There are many situations, in science and in life, in which we want to know what caused some particular event. The current coronavirus pandemic. The recent Australian bushfires. Hurricane Katrina. A specific political candidate winning (or losing) a specific election. Diederik Stapel becoming a serial scientific fraudster. Etc.
Some proposed explanations will appeal to “structural” causes. By “structural” causes, I mean background conditions that “set the stage” on which other contributing factors play out. For instance, global warming likely makes hurricanes more common and increases the intensity of the average hurricane, and so in that sense “set the stage” for the very severe Hurricane Katrina. As another example, here’s an article reviewing “structural” explanations for why Bernie Sanders did not win the 2020 Democratic Presidential primary (e.g., only 20-30% of Democratic primary voters self-identify as “very liberal”). And here’s my review of the structural factors thought to drive scientific misconduct.
But “structural” causes aren’t mutually exclusive with non-structural causes, meaning causes specific to the particular event in question. A meteorologist asked why Hurricane Katrina was so intense wouldn’t just say “because global warming makes hurricanes more common, and more intense on average” and leave it at that. After all, those same structural causes also apply to many other hurricanes that weren’t so intense as Hurricane Katrina. So the meteorologist would probably also talk about, e.g., water temperatures and atmospheric conditions at the particular place and time where Hurricane Katrina formed. As another example, someone asked to explain why political candidate X lost election Y might mention structural factors like the state of the economy in the run-up to the election, but might also suggest that candidate X would have won by campaigning differently (example). Someone asked to explain why Diederik Stapel became a fraudster wouldn’t just appeal to structural factors like pressure to publish, since after all those same structural factors apply to many people who don’t become fraudsters. Etc.
Because structural and non-structural causes aren’t mutually exclusive, we’re often interested in apportioning responsibility between them.* Which is difficult. Further, it’s not just difficult for the familiar reasons that most every causal attribution problem is difficult, such as that it’s often hard to tell correlation from causation without experimental data. Apportioning responsibility between structural and non-structural causes also is difficult for conceptual reasons. Frankly, I just don’t know what it means to “apportion causal responsibility” between structural and non-structural causes. It’s not merely that I don’t know of any practical way to calculate “Event X was due 80% to structural causes, 10% to non-structural causes, and 10% to the interaction of structural and non-structural causes.” It’s that I don’t even understand how the question “How do we apportion responsibility between structural and non-structural causes of event X?” could possibly have an answer at all. There are many contexts in which I understand very well how to quantitatively apportion responsibility–but this isn’t one of them.
Which is where you come in! Because I am of course far from the first person to wonder about this issue. So you tell me: what are the best discussions of partitioning the contributions of structural and non-structural causes of some particular event? I’m looking for both general philosophical discussions of this problem from philosophers and statisticians, and discussions in the context of particular examples (e.g., the contributions of global warming vs. non-structural causes to the intensity of specific hurricanes). Looking forward to learning from your comments.
p.s. Thank you in advance to anyone who reads this or comments in the middle of the coronavirus outbreak. Because, (i) coronavirus outbreak, and (ii) boy howdy, does the title of this post suck! I couldn’t think of a better title, sorry. Clearly, my ability to pick decent post titles is one casualty of the coronavirus outbreak. Now don’t all trip over each other to comment that, “No, your post titles have always sucked”. 🙂
*There’s probably a whole ‘nother post–or book!–to be written on why we want to apportion responsibility between structural and non-structural causes of events. The motives often go beyond disinterested scientific curiosity. Depending on the context, you might find structural explanations for event X much more, or less, congenial than non-structural ones. For instance, if you’re hostile to the idea of anthropogenic global warming, you’ll probably be hostile to any attempt to partially blame Hurricane Katrina on global warming. Conversely, back when it looked like I wasn’t going to get a faculty position, I wasn’t bitter or upset. In part because I focused on structural explanations for my apparent failure to land a faculty position, rather than on explanations that would imply some personal failing on my part. Personally, I much preferred to focus on the fact that the faculty job market’s tough for everyone, rather than on the possibility that I’d screwed up by not branching out more during my postdoc. More broadly, it’s my anecdotal impression that lots of unproductive arguments in lots of areas of life are about apportioning responsibility between structural and non-structural causes of particular events. If there were something general that could be said about how to do that apportioning–or about why it’s impossible to do that apportioning–it might help settle or defuse some** of those arguments.
**”some” meaning, like, one or two of those arguments. At most. People are people, they’re always going to find things to argue about. 🙂
This is an important but terribly hard question. I see it right now in questions regarding why Covid-19 has spread much more in some parts of Stockholm, and not others. People are very quick to make assumptions based on the structural-causes idea, whereas I think that in this particular situation, at least during the phase of exponential spread, pure random events have an unusually strong effect. Maybe that is one way of approaching the question: How much difference to an effect does pure random noise in inputs have?
Yeah, I’m thinking of using various wild speculations about geographic variation in the severity/timing of COVID-19 outbreaks as an example for my intro biostats students, illustrating why correlation is not causation.
Chris Whitty, the UKs chief medical officer, was asked this near the very early start of the European outbreak. And his answer was along the lines of: We got lucky, Italy got unlucky.
Yes, these are hard problems. They are grappled with directly using the concept of counterfactuals in the field of casual inference. Be advised though, this no place for the the intellectually meek. There is no free lunch. Any answer you gets depends on your causal model of the phenomenon. Thus, the credibility of the your answer depends on the credibility of your causal model. If you are not willing to posit a possible causal model, you can make no progress at all. To dive in, look at the post-2000 work of Judea Pearl and the thousands of papers his work has generated. Have fun.
Yeah, I have various tentative thoughts about stuff to read to better understand this issue. “Read Judea Pearl” and “read Don Rubin” are both on the list…
A tentative thought: one way to address the issue raised in the post is to subtly change the question. Instead of trying to apportion causal responsibility for a particular event, apportion responsibility for events (plural) that are sufficiently like the particular event of interest in relevant respects. This is a standard move in certain versions of Bayesian statistics; I think of it as the “Nate Silver” move because it’s his go-to move when it comes to predicting particular events. This move allows you to study the probability distribution of events that are sufficiently like the particular event of interest, which I think makes it easier to put structural and non-structural causes into the same framework.
This is like the central, ongoing, eternal debate in fields like history or electoral political science (and fairly explicitly so). But it is interesting to see that as ecology gets involved in global change it is coming to our field too. Why did species X go extinct? Well its complicated (although I do think ecologists have answered this question pretty accurately as a mix of systemic factors that drive abundance and range down, then random highly specific factors that knock a population out).
I wonder if the extinction example doesn’t point to a temporal sequencing of structural first then specific? That doesn’t answer the partitioning question but might provide a useful attack.
In population biology ecologists have long modelled both deterministic and stochastic forces. I wonder how fair it is to equate those to structural and specific respectively?
More generally, I wonder if it is even sensible to partition these structural vs specific because both are necessary conditions (event doesn’t happen without both). Or at least that is my world view of this although people could argue with it. Nazi Germany doesn’t happen without both oppressive settlement conditions leading to an extended weak economy AND a charismatic evil genius?
Or more prosaically, my car won’t start unless there is gas sitting in the tank waiting and I choose a moment in time to turn the ignition so what sense does it make to partition causality between the two?
“This is like the central, ongoing, eternal debate in fields like history or electoral political science (and fairly explicitly so).”
Yeah, this post definitely bites off more than one post should really try to chew!
“Or more prosaically, my car won’t start unless there is gas sitting in the tank waiting and I choose a moment in time to turn the ignition so what sense does it make to partition causality between the two?”
I’m glad to hear that it’s not just me who has trouble even making sense of the notion of apportioning blame between structural and non-structural causes of some specific event!
The problem with the deterministic-stochastic apportionment is that it may, more than anything else, reflect our ignorance at the time of the analysis. That is, what we apportion into the stochastic category often is not strictly without cause — it’s just that we don’t yet understand the causal relationship, and/or the relationship operates at one or more levels different from the main phenomenon of interest (i.e. macro vs. micro).
This is also like the long-lasting but ultimately futile debate about nature vs nurture affecting phenotypic variation (especially behavior). There is the old metaphor of a cake recipe about how much influence on the final product is the flour vs baking powder. The answer, if you want an eatable cake, is pretty much 100% for both. But I think a more useful way to think about this is statistical; usually these are problems of factors varying at different hierarchical levels with potential interactions between levels. So the nature-nurture debate is defused in large part by thinking about genetic variation for environmental influences on phenotypes. In the case of Katrina (always hard to analyze a one-off event though), factors on a global scale interact with local (both spatially and temporally) to produce specific events. Again, the answer is to not ask how much influence each factor has when there are interactive effects, but to ask how they have an effect, seems to me.
Elisabeth Lloyd has a few articles on the best way to attribute extreme events to climate change, sometimes with other writers (Naomi Oreskes and Michael Mann). Drawing on some proposals from climate researchers, she argues that we should often replace the traditional risk-ratio approach (“What is the probability of a certain class of weather events, given global climate change, relative to a world without?”) with a storyline approach (“How much did climate change affect the severity of a given storm?”).
These are indeed different questions, but one reason the storyline approach works is that there are well-understood physical relationships (e.g. the Clausius-Clapeyron relation) between air temperature and variables like moisture content — and further, between moisture content and storm severity. On the other hand, we don’t really understand and can’t predict the atmospheric dynamics the cause storms to form in a given place and time. So the proposal is just to shift attention away from the probability of whether a storm forms (where we can’t tease apart the ‘structural’ factor of climate change) towards understanding the severity of the storm (where we can).
Thanks for the pointers to this literature, Shan.
I’m not sure if I 100% understand the distinction between structural and non-structural causes, but the comments and examples here seem to point to two properties. 1) Structural causes affect more things than nonstructural causes. For example, compared to atmospheric pressure at a specific time and place, global warming influences much more than the probability of a large hurricane, like heatwaves and sea level rise. 2) Structural causes are more predictable. Sticking with the hurricane example, we’ve known about climate change since the 19th century, but we can only predict fluctuations in atmospheric pressure a week in the future. These two properties are not binary variables, but looking at the extreme cases points to a more complicated taxonomy of causes
Properties 1 and 2: structural causes
1, but not 2: historical contingencies. An unpredictable event that has many effects. Examples: “the rise of a charismatic evil genius”, or the assassination of Archduke Franz Ferdinand, or the Alvarez asteroid.
2, but not 1: Predictable things tend to be large-scale trends (e.g. climate change, the state of the economy, human population growth), so it’s hard to imagine a very predicable cause that affects very few things.
Neither 1 nor 2: stochasticity. It is not worth our time to identify unpredictable causes with isolated effects, so we model them as noise.
Jeremy makes a good point about the non-mutual exclusivity of structural and non-structural causes. I think this can be accounted for and measured with Judea Pearl’s causal diagrams (aka DAGs) through direct and indirect effects. For example, climate change has a direct effect on atmospheric pressure (among other things), and atmospheric pressure has a direct effect on the probability of hurricane formation, so climate change has an indirect effect on hurricanes, mediated through atmospheric pressure. A good (beginner’s) resource for how to actually measure these things is chapter 9 of Pearl’s “The book of why?”
Frankly, I’m not sure I’m 100% clear on the distinction between structural and non-structural causes either!
In some cases, I think “structural cause” means something like “pre-existing condition or state of affairs that is difficult or impossible to change, and in which there are many possible futures in which event X occurs and only a few in which event X doesn’t occur”. For instance, “only 20-30% of the US Democratic primary electorate identifies as ‘very liberal'”. That’s plausibly a pre-existing state of affairs that’s very hard for any candidate for the Democratic presidential nomination to change. And given that pre-existing condition, it’s plausible that at lot of other things would have to happen in just the right way in order for a very liberal candidate to win the nomination.
I wonder if it might be helpful to think of underlying symmetries (referring to Steven Frank’s paper and your post on it) as structural causes in this context. They are both different of course, but there is some degree of similarity in the concepts. This could help in partitioning because now you can divide the event into sub-events, and some central/defining sub-event (analogous to the power law increase in fitness in that example) can be explained by the structural causes while other sub-events (analogous to the specific parameter values, the variance between the lines, other pieces of data that need to be explained) can be explained by non-structural causes. This doesn’t give a quantitative partitioning, but I’m not even sure how to do that in a situation when all causes are necessary and only together sufficient to create the event. But now you do have different aspects of the event being attributed to different causes.
Very interesting suggestion! I’ll need to think more about that. I hadn’t made that connection.