Papers

Question 1

Arellano and Bond (1991).

Answer 1

One could also use additional instruments where there are available according to the GMM estimator

Using further lagged variables can make use of exogenous variation in ∆x1it to be extracted and gives more efficient estimators

Question 2

Anderson and Hsiao (1981)

Answer 2

First Differenced IV estimator in the

case of lagged dependent variables

Question 3

Train (2009)

Answer 3

Discrete choice models based on utility maximisation behaviour can also be used to represent other forms of decision making.

Question 4

Cameron and Trivedi (2005)

Answer 4

Conditional log likilihood = Log likelihood involve ignoring the marginal likelihood

No issue under no endogeneity

Better to use sample average of ME, not ME at sample average of regressors.
- Particularly for arguing Logit = Probit…not very different empirically, only at tails of distribution.

Question 5

Harrison (2011)

Answer 5

OLS in LDV is not an issue, logit and count models are the problem.

Question 6

Angrist and Pishke (2009)

Answer 6

LDV ness less important, OLS is okay to use and estimate AME.

Question 7

Cunha, Heckman and Navarro (2007)

Answer 7

Several classes of models including dynamic schooling choice - monotone step function

Question 8

McFadden (1974)

Answer 8

Models of discrete unordered choices applied to San Fran. BART (transport) choices

Question 9

Heckman (1976)

Answer 9

Heckit 2 step procedure to model sample selection issues. Can also estimate via MLE if we are willing to make full distributional assumptions: robust vs. efficient

“Control function approach”

Consistent estimators, inconsistent SEs –> due to heteroskedasticity and IMR is a generated regressor –> adjust SEs

Question 10

Staiger and Stock (1997)

Answer 10

Rule of thumb: F-test > 10 in the “SLS first stage equation

H0: All instruments uninformative

Question 11

Stock and Yogo (2005)

Answer 11

Alternative critical values for the F test controlling for bias in 2SLS

Question 12

Sargan (1958)

Answer 12

Over ID restrictions in 2SLS

Question 13

White (1980)

Answer 13

White’s heteroskedasticity robust SE calculations: we can replace ui with sample residuals from a ocnsistent estimator under reasonable assumptions
“Passive response to heteroskedasticity”

White’s test for heteroskedasticity.
- regress squared residuals on regressors, squares and corss products

Question 14

Ramsey (1969)

Breusch-Pagan (1979)

Answer 14

Other tests for heteroskedasticity

Question 15

Johansen (1996)

Answer 15

Granger Johansen representation of the long run matrix in cointegration

Question 16

Amemiya (1985)

Answer 16

ML efficiency counterexample

Question 17

Casella and Berger (2002)

Answer 17

MLE Invariance property

Question 18

Harbo et al. (1998)

Answer 18

If tests on beta and alpha in cointegration analysis alter the rank of pi then the test statistics follow non-standard inference.

Question 19

Hansen and Johansen (1999)

Answer 19

Battery of tests of parameter non-constancy

• Full model tests comparing likeilhood values
• Cointegrating relation constancy tests based on eigenvalues

Useful for identifying structural breaks.

Question 20

Johansen (1996)

Answer 20

Test procedures on the long ru cointegration matrix of pi, tests on beta and alpha.

Question 21

Hansen (1982)

Answer 21

GMM Over ID restrictions

Question 22

Pesaran and Smith (1995)

Answer 22

Mean groups estimator