Notes from conversation with Nadav

[ijyliu]

Preserving a record of this here

Nadav Project Notes:

military spending:

• LASSO: worried about sparsity assumption, since one country's spending can affect another's which can affect another
• take look at sparsity book (they have chart on graphs)
• you might find some data evidence for this, but most of these issues come from assumptions made

Population Dist. and CO2

• it's a hard pitch
• 2 problems: don't think it's stationary (not sure if data trained on 10 years ago is still predictive), ideally one would put it in a framework that can be justified ahead of time (like why would we use Random Forest and Population)

Political Instability:

• Nico's pitch: deal with attenuation bias by extracting "real value" using PCA instead of using IV
• Nadav says that it sounds interesting
• he recommends paper by Oster
• simulated vs real data: if use simulated data it would be good to give justifications for situation in paper

[ijyliu]

Has anyone tried to find the paper by Oster?

[paul-opheim]

Nadav TA Session (14-May) Notes:

could be nice to show when it helps and when it doesn't (in the simulation)

fine if the empirical results don't show any real effect (or if they don't match the predicted effects of using this PCA correction)

might be nice to relate what we're doing to a specific paper, but not needed

[ijyliu]

See #40

[ijyliu]

Oh, i found the paper he was talking about and seems interesting: https://static1.squarespace.com/static/5dedbd2dcd6b4e477ad8eefb/t/5dfa8d78b5d24a35631e25ba/1576701305153/Unobservable+Selection+and+Coefficient+Stability+Theory+and+Evidence.pdf

Notes on Oster: A common approach to ovb eval is looking at coef movements- but this works only if selection on obs tells about unobs. Connect bias specifically to coeff stability. Need to account for coef and R2 movement. There are often incomplete proxies. Look at sensitivity. Bias from observed/imperfect often assumed to tell about bias from unobs. Even if observables and treatment effect relation is assumed to be representativie, coeff movements don't indicate bias. Consider there being two orthogonal ability components and there is controlling, but one has more variance. If the lower variance item is seen than the coeff will appear much more stable with it's inclusion. Quality of control depends on the share of the variance explained by it. Rare R2 discussion in econ, often coefficient stability... only a third of the papers are actually said to be robust.

Comments- I guess this relates somewhat to use if we are making statements about coefficient stability. The paper seems to be more about omitted variable bias than measurement error. In our sims, the measurements have identical variances I think. And of course we normalize.

Verdict- No need to discuss this paper, because 1) we aren't making very causal claims and 2) we're normalizing the variables/getting the same variance so I don't know if the problem applies.