Gruber 3: Empirical Tools of Public Finance. What do the data tell us is true?
I. Correlation & Causality
i. Gold standard: Randomized trials, with (treatment group, control group)
1. Difficult for most economics questions
2. Much work lately like this in Development. The advantages of having really poor subjects
II. Types of “observational” data, not generated by controlled experiments: Cross sectional, time series, panel.
III. Problem of “bias”
i. Caused by differences between the control and treatment groups that is not due to the treatment.
ii. As such, estimated regression coefficients are “biased” in that they don’t represent the true treatment effect but are instead picking up something from the differences between the two populations.
iii. This is usually the result of some factor that caused individuals to choose to be in one group of the other.
iv. This effect is hopefully avoided by randomized trials.
v. Examples:
1. Ph.D. student completion rate study, TA versus RA.
2. Krueger piece on Job Corps, and attrition bias.
IV.Avoiding bias
i. Structural versus reduced form modeling
1. Structural: Estimating individual behavior, parameters of individual utility functions (discount rates, risk aversion, marginal utility, elasticity) or parameters of cost functions.
2. Reduced form: Market behavior, for example slopes and shift effects of supply and demand curves
ii. Quasi-experiments: “natural” experiments,
1. Events that that hit one of two similar groups
2. Instrumental variables (IV),
3. Differences within differences” estimators, allows fixed effects to drop out