I want to start by thanking Berk, Jeremy, and Johannes for hosting this incredibly interesting analysis over the last few posts. I can say that it's spurring a lot of really interesting debate at IDinsight and among the rest of the cozy SF development community.
I've been spending some time poring through the papers myself, and I was hoping that someone here could answer what I think is a more straightforward technical question on measurement of spillovers. In H+S '16 Table 3 (which is also Table 37 in the appendix), Column 1 reports spillover effects for the entire sample. Table 38, Column 4 in the appendix also reports spillover effects for the entire sample, so I would have expected the estimates to be exactly the same. And indeed they are for most outcomes, making it seem like they are running an identical specification.
However, there is one critical difference: the estimate of spillovers on assets. For this, the point estimate in Table 3/37 is 1.00, while in Table 38 it is 104.56. Yes, that is a massive difference, going from a precisely-estimated zero to a big positive spillover. Quite possible I'm missing something, but does anyone else pouring through these tables understand this?
Then, fast-forward to H+S '18, Table 7, column 5. Here the point estimate for spillovers on assets is 1.34. So I'm not really sure how to interpret this compared to H+S '16. Does this mean that whatever is happening with asset spillovers, it's staying pretty constant (as suggested by Table 3 in H+S '16), or am I to believe that there were initially positive spillovers that have disappeared 3 years later?
Happy if someone has this figured out and can enlighten me. Thanks!