We’ve just had a nice intro to causality from Eliezer. What if we have some variable nodes that we think might be relevant, but which we cannot observe? Judea Pearl still has a lot to say about that: as long as you can observe the right subset of variables to d-separate the source and target node, you can identify causal effects. What if you can’t observe the right subset? Hope would seem to be lost, but if you aren’t worried about determining the exact strength of the causal effect, but instead just determining if the strength is non-zero, you still have a chance.
You may remember a bunch of newspaper articles a few years ago claiming Obesity is contagious. That study’s conclusion was based on modeling assumptions (review), but Shalizi and Thomas considered the causal picture and found it to be bleak: exactly identifying the causal effect is problematic because of latent homophily. I.e. humans can have many hidden attributes that affect who we become friends with and whether we have a tendency to become obese. Measuring all these human factors is hopeless and interventional experiments (where we randomly force some people to become obese and then observe their friends) is unethical.
I’m happy to add a little to this conversation with my recent paper showing that latent homophily does not prevent us from finding causal relationships in social networks. I show that even if you can’t measure all the relevant hidden attributes describing humans, you can lower bound the strength of causal effects. We’re mostly interested in whether the effects exist, and only secondarily how strong they might be. Of course we still have to make some structural assumptions (this is the essence of Pearl’s work: it allows us to infer causality given general, intuitive, structural assumptions among variables). E.g., there is no external dynamic force (like a deity) that causes Alice and Bob to become friends and then causes them to become obese at the same time. Less extravagantly, how about a systematic, widespread change in the consistency of diets? So, although the conclusion of the paper is: latent homophily alone cannot explain correlations in obesity, we still can’t say for sure that your friends make you fat.
This is based on older work which considered a little more general setting about tests to rule out broader classes of hidden variable models, like the Bell inequalities.
We’ve just had a nice intro to causality from Eliezer. What if we have some variable nodes that we think might be relevant, but which we cannot observe? Judea Pearl still has a lot to say about that: as long as you can observe the right subset of variables to d-separate the source and target node, you can identify causal effects. What if you can’t observe the right subset? Hope would seem to be lost, but if you aren’t worried about determining the exact strength of the causal effect, but instead just determining if the strength is non-zero, you still have a chance.
You may remember a bunch of newspaper articles a few years ago claiming Obesity is contagious. That study’s conclusion was based on modeling assumptions (review), but Shalizi and Thomas considered the causal picture and found it to be bleak: exactly identifying the causal effect is problematic because of latent homophily. I.e. humans can have many hidden attributes that affect who we become friends with and whether we have a tendency to become obese. Measuring all these human factors is hopeless and interventional experiments (where we randomly force some people to become obese and then observe their friends) is unethical.
I’m happy to add a little to this conversation with my recent paper showing that latent homophily does not prevent us from finding causal relationships in social networks. I show that even if you can’t measure all the relevant hidden attributes describing humans, you can lower bound the strength of causal effects. We’re mostly interested in whether the effects exist, and only secondarily how strong they might be. Of course we still have to make some structural assumptions (this is the essence of Pearl’s work: it allows us to infer causality given general, intuitive, structural assumptions among variables). E.g., there is no external dynamic force (like a deity) that causes Alice and Bob to become friends and then causes them to become obese at the same time. Less extravagantly, how about a systematic, widespread change in the consistency of diets? So, although the conclusion of the paper is: latent homophily alone cannot explain correlations in obesity, we still can’t say for sure that your friends make you fat. This is based on older work which considered a little more general setting about tests to rule out broader classes of hidden variable models, like the Bell inequalities.