A Year Ago on Equitable Growth: Fifteen Worthy Reads On and Off Equitable Growth for July 26, 2018
Looking Backwards from This Week at 24, 16, 8, 4, 2, 1, 1/2, and 1/4 Years Ago (July 24-30, 2019)

Very nice review of a very good—if very cranky, and very much subject to not-invented-here syndrome—book. I still am unclear about the "three rungs" of Pearl's Ladder, however, largely because I do not think you can talk about causation at all without counterfactuals, except by fooling yourself: Lisa R. Goldberg: Review of Dana Mackenzie and Judea Pearl: The Book of Why: The New Science of Cause and Effect: "Pearl’s work on causal inference is often described as 'influential', and it is especially appreciated by computer scientists, but it is not the only approach...

...Statisticians tend to rely on the Neyman (or Neyman-Rubin) potential outcomes model.... Propensity scores are matched in an attempt to balance inequities between treated and untreated subjects. Since no subject can be both treated and untreated, however, the required estimate of impact can be formulated as a missing value problem.... Many economists favor models by James Heckman, whose concept of “fixing” resembles, superficially at least, the do operator that Pearl uses.... It is this reviewer’s opinion that the differences among these approaches to causal inference are far less important than their similarities....

Consider the 2007 study by Nicholas Christakis and James Fowler claiming that obesity is contagious. The claim was controversial because the mechanism of social contagion is hard to pin down, and because the study was observational. In their paper, Christakis and Fowler upgraded an observed association, clusters of obese individuals in a social network, to the assertion that obese individuals cause their friends, and friends of their friends, to become obese. It is difficult to comprehend the complex web of assumptions, arguments and data that comprise this study. It is also difficult to comprehend its nuanced refutations by Russell Lyons and by Cosma Shalizi and Andrew Thomas, which appeared in 2011.

There is a moment of clarity, however,in the commentary by Shalizi and Thomas, when they cite Pearl’s theorem about non-identifiability in particular graphical models. Using Pearl’s results, Shalizi and Thomas show that in the social network that Christakis and Fowler studied, it is impossible to disentangle contagion, the propagation of obesity via friendship, from the shared inclinations that led the friendship to be formed in the first place.

The top rung of the Ladder of Causation concerns counterfactuals, like the one posed by Daniel Kahneman and Amos Tversky in their study of how we explore alternative realities...


#shouldread

Comments