Effect of events Flashcards
Our interest is in the causal effect of finding a
partner. Under which condition could we identify
this causal effect by comparing different persons?
Fundamental problem: we cannot observe change in the same person under the same conditions at two different time points
Between solution:
- statistical twins / need to share all characteristics (unit homogeneity)
- no self-selection (self-select due to omitted/unobserved variable is a problem)
Our interest is in the causal effect of finding a
partner. What is better than between estimation?
What is the within-transformation?
- differencing the data within each individual by subtracting their individual-specific mean or a fixed effect from each observation
- by doing so, the individual-specific time-invariant factors, including the random effects, are differenced out and no longer contribute to the analysis –> random effects assumption becomes irrelevant
- new focus: the time-varying changes within each individual and the factors that drive those changes.
What is the
a) Random-effects assumption
b) Contemporaneous exogeneity assumption
Random-effects assumption: no time-constant unobserved heterogeneity (nothing in the error term that confounds the relationship) –> cancels out in FE estimation as poeple become their own control
Contemporaneous exogeneity assumption: nothing changes over time within person -> no time-varying unobserved heterogeneity
What happens to time-constant factors in within estimations?
“A major motivation for using panel data
has been the ability to control for possibly
correlated, time-invariant heterogeneity
without observing it (Arellano 2004).“
Who needs to be in the treatment group?
The ones that actually experience the event (i.e. finding a partner) - there needs to be variance, i.e. so who is always with partner will not contribute to change bc there is no variance
Interpret the coefficient of par
also: u always want a causal explanation but it is harder to get but only holds if exogeneity assumption applies
Why do we use panel-robust SE in FE estimations?
in panel data, assumption about residuals are not as easily satisfied, especially that residuals have a constant variance and that they not auto-correlated → if we ignore this = underestimate standard errors & regard our estimates as more precise then they truly are
but not as important as getting the effect estimates right (biased effect with correct SE does not get u far)
What is the problem with comparing those two?
Ignores temporal order
How to model time & temporal order
(1) Define the sample: Include only those who can experience the event
(2) Define the event
* Anchor the event in time
* Decide whether to remove or keep reverse and repeated transitions (Keep reverse/repeated transitions if the interest is in finding/having a partner, removeif the interest is in finding/keeping a partner)
(3) Control for unobserved heterogeneity
* Use within-estimation (removes time-constant confounders)
* Use controls for the temporal profile of the outcome (time-varying confounders)
* Use a control sample
Which impact functions are there?
How to control for time-varying heterogeneity?
Control temporal profile (i.e. age, period, anticipation effects)
Why use a control sample?
We could estimate the age profile from the treated
subsample (enough pre- and post-treatment observations)
BUT this is inefficient and can lead to collinearity issues
–> Always keep a control sample
How to model age?
How to model other time-varying heterogeneity?
To estimate the effect of finding a partner on life satisfaction - should we control for income?
Selection on growth, control for it without conditioning it out
What are anticipation effects?
What happens if wie ignore anticipation effects?
Should we model antipation effects?
Should you be careful with anticipation effects?
Finding a partner - anticipation effect included?
How to model heterogeneous effects for events?
1) Seperate models (men vs women)
2) Cross-level interaction effects
Why does including gender not work in a FE model?
+ solution
Interpret
Hybrid model
+ / -
Divorce example of hybrid model: what does this tell us?
What can we do with the hybrid model?
