Empirical Tools Flashcards
Identification problem
Correlation (move together) vs causation (movement causes movement of another)
How to solve?
B) issue
Randomised trial - have a treatment vs control group
B) bias - how do we ensure both groups are the same WITHOUT the treatment
How to treat bias:
Larger sample sizes can eliminate consistent differences via law of large numbers: (odds fall as sample size grows)
Example of randomised trial going wrong
3 different class arrangements
Regular size, single teacher
Regular size, teacher+assistant
Small size, single teacher
Result:
Further problems with randomised trials (3)
Expensive
Time consuming
Unethical….
2 technical issues with randomised trials (even with gold standard)
External validity: only valid for sample tested, may be different from population at large
Attrition: individuals may leave before experiment complete, if not random can create bias estimates
Problems with time series analysis (comparing 2 variables movement overtime)
Does not mean causation, they may just move together
Other factors get in the way to test causation since they may be also correlated with the variables of interest
So when is time series useful
If there is a sharp break in the data, more likely to suggest causation
E.g in 1993 simultaneous changes in price of cigarettes and the fall in smoking rates
Cross-sectional regression analysis
Analysis between 2 or more variables, exhibited by many individuals at ONE point in time
Problems with cross-sectional regression analysis
Reverse causation - does X cause Y or Y cause X
Control variables
Account for differences between treatment and control groups that can lead to bias
Quasi-experiments (natural) uses what
Difference in difference estimator
As if we only look at the effect of treatmetn and control after treatment, may not be representative as treatment and control may not be comparable before the treatment (e.g before one city given benefits)
DID estimator
(Y tafter - Y cafter) - (Y tbefore - Y cbefore)
See if the difference between treatment and control has changed since in the policy change
Quasi-experiment problems
Never can be sure control variable has got rid off all bias
External validity
How do quasi-experiments try to mitigate this
Robustness checks e.g find alternative control groups, do a placebo comparing treatment
Best way to check validity of DD estimator
Plot time series to see if there is a clear break between the 2 groups
