Thread by @Moshe_Hoffman

The social-sciences has a real problem w/ causal inference.

(Thread)

I am not referring to problems w/ the way we write down regressions.

Or confounds in our experimental designs.

But another endemic problem: conflating “proximate” causes w/ “ultimate” causes.

Allow me to expound. W/ examples. And describe how this holds science back.

I am referring to “explanations” in history like:

“dehumanization leads to genocide”

Or in psych like:

“Ineffective altruism is caused by the limits of empathy”

Or in poli sci like:

“Poor folk vote for ↓ taxes b/c they are ideological voters, who believe in libertarianism.”

These explanations are in some sense true.

There are valid studies documenting each.

But incredibly misleading. In a way that deeply distorts our scientific understanding.

And, importantly, the current accepted methods are unable to handle this criticism.

Eg psych experiments, designed to get at “causality”, only demonstrate “causality” in the problematic way I am about to describe. Same re historical narrative. And empirical analysis in poli sci.

Why are such “explanations” misleading? And why are the current methods not able to handle this problem?

(I’ll start w/ dehumanization. But the same logic can be applied to the other examples, and much of the rest of the lit. I’ll briefly illustrate how toward end of thread.)

The problem is that dehumanization *is* part of the “causal” pathway. Whenever a dictator tries to commit genocide, he dehumanizes the out-group he wants to kill off.

And this does make it easier for his followers to kill them off.

So the narrative is right. And even controlled lab experiments (Eg have someone read a dehumanizing essay about a minority, and you might find they are more say bad things, or accept mistreatment of that minority.)

But like does that show dehumanization “causes” genocide? “Causes” in the sense we as social-scientists might care about?

Well, a bit.

To the extent we can actually interfere with a dictator’s dehumanizing rhetoric.

But to a large extent no.

To the extent that dictators intent on killing, invariably use such rhetoric. As well as other means to ease their follower’s minds. And have the means to propoganda that rhetoric. And can incentivize people to listen to & internalize the rhetoric.

And mostly what matters is: do the people who want that outgroup dead have the power to motivate others to do the killing for them.

That’s what will predict genocide. That’s where we need to intervene if we wanna change the trajectory of history.

And hence that’s the real “cause.”

But notice, dehumanization invariably falls along the historical trajectory. So if all you are doing is tracing historical trajectories, it will look like dehumanization is the cause.

Every time a people was genocide, they were first dehumanized. And when a people are dehumanized, they often end up genocide (or ow gravely mistreated.)

Which is why the method of historical narrative, imo, isn’t up to the task of distinguishing the “proximate” cause of genocide (dehumanization) from the ultimate cause (someone w/ a motive to kill a people’s, and the power to motivate others to commit the attribute).

To see what’s so misleading bout “proximate” causes, we can think about formal definitions of causation. Like Judea Pearl’s.

The issue is that if we perturb the proximate cause (get someone to read a dehumanizing story, or look for a “natural experiment” where some had access to dehumanization radio broadcasts) the outcome of interest (willingness to mistreat outgroup members) *does* in fact change.

So by (a naive reading of) Pearl’s definition, causality is ascertained.

Pearl argues this is the right way to define causality because when we think bout causality we wanna know what in the system we can perturb to influence the outcome.

Reasonable enough.

And this is how historical narratives are justified. And likewise experimental designs.

BUT

So many things misleading here.

For starters, this lever isn’t in fact one that can be easily pulled irl.

Irl dictators choose to dehumanize. If and when it’s part of their strategic goal to harm. They won’t really just let you come in and perturb this part of the causal change so easily.

(Technically, pearl would still call this a cause. Since ease of perturbation doesn’t enter his definition. But presumably he would also admit a cause that’s really hard to perturb isn’t terribly informative.)

Another problem: irl if you artificially decrease the amount of dehumanization, the dictator will find a way to correct for this. (Unlike in a lab experiment.) Eg by buying more radio waves. Or printing more der sturmer’s. Or tweeting on his own.

(The naive definition alluded to above, and implicit in historical narratives and experimental designs, ignores all the additional causal arrows that enable the system to adjust to your perturbation. Arrows broken in the lab.)

A third problem: irl the perpetrators, if they are incentivized enough by their leader to kill, will be on the lookout for ways to justify the killing.

...

If you don’t feed them the dehumanization messages, fine. That might slow them down today. But they are liable to find another justification soon enough. Another way to ease their conscience.

(Again the lab experiment doesn’t allow for this adjustment, because it’s focused on short term impact. It’s perturbing a lever in a causal system that’s only a simplified representation of the causal model found in the outside world.)

Finally: ascertaining this “proximate” cause might give you the misleading impression that you don’t need to look for upstream effects. That there isn’t a need a benefit to seeing what causes dehumanization in the first place.

This is particularly misleading. Cause you will think you “understand” what’s driving things. You have your grand thesis of genocides.

But if you don’t stop to ask what causes the dehumanization (a more “ultimate” cause), you will miss out on WHEN to expect the dehumanization.

(Not to mention, the other proximate causes that tend to go with it, and hence many strong predictors of when the genocide is liable to happen.)

As well as all sorts of levers that can perhaps more easily be controlled than the dehumanization lever.

(Cause, unlike the dehumanization lever there won’t be all sorts of negative feedback loops, as discussed above, meant to mitigate such a perturbation. Plus pulling this lever, will influence *all* the downstream proximate causes, including dehumanization.)

Hopefully the logic is clear so far?

Let me briefly articulate how this same logic can get applied to a few other cases. And discuss a bit more what’s going wrong with the empirical analysis we run, and the historical narratives.

Why do poor Americans vote for low taxes?

A standard explanation would be that they *believe* in the corresponding ideology. Like that taxes are theft. Or trickle down economics.

A pretty famous econ paper argues its b/c they, in America at least, *believe* in upward mobility. (So they will benefit from such low taxes.)

But...

*Why* do they believe this?

(Especially given none of the above is self-evidently true, is believed by all Americans, or believed by the corresponding people in Europe.)

Hence, we again have a “proximate” explanation. Not an ultimate one.

An “ultimate” explanation would look something like: what political coalitions exist, who is incentivized to join those coalitions, what pressures are there within those coalitions to tow the coalition’s stance, what will dictate that coalition’s stance.

(Maybe also: and what arguments would help justify this stance. And under what conditions might people internalize that ideology.)

And once again, the proximate explanation can be verified empirically: in fact poor Americans who vote republican do tend to state such beliefs. And when they don’t, they don’t vote republican. And non-American poor are less likely to. And...

But, again, this “empirical verification” is misleading.

Cause these downstream variables are going to be picking up on all the causal power coming from the ultimate sources.

(Eg the people who believe the libertarian ideology, are those most incentivized to join the coalition that pushes that ideology—big business(?).)

(And the American poor, unlike European poor, are more liable to find themselves in the coalition dominated by big business(?).)

So all the same problems raised above wrt dehumanization will apply here.

(Ie lever won’t be easy to perturb, and if perturbed, system will correct. And will miss out on ultimate causes, perhaps easier levers to pull, with downstream proximate effects and no correcting force.)

Where here, the corrective force might be something like: well if you don’t buy libertarian propoganda, fine, your boss talks you into trickle down.

And the belief “lever” is hard to perturb because, well, people hold steadfast to the beliefs their boss and neighbors and pastors push on them. Even in the face of countervailing evidence and arguments.

In contrast, if we change the coalitional structure, eg by expanding the vote, or reducing the power of big business to drive the coalition’s agenda (Eg campaign finance)...

That can have downstream effects on all the above ideologies (what happened to beliefs post citizens united? What are campaign finance laws like in Europe?)

And is liable to be easier to change than trying to directly change people’s beliefs and ideologies.

Hence, perturbing ultimate causes liable to be easier, has many downstream effects, lacks corrective forces found wrt proximate causes, and are liable to be missed if focus on proximate causes.

(Briefly, the story wrt empathy is just that that too is highly susceptible to “ultimate” forces, like whether there are social and reputational rewards for helping. And predicts when empathy gets turned on and is perhaps easier to perturb and ... I think you get the point.)

To conclude:

Proximate “causes” are misleading.

Despite standard social-science methods treating them as causal.

And imho this is preventing the s.s.’s from growing into a real science—yielding an *actual* model of the underlying causal structure of human social systems.

Eom