So much to like here

• Smart
• Policy relevant
• Rich theoretical questions and nice engagement with mechanisms
• Very clearly done and cleanly written

So much good news!

My comments mostly marginal

RD

Identification uses fact that:

“villages whose population exceeds one thousand should be allocated into their own GP” 1995/ 2015

RD

• Note heterogeneity on left (and in principle on right)
• Note some non defined potential outcomes (missing blues)

Innovation

Two strategies:

For leavers:

\[\mathbb E_x\mathbb E_{i\in A_x}[y_{i}(x) - y_i(1000)]\] where \(A_x\) is the set of villages with population around 1000 in a GP of size \(x\).

For remainers

\[\mathbb E_x\mathbb E_{i\in B_x}[y_i(x) - y_i(x-1000)]\]

where \(B_x\) is the set of villages with neighbor village around 1000 in a GP of size \(x\).

Comparison

• Strategy 1: is a drop from a given size to 1000
• Strategy 2: is a drop from a given size by 1000

Note:

• Strategy 1 captures the effects of being in a new smaller GP
• Strategy 2 captures the effects of being in an old, now smaller, GP

Average effect of dropping by 1000 seems more interpretable than effect of dropping to 1000. More like a derivative!

Observations

Even with this cleverness:

• Particular set of comparisons.
• Unclear comparability across comparison groups

To discuss

• How comparable are the average numbers between strategy 1 and 2? Quite different estimands.

• How interpretible are differences between strata? these are not as-if random

To discuss

• Substantive 1: How much of this could be novelty effects?

• Substantive 2: Are there fixed costs per unit? Are smaller units more expensive with very diffuse costs? Interpretation if this is a transfer

Estimand of most interest?

• Effect of reduction or optimal size?
• Can you predict levels? (Have you considered more structural analysis also?)

Note on comparing interactions strategy 1 and 2

Clarifying this might help with interpreting interactions:

Interaction inferences: Consider the difference between

• A: 1800 -> (800, 1000) and
• B: 2800 -> (1800, 1000)

For the splitter B is a bigger shock than A

For the remainer A is a bigger shock than B

So

Inquiry:

• Can you fill out the potential outcomes schedule?
• Can you get a handle on when small is too small?

And

• Can you challenge the novelty interpretation?

End