Scott Cumminghan office hour questions

[eriliawu]

Synthetic control best practice

• Selecting donors: number of donors when the donor pool is large
  • what's an appropriate number (when we select 100 donors, only 6 to 7 are heavily weighted)
  • is there a rule of thumb for trimming down a large donor pool? (we currently use mahalanobis distance)
  • do we consider geographical proximity when we choose which donors to include/exclude in the donor pool
• Anticipatory effects: for baseline lags, allow anticipatory effects in the 2 months leading to tax/labeling?
• Placebo test: when N is large, how to proceed with placebo test?
• A change of synthetic control parameters leads to changes in results (i.e. menu labeling suffolk results)
  • what months to optimize (do we optimize until right before implementation)?
  • ask if time allows

[eriliawu]

Notes from Lloyd:

Abadie JOE — best practices with SCs. “Synthetic Controls in Action”. Go to his GitHub/causal-inference-2/readings/abadie synthetic controls in action

Don’t have to do synth, we can do vertical regression w/ matrix completion IN THE FUTURE. They are able to consider patterns in computing a synthetic control — takes away researcher degrees of freedom. JASA 2021. Matrix completion with nuclear norm regularization. Terence Tao. SCs are a special case of their method. Takes care of staggered rollout; don’t need to trim donors.

If we don’t think donors don’t belong, then don’t include them.

Proximity as ‘robustness check’ instead of straight out making it the main effect/approach. Neighboring states or something.

Citing Imbens and Rubin’s book, Abadie, on why we’re approaching our method this way.

Did not manage to get through all the questions; from the second main bullet on anticipatory effects