Econometrics Flashcards
What matters for the relevance criterion and what is a rule of thumb?
Significance level (not magnitude)
F statistic of at least 20
How can you test the exclusion restriction?
You cannot, but you can test specific violations
Two steps of IV analysis
First stage: regress predictor on instrument
Reduced form: regress outcome on instrument
What do you do if you have covariates in your IV analysis?
Include the same covariates in the first stage and the reduced form
What technique can you use if you have multiple instruments?
Two-stage least squares: takes optimally weighted combination of instruments
What is the ITT in an encouragement design?
Causal effect of encouragement
What can bias the ITT in an encouragement design?
- Randomization failure
- Spillover
- Attrition correlated with instrument
In an encouragement design, what can we still know if the exclusion restriction is violated?
ITT is still valid
What additional requirement (besides the usual IV requirements) do we have in an encouragement design?
Need successful randomization (no failure by chance, manipulation, or attrition)
What is the LATE? What analysis does it result from? Who does it apply to? What assumptions are required? Why does this matter?
- Local average treatment effect
- IV estimator
- Compliers
- Independence and monotonicity
- Whether the compliers are a population we care about determines whether we have an external validity problem
What is the IV assumption of independence?
The instrument is effectively randomly assigned – it’s not related to potential outcomes or potential take-up.
What is the IV assumption of monotonicity?
There are no “defiers”
Four ways to test the validity of an instrument
- Regress potential confounders on instrument
- Estimate IV causal effect with and without covariates
- Check whether attrition is correlated with instrument
- Falsification tests
Assumptions of regression discontinuity
- Correct functional form between running variable and outcome
- Potential outcomes continuous at threshold with no manipulation and nothing other than treatment happening at threshold (identification assumption)
What is the trade-off for a smaller bandwidth? What is an alternative approach?
- Reduces the importance of functional form assumptions so less likely to find a fake discontinuity
- People in region tend to be more similar
- Less precise estimates
- Weight obs by distance from cutoff
Validity/robustness checks for RD
- Check stability of estimates using different bandwidths
- Show that covariates are balanced around cutoff
- Look for bunching in running var near cutoff
- Estimate false cutoff
- Look for evidence along causal pathway
- Look for effect in groups that shouldn’t be affected
- Run model with and without covariates
Main assumption of DID
Treatment and control groups would have developed in same way in absence of treatment
Violations of DID assumptions
- Invalid comparator; groups had different trends in pre period
- Mean reversion
- Other programs targeted same groups at same time
- Migration causing overestimation of impact
- Spillover causing underestimation of impact
Tests to assess validity of DID
- Placebo tests
- Test for parallel pre-trends
- Control for other time-group varying factors
- Look for evidence of spillover
- Look for evidence of migration in response to program
Rationale for standard error corrections in DID
- Treatment is at group level so residuals will be correlated, biasing SEs towards zero
- With 30+ groups, can use basic clustering
- With fewer groups, can use bootstrapping or permutation tests
Main assumption of ITS
- Trend in pre period would have predicted trend in post period if not for intervention
Advantages of ITS
- Only comparing units to themselves; no cross-sectional confounding
- Confounding from slow-changing factors captured in time trend
Disadvantages of ITS
- Event of interest may coincide with other things
- Change in data collection at same time as policy change may be problematic
Assumption of CITS
Trend deviations in control areas would have been same as trend deviations in treatment areas
What should we remember when modeling counts for an ITS?
Possibility of over-dispersion in a Poisson model
What may happen if you have uncorrected autocorrelation in an ITS?
Bias in standard errors
