1

Referencing this question and its smart answer:

I'm still confused and the confusion has been in part fueled by "recent" Pearl/Hernán tweets (tweet, tweet and tweet).

Pearl's Ladder of causation and my rephrasing of the third W'question.

  1. Association - What is?
  2. Intervention - What if?
  3. Counterfactuals - What would be if?

Wouldn't interventions that are (perfect) randomized experiments, where two or more groups are treated, each group emulating counter-to-fact causes, not be squarely placed on rung 3.) - all along the Neyman-Rubin potential outcomes framework!? Isn't the potential outcomes framework and a framework for counterfactual outcomes conceptually identical? Pearl says:

Interventions change but do not contradict the observed world, because the world before and after the intervention entails time-distinct variables.

I can understand that interventions of the sort "If we take aspirin, will our headaches be cured" are rung 2.); but then they are not experiments, are they? (Perfect) randomized experiments are set up to compare a factual cause to a contrafactual one. What would be if I'd taken aspirin in comparison to not have taken any - or counterfactual: I've taken aspirin and others have been randomized to not to, so what is the difference in headache averages?

A part of the rung 2.) / rung 3.) "problem" can perhaps be traced back to different taxonomy. The "Harvard group" defines causal inference as:

Causal inference has a central role in public health; the determination that an association is causal indicates the possibility for intervention. (Glass, Goodman, Hernán, & Samet, 2013)

So in their eyes (Hernán's trichotomy: description, prediction (probably rung 1. on Pearl's ladder) and counterfactual prediction=(randomized, perfect) intervention=causal inference (Hernán, Hsu, & Healy, 2019)) they would most probably see interventions on rung 3.) (as "counterfactual interventions" perhaps?).

Which brings me to an interesting very recent intervention by Imbens (Imbens, 2019):

I would have liked to have seen a fourth rung of the ladder, dealing with “why,” or reverse causality questions.

Wouldn't that then call for a reformulation of the Ladder of causation?

  1. Association - What is?
  2. Intervention - What if?
  3. Counterfactual prediction/intervention (experimental or observational, if we could perfectly de-confound) - What would be if?
  4. Counterfactual explanation - Why?; by this I mean all that the other nice things we can do with counterfactuals (Chapter 8 -> The Book of Why; attribution, mediation, generalizability, etc., etc.)

Cheers

nafets
  • 21
  • 3
  • 1
    Voting to close as opinion based This is the age-old epistemology and heuristics debate, claiming you have to know why something works to know how it works. – AdamO Sep 17 '19 at 14:14
  • Isn't that an *opinion* in itself, one associating "age old debate" "nothing really new", so rung 1.) on the **Ladder of causation**? Just wondering.... In effect: there are quite a few taxonomies around these days, not least fueled by the undeniable prominence of *The Book of Why*. So a reasonable discussion. – nafets Sep 17 '19 at 14:26
  • Not quite sure what the *opinion* filter here is all about. Would suggest reading **Causation&Explanation** (Stathis Psillos). Theories are sets/web of beliefs and each belief has content that is made up of concepts. Causation in this context *is* a theory, alot of factual concepts, but a set of beliefs nonetheless. So if its only facts your asking for, there will be no discussion. There is a quite a blunt question asking for expertise at the end of my query: "Is Imbens right or Pearl". See also Intervention by @CarlosCinelli in the referenced post. – nafets Sep 19 '19 at 11:18
  • If you reformulate your question to make it more technical, it might be possible to reopen. Regarding the ladder of causation, the criteria in the Book of Why is defined *mathematically*. More precisely, the rungs are defined based on the *type of information* on the causal model that you need to answer a query of that level. So, in that vein, what type of extra information do you need to answer your rung 4 query that is not in rung 3? – Carlos Cinelli Sep 24 '19 at 22:40
  • You would need to define the type of query (mathematically) and the type of information needed to answer that query (mathematically), then we can see if we can have 4 rungs instead of 3. Hernan is collapsing rung's 2 and 3 into one, for instance, which is ok, but you lose some interesting distinctions. Imbens "reverse causality" seems to fit squarely on rung 3, as questions of actual causation. – Carlos Cinelli Sep 24 '19 at 22:45
  • Thanks Carlos - hadn't looked back here for a while, somewhat disillusioned by the restrictive moderation. – nafets Oct 19 '19 at 15:26

0 Answers0