Prof. Bryan Caplan

http://www.gmu.edu/departments/economics/bcaplan

Econ 345

Fall, 1998

Week 10: Endogenous vs. Exogenous Variables

1. Endogenous vs. Exogenous
1. Statistics by itself just gives correlation, but usually causation is what is interesting.
1. Exception: Forecasting.
2. There are many techniques you can try to see if a relationship is causal or merely correlation.
1. Controlling for other variables.
2. Adding trends to the list of regressors.
3. Double-checking your results using multiple data sets.
3. However, all of these techniques are imperfect. To get truly clean results, you need double-blind controlled experiments.
4. When your data come from double-blind, controlled experiments, your independent variables will be "exogenous." Literally, this means that the independent variables are determined "outside the system" - i.e., by you, the researcher.
5. On the other hand, when your data do not come from double-blind controlled experiments, you independent variables may be "endogenous." Literally, this means that the independent variables are determined "inside the system." This means that the "dependent" variable may in fact be causing the "independent" variables, and not vice versa.
6. However, data not derived from double-blind controlled experiments don't have to be endogenous; the danger is merely that they could be. Even if you don't set up a double-blind controlled experiment yourself, there is the possibility that historical data supplies you with a natural experiment, or something close to a natural experiment.
7. Some examples of natural experiments:
1. The Great Contraction
2. German hyperinflation
3. Twin studies
2. Illustration of the Problem
1. Disease and treatment
2. Anticipated inflation and monetary policy.
3. War and economic growth
4. Elections and economic growth
5. Mathematical example #1: suppose that inflation=money supply growth -2% + N(0,2); suppose further that the Fed sets money supply growth so that inflation=2%. Then you observe no correlation between money and inflation, even though there is a direct causal link!
6. Mathematical example #2: suppose that disease fatality rate = 5%-treatment dosage+5%*severity. Severity ranges from 0 (not sick) to 5 (most sick). Suppose further that treatment dosage=2.5%*severity. Then it appears that disease fatality rate=5%+treatment dosage!
3. Picking Exogenous Variables
1. The best source: go to the historical record.
2. Non-human (e.g. weather)
1. Famines under Communism
3. Human (e.g. shifts in ideology)
1. The Great Contraction
2. Hyperinflations
3. Fed announcements and the Romer dummies
4. In general: the more latitude/discretion an agent has, the more you can learn about the causation.
4. Resolving the Problem of Correlation vs. Causation
1. First best: Do double-blind controlled experiment. (Sometimes actually done: experimental economics is a growing field).
2. Second best: Find approximate natural experiment.
3. Third best: Control for any plausible omitted variables and/or add trend variables.
4. Fourth best: Present with a warning.
5. A big plus: Have a good theory to start with.

Next 4 weeks: Applications.