar to the "causes are handles" sort of view, but it is subtly different. Rather than saying that causes are handles, Cartwright says that causes underwrite effective strategies for action. Her example is that even though people who buy life insurance tend to live longer, buying life insurance isn't a good strategy for prolonging your life because insurance purchases do not cause longevity. But there might be causes even where we cannot use them as handles.
posted by Jonathan Livengood at 11:04 AM on July 31, 2016 [10 favorites]
It's also not at all true that we have to be able to control things in order to discover causal relations, as Clark Glymour, Judea Pearl, and their colleagues and students have been demonstrating for the last quarter century. The two big books in this area are Spirtes, Glymour, and Scheines' Causation, Prediction, and Search and Pearl's Causality.
Incidentally, Cartwright is not a naive manipulability theorist. She is closer to being a primitivist about causation. Her excellent 1979 paper Causal Laws and Effective Strategies (pdf) sets out one way probabilities and causal claims are (or might be) mutually constraining. She points out that one important reason for talking about causation is that causal laws are guides for action. This looks similar to the "causes are handles" sort of view, but it is subtly different. Rather than saying that causes are handles, Cartwright says that causes underwrite effective strategies for action. Her example is that even though people who buy life insurance tend to live longer, buying life insurance isn't a good strategy for prolonging your life because insurance purchases do not cause longevity. But there might be causes even where we cannot use them as handles.
posted by Jonathan Livengood at 11:04 AM on July 31, 2016 [10 favorites]
His closing claim that we can't attribute causes to singular events in the past because we can no longer control them ("there are no handles on the past") is really extreme: way more extreme than most manipulability theorists would care to build into a theory (they'd permit hypothetical past manipulations, I believe). His theory would render litigation and criminal justice pretty much impossible. We attribute causes for all sorts of reasons beyond being able to manipulate causal outcomes, such as in disbursing responsibility. Of course, that's one of his goals: he doesn't want to blame himself or his ex when considering his failed relationship. But he's adopted a solution that means we also can't blame criminals for their crimes. A pragmatic conception of causation isn't obviously terrible, but he has a very limited vision of the pragmatic uses of the concept.
posted by painquale at 12:17 PM on July 31, 2016 [2 favorites]
posted by painquale at 12:17 PM on July 31, 2016 [2 favorites]
His closing claim that we can't attribute causes to singular events in the past because we can no longer control them ("there are no handles on the past") is really extreme: way more extreme than most manipulability theorists would care to build into a theory (they'd permit hypothetical past manipulations, I believe).
This is spot on. I don't know of any actual, contemporary researchers who defend such an extreme position.
Causal attribution is hard, don't get me wrong. But it's not like there has been no progress since Mill's time. Ask me how I know. ;)
posted by Jonathan Livengood at 12:33 PM on July 31, 2016 [1 favorite]
This is spot on. I don't know of any actual, contemporary researchers who defend such an extreme position.
Causal attribution is hard, don't get me wrong. But it's not like there has been no progress since Mill's time. Ask me how I know. ;)
posted by Jonathan Livengood at 12:33 PM on July 31, 2016 [1 favorite]
one could advocate for eliminating rape by killing all men. (Perhaps some already do!)
I know this misses the point, but women do rape, so that'd be one counter argument for that handle on the problem...
This is a fascinating discussion of the topic, though. In my view, Marxism takes exactly that kind of overreaching, overly logical and abstract approach to the problem of capital hoarding and rent-seeking by proposing to abolish property and ownership as a solution, even as it very accurately analyzes how and why capitalism fails to deliver the greatest utilitarian benefit. Some interesting ideas to chew on in this post and thread.
posted by saulgoodman at 2:13 PM on July 31, 2016
I know this misses the point, but women do rape, so that'd be one counter argument for that handle on the problem...
This is a fascinating discussion of the topic, though. In my view, Marxism takes exactly that kind of overreaching, overly logical and abstract approach to the problem of capital hoarding and rent-seeking by proposing to abolish property and ownership as a solution, even as it very accurately analyzes how and why capitalism fails to deliver the greatest utilitarian benefit. Some interesting ideas to chew on in this post and thread.
posted by saulgoodman at 2:13 PM on July 31, 2016
The one redeeming idea of this article is the author demonstrating how evaluativist thinking might work in practice. When he goes, consider two causes and here is the criteria you should use to pick between the two, that's interesting and largely comprehensible. But what I also see, along with some commenters above, is that throughout this attempt also bungles and obscures itself with an appeal to what's "tractable" or practical.
All that talk about handles triggers me in that coming from computer science we have definitions of "observability", a property/concept having to do with fault localization. So to me it looks a lot like Silicon Valley engineer's disease model of causality, in its overgeneralization, pandering to "practicality", and implicit refusal to consider the subtle weakness of its argument.
Just yesterday I was reading about a 40 year old paper in analytic philsophy that makes the same sort of mistake. A Marxist critic had a field day with it.
Part of what feeds the controversy is that it's so hard to converge. I still haven't gotten around to reading Judea Pearl. For most of us, these topics are very hard to access--if that says anything about tractability at all.
posted by polymodus at 3:53 PM on July 31, 2016
All that talk about handles triggers me in that coming from computer science we have definitions of "observability", a property/concept having to do with fault localization. So to me it looks a lot like Silicon Valley engineer's disease model of causality, in its overgeneralization, pandering to "practicality", and implicit refusal to consider the subtle weakness of its argument.
Just yesterday I was reading about a 40 year old paper in analytic philsophy that makes the same sort of mistake. A Marxist critic had a field day with it.
Part of what feeds the controversy is that it's so hard to converge. I still haven't gotten around to reading Judea Pearl. For most of us, these topics are very hard to access--if that says anything about tractability at all.
posted by polymodus at 3:53 PM on July 31, 2016
I know I certainly can't do anything about the sun or the atmosphere, so I have to decide either to limit my exposure to the sun, and if that is an impractical handle, then wear sunscreen and appropriate clothing.
But then society wouldn't have changed emission standards enough to reverse the growth of the ozone hole...
posted by clew at 3:53 PM on July 31, 2016
But then society wouldn't have changed emission standards enough to reverse the growth of the ozone hole...
posted by clew at 3:53 PM on July 31, 2016
His closing claim that we can't attribute causes to singular events in the past because we can no longer control them ("there are no handles on the past") is really extreme: way more extreme than most manipulability theorists would care to build into a theory (they'd permit hypothetical past manipulations, I believe).
Seems to me that if we accept the idea of a not completely determined universe, which does seem to be baked into quantum mechanics, we could attribute causes to undetermined singular events in the past.
For example, if we had Schroedinger's Cat in a box arranged so that it would only live if that radioactive element decayed in a fixed time interval chosen to give a 50% probability, and we opened the box after that interval, and the cat jumped out, ran across the room and killed a mouse, we could say looking back that the singular event of a radioactive decay in that interval caused the death of that mouse as accomplished by that cat.
posted by jamjam at 6:43 PM on July 31, 2016
Seems to me that if we accept the idea of a not completely determined universe, which does seem to be baked into quantum mechanics, we could attribute causes to undetermined singular events in the past.
For example, if we had Schroedinger's Cat in a box arranged so that it would only live if that radioactive element decayed in a fixed time interval chosen to give a 50% probability, and we opened the box after that interval, and the cat jumped out, ran across the room and killed a mouse, we could say looking back that the singular event of a radioactive decay in that interval caused the death of that mouse as accomplished by that cat.
posted by jamjam at 6:43 PM on July 31, 2016
EFFECT, n.
The second of two phenomena which always occur together in the same order. The first, called a Cause, is said to generate the other -- which is no more sensible than it would be for one who has never seen a dog except in the pursuit of a rabbit to declare the rabbit the cause of a dog.
--Ambrose Bierce, The Devil's Dictionary
posted by lazycomputerkids at 7:33 PM on July 31, 2016 [2 favorites]
The second of two phenomena which always occur together in the same order. The first, called a Cause, is said to generate the other -- which is no more sensible than it would be for one who has never seen a dog except in the pursuit of a rabbit to declare the rabbit the cause of a dog.
--Ambrose Bierce, The Devil's Dictionary
posted by lazycomputerkids at 7:33 PM on July 31, 2016 [2 favorites]
It's also not at all true that we have to be able to control things in order to discover causal relations, as Clark Glymour, Judea Pearl, and their colleagues and students have been demonstrating for the last quarter century.I've read Causality and Probabilistic Reasoning in Intelligent Systems, and I don't know what you're talking about here. There is no way to discover causal relationships without either
• interventional evidence (literally "controlling things"),
• Pearl's front door method (whereby you assume that the entire causal effect is through a set of known measurable mediators), or
• Pearl's back door method (whereby you assume that you have controlled for all confounders without creating new ones).
Also, the cited “takedown” is ridiculous. Race and gender can be controlled — not as in wetting the ground with a bucket of water — but as in a randomized experiment. (Randomized experiments count as interventions because they disconnect causal parents, i.e., the experimenter can be sure that all back door paths are closed.) Even if you can't randomize gender or race, you can often randomize the perception of gender or race, e.g., by changing the name on a resume.
posted by esprit de l'escalier at 7:26 AM on August 1, 2016
I've read Causality and Probabilistic Reasoning in Intelligent Systems, and I don't know what you're talking about here. There is no way to discover causal relationships without either ...
Interventional evidence is not the same as controlling statistically for things, and confusing the two is a serious error. The terminology seems to encourage the error. Colloquially, to control something means to intervene on it, to set it, or to make it have a value that you want it to have. By contrast, statistically controlling for something just means conditioning on what you observe. But conditioning and intervening are very different things.
Algorithms like the IC, IC*, PC, FCI, and GES algorithms (among many others) all infer causal structures from observational, not interventional, data. It's true that every method makes some assumptions. Some of these methods make the assumptions you mention. Other methods make different assumptions (for example, Spirtes, Glymour, and Scheines produce algorithms that search for causal structure using observational data without assuming either sufficiency or faithfulness, Pearl's "stability"). My claim was simply that Pearl, Glymour, and others have shown how to make causal inferences reliably without intervening -- a claim that is denied by the essay under discussion.
As to race and gender, you're right that the perception of race and gender can be manipulated. But perception of race and gender is not the same as race and gender. If you take seriously the line offered in the target essay, you would have to conclude that while perceptions of race and gender sometimes cause things, race and gender themselves do not. I infer from experiments that manipulate the name on a resume that race and gender sometimes cause hiring outcomes. Do you want to limit your claim to saying that perception of race and gender sometimes cause hiring outcomes? Do you want to say that race and gender never cause the perception of race and gender? Because those seem to me to be the implications of the essay under discussion.
posted by Jonathan Livengood at 8:17 AM on August 1, 2016
Interventional evidence is not the same as controlling statistically for things, and confusing the two is a serious error. The terminology seems to encourage the error. Colloquially, to control something means to intervene on it, to set it, or to make it have a value that you want it to have. By contrast, statistically controlling for something just means conditioning on what you observe. But conditioning and intervening are very different things.
Algorithms like the IC, IC*, PC, FCI, and GES algorithms (among many others) all infer causal structures from observational, not interventional, data. It's true that every method makes some assumptions. Some of these methods make the assumptions you mention. Other methods make different assumptions (for example, Spirtes, Glymour, and Scheines produce algorithms that search for causal structure using observational data without assuming either sufficiency or faithfulness, Pearl's "stability"). My claim was simply that Pearl, Glymour, and others have shown how to make causal inferences reliably without intervening -- a claim that is denied by the essay under discussion.
As to race and gender, you're right that the perception of race and gender can be manipulated. But perception of race and gender is not the same as race and gender. If you take seriously the line offered in the target essay, you would have to conclude that while perceptions of race and gender sometimes cause things, race and gender themselves do not. I infer from experiments that manipulate the name on a resume that race and gender sometimes cause hiring outcomes. Do you want to limit your claim to saying that perception of race and gender sometimes cause hiring outcomes? Do you want to say that race and gender never cause the perception of race and gender? Because those seem to me to be the implications of the essay under discussion.
posted by Jonathan Livengood at 8:17 AM on August 1, 2016
I never said “controlling for” things — I said “controlling” (the author's causal wording of intervening at) things.
how to make causal inferences reliably without intervening -- a claim that is denied by the essay under discussion.
Where does the essay say that? I thought his point was that it's only reasonable to talk about treatments at which you can intervene, which is reasonable to me.
Do you want to limit your claim to saying that perception of race and gender sometimes cause hiring outcomes? Do you want to say that race and gender never cause the perception of race and gender?
Gender is clearly being randomized. Egg fertilization as far as I know is a random process and there is unlikely to be any confounder between that process and getting hired somewhere. Yet it's clear that men and women have different hiring outcomes. The authors of the “takedown” say:
The question that people have is not whether gender causes hiring. The question is the extent each mediator affects hiring: was it the employer's perception of Susan's gender? —or was it the difference in Susan's access to education? Pearl writes:
posted by esprit de l'escalier at 8:47 AM on August 1, 2016
how to make causal inferences reliably without intervening -- a claim that is denied by the essay under discussion.
Where does the essay say that? I thought his point was that it's only reasonable to talk about treatments at which you can intervene, which is reasonable to me.
Do you want to limit your claim to saying that perception of race and gender sometimes cause hiring outcomes? Do you want to say that race and gender never cause the perception of race and gender?
Gender is clearly being randomized. Egg fertilization as far as I know is a random process and there is unlikely to be any confounder between that process and getting hired somewhere. Yet it's clear that men and women have different hiring outcomes. The authors of the “takedown” say:
The counterfactual, if Susan were male and had applied for the job, she would have gotten it, suggests a vague, miraculous transformation of Susan into some unspecified male (maybe one with the same qualifications, provided Susan did not attend any all-female schools)— but it makes no literal sense as a practical intervention.But randomized sampling counts as intervention, and Susan's gender was randomly sampled.
The question that people have is not whether gender causes hiring. The question is the extent each mediator affects hiring: was it the employer's perception of Susan's gender? —or was it the difference in Susan's access to education? Pearl writes:
For example, in coping with the age-old problem of gender discrimination (Bickel et al., 1975; Goldberger, 1984) a policy maker may be interested in assessing the extent to which gender disparity in hiring can be reduced by making hiring decisions gender-blind, compared with eliminating gender inequality in education or job qualifications. The former concerns the direct effect of gender on hiring while the latter concerns the indirect effect or the effect mediated via job qualification.To evaluate the effect of these mediators, you need to be able to hold certain things constant (a special kind of intervention) while changing other things (a regular intervention). This is where the distinction you were highlighting above between “controlling for” and “intervening” comes into play. If you merely control for qualifications and then verify that men and women have different hiring outcomes, you cannot conclude sexism in the hiring process because you have opened potential back doors.
posted by esprit de l'escalier at 8:47 AM on August 1, 2016
I never said “controlling for” things — I said “controlling” (the author's causal wording of intervening at) things.
In your initial comment, you said, "Race and gender can be controlled." I took you to be referring to statistical control because I don't think there are sensible interventions on race or gender. If you really meant interventionist control, then we'll need to back up to discuss whether there are, in fact, sensible interventions with respect to race and gender.
how to make causal inferences reliably without intervening -- a claim that is denied by the essay under discussion.
Where does the essay say that? I thought his point was that it's only reasonable to talk about treatments at which you can intervene, which is reasonable to me.
Okay, looking back, you're right: I was over-interpreting on this one. I moved incorrectly from the claim that the difference between "causes" and "conditions" comes down to our ability to control things -- that only what we can control counts as a cause -- to the claim that causes cannot be discovered without exercising such control. But you're right. The article doesn't make that move. For all it says, we could figure out what we can control and what effects controlling something might have without ever intervening.
You say, "Randomized sampling counts as intervention, and Susan's gender was randomly sampled." What do you mean by "counts as"? On the one hand, I may agree with you that it's good enough for inference. Although I wonder whether it lets us distinguish between genuine experiments, where we intervene on a system to set its values and quasi-experiments or natural experiments, where nature sets the values. Maybe your point is that in this case we have evidence that nature acts in such a way that it is like randomized assignment of treatment in a genuine experiment? That seems reasonable. But then why would you expect Glymour (or me) to dispute it? Seems rather to support Glymour's point, no?
On the other hand, do you think the author of the article we're discussing agrees that randomized sampling counts as intervention? If not, then the point about race and gender can still be pressed to loosen up the idea that causes just are the conditions we can use as handles. Right?
The question that people have is not whether gender causes hiring. The question is the extent each mediator affects hiring: was it the employer's perception of Susan's gender? —or was it the difference in Susan's access to education?
I think that depends on the people. Some people want to know what the facts are, regardless of whether such knowledge has practical consequences. You're right that the policy question is about what we can control. But it sort of has to be, since policies are rules we set for action. If we want our policies to be effective, they have to be the sorts of things we can implement.
To evaluate the effect of these mediators, you need to be able to hold certain things constant (a special kind of intervention) while changing other things (a regular intervention).
I don't think this is correct in general. Earlier, I urged the claim that causal structure is sometimes recoverable without interventions. Do you agree with that? If so, then can't we estimate the effects of various mediators in two stages by first figuring out the causal structure and then applying ordinary statistical techniques? I thought that Robins had some results along these lines (pdf). But maybe I'm missing something important here.
posted by Jonathan Livengood at 10:18 AM on August 1, 2016
In your initial comment, you said, "Race and gender can be controlled." I took you to be referring to statistical control because I don't think there are sensible interventions on race or gender. If you really meant interventionist control, then we'll need to back up to discuss whether there are, in fact, sensible interventions with respect to race and gender.
how to make causal inferences reliably without intervening -- a claim that is denied by the essay under discussion.
Where does the essay say that? I thought his point was that it's only reasonable to talk about treatments at which you can intervene, which is reasonable to me.
Okay, looking back, you're right: I was over-interpreting on this one. I moved incorrectly from the claim that the difference between "causes" and "conditions" comes down to our ability to control things -- that only what we can control counts as a cause -- to the claim that causes cannot be discovered without exercising such control. But you're right. The article doesn't make that move. For all it says, we could figure out what we can control and what effects controlling something might have without ever intervening.
You say, "Randomized sampling counts as intervention, and Susan's gender was randomly sampled." What do you mean by "counts as"? On the one hand, I may agree with you that it's good enough for inference. Although I wonder whether it lets us distinguish between genuine experiments, where we intervene on a system to set its values and quasi-experiments or natural experiments, where nature sets the values. Maybe your point is that in this case we have evidence that nature acts in such a way that it is like randomized assignment of treatment in a genuine experiment? That seems reasonable. But then why would you expect Glymour (or me) to dispute it? Seems rather to support Glymour's point, no?
On the other hand, do you think the author of the article we're discussing agrees that randomized sampling counts as intervention? If not, then the point about race and gender can still be pressed to loosen up the idea that causes just are the conditions we can use as handles. Right?
The question that people have is not whether gender causes hiring. The question is the extent each mediator affects hiring: was it the employer's perception of Susan's gender? —or was it the difference in Susan's access to education?
I think that depends on the people. Some people want to know what the facts are, regardless of whether such knowledge has practical consequences. You're right that the policy question is about what we can control. But it sort of has to be, since policies are rules we set for action. If we want our policies to be effective, they have to be the sorts of things we can implement.
To evaluate the effect of these mediators, you need to be able to hold certain things constant (a special kind of intervention) while changing other things (a regular intervention).
I don't think this is correct in general. Earlier, I urged the claim that causal structure is sometimes recoverable without interventions. Do you agree with that? If so, then can't we estimate the effects of various mediators in two stages by first figuring out the causal structure and then applying ordinary statistical techniques? I thought that Robins had some results along these lines (pdf). But maybe I'm missing something important here.
posted by Jonathan Livengood at 10:18 AM on August 1, 2016
You say, "Randomized sampling counts as intervention, and Susan's gender was randomly sampled." What do you mean by "counts as"?
Pearl himself draws an equivalence between "interventions" and "controlled experiments in which [the treatment] is randomized" (Causality, p. 184). What is the cause of gender besides some opaque biological process?
Maybe your point is that in this case we have evidence that nature acts in such a way that it is like randomized assignment of treatment in a genuine experiment?
Right.
If not, then the point about race and gender can still be pressed to loosen up the idea that causes just are the conditions we can use as handles. Right?
That was the distinction the author wanted to draw. Since you and I seem to prefer the statistical language, I think we should avoid his preferred language. The authors "causes" and "conditions" are both causes, the author's "causes" are the targets of interventions. It is true that we find it easiest to learn causes whose treatments we can intervene at.
To evaluate the effect of these mediators, you need to be able to hold certain things constant (a special kind of intervention) while changing other things (a regular intervention).
I don't think this is correct in general. Earlier, I urged the claim that causal structure is sometimes recoverable without interventions. Do you agree with that? If so, then can't we estimate the effects of various mediators in two stages by first figuring out the causal structure and then applying ordinary statistical techniques? I thought that Robins had some results along these lines (pdf). But maybe I'm missing something important here.
It is correct, and you can look at the paper I linked if you're curious.
You are right that you can recover causal structure without interventional evidence — but only if you make causal assumptions. The only two methods I know to do that are the front door method and the back door method, which I described.
However, mediation analysis is a harder problem than just recovering structure. Mediation analysis is quantifying how much of an effect is due to a particular mediator — while holding all other mediators constant. Hence, you need a holding operator, which goes beyond the intervention operator of causal models. Pearl recently wrote up some notes about this.
posted by esprit de l'escalier at 11:19 AM on August 1, 2016
Pearl himself draws an equivalence between "interventions" and "controlled experiments in which [the treatment] is randomized" (Causality, p. 184). What is the cause of gender besides some opaque biological process?
Maybe your point is that in this case we have evidence that nature acts in such a way that it is like randomized assignment of treatment in a genuine experiment?
Right.
If not, then the point about race and gender can still be pressed to loosen up the idea that causes just are the conditions we can use as handles. Right?
That was the distinction the author wanted to draw. Since you and I seem to prefer the statistical language, I think we should avoid his preferred language. The authors "causes" and "conditions" are both causes, the author's "causes" are the targets of interventions. It is true that we find it easiest to learn causes whose treatments we can intervene at.
To evaluate the effect of these mediators, you need to be able to hold certain things constant (a special kind of intervention) while changing other things (a regular intervention).
I don't think this is correct in general. Earlier, I urged the claim that causal structure is sometimes recoverable without interventions. Do you agree with that? If so, then can't we estimate the effects of various mediators in two stages by first figuring out the causal structure and then applying ordinary statistical techniques? I thought that Robins had some results along these lines (pdf). But maybe I'm missing something important here.
It is correct, and you can look at the paper I linked if you're curious.
You are right that you can recover causal structure without interventional evidence — but only if you make causal assumptions. The only two methods I know to do that are the front door method and the back door method, which I described.
However, mediation analysis is a harder problem than just recovering structure. Mediation analysis is quantifying how much of an effect is due to a particular mediator — while holding all other mediators constant. Hence, you need a holding operator, which goes beyond the intervention operator of causal models. Pearl recently wrote up some notes about this.
posted by esprit de l'escalier at 11:19 AM on August 1, 2016
Pearl himself draws an equivalence between "interventions" and "controlled experiments in which [the treatment] is randomized" (Causality, p. 184).
Right. But there is a difference between randomized sampling and randomized control. I'm happy to say that randomized control -- assigning a value to a treatment variable on the basis of a random process -- is a variety of intervention. What I'm saying is that this is not what happens in the case of race and gender. We aren't randomly assigning race or gender. But you may be right that we have good reason to think that nature assigns the values at random and hence that we can draw causal inferences as if we had experimentally set the values. Again, that seems to support, rather than counter-support the point that Glymour wants to make.
That was the distinction the author wanted to draw. Since you and I seem to prefer the statistical language, I think we should avoid his preferred language. The authors "causes" and "conditions" are both causes, the author's "causes" are the targets of interventions. It is true that we find it easiest to learn causes whose treatments we can intervene at.
I agree with all of this.
It is correct, and you can look at the paper I linked if you're curious.
I haven't had a chance to read the paper carefully, but based on a very quick skim, I suspect that we are talking past one another. In the abstract, Pearl writes:
You are right that you can recover causal structure without interventional evidence — but only if you make causal assumptions. The only two methods I know to do that are the front door method and the back door method, which I described.
I am not and have never been denying that causal assumptions are required. The exact assumptions vary depending on the algorithm. Spirtes, Glymour, and Scheines' FCI algorithm doesn't require the sufficiency assumption -- that we have measured all of the common causes -- but it does make some causal assumptions: the causal Markov condition for sure. This may be the source of our dispute. When I talk about making causal inferences from observation or from data, what I mean is that we make inferences from non-experimental data -- data where no experimenter exercises control -- by way of some assumptions about how statistical and causal facts constrain each other. I think it's fair to call those assumptions "causal assumptions."
However, the causal assumptions required are very general ones, many of which also seem to be required to draw causal inferences from experimental data. For example, the causal Markov condition justifies the usual procedure for randomized controlled trials. Perhaps such procedures could be justified with some other assumption. But if so, then I would expect that there are automated search procedures that will be able to recover causal structure in observational cases using those assumptions. We just won't be guaranteed to recover causal structure in every case.
However, mediation analysis is a harder problem than just recovering structure. Mediation analysis is quantifying how much of an effect is due to a particular mediator — while holding all other mediators constant. Hence, you need a holding operator, which goes beyond the intervention operator of causal models. Pearl recently wrote up some notes about this.
I agree that mediation analysis is a harder problem than structure discovery. And I agree about what it is aimed at saying. I am not yet convinced that estimating direct and indirect effects requires interventions. I'm not sure what a "holding" operator is supposed to be. In what way does such a thing go beyond the do() operator? I also don't see any mention of a "holding operator" in those notes or any discussion of the sorts of problem we're talking about here. (He says that recent discussion of mediation is motivating his remarks on the "causal hierarchy," but he doesn't connect what he says back to anything specific about mediation.)
I have to run, and I probably won't have a chance to look back at this thread for a while. I'll try to look back in this evening, though. I think this exchange has been interesting and useful. Thanks!
posted by Jonathan Livengood at 11:59 AM on August 1, 2016
Right. But there is a difference between randomized sampling and randomized control. I'm happy to say that randomized control -- assigning a value to a treatment variable on the basis of a random process -- is a variety of intervention. What I'm saying is that this is not what happens in the case of race and gender. We aren't randomly assigning race or gender. But you may be right that we have good reason to think that nature assigns the values at random and hence that we can draw causal inferences as if we had experimentally set the values. Again, that seems to support, rather than counter-support the point that Glymour wants to make.
That was the distinction the author wanted to draw. Since you and I seem to prefer the statistical language, I think we should avoid his preferred language. The authors "causes" and "conditions" are both causes, the author's "causes" are the targets of interventions. It is true that we find it easiest to learn causes whose treatments we can intervene at.
I agree with all of this.
It is correct, and you can look at the paper I linked if you're curious.
I haven't had a chance to read the paper carefully, but based on a very quick skim, I suspect that we are talking past one another. In the abstract, Pearl writes:
We show that the conditions [necessary for the identification of natural direct and indirect effects] usually cited in the literature are overly restrictive, and can be relaxed substantially, without compromising identification. In particular, we show that natural effects can be identified by methods that go beyond standard adjustment for confounders, applicable to observational studies in which treatment assignment remains confounded with the mediator or with the outcome.Emphasis added. Maybe I'm missing something or simply not understanding the terminology here, but I thought that when Pearl talks about observational studies, he's talking about cases where "treatment" is not assigned by an experimenter at all. (It's more like an "exposure" than a treatment.) Is that not correct? If it is correct, then Pearl is saying exactly what I've been urging: that we can get estimates of these things without intervening. I will give the paper a careful read in the pretty near future ... but I'm about to head out the door, so it will be later tonight at the earliest.
You are right that you can recover causal structure without interventional evidence — but only if you make causal assumptions. The only two methods I know to do that are the front door method and the back door method, which I described.
I am not and have never been denying that causal assumptions are required. The exact assumptions vary depending on the algorithm. Spirtes, Glymour, and Scheines' FCI algorithm doesn't require the sufficiency assumption -- that we have measured all of the common causes -- but it does make some causal assumptions: the causal Markov condition for sure. This may be the source of our dispute. When I talk about making causal inferences from observation or from data, what I mean is that we make inferences from non-experimental data -- data where no experimenter exercises control -- by way of some assumptions about how statistical and causal facts constrain each other. I think it's fair to call those assumptions "causal assumptions."
However, the causal assumptions required are very general ones, many of which also seem to be required to draw causal inferences from experimental data. For example, the causal Markov condition justifies the usual procedure for randomized controlled trials. Perhaps such procedures could be justified with some other assumption. But if so, then I would expect that there are automated search procedures that will be able to recover causal structure in observational cases using those assumptions. We just won't be guaranteed to recover causal structure in every case.
However, mediation analysis is a harder problem than just recovering structure. Mediation analysis is quantifying how much of an effect is due to a particular mediator — while holding all other mediators constant. Hence, you need a holding operator, which goes beyond the intervention operator of causal models. Pearl recently wrote up some notes about this.
I agree that mediation analysis is a harder problem than structure discovery. And I agree about what it is aimed at saying. I am not yet convinced that estimating direct and indirect effects requires interventions. I'm not sure what a "holding" operator is supposed to be. In what way does such a thing go beyond the do() operator? I also don't see any mention of a "holding operator" in those notes or any discussion of the sorts of problem we're talking about here. (He says that recent discussion of mediation is motivating his remarks on the "causal hierarchy," but he doesn't connect what he says back to anything specific about mediation.)
I have to run, and I probably won't have a chance to look back at this thread for a while. I'll try to look back in this evening, though. I think this exchange has been interesting and useful. Thanks!
posted by Jonathan Livengood at 11:59 AM on August 1, 2016
As spacewrench and Boswell point out this is a pragmatic heuristic for trying to suss out what we can and should do about a (usually very complex) problem. It is not an analysis of causation.
The article is really confusing, maybe because it's so short it doesn't develop the understanding and context needed by laypeople and outsiders. I had a billion questions. Stuff I don't understand:
1) What's all this business about physics having a different model of causality. The article's explanation is vague, and I don't want to read Russell to understand what the issue is. Is there a modern philosopher who explains this supposed problem?
2) An agency theory of causation is itself a candidate cause. Then, where is its handle? Is its handle justified? Tractable? How would this be addressed following the argument of the article? Similarly, the claim "Single events have multiple causes" itself is a cause (I don't know, call these meta-causes?), so how does that candidate cause fare among its peers? Meanwhile the author just keeps saying he feels it's the best explanation. Is that satisfying?
3) Most real people already have intuitions about justification, responsibility, and tractability in our daily lives and interactions with other people, institutions, cultures, etc. It also means people disagree, very much, on justification, responsibility, and tractability, on myriad issues. The author doesn't explain how that level of conflict gets resolved, only that those three criteria inform how people should evaluate causes. But isn't that level where the real, subjective conflict's at?
4) Agency theory of causation is an explanatory theory. What allows the author to wield it as an "ought" philosophy for use in our daily lives/discourse?
posted by polymodus at 12:51 AM on August 2, 2016
The article is really confusing, maybe because it's so short it doesn't develop the understanding and context needed by laypeople and outsiders. I had a billion questions. Stuff I don't understand:
1) What's all this business about physics having a different model of causality. The article's explanation is vague, and I don't want to read Russell to understand what the issue is. Is there a modern philosopher who explains this supposed problem?
2) An agency theory of causation is itself a candidate cause. Then, where is its handle? Is its handle justified? Tractable? How would this be addressed following the argument of the article? Similarly, the claim "Single events have multiple causes" itself is a cause (I don't know, call these meta-causes?), so how does that candidate cause fare among its peers? Meanwhile the author just keeps saying he feels it's the best explanation. Is that satisfying?
3) Most real people already have intuitions about justification, responsibility, and tractability in our daily lives and interactions with other people, institutions, cultures, etc. It also means people disagree, very much, on justification, responsibility, and tractability, on myriad issues. The author doesn't explain how that level of conflict gets resolved, only that those three criteria inform how people should evaluate causes. But isn't that level where the real, subjective conflict's at?
4) Agency theory of causation is an explanatory theory. What allows the author to wield it as an "ought" philosophy for use in our daily lives/discourse?
posted by polymodus at 12:51 AM on August 2, 2016
Maybe I'm missing something or simply not understanding the terminology here, but I thought that when Pearl talks about observational studies, he's talking about cases where "treatment" is not assigned by an experimenter at all. (It's more like an "exposure" than a treatment.) Is that not correct? If it is correct, then Pearl is saying exactly what I've been urging: that we can get estimates of these things without intervening.
If you don't intervene at the treatment, then you must make causal assumptions. In this paper, there are interventions at the mediators and holding operators (other kinds of interventions). If you want to work with purely observational evidence, then you must make even stronger causal assumptions (that are harder to justify).
In what way does such a thing go beyond the do() operator?
My understanding is that the holding operator holds one value across two groups. For example, in Pearl's gender discrimination example, you would have female applicants with "male resumes" and male applicants with "female resumes". When compared with male applicants with their own resumes, the former would isolate sexism in the hiring process; the latter would isolate differing hiring due to differing qualifications.
posted by esprit de l'escalier at 8:54 PM on August 2, 2016
If you don't intervene at the treatment, then you must make causal assumptions. In this paper, there are interventions at the mediators and holding operators (other kinds of interventions). If you want to work with purely observational evidence, then you must make even stronger causal assumptions (that are harder to justify).
In what way does such a thing go beyond the do() operator?
My understanding is that the holding operator holds one value across two groups. For example, in Pearl's gender discrimination example, you would have female applicants with "male resumes" and male applicants with "female resumes". When compared with male applicants with their own resumes, the former would isolate sexism in the hiring process; the latter would isolate differing hiring due to differing qualifications.
posted by esprit de l'escalier at 8:54 PM on August 2, 2016
Anyway, about the article:
It's normal after any failure (such as a breakup) to examine everything you did and ask yourself what you could have done better. It's definitely more productive to do that than it is to ask yourself what the other person could have done better. However, most people don't break up with you because of what you did or didn't do, but rather because of who you were. And the kind of spiritual transformation that happens inside us after failure is not, in my opinion, brought about by a mental search for the "handles of control". It is caused by experiencing the failure without mental gymnastics and other escapism.
posted by esprit de l'escalier at 8:59 PM on August 2, 2016
It's normal after any failure (such as a breakup) to examine everything you did and ask yourself what you could have done better. It's definitely more productive to do that than it is to ask yourself what the other person could have done better. However, most people don't break up with you because of what you did or didn't do, but rather because of who you were. And the kind of spiritual transformation that happens inside us after failure is not, in my opinion, brought about by a mental search for the "handles of control". It is caused by experiencing the failure without mental gymnastics and other escapism.
posted by esprit de l'escalier at 8:59 PM on August 2, 2016
If you don't intervene at the treatment, then you must make causal assumptions. In this paper, there are interventions at the mediators and holding operators (other kinds of interventions). If you want to work with purely observational evidence, then you must make even stronger causal assumptions (that are harder to justify).
Again, I do not deny that one needs to make causal assumptions of some kind. But making causal assumptions is not the same thing as making interventions. As before, my point is that one can make inferences about what causes what in a given system without intervening on that system. The way to do that is to assume some very general things about how causation works and then apply those assumptions to what you observe about the system.
Take a look at Corollary 1 in the Pearl paper you linked earlier.
posted by Jonathan Livengood at 12:24 PM on August 3, 2016
Again, I do not deny that one needs to make causal assumptions of some kind. But making causal assumptions is not the same thing as making interventions. As before, my point is that one can make inferences about what causes what in a given system without intervening on that system. The way to do that is to assume some very general things about how causation works and then apply those assumptions to what you observe about the system.
Take a look at Corollary 1 in the Pearl paper you linked earlier.
If conditions A-1 and A-2 are satisfied by a set W that also deconfounds the relationships in A-3 and A-4, then the do-expressions in (13) are reducible to conditional expectations, and the natural direct and indirect effects become: ...Conditions A-1, A-2, A-3, and A-4 are causal assumptions. They make claims about confounding and about whether "treatment" is affected by measured covariates. So, causal assumptions are being made. But if we make those assumptions, then we can estimate the direct and indirect effects without intervening on the target system. That is why reducing do-expressions (which represent interventions) to conditional expectations (which represent non-interventional observations) was important enough for Pearl to state in a corollary.
posted by Jonathan Livengood at 12:24 PM on August 3, 2016
Okay, but I don't see why you're arguing. I specifically said: "If you want to work with purely observational evidence, then you must make even stronger causal assumptions (that are harder to justify)."
posted by esprit de l'escalier at 2:55 PM on August 3, 2016
posted by esprit de l'escalier at 2:55 PM on August 3, 2016
« Older "It appears this is my first entry in two years." | The Guantánamo Failure Newer »
This thread has been archived and is closed to new comments
posted by CMcG at 6:12 AM on July 31, 2016