Econometrics - MCQ Flashcards - Quizlet

5/18/2021 Econometrics - MCQ Flashcards | Quizlet
Home / Social Science / Economics / Econometrics
Econometrics - MCQ
Terms in this set (92)
- Discrete: sum of x*f(x)

Expected Value - Continuous: integral of xf(x)dx
- E(x) = μ
Variance Var(X) = E(x^2) - [E(x)]^2 = E((x-μ)^2)
Covariance Cov(X,Y) = E(XY) - E(X)*E(Y) - E[(X-μx)(Y-μy)]
Corr(X,Y) = Cov(X,Y) / [sd(x)*sd(y)]

Correlation
-1 ≤ Corr(X,Y) ≤ +1
X's aren't constant nor perfect linear combos of

No Perfect Collinearity (TS.2)
each other
- E(u|X) = 0
Zero Conditional Mean (TS.3)
- Weak Exog. vs. Strict Exog.
- Contemporaneous Exogeneity
- Errors are unrelated to X of the same period
Weak Exogeneity
--> beta hats are consistent
Econometrics - MCQ
https://quizlet.com/199464364/econometrics-mcq-flash-cards/ 1/19
Errors are unrelated to X's of all periods
Strict Exogeneity
--> beta hats are unbiased regardless of S.C. or
Heteroskedasticity (BLUE)
Var (u | X) = Var(u) = σ^2
Homoskedasticity (TS.4)
Variance of errors are uncorrelated with the X's
AND are constant over time
- Corr(u1, u2 | X) = 0
No Serial Correlation (TS.5)
- Errors are uncorrelated with each other
u ~ N(o, σ^2)
Normality (TS.6)
Usual t and F-tests are valid
"Best Linear Unbiased Estimator"

BLUE
(Smallest variance and small distribution)
(1) Mean = constant

Covariance Stationary (2) Variance = constant
* Cov(Xt, Xt+h) depends on only h, NOT t
(1) Mean = constant

(2) Variance = constant
Stationary (3) Covariance = constant --> Corr = constant
* If it is stationary, it is also covariance stationary.
- Corr(Xt, Xt+h) --> 0 as h --> infinity

Weakly Dependent - Asymptotically uncorrelated... similar to "no serial
correlation" but for the X's, not the u's
Econometrics - MCQ
MA(1): Xt = Et + α1*Et-1
Moving Average Process * Depends on PAST ERRORS

- Covariance Stationary
AR(1): Yt = ρ1*Yt-1 + Et
Autoregressive Process * Depends on PAST VALUES (i.e. lags)

- Corr(Yt, Yt+h) = ρ1^h
- If ρ1 < 1 --> STABLE and Covariance Stationary
Cov(u, X) = 0
Unbiasedness - Sampling distribution centered around true value

** Need strict exogeneity
** Does not imply consistency!!
As n --> infinity, distribution piles up onto true value

(kind of like asymptotic unbiasedness)
Consistency
** Need weak exogeneity
Need TS.4 and TS.5 --> Homoskedasticity & No

Serial Correlation:
Var(u | X) = constant, Cov(u1, u2) = 0
Efficiency
- Small variance
- Narrow on graph
- Ratio between the errors
Only past returns are considered as relevant info to

Efficient Markets Hypothesis
predict returns in week t
Econometrics - MCQ
(1) Stationary around trend

(2) Weakly Dependent
Trend Stationary
--> Satisfies TS.1' (linearity & weak dependence)
- Unit root, I(1)

- ρ1 = 1
Highly Persistent - "strongly dependent"
- Makes OLS methods invalid --> Use Order of
Integration Method (Take 1st difference)
First Differences Method

Order of Integration Method
ΔY = Yt - Y(t-1)
- Problematic Case of AR(1)

- Highly persistent (not weakly dependent)
Random Walk - Value today = sum of all past shocks + initial value -
-> Effect of shock contained forever in the series
- Not Covariance Stationary
Deterministic increase/decrease in each period

(similar to time trend for random walk)
Yt = α0 + Yt-1 + Et
Drift
- Not Covariance Stationary

- Not weakly dependent
--> Take First Differences
Econometrics - MCQ
Corr (u1, u2 | X) ≠ 0
Errors are correlated with each other
- Unbiased and consistent BUT inefficient

Serial Correlation - Correctness of R^2 is independent of S.C.
- Invalid SE's and tests
Ex: if error is positive in period 1, it will be positive in

period 2
TEST for Serial Correlation in AR(1) with strict

exogeneity and small samples
DW = 2(1 - ρhat)
Durbin-Watson
- Works with upper and lower bounds --> area in
between is inconclusive
- Test Ho: ρ = 0 against Ha: ρ > 0
- Reject if DW < dL
- Fail to reject if DW > dU
TEST for serial correlation in AR(q) without strict

exogeneity
Breusch-Godfrey
Ut = α0 + α1Xt1 + ... + αkXtk + ρ1Ut-1 + ... + ρqUt-q + ...
Test Ho: ρ1 = ... = ρq = 0, (F-test)
Econometrics - MCQ
"Feasible Generalized Least Squares"
- CORRECTION for serial correlation identified by

Durbin Watson
- Used when ρ is unknown
- Need: (1) strict exogeneity --> consistency and (2)
WLS/FGLS
Cov(Xt, Ut) = 0 and Cov(Xt-1 + Xt+1, Ut) = 0
- Asymptotically efficient (More efficient than OLS)

- Takes into account that variance of errors is not
constant over time
- Consistent and unbiased asymptotically
- Used in FGLS to estimate ρ with iterative scheme

Cochrane-Orcutt
- Omits 1st observation
- Used in FGLS to estimate ρ with iterative scheme

Prais-Winsten - Adds a transformed 1st observation
* Use in small samples for efficiency
- CORRECTION factor used in Serial Correlation

Robust Errors (w/o strict exogeneity) identified by
Breusch-Godfrey
Newey-West
- integer 'g' = how much serial correlation we allow
(g = 2,3)
- higher g means more serial correlation allowed
* Only fixes SE - NOT efficiency
Var(u) ≠ 0 (not constant)
Heteroskedasticity
- automatically corrected with serial correlation
robust SE (Newey West)
Econometrics - MCQ
- TEST for Heteroskedasticity

- Assumes absence of serial correlation --> so test
for S.C. first
Breusch-Pagan
Ut^2 = δo + δ1Xt1 + ... + δkXtk + Vt
Test Ho: δ1 = δ2 = ... = δk = 0, (F-test)
- Serial correlation in variance of errors

- Consistent but inefficient
- Static models (no lagged dependent vars) -->

Autoregressive Conditional BLUE
Heteroskedasticity (ARCH) - Dynamic models --> not BLUE
Can...
(1) Asymptotically improve efficiency
(2) Forecast trends in variance (i.e. stocks)
- TEST for Unit Root/High Persistence/I(1)

* Can use for cointegration tests, too.
Yt = α + ρYt-1 + Et
Dickey-Fuller
Ho: ρ = 1, Ha: ρ < 1
ΔYt = α + θYt-1 + Et
Ho: θ = 0, Ha: θ <0

Augmented Dickey-Fuller - Includes lagged differences --> allows for more

serial correlation
ΔYt = α + θYt-1 + γ1ΔYt-1 + ... + γpΔYt-p + Et

Econometrics - MCQ

Dickey-Fuller with Time Trend ΔYt = α + δt + θYt-1 + γ1ΔYt-1 + ... + γpΔYt-p + Et
Ho: I(1), Ha: trend-stationary
- Regression of I(1) series on another I(1) series

- Produces very high R^2 and t-stats and very
Spurious Regression misleading results
- Correct using Cointegration test (DF)
- When two time series are individually I(1) but I(0)

together (i.e. stable together)
- Models eqbm between non-stationary variables
Cointegration
- Short-term dynamics are related
Yt = α + Xt + Et, Et ~ I(0)
- TEST for Cointegration when parameters are

Engle-Granger unknown
(Estimate parameters then use DF)
Error Correction Model ΔYt = α + γΔXt + δ[Yt-1 - β*Xt-1] + Ut
- Predicting future outcomes using available info

Forecasting
- in-sample criteria vs. out-of-sample criteria
- Measuring model up to and including t to forecast

In-Sample Criteria
past and today
Econometrics - MCQ
- Measuring model up to and including t to forecast

the future
Out-of-Sample Criteria
(Better than in-sample criteria)
Vector Autoregressive Models - Collection of TS as linear functions of their own

(VAR) past values and of past values of other TS
- Correlation in VAR models
Granger Causality
- used for simultaneous TS of systems of variables
that are cointegrated
Ek+1 = Yk+1 - fn
SE[Ek+1] = {[SE(fn)]^2 + σ^2}^0.5

Forecast Error
- SE(fn) = how uncertain forecast is

- σ^2 = shock in future period (unpredictable)
(1) Same population

(2) Different points in time
- larger n => more predictive power than single

Pooled Cross Sections
cross section
- allows for structural change
- Beware of omitted variable bias/changing
distribution
TEST for structural change over time
Chow Test
F = (SSRu - SSRr)/SSRu [n-2(k+1)]/k
k = # of restrictions
Econometrics - MCQ
Omitted Variable (error) is correlated with X's
- Overestimates effect go X1 (β) on Y

Omitted Variable
Bias/Endogeneity
* Test using Difference-in-Differences
* Correct with (1) proxies or (2) IV
"Average Treatment Effect"

- TEST for omitted variable bias/endogeneity
- treatment = dummy variable

Difference-in-Differences - quasi-experimental setting (Treatment applied as if
randomly assigned)
- Must satisfy parallel assumption => Does not

account for changing relationship effect
(1) Same population AND sample (same exact

Panel Data people)
(2) Different points in time
- Use with panel data and random walks

- Can help specifically with time-variant omitted
variables
- Takes First Difference!
Fixed Effects Model
Yit = βo + δod2t + β1Xit + ai + uit
ai = fixed effect
uit = time-varying error
vit = ai + uit = composite error
Econometrics - MCQ
1. Linear in Parameters
2. RANDOM SAMPLING
3. No perfect collinearity
Fixed Effects Model/First 4. WEAK exogeneity --> ai can be core. with xit or ut
Difference Assumptions 5. Homoskedasticity
6. No serial correlation
7. Normality
* Alternative in presence of ait

- Use if serially correlated
- Use if X varies over time and E(ai | Xi) ≠0
- Uses WITHIN estimation

- Pooled OLS on de-meaned variables
Fixed Effects ESTIMATION
- Equivalent to dummy variable regression except
with lower R^2 here
df = N(T-1)K
Need to correct SEs and t-tests by multiplying SEs
by sqrt[ (NT-K) / N(T-1)K]
- Relies on variation within individuals (FE)
Within Estimator
- FE heterogeneity doesn't bias 'within' estimator but
cannot be used to estimate FE over time
- Relies on variation between individuals (OLS)
Between Estimator - Not the most efficient

- Biased if ai is correlated with avg. x --> Use RE if
this is the case
Econometrics - MCQ
1. Linearity
2. RANDOM SAMPLING
3. No perfect Collinearity (no constant X's)
4. STRICT exogeneity
5. Homoskedasticity
6. No serial correlation
FE Estimation Assumptions
* 1-4 = unbiased and consistent
* 1-6 = BLUE
* If 5 or 6 doesn't hold, fix SEs (they will be
understated) --> also may need to test for clustering
SEs
- Potential selective attrition

Unbalanced Panel
- FE valid asymptotically (as t increases)
- Combination of 'between' and 'within'
- Use with representative sample (not whole

population)
- Quasi-demeaned data with θ multiplier
If θ=0: RE = pooled OLS
Random Effects
If θ=1: RE = FE
- Use if X is constant over time

- Use if E(ai | Xi) = 0 (If ai is uncorrelated with all X's
for all T, then it is inefficient to eliminate it)
- Vit can be correlated over T
Econometrics - MCQ
- TEST for Exogeneity (of IVs) and to tell whether to

use RE or FE
Hausman Test Ho: Corr(ai, Xit) = 0

- If reject Ho, use FE (endogenous)
- If you don't reject, use RE (exogenous, betas are
similar and valid --> use both in 2nd stage of 2SLS)
* Alternative to FE, allowing ai to be correlated with

uit
- Used to include time-invariant X's
Correlated Random Effects - Think of ai's as random variables

(CRE) - Plug ai into Yit eqn --> RE model is now
uncorrelated with ri and the time averages (x bar)
are now additional regressors
- βCRE = βFE
- Used to CORRECT Endogeneity (2nd to proxies)
Requirements:
(1) Instrument exogeneity: Cov(z, u) = 0
(2) Instrument relevance: Cov(z, x) ≠ 0
- SEs will always be larger than OLS (biased BUT

Instrumental Variable
consistent)
-Reduces variation in X's --> less efficient than OLS
- IV estimate should be < OLS estimate (if not, use
OLS)
- Valid IV should not explain y other than through x
βOLS = βIV if Corr(x, u) = 0 and Corr(z, u) = 0
Econometrics - MCQ
(1) Just barely correlated with x

(2) Really large SEs --> wide CI
(3) βIV is not well approximated by normal
Weak IVs
distribution
Exactly Identified M=K
Over-identified M>K
Under-identified M<K
* Use if M>K
1. reg x on z => obtain x hat

Two Stage Least Squares 2. reg y on x hat => obtain β2sls
(2SLS) 3. correct SEs (make smaller)
* Run both regs in stage 1 SEPARATELY and bring

together to solve for βs in stage 2
Structural Equation Contains parameter of interest (β)
Reduced-form Equation Equation with x's and z's (1st stage of 2SLS)
Econometrics - MCQ
TEST for Exogeneity (alternative to Hausman Test)
- M>K
Test of Over-identifying
1. 2SLS --> obtain u hat
Restrictions
2. reg u hat on all non-problemative regressors
(exogenous X's and IVs) --> obtain R squared
3. Low R^2: exogenous, High R^2: endogenous
nR^2 ~ Chi-squared(q), q = extra restrictions = M - K
- Heterogeneous effect of instrument

- Correlated with endogenous variable but not
whole population
Local Average Treatment Effect
- Uses weights to determine how much IV (z)

influences explanatory variable (x)
- binary/dummy
- non-negative
Limited Dependent Variables - integers
- count/censored
- lots of zeros
E(y) = P(y=1)
E(y | x) = P(y=1 | x) = p(x)
ΔP(y=1 | x) = β*Δx
Linear Probability Model (LPM)

- Computationally simple and easy to interpret
(percentage points)
- Predicts outside 0-1 range, constant partial effects,
and heteroskedastic
Econometrics - MCQ
- Cummulative distribution functions used to

observe latent variables (i.e. score computed to
take a loan)
- Estimated using MLE
Logit & Probit Models - Similar distributions (logic and pro bit)
- Neither restrict marginal effects to be constant
y* = βo + xβ + e, e ~ N(0,1)
y = 1[y>0] => y = 1 if y>0, and y = 0 if y*≤0
- cdf of normal distribution (assumes normality)
Probit Model
P(y=1 | x) = Φ(z) = ∫φ(v)dv = ∫(1/(2π))e^(-z^2/2)*dz =
P(z≤xβ) = P(e > -(βo + xβ) | x)
- cdf of logistic distribution

- easier to compute than Probit
Logit Model
f(xβ) = (e^z) / (1+e^z)
- Most likely to reproduce the data

- Used to estimate logit, probit, and tobit models
- Consistent and asymptotically normal and efficient
Maximum Likelihood Estimator

1. Likelihood function: L = Π p^x * (1-p)^(1-x)
(MLE)
2. Max L
(a) take log
(b) differentiate with respect to p
3. βo, ..., βk = argmax L(β)
nonlinear, depend on x
Partial Effects
∂P(y = 1 | x) / ∂x = g(xβ)*β
Econometrics - MCQ
- Used for corner solutions (i.e. prices)

- Estimate using MLE
- Only use if the effects of x on y are the same for
Tobit Model
the change in y (i.e. if x predicts y to be positive, it
will also predict y to increase)
- Scientific procedure undertaken to make a

discovery, test a hypothesis, or demonstrate a
Experiments known fact
- Treatment, control, random assignment
- Outcome for an individual under a given treatment
Potential Outcome
- We never observe both outcomes (hence
potential)
- Difference between outcome with and without

treatment
Treatment Effect
- Cannot be observed!
- Average effect of the treatment in the population

- We would like to observe E(Y1-Yo) but cannot
observe Y1 and Yo simultaneously
Average Treatment Effect Under Random Assignment: E(Y) = E(Y1 | X=1) - E(Yo |
X=0)
- Observe using the difference estimator (i.e.

dummy var)
Econometrics - MCQ
(1) Threats to internal validity

- failure to randomize
- failure to follow the treatment
- attrition
Issues with Experiments - experimental effects
- small powers
- non-representativeness of sample
- non-representative program or policy
- general eqbm effects
(1) Threats to Internal validity

-failure of randomization
-failure to follow the the treatment
-attrition
Issues with Quasi-
-experimental effects
Experiements
-instrument validity in quasi-experiments
(2)Threats to external validity

-same as usual regression analysis
- Randomness introduced as 'in' an experiment by

other elements (such as unexpected changes in
legal institutions, location, timing of policies) and
natural randomness (birth dates, rainfall)
- Researcher does not control randomization

Quasi-Experiments
- Ex: difference-in-differences estimations
Heterogenous population => treatments may have

different causal effects for different types of
individuals
Econometrics - MCQ
- Treatment assigned based on a variable being

above or below a threshold
(i.e. acceptance to uni based on grades)
Regression Discontinuity
- Treatment may depend fully (sharp discontinuity)
Estimator
or partially (fuzzy discontinuity) on being above or
below the cut-off
- If fuzzy discontinuity, use as instrument variable.
TS.1' : Linearity AND WEAK DEPENDENCE

TS.2' : No perfect collinearity
TS.3' : Zero Conditional Mean - WEAK EXOGENEITY
TS.4' : Homoskedasticity
TS.5' : No serial correlation
Asymptotic Properties of OLS
(TS' Assumptions)
- Under Ts.1' - TS.3' we have consistency but
baseness
- Under TS.1' - TS.5' we have asymptotic normality --
> usual OLS SEs, t-stats and F-states are
asymptotically valid
1. Linear Parameters
2. No perfect collinearity
3. Zero conditional Mean - STRICT exogeneity
4. Homoskedasticity
Classical OLS Assumptions 5. No serial Correlation
6. Normality
- Under 1-3: unbiasedness

- Under 1-5 (Guass Markov): efficiency

Econometrics - MCQ Flashcards - Quizlet

Uploaded by

Copyright:

Available Formats

Econometrics - MCQ Flashcards - Quizlet

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Econometrics - MCQ Flashcards - Quizlet

Uploaded by

Copyright:

Available Formats

5/18/2021 Econometrics - MCQ Flashcards | Quizlet

Home / Social Science / Economics / Econometrics

- Discrete: sum of x*f(x)

Variance Var(X) = E(x^2) - [E(x)]^2 = E((x-μ)^2)

Covariance Cov(X,Y) = E(XY) - E(X)*E(Y) - E[(X-μx)(Y-μy)]

Corr(X,Y) = Cov(X,Y) / [sd(x)*sd(y)]

X's aren't constant nor perfect linear combos of

--> beta hats are consistent

Errors are unrelated to X's of all periods

Var (u | X) = Var(u) = σ^2

"Best Linear Unbiased Estimator"

(1) Mean = constant

(1) Mean = constant

* If it is stationary, it is also covariance stationary.

- Corr(Xt, Xt+h) --> 0 as h --> infinity

Moving Average Process * Depends on PAST ERRORS

Autoregressive Process * Depends on PAST VALUES (i.e. lags)

Unbiasedness - Sampling distribution centered around true value

As n --> infinity, distribution piles up onto true value

** Need weak exogeneity

Need TS.4 and TS.5 --> Homoskedasticity & No

Only past returns are considered as relevant info to

(1) Stationary around trend

- Unit root, I(1)

First Differences Method

- Problematic Case of AR(1)

Deterministic increase/decrease in each period

- Not Covariance Stationary

- Unbiased and consistent BUT inefficient

Ex: if error is positive in period 1, it will be positive in

TEST for Serial Correlation in AR(1) with strict

TEST for serial correlation in AR(q) without strict

Test Ho: ρ1 = ... = ρq = 0, (F-test)

"Feasible Generalized Least Squares"

- CORRECTION for serial correlation identified by

- Asymptotically efficient (More efficient than OLS)

- Used in FGLS to estimate ρ with iterative scheme

- Used in FGLS to estimate ρ with iterative scheme

- CORRECTION factor used in Serial Correlation

Var(u) ≠ 0 (not constant)

- TEST for Heteroskedasticity

- Serial correlation in variance of errors

- Static models (no lagged dependent vars) -->

- TEST for Unit Root/High Persistence/I(1)

- TEST for Unit Root/High Persistence/I(1)

Augmented Dickey-Fuller - Includes lagged differences --> allows for more

ΔYt = α + θYt-1 + γ1ΔYt-1 + ... + γpΔYt-p + Et

- TEST for Unit Root/High Persistence/I(1)

Dickey-Fuller with Time Trend ΔYt = α + δt + θYt-1 + γ1ΔYt-1 + ... + γpΔYt-p + Et

Ho: I(1), Ha: trend-stationary

- Regression of I(1) series on another I(1) series

- Correct using Cointegration test (DF)

- When two time series are individually I(1) but I(0)

- TEST for Cointegration when parameters are

Error Correction Model ΔYt = α + γΔXt + δ[Yt-1 - β*Xt-1] + Ut

- Predicting future outcomes using available info

- Measuring model up to and including t to forecast

- Measuring model up to and including t to forecast