Keep in mind that the timeline from medical research being published to entering practice can sometimes be measured in decades, and causal inference is swimming against the RCT-heavy currents of medical research. Several journals I publish in, for example, expressly forbid causal language for non-RCTs (much to my consternation as a mathematical modeler).
Many of the tools of causal inference were also relatively inaccessible until fairly recently.
I'd argue that the potential outcomes frameworks are really useful from a philosophical standpoint in teaching students, and things like the target trial framework has been doing useful work in conceptualizing observational studies in the context of a hypothetical trial (and recognizing that trials are themselves essentially a special case of cohort studies).
The major source of potential is the places where you can't ethically randomize interventions, yet still need to produce evidence.
Thanks for this. I agree that potential outcomes is very useful, at least on a pedagogical level. I have been reading Imbens and Rubin's book and loving it.
I think causal inference is succesfull where it is impossible do an intervention to force a randomized trail on your population. Classical statistics like in agriculture that you set up a design and field trail to find out the interactions and additive effects is sometimes not possible. Say you want to check the effect of some economic policy change or medical treatment that would be unethical to refuse to some part of the population.
I'm aware that causal inference is a popular technique in econometrics, and other places where we cannot conduct experiments.
What I'm not aware of, is if these techniques have produced highly useful and reliable inferences. (Putting on my counterfactual hat) Are there examples of observational studies that would have failed to change public policy without the techniques of causal inference?
Have a look at the literature on how smoking was established as a cause for cancer. You can't ethically intervene to have non-smokers smoke long enough to develop lung cancer. A lot of money and intellectual effort was spent on correlation not equaling causation in this case.
I'm no expert on the literature here, but Peter Norvig mentions the smoking-cancer example in his article on experiment design [0]. He gets to the same place the causality people do; observational studies.
The core idea behind a RCT is that the characteristics of a "unit" (a patient) can't affect which treatment is selected. On average, people who got treatment A are statistically the same as those who got treatment B. So you can assume any difference in outcome is a result of the treatment.
One of the simpler ways to do causal inference is by pairwise matching:
You try to identify what variables make patients different. Then find pairs of units which are "the same" but received different treatments. After the pairing process, your treatment and control groups should ("should" is doing some heavy lifting here) now be statistically "the same" by construction. Recall, that this is what we were going for in an RCT. If you did everything right, you can now apply all the normal statistical machinery that you would apply to an RCT.
The challenge is:
1. Identifying all the variables that make units alike.
2. You tend to throw away a lot of data, which reduces your statistical power. Even when the treatment classes are balanced, a given unit in class A may not pair up well with any unit from class B.
3. (Related to 2) Finding globally-optimal pairs of closest matches can be hard.
4. (Also related to 2) You need at least some people in each group. Sometimes the treatment and control are just so different that nobody pairs up very well.
In some sense, the pairing process is just a re-weighting of your data. People who are similar to someone in the other group have a large weight. People who are unlike the other group have a low weight.
You can generalize that idea a bit and reinvent what's called Inverse Propensity Score Weighting. In this case, you try to model a unit's propensity to receive a treatment, and then use 1/propensity as that unit's weight.
The intuition is: If the model says you were likely to receive treatment B (you have a low propensity for A) and you actually received treatment A, then you are likely to pair up with someone who actually received B. So we should up-weight you.
I'm currently working on my Master's thesis related to implementing propensity score matching for program evaluation in the child protection service system.
I cannot stress enough how important #1 is above. The most important part of making causal inferences in an observational experimental setting is identifying and collecting the variables associated with the treatment and outcome. It is easy to conceptualize but much harder to do in practice.
This is true only for a small subset of Causal DAGs even within this 'Causal Calculus'. It can't account for circular causality or discontinuous relationships. That's not to diminish your suggestion, only to contextualise it.
You can't really, at least not in the sense that I think most people think of it.
You basically need to make some assumptions that are broadly equivalent to assuming you've already correctly guessed certain parts of the underlying causal structure. So in a certain sense you're kind of begging the question, in a way that you wouldn't need to do if you had the ability to do interventions/randomized trials.
That being said causal inference techniques are still very valuable in making explicit exactly what assumptions you're making and how those affect your final conclusion and therefore how to minimize the impact of those assumptions.
The rules also provide a framework within which you can rule out some causal relationships. So they at least go some way to confirming which hypotheses can't be correct given the data.
Propensity score matching is probably the one area with the greatest utility. There's a lot of good literature on the subject and it's fairly popular in econometrics and health analytics.
Multiarmed bandits and contextual bandits are essentially causal inference with a cooler name. You can formulate both with a potential outcomes/couterfactual framework, and contextual bandits typically is presented that way.
(Bandits are often presented as a loop where you control the policy collecting data and update it frequently, but that does not have to be the case.)
Recommender systems and search/counterfactual learning-to-rank can be thought of as an extension to counterfactual bandits as well.
If it helps, one connection I see is that recommender systems often involve causal questions like: “how will user behavior change if we change the order these results appear in, or if we change which results appear in the first page of results, etc.”. Additionally, since we can only show one ranked set of answers for each query, counterfactual questions also rapidly arise about what would have happened if we had answered past queries differently.
I'm pretty sure the link between smoking and cancer was established before causal inference came about.