Applied Econometrics Assignment
1. You are interested in the effects of decriminalising illicit drugs. Between 2001 and 2002 the
London borough of Lambeth experimented with such a strategy: the borough decriminalised the
possession of cannabis in July 2001. (The policy was subsequently reversed in July 2002, but in this
question we are simply concerned with the immediate effects of the July 2001 reform.) Apart from
boosting demand for cannabis in Lambeth during the experiment period, one other plausible
effect of such a policy is that the police may reallocate effort to other types of crime, such as
burglary.
(a) You decide to investigate whether there is an increase in police effectiveness for non-drug
related crime by estimating the impact of the policy on burglaries. You do this by comparing
outcomes for Lambeth (the ‘treatment’ borough) with the 31 other boroughs (the ‘control’
boroughs). You have monthly borough-level data available and using data for the month of June
2002, you estimate the following model:
where for each of the 32 London boroughs, , y is burglaries per 1,000 residents,
= 1 for Lambeth
and is 0 for the other 31 boroughs, x is the unemployment rate in percent and w is average
household income in pounds.
What is the parameter of interest in this model? Why is it useful to include x and w in the model?
Carefully interpret the effect of a change in unemployment and household income in this
regression. [5 marks]
(b) Do you expect this regression analysis to yield the causal effect of the decriminalisation policy?
Explain in detail why or why not. [10 marks]
(c) You then notice that you also have ‘pre-treatment’ data from May 2001. Carefully explain how
you would use the data from the two time periods (May 2001 and June 2002) in order to exploit
2
this ‘within-borough variation’ in order to implement a Difference-in-Difference model. (Hint: As
before, Lambeth will be your treatment borough; the other 31 boroughs will be the control
boroughs.)
In your answer, explain what is meant by ‘within-borough variation’ in the data; write
down the Difference-in-Difference model; state the key assumption(s) which must be satisfied in
order that the Difference-in-Difference identifies the causal effect of the policy on burglary rates.
In your answer also explain how this empirical strategy improves upon the strategy outlined in (a)
above.
[15 marks]
Part B (70 marks)
Your answer to part B should be typed (you may write or draw equations and graphs by hand,
if they are needed); it should be no more than 2 pages, using 12pt font.
Read the paper “Do CEOs Set Their Own Pay? The Ones Without Principals Do” by Bertrand and
Mullainathan (NBER WP 7604, 2000). You can ignore section 4 (“Charge for Options”).
[Note that this paper covers a lot of ground. You have limited space for answers and so you should
stick as close as possible to the specific tables and regressions mentioned in the questions below.]
(a) CEO’s are very well rewarded and we would like to think that this is for their high value added
to the firm. Bertrand and Mullainathan want to learn whether they are in fact rewarded simply
for good ‘luck’.
(i) Very briefly, explain the intuition (generated by the simple principal-agent model) that
CEO pay should be related to the firm’s performance. (ii) Thinking about the data the authors use,
how do the authors propose to measure firm performance? How do they measure CEO pay? (iii)
For the oil industry study what is the authors’ measure of luck? (iv) In theory, how should luck
affect CEO pay? [15 marks]
(b) What does the suggestive evidence presented in Figures 1 and 2 tell us? How might regression
analysis improve on this graphical analysis? [5 marks]
(c) Before investigating the role of luck, the authors test the relationship between company
performance and CEO pay.
(i) Suppose you are also interested in such a relationship. You have data from a single year,
say 1997, on firm performance, CEO pay, as well as firm- and CEO-specific variables such
as firm size and CEO tenure. Thus you can estimate an equation such as model (2) on p.10
of the paper. However, a key difference is that you have only one year of data. How does
this limit your empirical approach or strategy? [10 marks]
(ii) Now suppose that you have two years of data, from 1990 and 1997, say. Explain how
you would now estimate the pay-performance model using first differences. In particular,
explain how you would deal with firm fixed effects. What are the advantages of this model
compared to the one you estimate in (i) above? Finally, and briefly, do you include year
effects in your regression? If yes, what do they tell us? [10 marks]
(d) Let’s now think about the model estimated by the authors in column (1) of Table 1 (the
‘General’ specification).
(i) First, how does this estimated model differ from your first differences model discussed
in c(ii) above? Second, precisely interpret the effect of age and tenure in this model.
[5 marks]
(ii) The key parameter in this model is the coefficient on the accounting rate of return.
Precisely interpret this effect and comment on the economic significance of this finding.
[5 marks]
(e) One of the most important results in the paper is presented in Table 1, column (2). Explain
precisely the key question being addressed here, the intuition behind the empirical approach
adopted by the authors (comment on the IV or two-stage least squares approach) and the main
conclusions which follow from this one regression. [10 marks]
(f) The authors are also interested in how the quality of the firm’s governing board and CEO tenure
influence pay for luck. Provide intuition behind the regressions in Table 4, columns (1) and (2) (on
governance) and Table 5, columns (1) to (4) (on CEO tenure). What do you conclude about the
effect of these two mechanisms on pay for luck? [10 marks]