Earlier this week I gave a presentation based on The Book of Why at the Simons Institute for the Theory of Computing at UC Berkeley, where I have been a journalist in residence for the last eight weeks. It was very similar to my presentation at the Aptos Library in Santa Cruz, with one important difference: this one was video recorded! You can watch the whole talk at this link on YouTube, or you can watch it on the embedded viewer below. In two days, it has already gotten more than 400 views!
Just one word of warning: The first three minutes of the video is devoid of content; we were just chatting and waiting for latecomers to settle down. Beginning around the 3:00 mark, Shafi Goldwasser introduces me. If you want to skip the introduction and just listen to the actual lecture, it starts around 4:30.
Some good questions were asked during the lecture. I’d like to say a little bit more about two of them. Around the 39:30 mark, one listener asks whether an inference of causation might require a higher level of significance (such as p < 0.01) in the presence of a confounder. The answer I gave was correct but I could have explained it a bit better. I should have reminded him of the earlier example of the walking study (from about 23:00 to 23:20). The association between walking and reduced mortality was very strong and would probably have passed the test he suggested. Nevertheless, the association would be completely spurious if, say, the people in the intense-walking group were 10 years younger than the casual walkers. In that case the association would be completely explained by the age difference. So a low p-value, in the presence of a confounder, does not give us any guarantees about causality.
Secondly, one listener said something that I unfortunately talked over, and it’s hard to make out what he said. At 52:30, he suggested that one way a researcher could present a causal finding would be this: “Either you believe that smoking causes lung cancer, or there exists a confounding factor. Given our experience, we do not believe there is such a confounder.” I like this way of putting it because it puts the responsibility on a skeptic: if you don’t believe in this result, you have to say what specifically is wrong about our causal diagram. Perhaps we have omitted a confounder, or assumed two variables do not have a direct causal relation when in fact they do. In this way, we direct the discussion precisely where it should be directed: What do we know about the web of causes and effects, what is it reasonable to assume, and what do we not know yet? If we disagree about some aspect of the diagram, how sensitive are the results to our choice?
Not all of the lecture is quite this serious and academic! Keep an eye out for cute puppy pictures at 14:39 and a cute bunny picture at 50:13!
