1.

Answer the following questions using step 2 of the case

(a) How large is the explanatory power of the model? Is it significantly different from zero?
What can you learn from this about the statistical properties of the estimator?
Explanatory power of the model = R2 or the goodness of fit of the model. It tells what
proportion of the variation in the dependent variable, is explained by the explanatory variable.
This lies between 0 and 1. The closer to 1 (100%), the bigger the explanatory power. We have
40,37% or 36,53% if we adjust for a multivariate model. This is relatively low. F-test decides
whether this is significantly different from 0, if p<0,05  significant. We have p=0,004 for our F-
test which means we can reject H0 and conclude that R2 is statistically significant.

R2 doesn’t tell us whether Gauss-Markov assumptions are fulfilled and therefore we know
nothing about the statistical properties of the estimator.

(b) What is the estimated impact of … on …, and how do you interpret this impact?
 ß0: als dBBP en dRCOMP gelijk zijn aan 0 dan is er een daling van dEMP van -0,652
procentpunten.
 ß1: als het BBP met 1 procentpunt stijgt, neemt de groei in employment toe met 0,64
procentpunten
 ß2 als de reële looncompensatie toeneemt met 1 procentpunt neemt de groei in employment af
met 0,098 procentpunten.

(c) Do your estimation results support the a priori theoretical expectations concerning the
impact of … on …?
We expected dBBP to have a positive influence on dEMP and dCOMP to have a negative effect
on dEMP. We expected the intercept to be different from 0.

(d) Assume that the propensity to invest and the interest rate are situated at their respective
average value over the time period 1977Q4-2006Q4. By how many percentdoes the
propensity to invest change when the interest rate increases by 10 percent?
When … increases by 10 percent, dEMP will rise/decrease by … percent.

2. Answer the following questions using step 3 of the case

(a) Gauss-Markov assumptions: no autocorrelation in the error terms
(i) What are the implications of possible deviations from this assumption for the properties of the OLS
estimator?
The estimated regression parameters are still normally distributed and unbiased estimators.
The formula of the variance must be changed.
The OLS estimator is not efficient anymore, GLS has a lower variance.

(ii) Is this assumption satisfied? Give the assumptions and statistical properties of possible tests. In case
you used more than one test, only report the resultsof the most appropriate test according to you, and
explain why you think it is appropriate.
No, we have autocorrelation. Possible tests: Runs, Durbin Watson and Breusch Godfrey LM test.
 - The Durbin Watson test has assumptions which are not fulfilled and therefore we cannot
trust this test in our case. This test tells us whether autocorr is positive or negative.
o Regression model has an intercept (ok)
o X’s are deterministic (NOT OK)
o AR(1) pattern in the error terms (NOT OK)
o Error terms normally distributed (ok)
o Regression model does not contain lagged dependent variables (e.g. Y t-1) (ok)
- The Breusch Godfrey LM test is an asymptotic test and we need to choose p. This test
has the advantage over Durbin Watson test concerning the assumptions.
Test for p=1, p=2, p=3, p=4 and reject H0 every time. From p=5 the p-value is not
signifant anymore.
- The Runs test is a nonparametric test for autocorrelation, which means we don’t have to
assume a specific pattern
runs(x)

E(R) = 18 var(R) = 8,2424 stand error R = 2,871

Conf interval 95%: [12,37284 ; 23,62716]
R = 14 lies within the confidence interval meaning we cannot reject H 0 with Runs test.

Note: answer the following two questions even if you didn’t find autocorrelation under 2(ii). Hence, this
answer is purely hypothetical, i.e. under the assumption that we found autocorrelation
(iii) Which alternative estimation method and/or specification do you think is most appropriate to correct
for possible problems?
Our estimator is still unbiased and normally distributed but no longer efficient. GLS will have a
lower variance. EGLS method will not necessarily have the usual optimum properties of the
classical model, such as BLUE, especially in small samples like in our case.
Formula of variance of OLS must be adjusted. We do this with the Newey-West correction but is
not appropriate in small samples. This solves an autocorrelation problem as well as a
heteroscedasticity problem.
First difference method to make our data stationary.

(iv) What are the statistical properties of this alternative estimation method?
p is unknown and has to be estimated. EGLS estimator is consistent and asymptotically correct.
HAC Newey-West correction is only asymptotically correct. We have to be carefull with our
small sample size with both of these.

3. Give the correct answer (and complete if necessary). Only one possibility can be
circled. There is no negative grading.
Using the initial specification and estimation method from step 2.
(a) There is a problem of multicollinearity (1)
- Yes, because there is a perfect linear relation between the interest rate and the output gap
- No, because there is no perfect linear relation between the interest rate and the output gap
- Yes, because the variance inflating factor is
- No, because the variance inflating factor is less than 10

(b) The error terms are normally distributed

- This assumption cannot be rejected on the significance level
- Not sure, because the error terms tend to the normal distribution but the sample size is not
very large
- No, because the exact significance level on which this assumption can be rejected is
- Approximately because the error terms tend to the normal distribution, we don’t see any
pattern in the residual vs fitted plot. Jarque-Bera normality test cannot reject its H 0.

(c) The OLS estimator is normally distributed

- Certainly not, because the regressors are stochastic
- Certainly not, because the error terms are not normally distributed
- Approximately, because the OLS estimator tends to the normal distribution
- Certainly, because the error terms are normally distributed

Using the specification from step 4, namely: add an extra variable dEMP_1 to the model.
(d) Given the theoretical framework and the available data, the 2SLS estimator results
from
- replacing the endogenous variable(s) by a linear combination of the exogenous variable(s)
- replacing the endogenous variable(s) by the exogenous variable(s)
- estimating the model in its structural form and from this, calculating the parameters of the
reduced form
- estimating the model in reduced form and from this, calculating the parameters of the
structural form

(e) There is no endogeneity problem (1)

- Correct, because this assumption can be accepted on the signifi- cance level
- Correct, because the exact significance level on which this assumption cannot be rejected is
- Correct, because this assumption cannot be rejected on the 5% significance level
- False, because the exact significance level on which this assumption can be rejected is

4. Briefly discuss the following statements. Careful: these might be false!

(a) In a regression of time series data, xt−1 can be used as an instrument for the endogenous
variable xt.
WRONG: Yt=Yt−1+μt Xt=Xt−1+εt

(b) In a regression of yt on a constant and yt−1, the OLS estimator is biased but consistent.
WRONG: non-stationary data with autocorrelation, biased and inconsistent.
(c) When an estimator is unbiased and efficient, it implies that the parameters can be
estimated in a precise manner.
WRONG: when the degree of multicollinearity is high, parameters cannot be estimated in a
precise manner, eventhough OLS is unbiased and efficient. This would mean we are dealing with
a practical sample problem.

Question 1
Met hoeveel … neemt … toe als … stijgt met … % indien de andere variabelen gelijk
blijven?
ß0: als dBBP en dRCOMP gelijk zijn aan 0 dan is er een daling van dEMP van -0,652
procentpunten.
ß1: als het BBP met 1 procentpunt stijgt, neemt de groei in employment toe met 0,64
procentpunten
ß2 als de reële looncompensatie toeneemt met 1 procentpunt neemt de groei in employment af
met 0,098 procentpunten.

What is the average value of the estimated error terms? What does this mean for the
properties of the OLS estimator?
The numeric properties of the OLS estimator are:
- The sample regression function goes through the sample averages of X and Y
- The estimated error terms have an average of 0
- The estimated error terms are not correlated to Xi or Yi
The statistical properties of OLS estimator are unbiased and efficient (and linear). These are
fulfilled when all the Gauss-Markov assumptions are fulfilled.

Are the estimated parameters fulfilling the a priori expectations? Explain with the help of a
statistical test.
We expected dBBP to have a positive influence on dEMP and dCOMP to have a negative effect
on dEMP. We expected the intercept to be different from 0.
We test this with the t-test: H0: ß0 = 0 H 1: ß0 ≠ 0 or H0: ß1 ≤ 0 H1: ß1 >/< 0
With ß0 and ß2 we do not reject. ß1 we can reject H0, meaning that dBBP has a significant
effect on dEMP. But all the estimated values of the parameters comply to our a priori
expectations.

Question 2
Are the error terms normally distributed? What does this mean for the OLS estimator?
Yes, they are normally distributed. Assumption 11 deals with this hypothesis. Jarque-Bera
normality test p-value=0,959. We cannot reject the null hypothesis. The OLS estimator is BUE
and normally distributed.

How can you detect heteroscedasticity and which test is the best?
This means that the variance of the error terms is not constant. We can detect this with:
- Informal methods (graphic and intuitive)
- Goldfeld-Quandt test (Assumption: σi2 is positively correlated to one of the Xi)
 - White’s General heteroscedasticity test (no assumptions, asymptotically correct)
What impact does heteroscedasticity have on the properties of OLS and the hypotheses?
Causes Impact on OLS properties
Properties population and data collection Estimated regression parameters are unbi-
ased and asymptotically normally distributed
Specification errors OLS estimator not efficient anymore (GLS)
What remediation would be the most effective? Why?
2 2
σi is known: GLS is BUE or σi can be estimated consistently: EGLS is consistent and
asymptotically efficient  not much happens
2
σi is not known: (E)GLS (ad hoc transformations) / OLS (White’s heteroscedasticity consistent
variance: adjusted variance formula)  adjust specifications because this is the reason for the
heteroscedasticity.

What has changed? Does this fulfill the expectations?

Question 3
What is the specification is STEP 4? Why have you chosen this?
The same model, but with one more variable added: emp-1. This solves our autocorrelation and
specifiation error.

Question 4 TRUE or WRONG

If VIF>10, this will not only have an impact on the statistical properties of the OLS
estimator, but also on the practical consequenses of the hypotheses.
WRONG: first not, second yes. We’ll have a broader confidence interval and lower t-statistics.
The chance for type II error increases. But there is no impact on the OLS estimator properties.

Unbias and efficiency assure a precise estimation.

WRONG: this is possible but not assured. When the degree of multicollinearity is high, these OLS
properties lead to an unprecise estimation.

The OLS estimator is biased and inconsistent when leaving out a relevant variable.
TRUE: if r23 =/= 0
When a irrelevant is included, OLS remains unbiased and consistent.

The OLS estimator is biased and inconsistent when measuring error occurs in dependent
variable.
WRONG: this is the case for explanatory variables. For dependent variable the OLS estimator is
unbiased and consistent.
 Als p met 3% veranderd, wat is effect gemiddeld op q
● Wat is het gemiddelde van u hoed, wat zegt dit over statistische eigenschappen
● Is u normaal verdeeld, wat zegt dit over statistische eigenschappen
● Zijn uw schattingen van p in lijn met je verwachtingen, gebruik statistische test
● Test heteroscedatisiteit, verklaar waarom je die test gebruikt
● Hoe zou heteroscedatisiteit impact hebben op uw schattingen, geef 2 mogelijke verklaring
● Hoe hebt je dit geremedieerd
● Statistische eigenschappen van die nieuwe schatting
● Hoe is je schattingsresultaat veranderd, ligt dit in lijn met je verwachtingen
● Theoretische redenen waarom p zou gecorreleerd zijn met storingsterm
● Gevolgen voor eigenschappen
● Beschrijf de methode om te remediëring
● Nieuwe eigenschappen door die methode
● Effect op schattingsresultaat en ligt dat in lijn met je verwachtingen
● Heb je te maken met een zwak instrument wat betekent dat en wat zijn de gevolgen
● Stap 4: geef u nieuwe specificatie
● Stel we voegen ln( p(t-1)) bij u specificatie. Je hebt heteroscedasticiteit en autocorrelatie,
geen specificatie fout. Is u specificatie beter na dat te hebben toegevoegd? Waarom. En
heb je voldoende info om het toe toevoegen (tip: p t-1 is endogeen)
● Gevolgen vif groter als 10
● Gevolgen weglaten relevante verklarende variabele
● Gevolgen meetfout in afhankelijke variabele
● Nauwkeurig en onvertekend kan je dan iets precies schatten

Van andere versie

Eerste vraag: Prijs stijgt 5%, wat met vraag, ook R2 uitleggen en significantie testen met F-test.
Tweede vraag was over autocorrelatie: welke test gebruik je en verklaar waarom je die kiest,
ook de stappen van
EGLS uitleggen toegepast op de case en nog een correctie uitleggen die ik niet kende.
7 punten endogeniteit ook, vrij algemeen: wat is het probleem in woorden. Hoe los je het op,...
Dan je nieuwe specificatie zeggen en uitleggen waarom je die gekozen hebt + statistische
eigenschappen ervan.
Theorie herinner ik mij:
* iets van: is zuiver altijd ook nauwkeurig?
* en de stelling: Yt regresseren op Yt-1 levert een vertekende en inconsistente