Econometrics - Review Sheet

` (Main Concepts)
May 2020
This list tries to be exhaustive but probably isn’t.
Final Exam is cumulative

1. Time Series Econometrics

(a) Started looking at stationary linear processes, ARMA(p,q).

(b) Introduced the definition of stationarity, autocovariance, lag operators,
(c) ACF useful for identifying MA Processes, PACF for AR Processes
(d) Introduced Unit roots processes, random walks, Order of integration,
spurious regressions
(e) More stationary models - Distributed Lag Models, Koyck Models, ARX(p),
ARDL(p,q) and dynamically complete models.
(f) Testing for serial correlation (Breusch-Godfrey Test) - What are the
consequences on OLS?
(g) What to do with OLS when error term is serially correlated - HAC
correction
(h) GLS procedures - Prais Winsten procedure
(i) Testing for unit roots - the DF and the ADF.
(j) Also tested for cointegrating equation where the variables are stationary
- the Engle Granger test which test for unit roots in the residual of the
cointegrating equation.

2. Panel Data Model

(a) The model yit = β0 + β1 x1it + . . . + βk xkit + αi + uit

(b) What are the assumptions behind a Pooled Regression, FE, FD and RE
estimator?
(c) A choice between i) E(αi | xit ) = 0 versus ii) E(αi | xit ) =
6 0
(d) If you assume i), use Pooled and RE. The RE is the GLS estimator.
(e) If you assume ii), use FE or FD. Bothe of these are more robust.
(f) Assumption ii) is less restrictive.
(g) How to you estimate the model using these estimators?

1
 (h) FE - demeaned regression.
(i) RE - Quasi-demeaned regression
(j) Hausman Test

3. Omitted Variable Bias: Suppose we omit x2 from the model y = β0 + β1 x1 +

β2 x2 + u, the bias on β̃1 is

cov(x1 , x2 )
E(β̃1 − β1 ) = β2
V ar(x1 )

Violates Assumption MLR 4.

(a) OLS is biased and inconsistent.

(b) Solution 1: find data on the omitted variables. Not always possible.
(c) Solution 2: find an instrument z which is correlated with x, but uncor-
related with all the omitted variables. Compute instrumental variables
(IV) estimate as βIV = cov(z, y)/cov(z, x). Multivariate version is 2SLS.
(d) You should know what is the first stage regression, structural equation
and the reduced form regression.
(e) The IV estimate is consistent, but not unbiased. Also using a large
number of instruments can increase the bias significantly (see Angrist
and Krueger (91)).
(f) Recall the plim of OLS and IV

σu corr(z, u) σu
plimβ̂OLS = β + corr(x, u) plimβ̂IV = β +
σx corr(z, x) σx

(g) Know the conditions that a valid IV need to satisfy.

(h) What is the First Stage regression, Reduced Form regression?
(i) Two Stage Least Squares (2SLS).
(j) How do you test for endogeneity? Idea behind Hausman test.
(k) Why is weak instrument a concern? When does an instrument become
weak?
(l) What is a Wald Estimator?
(m) Know the Angrist(89) - Vietnam Draft Paper, and Angrist and Krueger(91)
- Quarter of Birth IV Paper). What does a natural experiment buy you?

2
4. Policy Evaluation - Difference-in-Difference (DiD) Estimators.

(a) The model

yit = β0 + β1 T reatit + β2 P ostit + β3 T reatit P ostit + uit

(b) Minimal setup need i ∈ {T reatment, Control} and t ∈ {P re, P ost}

(c) What is the DiD estimator?
(d) What assumptions do you need for DiD to work?

5. Heteroscedasticity: ui no longer have a constant conditional variance (not the

residuals!). Violates Assumption 4. This can happen when the dependent
variable is binary .

(a) OLS is still unbiased and consistent.

(b) OLS is no longer BLUE (i.e. is inefficient).
(c) Usual formulas for standard error are not correct ⇒ confidence intervals
and tests can give misleading results.
(d) Solution: Do OLS, but report White’s corrected standard errors (robust
option in STATA). White’s correction DOES NOT make error term
homoskedastic. Know the White’s formula for bivariate regression.
(e) How do you test for heteroskedasticity? BP test
(f) Linear Probability Model - Note the heteroskedastic Standard Errors.

6. Regression with dummy variables - allows intercept and slope to vary by

qualitative variables).

(a) Make sure you know what the different dummy variable coefficients
capture.
(b) What is the Dummy Variable trap?

Adjusted R2 .

7. BELOW - MATERIAL COVERED BEFORE MIDTERM

3
 8. Started with a Review of Probability and Statistics. Covered Rules
of Probability, Rules of Expectation and Variances, How do you conduct a
Hypothesis Test, How to construct a Confidence Interval.
Appendix A, B, C.

9. Basic idea of this course: to quantify the relationships between variables (say
X and Y ). Its a good idea to keep some concrete example of X and Y at the
back of your mind e.g. education and wages, access to schools and length of
schooling, English language skills and wages, disability payments and labor
force participation, strength of national institutions and GNP per capita, firm
performance and CEO compensation, parents divorce and childrens lives

Bivariate Regression: Simplest model of relationships between Y and X

is the bivariate linear model

Y = α + βX + u

Population regression coefficients are:

Cov(X, Y )
β= ; α = E(Y ) − βE(X)
V (X)
Sample regression coefficients are:
Pn
(xi − x̄)yi
β̂ = Pi=1n 2
, α̂ = ȳ − β̂ x̄
i=1 (xi − x̄)

These are called OLS (ordinary least squares) estimates of α and β. OLS
estimates minimize the sum of squared errors ni=1 (yi − a − bxi )2 over all
P

possible values of (a, b).

- we also looked at the method of moment derivation of the bivariate OLS

estimator. That is, we have two moment condition for two unknowns: Eui =
0, and Eui Xi = 0. We take the sample analog of these two conditions, and
we have two equations and two unknowns.

- Know the bivariate model well - what happens when variables are scaled.
What does the estimator look like when we do not have an intercept.
Chapter 2 - everything.

- We generalized this to the multivariate context.

10. Multivariate Regression properties under Classical Assumptions:

4
 (a) α̂ and β̂ are unbiased for α and β. Relies on Assumption MLR1 to
MLR4
(b) Sampling variance (standard error2 ) of β̂ is σ 2 / ni=1 (xi − x̄)2 , in the
P

bivariate case and for β̂j in the multivariate case is

σ 2 /[ ni=1 (xji − x̄j )2 (1 − Rj2 )], These relies on Ass. MLR 1 to MLR 5.
P

(c) β̂ is the Best Linear Unbiased Estimator(BLUE) - Gauss Markov The-

orem . (Ass. MLR 1 to MLR 5)
(d) β̂ is normally distributed. Test statistic for H0 : β = β0 , is t = (β̂ −
β0 )/s.e.(β̂). This is distributed as tn−2 ( N (0, 1) for large samples). Be
sure about the asumptions needed for the t-statistic to be valid. (Ass.
1 to 5)
(e) How do you construct confidence interval, test multiple restrictions, test
a linear restriction on parameters (eg. β1 − β5 = 0)? All these requires
Ass. 5.
(f) When do you have to use a t-test ? When is an F test appropriate?
(g) Functional form - how do you interpret the slope coefficients in log-
linear, log-log or linear-log models? Understand the interpretation of
coefficients estimates especially when you have log transformations.
(h) Properties of OLS residuals: ni=1 ûi = 0, ni=1 ûi xi = 0, and ni=1 ûi ŷi =
P P P

0. Where do these come from? The last equation allows us variance de-
composition, V ar(yi ) = V ar(ŷi ) + V ar(ûi ) = SSR + SSE. What is the
R2 ?
Chapter 3 - ignore part of 3.2 on ‘A ”Partialling Out” Interpretation
of Multiple Regression’ (page 78 of 5th edition). Everything else is fair
game.
Chapter 4 - everything

11. Know your STATA regression output.

12. (a) Asymptotics: With the Gauss-Markov Assumptions 1 to 5 and n → ∞,

β̂j −βj
we get asymptotic normality, i.e. s.e.(β̂ )
∼ N (0, 1).
j

(b) What does consistency mean?

(c) Chapter 5 - Section 5.1 only, on definitions of consistency