ECM9/ECM10/ECM11/MGM3

MPhil in Economics
MPhil in Economic Research
MPhil in Finance and Economics
MPhil Finance

Wednesday 15 May 2024 10:00am to 12:00pm

R301B
CROSS SECTION AND PANEL DATA

Candidates are required to answer three questions from Section A

and one question from Section B.
Section A is weighted as 60% and Section B as 40%.

Write your candidate number (not your name) on the cover of each booklet.

Write legibly.

If you identify an error in this paper, please alert the Invigilator, who will notify
the Examiner. A general announcement will be made if the error is validated.

STATIONERY REQUIREMENTS
20 Page booklet x 1
Rough work pads

SPECIAL REQUIREMENTS TO BE SUPPLIED FOR THIS

EXAMINATION
Calculator - students are permitted to bring an approved calculator
New Cambridge Statistical Tables

You may not start to read the questions printed

on the subsequent pages of this question paper
until instructed that you may do so by the
Invigilator.

1 of 9
 SECTION A

A1 Consider the model:

D = 1(xe βe + xo βo + ε ≥ 0) (1)
xe = 1(z ′ θ + e ≥ 0) (2)

where z = (xo , z1 ) is a vector of exogenous instruments. xo is a vector of exoge-

nous included instruments; z1 is a vector of exogenous excluded instruments.

(ε, e) is independent of z and distributed bivariate normal, with ρ = Corr(ε, e).

(a) Given (1) and (2), evaluate whether a plug-in 2sls estimator would deliver
a consistent estimate of βe .
(b) Evaluate the observation that
for a linear structural model with a non-linear reduced form, and
where an endogenous variable is binary, the use of a control func-
tion estimator is appropriate. However, the use of a plug-in 2sls
like estimator is an example of a forbidden regression.
A2 Denote by Uij the unobserved utility perceived by individual i who chooses al-
ternative j. This utility may be modelled as follows:

Uij = x′ij βi + ϵij , (3)

where i = 1, ..., n, and j = 1, ..., J, index, respectively, individuals and alterna-

tives. xij is a k-dimensional vector of explanatory variables, βi is a k-dimensional
parameter vector and ϵij is a random shock known to individual i.
(a) Assume that the ϵij are distributed type 1 extreme value, that βi = β ∀i,
and that the choice set comprises three kinds of vehicles: large gas cars
(lgc), small gas cars (sgc), and small electric cars (sec). The probability
that a given household will choose each of these is, respectively, Plgc = 0.66,
Psgc = 0.33, and Psec = 0.01.

Suppose a subsidy for electric cars is introduced and that this subsidy raises
Psec from 0.01 to 0.10. Given the error assumptions, what are the proba-
bilities for lgc and sgc after the subsidy has been introduced.

To what extent are these probabilities unexpected?

(b) Consider the following discrete choice model

Uij = x′ij β̄ + x′ij β̃i + ϵij , (4)

2 of 9
 where individual i = 1, ..., N and j = 1, ..., J indexes individuals and
choices, respectively. Individual faces utility (U ) over two choices, deter-
mined by a single covariate xij .

Letting βi ∼ N (β̄, Σβ ) and ηij = x′ij β̃i + ϵij , where β̃i is mean zero, show
that the covariance matrix of the errors, Cov(ηi1 , ηi2 ), is given by
 2   
xi1 xi1 xi2 1 0
Ωη = σ β + σϵ .
xi1 xi2 x2i2 0 1
where σβ denotes the variance of βi .

Are all parameters in Ωη identified? Explain your answer.

A3 Consider the following unobserved effects panel data model
yit = α + β ′ xit + µi + εit (5)
where the µi ∼ i.i.d(0, σµ2 ) and εit ∼ i.i.d(0, σε2 ). xit denotes a K × 1 vector of
exogenous variables.
The random effects (re) estimator may be written as
" N N
#−1
1 X
′
X
′
βbre = X QXi + ψ (xi − x)(xi − x) (6)
T i=1 i i=1
" N N
#
1X ′ X
× X Qyi + ψ (xi − x)(y i − y) ,
T i=1 i i=1

where ψ = σε2 /(σε2 + T σµ2 ), Q = IT − T1 ii′ , i is a T × 1 unit vector, and IT is a

T × T identity matrix. Xi is T × K and yi is T × 1.
σµ2 and σε2 denote, respectively, the variance of µi and εit .
a) Making reference to (6), evaluate the statement that in assuming µi ∼
i.i.d(0, σµ2 ), βbre provides a solution intermediate between treating αi ’s as
different and treating them all as equal.
b) Let κ = βbre,s − βbfe , where s indexes time-varying regressors.

State the implied null and alternative hypothesis that follow directly from
b = κ′ [V ar(κ)]−1 κ.
the use of the test statistic H

Explain why one might want to consider a robust form of this test statistic.
c) A different formulation of (5) may be written as
yit = α + β ′ xit + ζ ′ x̄i + ai + εit , (7)
where ui = ζ ′ x̄i + ai , with ai ∼ i.i.d(0, σa2 ).

3 of 9 [TURN OVER]
 i) In what sense can we think of (7) as unifying alternate forms of linear
unobserved effects panel data models.
ii) How might you interpret the fe estimator of β given that this can
obtained using (7) as the estimating equation.
A4 Consider the following observational rule:

yi∗ = x′i β + εi
yi = τ (yi∗ ) = 1(yi∗ > 0)
εi ∼ N (0, σ 2 )

(a) Show that by the use of the e and m step of the em algorithm we can
solve for β using the normal equations

X X
x′i xi β 0 = x′i vi
i i

where β 0 is an initial estimate of β and vi = E[yi∗ | xi , yi , β 0 ]

(b) Consider the case where the variable of interest is one element of xi ,
say x1i . x1i is a binary variable and endogenous. Explain how the use
of a causal random forest can mitigate the problems encountered with
the use of both control function estimation and a 2-stage least squares
type estimator.
A.5 Evaluate four of the following statements
(a) A regression tree is honest if, for each training observation i, it only
uses a given observation to estimate the within-leaf treatment effect or
to decide where to place the splits, but not both.
(b) In the case of generalised random forests asking for individual tree pre-
dictions, as one would for a traditional random forest, is inappropriate
as it is not individual trees that make up predictions, but rather the
neighborhoods they induce.
(c) Causal trees and forests can estimate treatment effect probabilities
when the outcome is binary, using a nonparametric voting process
rather than modeling probabilities parametrically through something
like a probit model.
(d) Relative to the linear probability model, the advantages of the threshold
nonlinear binary response model are less when the design matrix is
sparse, and/or when there are endogenous binary regressors.

4 of 9
(e) The error components representation of a mixed logit model and the
random effects panel data estimator both generate non-iid covariance
matrices which may be exploited to solve different problems.

5 of 9 [TURN OVER]
 SECTION B
B.1 Denote by Uijt the unobserved utility of individual i who chooses alternative
j at time t. We write Uijt as

Uijt = x′ijt βi + ϵijt ,

where i = 1, ..., n, j = 1, ..., J, t = 1, ..., T index, respectively, individuals,

alternatives and time periods. xijt is a k-dimensional vector of explanatory
variables, βi is a k-dimensional parameter vector and ϵijt is a random shock
known to individual i, that has a Type 1 extreme value distribution.

Variables xijt are denoted collectively (for all time periods and alternatives)
by xi . βi ∼ g(β|θ) in the population, where g(.) is a mixing distribution
with θ a vector of hyperparameters.
We observe yi = (yi1 , ..., yiT )′ where yit = j if individual i chooses alterna-
tive j at time t.
(a) Write down an expression for the conditional and unconditional proba-
bility of individual i’s sequence of choices, respectively Pij = P (yi |xi , βi )
and Pij = P (yi |xi , θ).
(b) i. Without imposing any restrictions, what is the dimension of θ?
ii. One analyst proposed the use of k(β|θ ′ ), based on the deployment
of independent mixing. θ ′ denotes a different vector of hyperpa-
rameters.

In what sense is this a solution to a dimensionality problem?

If such a solution is imposed, outline how you would test the validity
of this approach.
(c) Estimating the parameters of the distribution g(β|θ) represents a sig-
nificant constraint on learning about where an individual customer’s β
lies in this distribution.

By considering the joint density of β and yi , what distribution should

the analyst estimate to resolve this and how does this relate to g(β|θ)?
(d) In Table 2 we present parameter estimates based on a stated prefer-
ence experiment designed to elicit consumer preferences over energy
suppliers.

The suppliers were differentiated based on price, the length of the con-
tract in years, whether the supplier was local, whether the supplier was

6 of 9
well known, whether the supplier charged time-of-day (tod) rates, and
whether the supplier charged seasonal rates.

A mixed logit model was estimated under the assumption that the
parameters are independently normally distributed in the population.
i. Interpret the results.
ii. Demonstrate how these results can be used to determine preference
heterogeneity for the length of contract
iii. Test whether a mixed logit model is to be preferred to a standard
Logit specification.
Table 2: Mixed logit model of choice among energy suppliers

θb σ
bθ

Price coeff. −0.8827(0.050) −

Contract −0.2125(0.026) 0.3865(0.028)

Local 2.2277(0.127) 1.7514(0.137)

Well-known 1.5906(0.100) 0.9621(0.098)

TOD 2.1328(0.054) 0.4113(0.040)

Seasonal 2.1577(0.051) 0.2812(0.022)

Log Likelihood Logit: −4959

Log Likelihood Mixed Logit −3619

Standard errors in parenthesis

θb denotes the mean and σ
bθ the standard deviation of each parameter estimated.
σ
bθ has been set to zero for price.

7 of 9 [TURN OVER]
B2 (a) An analyst has data on two covariates, X1 and X2 , and Y . A tree
partitions the covariate space X as shown in Figure 1.

Figure 1: An example of a tree

Show the equivalent partition on ℜ2

(b) A random forest for predicting the conditional mean

µ(x) = E[Yi | Xi = x]

can be summarised as aggregating each tree’s prediction to form

the final point estimate using
B n
1 XX
µ
b(x) = αbi (x) (8)
B b=1 i=1

where αbi (x) = 1({X|Li ∈L b (x)})

b (x)|
, and Lb (x) denotes the set of training
observations that fall into the same terminal node as x.

With reference to (8), carefully explain the distinction between ran-

dom forests (rf) and generalised random forests (grf).
(c) Let θ(x) denote a quantity of interest, defined by a local estimating
equation of the form

E(ψθ(x) (Oi )|Xi = x) = 0, (9)

where Oi refers to observable data, and ψθ(x) is a scoring function.

In the case where Oi = Yi , and θ(x) = µ(x) = E(Yi |Xi = x), the
conditional mean function, demonstrate the equivalence of predic-
tions for θ(x) using a rf and grf.

8 of 9
(d) Consider the following structural model

Yi = µ(Xi ) + τ (Xi )Wi + εi . (10)

Yi ∈ {0, 1} denotes an outcome and Wi ∈ {0, 1}. τ (Xi ) denotes the

causal effects of Wi on Yi . Given the assumed endogeneity of Wi ,
we also have access to an instrument Zi ∈ {0, 1}.
i. For Oi = {Xi , Wi , Zi } write down the moment functions which will
allow us to estimate τ (Xi ).
ii. In what sense does the estimation of a local linear instrumental
variable forest for (10) represent an improvement over a parametric
probit model when seeking to make causal inference on a binary
endogenous variable?

END OF PAPER

9 of 9