Ninon Moreau-Kastler
@nmoreaukastler.bsky.social
Researcher @taxobservatory.bsky.social @pse.bsky.social. International Trade, Development and Public Econ. PhD @ENS_ParisSaclay and @ENSdeLyon. Member of @WEPSecon initiative.
https://ninonmoreaukastler.com
https://ninonmoreaukastler.com
Congrats Kenza !! 💐
September 25, 2025 at 9:11 AM
Congrats Kenza !! 💐
May 7, 2025 at 8:26 AM
Proposed counterfactual estimator to recover RoR:
1. Estimate FE on untreated units
2. Predict counterfactual outcomes for treated
3. Compute average effect in levels
4. Scale by average counterfactual outcome
Matches RoR interpretation; generalizes 2×2 PPML.
Event study simulation:
8/
1. Estimate FE on untreated units
2. Predict counterfactual outcomes for treated
3. Compute average effect in levels
4. Scale by average counterfactual outcome
Matches RoR interpretation; generalizes 2×2 PPML.
Event study simulation:
8/
May 7, 2025 at 8:21 AM
Proposed counterfactual estimator to recover RoR:
1. Estimate FE on untreated units
2. Predict counterfactual outcomes for treated
3. Compute average effect in levels
4. Scale by average counterfactual outcome
Matches RoR interpretation; generalizes 2×2 PPML.
Event study simulation:
8/
1. Estimate FE on untreated units
2. Predict counterfactual outcomes for treated
3. Compute average effect in levels
4. Scale by average counterfactual outcome
Matches RoR interpretation; generalizes 2×2 PPML.
Event study simulation:
8/
Aggregating smaller scale correct RoRs recovers a different intermediate quantity:
* \= original RoR with heterogeneity (don’t compare to PPML)
* Interpretation depends on economic meaning of cohorts
* Divergence with PPML and log-OLS depends on correlation of outcomes (y) & treatment effects (δ)
7/
* \= original RoR with heterogeneity (don’t compare to PPML)
* Interpretation depends on economic meaning of cohorts
* Divergence with PPML and log-OLS depends on correlation of outcomes (y) & treatment effects (δ)
7/
May 7, 2025 at 8:19 AM
Aggregating smaller scale correct RoRs recovers a different intermediate quantity:
* \= original RoR with heterogeneity (don’t compare to PPML)
* Interpretation depends on economic meaning of cohorts
* Divergence with PPML and log-OLS depends on correlation of outcomes (y) & treatment effects (δ)
7/
* \= original RoR with heterogeneity (don’t compare to PPML)
* Interpretation depends on economic meaning of cohorts
* Divergence with PPML and log-OLS depends on correlation of outcomes (y) & treatment effects (δ)
7/
🖼️ Imagine firms A & B in regions 1 & 2, firm C also in region 1. Employment in C rises from 2 to 3.
* % change of the average & avg % change: +60%, +66%.
But one could also be interested in avg % change by region: +50%.
No reason for these to coincide—distinct quantities under heterogeneity.
6/
* % change of the average & avg % change: +60%, +66%.
But one could also be interested in avg % change by region: +50%.
No reason for these to coincide—distinct quantities under heterogeneity.
6/
May 7, 2025 at 8:15 AM
🖼️ Imagine firms A & B in regions 1 & 2, firm C also in region 1. Employment in C rises from 2 to 3.
* % change of the average & avg % change: +60%, +66%.
But one could also be interested in avg % change by region: +50%.
No reason for these to coincide—distinct quantities under heterogeneity.
6/
* % change of the average & avg % change: +60%, +66%.
But one could also be interested in avg % change by region: +50%.
No reason for these to coincide—distinct quantities under heterogeneity.
6/
💡To solve this, linear robust DiD estimators aggregate cohort-level correct treatment effects estimates.
🤔 But I show this approach has problems in non-linear (PPML) settings:
Different averages yield distinct targets, complicating interpretation and comparability.
5/
🤔 But I show this approach has problems in non-linear (PPML) settings:
Different averages yield distinct targets, complicating interpretation and comparability.
5/
May 7, 2025 at 8:10 AM
💡To solve this, linear robust DiD estimators aggregate cohort-level correct treatment effects estimates.
🤔 But I show this approach has problems in non-linear (PPML) settings:
Different averages yield distinct targets, complicating interpretation and comparability.
5/
🤔 But I show this approach has problems in non-linear (PPML) settings:
Different averages yield distinct targets, complicating interpretation and comparability.
5/
What happens with staggered treatment and heterogeneous treatment effect?
In a simple example (N=2, T=3), I show that TWFE PPML is biased (similarly to TWFE OLS). It "downscales" treatment effects, analog to the negative weights problem.
4/
In a simple example (N=2, T=3), I show that TWFE PPML is biased (similarly to TWFE OLS). It "downscales" treatment effects, analog to the negative weights problem.
4/
May 7, 2025 at 7:54 AM
What happens with staggered treatment and heterogeneous treatment effect?
In a simple example (N=2, T=3), I show that TWFE PPML is biased (similarly to TWFE OLS). It "downscales" treatment effects, analog to the negative weights problem.
4/
In a simple example (N=2, T=3), I show that TWFE PPML is biased (similarly to TWFE OLS). It "downscales" treatment effects, analog to the negative weights problem.
4/
Ex: Take firms A & B and their change in employment.
Pre-treatment: A=1, B=2 employees.
Post-treatment: +1 each.
* The average % change: (100%+50%)/2 = +75% (≈TWFE log-OLS)
* The change of average employment: ((2+3)-(1+2))/(1+2)=+66% (=TWFE PPML)
3/
Pre-treatment: A=1, B=2 employees.
Post-treatment: +1 each.
* The average % change: (100%+50%)/2 = +75% (≈TWFE log-OLS)
* The change of average employment: ((2+3)-(1+2))/(1+2)=+66% (=TWFE PPML)
3/
May 7, 2025 at 7:51 AM
Ex: Take firms A & B and their change in employment.
Pre-treatment: A=1, B=2 employees.
Post-treatment: +1 each.
* The average % change: (100%+50%)/2 = +75% (≈TWFE log-OLS)
* The change of average employment: ((2+3)-(1+2))/(1+2)=+66% (=TWFE PPML)
3/
Pre-treatment: A=1, B=2 employees.
Post-treatment: +1 each.
* The average % change: (100%+50%)/2 = +75% (≈TWFE log-OLS)
* The change of average employment: ((2+3)-(1+2))/(1+2)=+66% (=TWFE PPML)
3/
Reminder: in the 2×2 settings with treatment heterogeneity, TWFE PPML estimates the multiplicative DiD, or Ratio-of-Ratios (RoR), targeting % change of average in treated group.
TWFE log-OLS instead captures the DiD in log points, approximating average % unit changes (for small effects).
2/
TWFE log-OLS instead captures the DiD in log points, approximating average % unit changes (for small effects).
2/
May 7, 2025 at 7:49 AM
Reminder: in the 2×2 settings with treatment heterogeneity, TWFE PPML estimates the multiplicative DiD, or Ratio-of-Ratios (RoR), targeting % change of average in treated group.
TWFE log-OLS instead captures the DiD in log points, approximating average % unit changes (for small effects).
2/
TWFE log-OLS instead captures the DiD in log points, approximating average % unit changes (for small effects).
2/