lecture5_140b_2011

# lecture5_140b_2011 - Lecture 5: Evaluating Regression...

This preview shows pages 1–7. Sign up to view the full content.

Olivier Deschenes, UCSB, Econ 140B, Winter 2011 Lecture 5: Evaluating Regression Studies 3 reasons to regress Causal relationships Observational studies, controlled randomized experiment, natural experiment Threats to internal and external validity Chapter 9 in S&W

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Olivier Deschenes, UCSB, Econ 140B, Winter 2011 Three reasons to regress: 1. Summarize the data (i.e. calculate M/F wage differential) 2. Forecast future values of Y (i.e. predict Urate in 2012) 3. Predict the impact of interventions / policy changes There is nothing special about (1) and (2). Just regress, fit the data as best as possible However (3) is about causal inference: what is the effect of an exogenous change in X on Y The question is when does a regression estimate a causal effect? Also when does it not?
Olivier Deschenes, UCSB, Econ 140B, Winter 2011 Causal relationships „Causality‟ is a matter of deep philosophical questioning For our purposes, we say that an action (T) causes an outcome (Y), if Y is the direct result of that action We will call T the treatment and Y the outcome Examples of causation: Pressing the “on” button on the remote causes the TV starts Putting your hand in a fire causes your hand to get burned Counter-examples (correlation causation): Walking with an umbrella does not cause rain to fall (but it is correlated with rainfall)

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Olivier Deschenes, UCSB, Econ 140B, Winter 2011 Simple example: Y i = 0 + 1 T i + u i T i = 1(observation i receives treatment) 1 = (average) causal effect of T of Y = E[Y i |T i =1]- E[Y i |T i =0] Can we estimate the causal effect of T on Y from data? Key issue is whether the assumption E[u i |T i ]=0 is satisfied or some sort of exogeneity assumption Earlier, we simply assumed it. From now on we will always discuss if this assumption is credible
Olivier Deschenes, UCSB, Econ 140B, Winter 2011 Causal inferences are made from: Observational studies Natural experiments Controlled experiments In economics, we almost exclusively deal with observational data With observational (non-experimental) data, a key issue is confounding (or omitted variables bias) Confounding arise if u i is correlated with T i , even after we condition on other variables (X i )

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Olivier Deschenes, UCSB, Econ 140B, Winter 2011 Controlled randomized experiment In a controlled experiment, treatment assignment is random (ex: coin toss) Up to some random error, the coin toss balances the two groups (treatment group T=1, control group T=0) with respect to all factors except the treatment This makes the assumption E[u i |T i ]=0 highly credible since T i is randomly assigned, and thus in principle independent of u i Causal inferences based on controlled experiments are the most credible However, in social sciences, controlled experiments are rare (because unethical, unpractical, expensive, etc)
This is the end of the preview. Sign up to access the rest of the document.

## This note was uploaded on 09/04/2011 for the course ECON 140b taught by Professor Staff during the Winter '08 term at UCSB.

### Page1 / 40

lecture5_140b_2011 - Lecture 5: Evaluating Regression...

This preview shows document pages 1 - 7. Sign up to view the full document.

View Full Document
Ask a homework question - tutors are online