FM437MockExam AT2023
FM437MockExam AT2023
FM437MockExam AT2023
Instructions to candidates
This paper contains seven questions, in two sections:
Section A: 5 short-answer questions
Section B: 2 analytical or contextual questions
Answer all questions.
An electronic calculator may be used.
You are not supplied with any additional materials.
If at any point in this examination you feel that anything is unclear, please make any addi-
tional assumptions that you feel are necessary and state them clearly.
Good luck!
INSTRUCTIONS (read carefully): For true/false questions, include a short (1-2 sentences)
explanation. A good verbal argument can earn full marks, but often a mathematical ar-
gument is clearer. If you find that any statement is unclear or does not provide sufficient
information, please make an assumption and solve the problem based on your stated as-
sumption.
(a) (6 marks) You can generally achieve a higher statistical power by picking a lower
significance level.
Answer. False. Power is the probability of successfully rejecting the null hypothesis if
the null is false. Picking a low significance level decreases the probability of
Type I error, which is the probability of rejecting the null when it is true. Doing
so generally increases the probability of Type II error of not rejecting the null
when it is false.
(b) (6 marks) The ordinary least square estimator of the standard linear regression
is a particular case of over-identified GMM.
Answer. False. Consider the standard linear regression
Yt = Xt′ β0 + et .
and this is simply amoment condition i.e. a particular case of GMM, and it
can be estimated with the sample analogue
T
1X
0= Xt (Yt − Xt′ β̂T )
T t=1
and this is simply the FOC of OLS. Nevertheless, number of moments = num-
ber of parameters, so this is an exactly identified GMM, not an over-identified
GMM.
Answer. Take a log of both sides, logY = logA+β1 logK +β2 logL. Then we can use E[logA]
as the intercept and logA − E[logA] as the error term in the linear regression
model.
We can conduct a Wald test to see if β1 + β2 = 1. Denote
!
β̂1
R= 1 1 r=1 β̂ =
β̂2
Because ei and xi are independent of each other, E(x3 ei ) = 0 so the second term
is zero.
where
V = E(xi x′i )−1 E(xi x′i e2i )E(xi x′i )−1 = σe2 E(xi x′i )−1
q
by conditional heteroskedasticity. The standard error of β̂ is given by Vn . It
decreases with the number of observations and decreases with the variance of
xi relative to the variance of the error term.
5. (6 marks) Consider using the data of the past two decades to forecast monthly interest
rate yt+h with a stochastic process: yt = ρyt−1 + et . Regressing yt on its lag by OLS
yields ρ̂ = 0.95. Your forecast of interest rate in period t + 24 is then given by ŷt+24 =
ρ̂24 yt . Do you expect the forecast to have a mean forecast error of zero?
6. The Consumption-CAPM with log utility implies the following unconditional Euler equa-
tion
h i
E CCt+1t e
Ri,t+1 = 0 for any i,
(a) (6 marks) How would you test whether the above model is consistent with the
data? Be precise in describing the procedure and test statistic, and provide an
economic interpretation of your test.
(b) (6 marks) What is the minimum number of assets that you need to perform such
a test?
(c) (6 marks) Does the approach require multiple steps to be implemented?
Answer: Note that in the above model there is actually no parameter to be estimated. Never-
theless, we can use the GMM method to construct the Hansen J-test of the above
specification.
where Yipost is the average scaled citations of patents in firm i in the five years following
(a) (6 marks) Suppose you are interested in the causal effect of IPO completion on
innovation. Present an endogenous concern that gets resolved by the IPO filing
year fixed effects. Explain how the fixed effects can handle the concern.
Answer: Consider two firms in the same industry with identical Yipre and Xi . They file
for IPO in two different years, one in an economic expansion and the other in a
recession. The one that files in expansion might end up more innovative than the
other because the former has an easier time finding the right skilled labor in a
stronger labor market. At the same time, it might also be easier for the former
firm to complete the IPO because investors have more capital in good times.
In other words, the correlation between Yipost and IP Oi might simply reflect the
effects of the state of the economy on both instead of the causal effect of IP Oi
on Yipost . The state of the economy could cause a positive OLS bias under the
argument above.
Adding the IPO filing year fixed effects to the regression, we are able to control
for any factors that (1) vary year by year; (2) are common to all firms; and (3) af-
fect innovation and correlate with the chance of completing IPO simultaneously.
Summarize these factors with µt . With the fixed effects µt , we can take the mean
of equation (1) over all firms in industry k that file in year t, which yields
Yipost − Ȳ post = β1 (IP Oi − IP¯O) + γ1 (Yipre − Ȳ pre ) + (Xi − X̄)′ δ1 + (ϵ1i − ϵ̄1 )
µt does not show up in this regression. With the filing year fixed effects, we are
able to identify the causal effect by comparing two firms in the same industry not
only with identical characteristics, Yipre and Xi , but also filing for IPO in the same
year.
(b) (6 marks) Consider two firms in the same industry that file for IPO in the same
year. Suppose on average the firm with better future innovative opportunities is
better able to secure funding from private equity and consequently more likely to
withdraw the IPO filing than the other. In which direction would this bias the OLS
estimate β̂1OLS ? Show your analysis with the variables defined in equation (1).
Answer: The argument suggests that E(ϵ1i IP Oi |Yipre , Xi , νk , µt ) < 0. The OLS estimate is
therefore biased downward. This is an example of endogeneity bias caused by
reverse causality.
1
scaled citations of a firm in a year is citations of the firm divided by the industry average of citations
excluding the firm’s in the same year
Suppose the OLS estimate β̂2OLS is equal to 0.704 and its robust standard error
is 0.102. The F-statistic in regression (2) is 47.79. What can you conclude about
the instrument?
Answer: The F-statistic is much larger than 10, so the instrument is relevant for IP Oi .
(e) (6 marks) The author compares the observable firm characteristics between firms
that experience a NASDAQ drop and other firms that file to go public in the same
year (see Table III). By doing so, what point does the author try to make about the
instrument?
Answer: Based on the table, the firms that had NASDAQ drops and those that didn’t were
similar to each other in observable characteristics. This suggests that the in-
strument is unlikely to correlate with ϵ1i . Note that this is not a formal test for
exogeneity but an indication at best.
(f) (6 marks) What is the name of the standard error you would like to use for in-
ference on the two-stage-least squares estimate β̂1 ? Use a sentence or two to
justify your choice.
Answer: We can use the heteroskedasticity-robust standard error for β̂1 . Assuming we
have a large sample, it takes care of the potential correlation between ϵ1i and the
regressors.
(g) (6 marks) Column (1) and (3) of Table V show the OLS estimate and the two-
stage-least-squares estimate of β1 . Interpret the results, and comment on its
relationship to the bias we conjectured in (b).
Answer: Column (1) shows that innovation is negatively correlated with the chance of
completing IPO for a firm although the correlation is insignificant. Column (3)
shows that completing IPO reduced a firm’s scaled citations by 0.83%, which is
significant at the 5% leve. This indicates that going public had a significantly neg-
ative impact on innovation novelty. The TSLS estimate is more negative than the
OLS estimate. The direction of the bias is the opposite of what we conjectured
in (b). This suggests other factors in the error term might have biased the OLS