F11 Panel data I Flashcards
What is panel data?
Multiple units (i = 1, 2, 3…) observed in multiple time periods (t = 0, 1, 2…).
What is the upside to panel data?
Causal inference is possible under specific conditions: Units are the same for different time periods and both observed and unobserved confounders are time-invariant.
What is time series analysis?
Analyzing one unit over time
What is a balanced panel?
Complete data for all time periods (no attrition). If there is attrition then the panel is unbalanced
Why is it important to differentiate between time variant and time invariant confounders?
Time invariant confounders are constant within a unit over time and can be discarded through fixed effects.
If a confounder is timevariant pose a risk for the causal estimate.
Draw the DAG for DiD
Good luck
How is timevariant confounders indicated in panel data models?
If the subscript is it and not only i
What is fixed effects?
Each unit have an intercept that is fixed for all time periods. The intercept absorbs all timeinvariant confounders.
It allows for different baselines.
How can you handled panel data? Two ways
Fixed effects (unit-specific intercept) or pooled regression (global intercept). The latter is problematic as you ignore the potential upside with panel data structure from a causal inference point of view.
What does pooled regression not include?
Unit-specific fixed effects, period-specific fixed effects and no time trends
How can you see whether an intercept is global or unit-specific?
Global: α (no subscript)
Unit-specific: u_i (subscript for unit)
What challenges do panel data face?
Reverse causality
Unobserved time variant confounders
Degrees of freedom (you need 10-15 observations per unit for FE)
What is twoway fixed effects?
A baseline approach for panel data in political science. Fixed effects for both unit and time
Controlling for both unit invariant confounders and exogenous time shocks for all units.
Could also include a time trend
Why is degrees of freedom important with panel data?
For each unit- or time specific fixed effect estimated we lose degrees of freedom.
Therefore, leave out the global intercept. Year fixed effects could be left out if there are strong theoretical arguments.
What is degrees of freedom?
The number of observations in relation to number of estimated parameters.
df = n - parameters
