POTENTIËLE VRAGEN OVER ONZE CASE, GEBASEERD OP VRAAGSTELLING VORIGE JAREN

1. Answer the questions below based on your OLS estimates of the baseline specification,
Fertct= !! + !"*ln(AgriYct)+ !#*ln(ManuSct)+ µct, obtained in step 2 of the case.

- Test whether the estimation results support your a-priori theoretical expectations about
the impact of ln(AgriYct) on Fertct . (1)
A priori expectation:
• H0: β( ³ 0
• H1: β( < 0 an increase in Agricultural yield leads to a decrease in Fertility
This expectation seems to be correct, as increasing ln(AgriY) statistically results, on average, in a
relative decrease of Fertility of -0.1316 compared counties with the same manufacturing share . This is
statistically the case, because of the t-statistic of -34.34 (< -1.645) and hence the p value of 0.000 (< 0.05).

- What is the explanatory power of the model? Is it statistically significantly different from zero? What does
this teach you about the statistical properties of the estimator used? (1)
The explanatory power or R² of this model is 0.3364 = 33.64%.
One uses an F-test (F = 14442.017; p = 0,000) to see if R² is significantly different from 0. We thus reject the
H0 that states that the R² = 0. The fact that the R² is low does not directly teach us something about the
statistical properties of the estimator. However, it could indicate that there is a specification error, which
has important implications for the statistical properties (of the OLS estimator).

- What are the 4 determinants of the variance of the OLS estimator, indicate whether their impact is positive
or negative and quantify these measurements with the results of your case. (1,5)
o Increase in variance of the population error terms à variance OLS estimator increases
o Decrease in the variance of X (more variance = more info) à variance OLS estimator increases
o Decrease in sample size (more data = more info) à variance OLS estimator increases
o Higher degree of multicollinearity (= VIF) à variance OLS estimator increases
In our case:
(Depends on the estimator for which Gerdie poses this question: look at the descriptive table for Var(X),
residual variance can be found in the output.) Suppose this question is for variable AgriY Niet aangepast!
o Estimated variance of population error terms: #$ = 18.278 (df = 380 355)
o Variance of X = 0.241
o Very large simple size of 380,362 à low variance OLS estimator
o Faint degree of multicollinearity, as VIF are sufficiently low (max. 1.332) and R² of auxiliary regressions
does not exceed 50%

- Based on the previous question, is there multicollinearity? (0,5)

Faint degree of multicollinearity, as VIF are sufficiently low (not that much higher than the neutral (=
minimum) value of 1, both AgriY and ManuS (1.0109) and R² of auxiliary regressions does not exceed 50%
(0.0108 both AgriY and ManuS ).
o Does this change anything for your answer in the first question (about power & size of t-test etc.)? (1)
As the variance of the estimators will be higher, the confidence intervals will be wider and the t-test
statistics will be lower, leading to less significant variables (cfr. formula t-test).

Matteo, Yaël, Sebastiaan, Ward en Tom

Edited by Rie
- AgriY has a negative impact on y. Prove using a T-statistic.
o Hypothesis
%2 ³ 0
H0: β
%2< 0
H1: β
o Calculation t-statistic & critical values (e.g. variable AgriY)
*(+ )!
) +0.1316 + 1
&= -
."
= 1.1038
= −34.632
#!

(this value differs a little due to roundingUsing a significance level of 5% à critical value -t5691,0.05 = -1.645
o Conclusion
As the t-statistic is smaller than the critical value (or larger in absolute values), we reject the null
hypothesis that β%2 ³ 0.
AgriY has, in other words, a significantly negative impact on y at the 5% significance level.

- Calculation of a p-value.
Take your t-statistic and degrees of freedom, look for the interval in which this t lies (in table of t-distribution).
E.g. for t = -2.54 with 8 degrees of freedom: 2% < p < 5% because the absolute value of -2.54 = 2.54 lies
between 2.306 and 2.896 for 8df.

- Interpretation + calculations for an explanatory variable increasing by e.g. 2 percentage points.

Unit increase in Fert for 2% increase of AgriY?
If we use a percentage change we can say if the Agricultural yield goes up by 2% then the Fertility will go
down by -0.1316/50 ( β2/50)
- Calculations: -> Just fill in absolute change here

-> Fill in percentage here for percentage change, relative

change means a 1 unit (100%) change

For ManuS we use the exact same calculations BUT a relative change of 1 unit 100% can not be used because a
manufacturing share can not surpass 100% so keep that in mind.

- Continuation, it could be possible for Gerdie to ask about dummy variables and their interpretation, like in
the 2019 example, what would in our case happen to Fertility, how much would it, on average, change if we
move from 1880 to 1890?
− To compare the average value of Fert between 1880 and 1900, you can subtract the coefficient for
Dummy1900 (0.2479) from the coefficient for Dummy1880 (0.3358). This gives a difference of 0.0879,
indicating that the average value of Fert is 0.0879 higher in 1880 compared to 1900.
− Similarly, you can compare the average value of Fert between any two years by subtracting the
coefficients for their respective dummy variables. Here are the differences in the average value of Fert
between each pair of years:
o 1880 and 1900: 0.3358 - 0.2479 = 0.0879
o 1880 and 1910: 0.3358 - 0.2001 = 0.1357
o 1880 and 1920: 0.3358 - 0.0828 = 0.2530
o 1900 and 1910: 0.2479 - 0.2001 = 0.0478
o 1900 and 1920: 0.2479 - 0.0828 = 0.1651
o 1910 and 1920: 0.2001 - 0.0828 = 0.1173

Matteo, Yaël, Sebastiaan, Ward en Tom

Edited by Rie
- Determine joint significance of e.g. 2 variables.
We calculate the joint significance of AgriY and ManuS:
!"" 40.4442
#$ 2
= 79.7937 = 1442.015 = F!,5690 > 3 (789:; < − 7;>7)
%""
#$ 5690
è Jointly significant

- Calculate p-value of a one-sided hypothesis test

- Does β% 2 measure the full impact of AgriY on Fert? Explain. (1)HS7 SLIDE 16 -> ZEKER BEKIJKEN
- Measure if there is an indirect impact of AgriY on Fert. 3 possibilities:
• Causal impact from AgriY to ManuS only: Probably
• Causal impact from ManuS to AgriY only
• Causal impact both directions à probably the case

o Assume that we are only interested in the impact of AgriY on Fert and we estimate the following
specification: Fert = π0+π1AgriY+μi. What are the statistical properties of the OLS estimator _$1 for the
population parameter β2?
The OLS estimator will be biased and inconsistent if the correlation between AgriY and the other
control variables is ≠ 0. The variance of the error terms will be larger, as the omitted variables have
been pushed into the error terms. Also, the variance of the OLS estimator will be impacted, even if the
correlation equals 0. The variance of the new estimator will be smaller if the VIF (in the original
regression) is high and the increase in variance of the error terms in the new regression is low and vice
versa. Probability of a type II error increases, increasing in the degree of multicollinearity.

- Are the error terms normally distributed? What are the consequences for the OLS estimator? What test did
you use for this? (1)
By computing the Jarque-Bera test for the estimated residuals µ$ I to test H0: µi ~ N, we conclude that the error
terms are not normally distributed, as the test value of 37.72 exceeds the 5% critical value, p=0.000 This
test is only asymptotically valid, but given the sample size of 5693, we deem this size to be sufficient.
The assumption of normality is under H0, which means that the fact that it got rejected indicates that there
is significant evidence against the normality assumption. Consequences: as the error terms are not normally
distributed, the distribution of the OLS estimator will not follow a normal distribution. However, due to the
CLT (central limit theorem), the distribution of the OLS estimator will converge to a normal distribution.
In short, the OLS estimator will asymptotically follow a normal distribution. OLS will be BLUE instead if
BUE. Asymptotically, if the number of samples goes to infinity, the OLS estimator will become normally
distributed again. INDEPENDENT OF SAMPLE SIZE ASYMPTOTIC JUSTIFICATION FOR USING OLS BUT IT
REMAINS AN APPROXIMATION IN FINITE SAMPLES

Matteo, Yaël, Sebastiaan, Ward en Tom

Edited by Rie
2. Answer the questions below based on your analysis in step 3 of the case (evaluation of the
Gauss-Markov assumptions and individual remediation of deviations).

- Gauss-Markov assumption: no heteroskedasticity

o What are the implications of potential deviations from this assumption for the statistical properties of
the OLS estimator?
The OLS estimator is still unbiased and is asymptotically normally distributed. In smaller samples, the
estimators are not normally distributed. The variance formula of the OLS estimator must be adjusted:
# ∑ L7 ( #7 (
G:8HI 1( J = → G:8HI1( J = (see formula sheet)
∑ L7 ( (∑ L7 ( )(
Also, the OLS estimator is not efficient anymore, as this variance is higher than the variance resulting
from the Generalized Least Squares (GLS) estimator.

O Is this assumption valid? Also list the conditions and statistical properties of the potentially used tests.
If you used more than one test, report the results of the one you deem as being the most adequate
and explain why. (2)
This assumption is not valid, there is heteroskedasticity present in the data. As indicated by White’s
general heteroskedasticity test: it has a test value of nR² = 215.765 ∼ χ², resulting in a p-value of
approximately 0, as the test value exceeds the critical value of 11.071 (α = .05, df = 5) à we reject H0
of no heteroskedasticity. We pick this test, as the alternative (Goldfeld-Quandt test) assumes that the
error terms are normally distributed, which is not the case. This test also, in contrast to the GQ-test,
does not assume the heteroskedasticity pattern to be increasing or decreasing, resulting in a more
flexible test. White’s general heteroskedasticity tests’ test-statistic being significant could also be an
indicator for a specification error, which we assume is the case here.

(Note: also answer the next question if no heteroskedasticity was detected. Your answer is in that case purely
hypothetical, assuming that heteroskedasticity was detected.)
o Which alternative specification and/or estimation method do you deem the most adequate to
remediate the potential problem? What are the statistical properties of this alternative estimation
method? (1)
By adjusting the specification, as the heteroskedasticity most likely stems from a specification error.
Our suspicion goes to the omission relevant variables (like dummies for the different years) or the
use of a wrong functional form. Our remedy would be to include these variables in the model or
change the functional form. If not corrected , the estimator for the population variance is biased and
inconsistent. -> wrong inference
à by changing the specification to the right one, OLS becomes unbiased and consistent again.

Matteo, Yaël, Sebastiaan, Ward en Tom

Edited by Rie
- Gauss-Markov assumption: no autocorrelation in the error terms
O What are the implications of potential deviations from this assumption for the properties of the OLS
estimators? (2)
The OLS estimators are still unbiased and normally distributed, but no longer efficient. There is an
estimator (GLS) that has a lower variance. The formula for the variance of I\( should also be
adjusted when using OLS assuming you have an AR(1) pattern:
# # ( (1 + 8^)
1( J =
G:8HI → 1( J = 7 ∗
G:8HI
∑ L7 ( ∑ L7 ( (1 − 8^)

O Is this assumption valid? Also list the conditions and statistical properties of the potentially used tests.
If you used more than one test, report the results of the one you deem as being the most adequate
and explain why. (2)
This assumption is not valid. The Durbin-Watson d-test assumes that Xi is deterministic, assumes an
AR(1) pattern in the error terms, the error terms to be normally distributed and no lagged independent
variables (e.g. Yt-1) included. The Breusch-Godfrey test allows Xi to be stochastic, the error terms can
follow an AR(p) pattern and lagged independent variables are allowed. This test is only asymptotically
valid, however. The Runs test is another possible test for autocorrelation. Under the H0 of no
autocorrelation and for the sample size going to infinity, the number of runs is normally distributed.
We deem the Runs test as most adequate as this test does not assume any pattern whatsoever.

Given that the number of runs (1341) does not lie within’ the CI[2745.28;2891.64], we
reject the H0 of no autocorrelation. Since the number of runs is lower than the lower limit of the CI,
we can deduct that there is positive autocorrelation.

(Note: also answer the next question if no autocorrelation was detected. Your answer is in that case purely
hypothetical, assuming that autocorrelation was detected.)
o Which alternative specification and/or estimation method do you deem the most adequate to
remediate the potential problem? What are the statistical properties of this alternative estimation
method? (1)
Our data is made up of data measurements for counties in 5 different years, as we plot the data
ordered by year a pattern of decreasing over the 5 years is could be detected, that's why we
say the autocorrelation detection suggests a specification error. Our suspicion goes to the omission
of dummy variables or the use of a wrong functional form, so our remedy would be to include these
variables in the model or change the functional form. If not, the estimator is biased and
inconsistent.

- One of the Gauss-Markov assumptions states that variables that are not explicitly included in the model do
not have a systematic influence on Yi, i.e. E(µi|Xi) = 0 for stochastic explanatory variables Xi. Use one
autocorrelation test to verify this assumption.
We would use a Breusch-Godfrey LM test since a specification error results in autocorrelation as well.

o Justify why you prefer this test and why it can be used to test E(µi|Xi) = 0.
We prefer this test since it has less strict assumptions about the explanatory variables - they do not
need to be deterministic compared to the Durbin Watson d-test, lagged dependent variables are
allowed and not only AR(1) patterns are allowed. Furthermore, specification errors often induce both
autocorrelation and heteroskedasticity. Because there is no logical ordering in the data (as would be
the case with pure autocorrelation) we can assume that the autocorrelation is due to a specification
error. In this case, there will be variables that are moved into the error terms but have a systematic
influence on the model, causing the autocorrelation.

Matteo, Yaël, Sebastiaan, Ward en Tom

Edited by Rie
- Suppose that there is autocorrelation in the error terms.
o Does your answer under the last question have an impact on the statistical properties of the OLS
estimator and the hypothesis test you used in the first question (about a priori expectations test)?If
there is autocorrelation, OLS will no longer be efficient. Furthermore, the variance of the OLS
estimator needs to be adjusted. OLS estimators will, however, remain normally distributed.
The hypothesis test we have performed is no longer valid since the current variance of the OLS
estimator is biased and inconsistent.

o How would you remediate the autocorrelation in the error terms in the best possible way?
In our case, the autocorrelation is due to a specification error. Therefore, the best possible way to
remediate it is by adapting the specification.

- The VBO estimates a Durbin Watson statistic of 2,02 and says that there is no autocorrelation. Is the VBO
correct? Explain why (not). (1)
We have 2 regressors (not including intercept) and n > 200. When looking at the table for DL and DU we find
that 2.02 is not included in the zone of indifference meaning that there is indeed no autocorrelation.

Matteo, Yaël, Sebastiaan, Ward en Tom

Edited by Rie
- Gauss-Markov assumption: no specification errors
o What are the implications of potential deviations from this assumption for the statistical properties of
the OLS-estimator?
Causes here:
1. Omitting a relevant variable (= Underfitting):
a. OLS is biased and inconsistent if the dropped relevant variable is correlated with the
remaining explanatory variables (which, in practice, will always be the case).
Else the OLS estimator is unbiased and consistent.
b. Estimator variance error terms will be higher because a relevant factor is pushed into the
error terms.
c. The variance of the model with the specification error will be higher if the VIF is low and the
variance of the stochastic error term is way higher (in case of dropping the relevant variable)
and vice versa.
We assume that this is the case in our case study as we assume that the omitted dummy fir the
years are still relevant

2. Wrong functional form

a. OLS estimator is biased and inconsistent.
It is very likely that the functional form is incorrect.

Theoretically: OLS will be biased when underfitting, using a wrong functional form or when
measurement errors in explanatory variables are present (endogeneity). OLS will be unbiased when
overfitting or when measurement errors in dependent variable are present. The variance of the OLS
estimator will increase in all 5 cases and will result in OLS being an inconsistent estimator when bias is
present (in small samples): underfitting, wrong functional form and measurement errors in the
explanatory variables (recall: endogeneity) lead to inconsistent and biased estimators. Overfitting
and measurement errors in the dependent variables don't cause bias and OLS is still consistent.

O Is this assumption valid? Also, list the conditions and statistical properties of the potentially used tests.
If you used more than one test, report the results of the one you deem as being the most adequate
and explain why. (2)
This assumption is not valid.
Ramsey RESET test: test statistic = 142.588; p = .0000. We use an F-test to test H0: β3 = β8 = 0

Lagrange multiplier test: test statistic = 23,342.270; p = .0000.

Under the H0 of no specification error and for n going to infinity, nR² ~ c² with k degrees of freedom.

Forecast Chi squared (on unordered data): test statistic = 150,073.900; p = .7.
Under the H0 that the specification is correct:

This method requires a sufficiently large sample.

We would choose the Ramsey RESET test since it is not an asymptotic test.

(Note: also answer the next question if no specification error was detected. Your answer is in that case purely
hypothetical, assuming that a specification error was detected.)
o Which alternative specification and/or estimation method do you deem the most adequate to
remediate the potential problem? What are the statistical properties of this alternative estimation
method? (1)
(personal team answer)

Matteo, Yaël, Sebastiaan, Ward en Tom

Edited by Rie
- Gauss-Markov assumption: no endogeneity

- Given the information in the case, are there any theoretical reasons that indicate that the assumption
E[µiAgriYi] = 0 is violated?
o Explain clearly.
The assumption given means that the error terms are correlated with the explanatory variable
Agricultural Yield. As stated in the case, the arrival of the bollweevil beetle had great effect on cotton
harvests and thus we can assume also the yield. The case also includes the impact of an increase in
cigarette consumption during World War I on local tobacco cultivation in the American South as a
second source of exogenous variation in agricultural production and thus yield. These two channels
are not included in the model and could lead to correlation between AgriY and the error terms. This
can also apply to ManuS, BollWeevil makes for example farming less attractive in cotton belt ->
shift to manufacturing

o What are the consequences for the OLS estimator when this assumption
is violated?OLS is biased and inconsistent.

o Which remediation do you propose to solve the violation of this assumption?

- Explain the necessary steps and required (additional) assumptions. Be complete!
We have to estimate the coefficients of the model with 2 stage least squares (2SLS):
1st stage: estimate proxies for AgriY and ManuS
à Estimate the reduced form with OLS using predetermined variables Tobacco and BollWeevil
à variable is split into 2 parts: endogenous part (estimated error terms) and exogenous part
(estimated proxy) (by assumption)

2nd stage: Use the proxies for both AgriY and ManuS
à OLS will be biased but asymptotically consistent

- Which instruments do you use?

BollWeevilIntensity, Tobacco

- Report the resulting coefficient estimate for the impact of AgriY (β2) along with its standard
error. β% 2 = -0.293 se^(β%2) = 0.031 Could probably ask for ManuS on exam eswell
o Next to the theoretical reasons outlined earlier, is there any statistical evidence that AgriYi is
endogenous?Clearly outline how you have tested this.
Yes, as the Hausman test leads to significant residuals for both AgriY (t=15.1821) and ManuS
(34.0037). This means that the 2SLS is significantly different from OLS and both AgriY and ManuS are
endogenous

- Check in your results for weak instruments and explain why this is a problem. (1)

There was an indication for weak instruments due to the low R^2 but to test wether there are weak
instruments we use a join significance test (F test), in both first stage regressions for AgriY (F=279.61) and
ManuS (F = 189.82) this is both higher than the critical value 10(Common rule of thumb)

Matteo, Yaël, Sebastiaan, Ward en Tom

Edited by Rie
3. Answer the following questions based on your analysis in step 4 of the case (general remediation
Gauss-Markov deviations).
- What is your preferred final specification? Explain how and why you choose this (this is independent of
whether and how it can be estimated). (2)
We noticed a downward trend in the error terms when ordered by year, so we chose to specify using
dummy variables, the Ramsey RESET test provides us with empirical proof that the specification error has
been resolved. We choose to use this one as the main reason because it is not an asymptotic test. The
heteroskedasticity and autocrrelation remained so we corrected using a Newey-West (HAC) consistent
standard error remediation. Which is still asymptotically valid/consistent

- Which estimation method is appropriate to estimate your adjusted specification? Clearly motivate why and
outline the statistical properties of your suggested estimator for the coefficients as well as the standard
errors (1)
We should perform 2SLS. Implications of this remediation are that the estimators will be biased, but
consistent (if we assume that the other GM assumptions are fulfilled).

Matteo, Yaël, Sebastiaan, Ward en Tom

Edited by Rie
TRUE OR FALSE STATEMENTS

CHAPTER 2-6: REGRESSION WITH 2 VARIABLES

- Unbiased and efficient estimators can estimate accurately. (1.5p)

Theory: unbiasedness means ‘on average you’re right’ and unbiasedness implies consistency.
Efficiency means that no other estimator has a lower variance.

Statement is false: neither of these 2 statistical properties of OLS guarantee that the variance is
low, however. The accuracy (variance) of the OLS estimator is determined by the variance of the
population error terms, variance in X, number of observations and degree of multicollinearity.

This phenomenon is especially visible with large degrees of multicollinearity, as the OLS
estimators will be efficient, but it will have a very large variance.

- As a result of the central limit theorem (CLT), we have that the error terms are by definition
normally distributed in a very large sample, except when we are dealing with stochastic
explanatory variables.

False: the CLT states that the error terms are normally distributed by definition, assuming that
they are made up of independent and identically distributed factors, whose sum will converge to
a normal distribution if the number of factors tends to infinity. Hence, this theory assumes
normality when the number of factors that make up the error terms (and are thus not included
in the model) tend to infinity, not the sample size.

This assumption of normality of the error terms is also valid when dealing with stochastic
explanatory variables. The stochastic nature of these variables only influences whether or not
the distribution of the estimators will be normal or not, not the distribution of the error terms.

CHAPTER 10: MULTICOLLINEARITY

- A variance inflation factor (VIF) bigger than 10 not only gives problems for the statistical
properties of the OLS estimator but also has a notable practical impact on hypothesis
testing.

A large VIF only has practical implications, as the OLS estimator(s) will remain unbiased and
efficient (BUE) and normally distributed. These practical issues are higher CI’s, a larger chance
of not rejecting H0 and thus a higher probability of making type II-errors (making the variables
less significant). The gravity of a “large” VIF also depends on the base variance of the estimators,
so a VIF of 10 could possibly not be an issue if the base variance is very small. VIF is a relative
measure and it should not be used as a benchmark.

Matteo, Yaël, Sebastiaan, Ward en Tom

CHAPTER 11: HETEROSKEDASTICITY
- Heteroskedasticity has important implications for the statistical properties of the OLS estimator
and inference based on it, such that it is advisable to remediate using EGLS. (2p)

First part of the sentence is true: β%( is not normally distributed anymore (only asymptotically),
∑ "! " ∗$!"
E" needs to be adjusted to Var (β%( ) =
the formula for the variance of ! " and OLS is no
(∑ "! " )
longer efficient, as GLS has a lower variance.

EGLS should only be used as a remediation if !'( can be estimated consistently and if the sample
size is large, as EGLS is only asymptotically efficient and biased in small samples.

If !'( is not known or cannot be estimated, White’s heteroskedasticity consistent variance is

advised. OLS is not efficient, but it has a consistently estimated variance using White. Beware
that in small samples, we encounter a bias.

CHAPTER 12: AUTOCORRELATION

- Autocorrelation affects properties of OLS estimate. Advised to always use (E)GLS?

Autocorrelation makes the OLS estimator inefficient (resulting in an estimator that does not
result in the lowest variance, as GLS has a lower variance) and not precise, but the estimator
remains unbiased.

One should not always use EGLS since it is biased but consistent and asymptotically efficient. In
small samples for example, EGLS is not a good choice. If you have pure autocorrelation, one could
use OLS because it is unbiased.

CHAPTER 13: SPECIFICATION ERROR

- Omitting a relevant variable always makes OLS estimators biased and inconsistent (1.5p)

We cannot omit a relevant variable if it is correlated with the other explanatory variables. If it
is not correlated with the other explanatory variables, OLS remains unbiased!

Matteo, Yaël, Sebastiaan, Ward en Tom

CHAPTER 18-20: ENDOGENEITY
- In time series data, we can use xt-1 as an instrument for the endogenous explanatory variable xt .
(2p)

• Instruments are selected based on the dual necessity of being correlated with the
endogenous variable but not with the error terms.
à lagged dependent variables in time series can be used as instruments when we assume
no autocorrelation is present
• 3 conditions to check whether xt-1 is an instrument for xt using Yt = B1 + B2*Xt + µt
o Order condition: Xt-1 is a predetermined variable, not in the above equation =
exclusion restriction
o Rank condition: Xt-1 impacts Xt à E(XtXt-1) ≠ 0 meaning there is correlation (this if often
the case in time series = persistence)
à Correlation with the endogenous variable
o E(Xt-1µt) = 0, or no correlation between the instrument and the error terms. This
means that no autocorrelation in the error terms is allowed
As the endogeneity issue E(Xtµt) ≠ 0 also means that E(Xt-1µt-1) ≠ 0
à if Xt-1 & µt-1 are correlated, and so are µt & µt-1, then Xt-1 is correlated with the error
terms µt à correlation with the error terms L
à if these 3 conditions are satisfied, Xt-1 is a correct instrument and 2SLS is consistent

- When regressing yt on a constant and yt-1, the OLS estimator is biased but consistent.

• If there is autocorrelation, there will be endogeneity (in combination with the lagged
dependent variable) which results in OLS being biased and inconsistent.
• However, if there is no autocorrelation, the properties of the OLS estimator will depend on
the fulfilment of the other GM assumptions (e.g. dependent variables deterministic and
error terms on average zero à unbiasedness).

Matteo, Yaël, Sebastiaan, Ward en Tom