Empirical Research in Business Part II

1
Empirical Research in Business Part II

• N. Meltem Daysal
• Tilburg University
• Fall 2009
• Lecture 7

2
Why study experiments?
3
Terminology experiments and quasi-experiments
4
Different types of experiments three examples
5
Experiments a brief outline
6
Potential Problems with Experiments

• Internal validity statistical inferences about
  causal effects are valid for the population being
  studied
• Failure to randomize choose T and C by last
  name
• Failure to follow protocol receive T when in C
  (and vice versa)
• Attrition problematic if people leave in a way
  correlated with T
• Experimental effects (Hawthorne effect) double
  blinding??
• Small samples not a problem for consistency but
  precision
• External validity inferences and conclusions can
  be generalized from the population and setting
  studied to other populations and settings
• Nonrepresentative sample certain groups,
  location etc
• Nonrepresentative program or policy scale and
  duration
• General equilibrium effects job
  training/employer provided training
• Treatment vs. eligibility effects ATE or ATET

7
Differences Estimator
8
Differences Estimator
9
Differences Estimator
10
Experiments

• True randomized controlled experiments can be
  expensive and they often raise ethical concerns
  (Homework Project STAR)
• Can we recover causal estimates using
  observational data?
• Identification strategy manner in which a
  researcher uses observational data to approximate
  a real experiment

11
Panel Data Models
12
Notation for panel data
13
Notation for panel data
14
Why are panel data useful?
15
Example Traffic deaths and alcohol taxes
16
U.S. traffic death data for 1982
17
U.S. traffic death data for 1988
18
Is the impact causal?
19
Panel Data with Two Time Periods
20
(No Transcript)
21
(No Transcript)
22
?FatalityRate v. ?BeerTax
23
Fixed Effects Regression
24
(No Transcript)
25
(No Transcript)
26
The regression lines for each state
27
(No Transcript)
28
Summary Two ways to write the FE model
29
Fixed Effects Regression Estimation
30
Method 1 n-1 binary regressors OLS regression
31
Method 2 Entity-demeaned OLS regression
32
Method 2 Entity-demeaned OLS regression
33
Method 2 Entity-demeaned OLS regression
34
Example Traffic deaths and beer taxes in STATA
35
Time Fixed Effects
36
Time fixed effects only
37
Two formulations for time fixed effects
38
Time fixed effects estimation methods
39
Combined Entity and Time Effects

• Dummy variable form
• Yit ?0 ?1Xit ?2D2i ?nDni ?2B2t
  ?TBTt uit
• There are n-1 entity binary indicators, T-1 time
  binary indicators, an intercept and other
  covariates (in this example only 1)
• Fixed effects model form
• Yit ?1Xit ai ?t uit

40
(No Transcript)
41
The Fixed Effects Regression Assumptions
42
Extension of LS Assumptions to Panel Data
43
Assumption 1 E(uitXi1,,XiT,?i) 0
44
Assumption 2 (Xi1,,XiT,Yi1,,YiT), i 1,,n,
are i.i.d.
45
Assumption 3 corr(uit,uisXit,Xis,?i) 0 for t
? s
46
Assumption 3 in a picture
47
What if Assumption 3 fails so
corr(uit,uisXit,Xis,?i) ?0?
48
Application Drunk Driving Laws and Traffic Deaths
49
Application Drunk Driving Laws and Traffic Deaths
50
Application Drunk Driving Laws and Traffic Deaths
51
Application Drunk Driving Laws and Traffic Deaths
52
Application Drunk Driving Laws and Traffic Deaths
53
Application Drunk Driving Laws and Traffic Deaths
54
Summary
55
Summary
56
Summary

• Limitations/challenges
• Need variation in X over time within states
• If there is measurement error in X, then
  attenuation bias is larger than cross sectional
  models very important
• Demeaning eliminates both good and bad variation
  so use caution when interpreting results as
  causal very important
• You should use heteroskedasticity- and
  autocorrelation-consistent standard errors if you
  think uit could be correlated over time