Dear Statalist members,
I am currently working with a panel dataset and would greatly appreciate your advice on how to appropriately model a causal relationship involving a continuous treatment variable and staggered adoption.
# Research context:
- Units: ~2,300 census tracts
- Years: 2010–2024 (balanced panel)
- Treatment variable: `gi_added_density`, a continuous measure of green infrastructure (GI) construction density added per year
- Outcome variable: `yearly_311_density`, measuring the density of 311 complaints
- Treatment timing is staggered: tracts start receiving GI at different years
- Treatment intensity varies across units and time — some tracts receive more GI than others, and some not at all
# Goals:
I aim to estimate the causal effect of **treatment intensity** (not just binary treatment status) on the outcome.
I’m also interested in examining:
1. Whether the treatment effect is *heterogeneous* across baseline levels
2. Whether the effect is *delayed* (i.e., lagged treatment effects)
#Methods I’ve tried:
- `wooldid`: Wooldridge’s 2021 TWFE-robust method with continuous treatment （from https://github.com/thegland/wooldid/）
# My questions:
1. What would be the recommended approach(es) to estimate treatment effects under *continuous intensity and staggered timing*?
2. Are there other Stata packages or model structures (e.g., event-study-like interaction designs) that are robust under this setting?
3. Is it still possible or appropriate to test the *parallel trends assumption* when the treatment is continuous and adopted at different times?
4. Would it make sense to model *lagged effects* directly (e.g., including `L1.gi_added_density`, `L2.gi_added_density`)?
Any guidance, references, or suggestions for best practices would be sincerely appreciated.
Thank you very much for your time and help!
Best regards,
Chang
I am currently working with a panel dataset and would greatly appreciate your advice on how to appropriately model a causal relationship involving a continuous treatment variable and staggered adoption.
# Research context:
- Units: ~2,300 census tracts
- Years: 2010–2024 (balanced panel)
- Treatment variable: `gi_added_density`, a continuous measure of green infrastructure (GI) construction density added per year
- Outcome variable: `yearly_311_density`, measuring the density of 311 complaints
- Treatment timing is staggered: tracts start receiving GI at different years
- Treatment intensity varies across units and time — some tracts receive more GI than others, and some not at all
# Goals:
I aim to estimate the causal effect of **treatment intensity** (not just binary treatment status) on the outcome.
I’m also interested in examining:
1. Whether the treatment effect is *heterogeneous* across baseline levels
2. Whether the effect is *delayed* (i.e., lagged treatment effects)
#Methods I’ve tried:
- `wooldid`: Wooldridge’s 2021 TWFE-robust method with continuous treatment （from https://github.com/thegland/wooldid/）
# My questions:
1. What would be the recommended approach(es) to estimate treatment effects under *continuous intensity and staggered timing*?
2. Are there other Stata packages or model structures (e.g., event-study-like interaction designs) that are robust under this setting?
3. Is it still possible or appropriate to test the *parallel trends assumption* when the treatment is continuous and adopted at different times?
4. Would it make sense to model *lagged effects* directly (e.g., including `L1.gi_added_density`, `L2.gi_added_density`)?
Any guidance, references, or suggestions for best practices would be sincerely appreciated.
Thank you very much for your time and help!
Best regards,
Chang
