Thank you for the lecture Prof! Super interesting, specially the part on visualisation! Prof., what is the relationship between binscatter and ML? It looked to me like bin-scatter is just a flexible way to estimate the CEF even if it is non linear. In that spirit, can't you use any ML function approximation to display the relationship between the treatment and the outcome (or bayesian ML if you confidence intervals)?
Hi Matheus, yes, for sure -- there's a clear connection between estimating E(y|X) (the conditional expectation) and binscatter. The trick is always how to visualize the "right" relationship. Say there are two covariates: x_1 and x_2. If you allow E(y|x1,x2) = g(x1,x2) to be fully flexible, it's challegning to visualize the relationship between x1 and y --- it really depends on your choice of x2 (this is true of logit, for example). What the binscatter models tend to assume is some additive linearity -- E(y|x1,x2) = g(x1) + x2*beta, such that you can residualize the effect of x2, and then just plot the relationship between x1 and y. The "On Binscatter" paper talks about this in more detail, but that's the crux of the main difference between more general flexible models and binscatter style approaches.
Dear Prof. Is it possible to get access to the recordings of the last five lectures too? I am very keen to go through the slides and other teaching material as well. Looking forward to your response. Regards Gaurav
Thank you, Prof! Very interesting, I hope to see more like this
Thank you for the lecture Prof! Super interesting, specially the part on visualisation! Prof., what is the relationship between binscatter and ML? It looked to me like bin-scatter is just a flexible way to estimate the CEF even if it is non linear. In that spirit, can't you use any ML function approximation to display the relationship between the treatment and the outcome (or bayesian ML if you confidence intervals)?
Hi Matheus, yes, for sure -- there's a clear connection between estimating E(y|X) (the conditional expectation) and binscatter. The trick is always how to visualize the "right" relationship. Say there are two covariates: x_1 and x_2. If you allow E(y|x1,x2) = g(x1,x2) to be fully flexible, it's challegning to visualize the relationship between x1 and y --- it really depends on your choice of x2 (this is true of logit, for example). What the binscatter models tend to assume is some additive linearity -- E(y|x1,x2) = g(x1) + x2*beta, such that you can residualize the effect of x2, and then just plot the relationship between x1 and y. The "On Binscatter" paper talks about this in more detail, but that's the crux of the main difference between more general flexible models and binscatter style approaches.
Dear Prof.
Is it possible to get access to the recordings of the last five lectures too?
I am very keen to go through the slides and other teaching material as well.
Looking forward to your response.
Regards
Gaurav
I'm slowly putting them up, 1-4 are now available.
@@paulg-p Thank you so much.