Description
Use repeated observations of the same units to remove all time-invariant unobserved differences between them, by demeaning within units or adding unit dummies.
With panel data, any confounder that differs between units but is constant over time can be eliminated. Pooled regression ignores this and is biased when unit traits correlate with treatment, with standard errors too small because of serial correlation. The within estimator demeans outcome and treatment by each unit's mean (equivalently, includes a dummy per unit); adding period dummies gives two-way fixed effects. Identification needs strict exogeneity conditional on the unit effect and within-unit variation in treatment, and time-invariant covariates drop out. Standard errors are clustered by unit, ideally with at least about thirty clusters. Comparing pooled and fixed-effects estimates shows how much unit-level confounding was removed: a marriage wage premium of 8.3% fell about 60% under fixed effects. It cannot fix reverse causality, feedback from past outcomes or time-varying confounders.
Attributes
- Discipline
- Experimentation & Causal Inference
- Loss type
- Unmigrated
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Draw a DAG and judge whether confounding is plausibly time-invariant, with no feedback from past outcomes to treatment.
- Structure the data as a panel of units (users, accounts, stores) by period.
- Estimate pooled regression as a baseline.
- Estimate the within estimator by demeaning or unit dummies; add period dummies for two-way fixed effects.
- Cluster standard errors by unit (at least about thirty clusters).
- Compare pooled and fixed-effects estimates and interpret the change as removed unit-level confounding; note remaining threats.
Sources
- Causal Inference: The Mixtape (Cunningham, 2021)
- Handbook of Child Psychology Vol 4 (Lerner et al., 2015)
Also known as
- Panel fixed effects (within) estimator
- Fixed effects regression
- Within estimator
- Two-way fixed effects
- Unit fixed effects
- First-difference and sibling-difference (fixed-effects) models
- Individual fixed effects
- Family fixed effects
- Sibling design
- Change models
- Panel fixed-effects estimator
Categories
- Analytics & Behaviour
- Experimentation & Causal Inference