UC San Diego Econometrics 120C
Department of Economics Winter 2022
STATA Assignment 2
Important instruction before you begin
The deadline to submit the assignment is Friday February 25, 11:59 PM. The dataset for this
problem set is analysis1.dta. As before, we will provide a basic do file called PS2 yourPID.do.
Please rename with your PID so the file you turn in resembles PS2 A34567890. Only the final do
file needs to be uploaded to Canvas. You will solve this problem set by modifying the do file
only. Ideally, your .do file should be in the same folder or directory as the data file so you do not
need to use cd to switch to the needed directory. If you need to switch directories, put the word cap
at the beginning of the command so it will not execute when we try to run your code.
The do file has comments indicating where your answers need to be written. For the questions
which require an answer in words, write the answer where indicated in the do file. Remember that in
Stata, commands beginning with a * or wrapped like /*my command*/ will not execute. Make sure
your code runs.1 If it does not run, then you will have 25% subtracted from your final
score. We highly recommend running your do file before submitting it to make sure it runs.
Background
This assignment is based on the paper “From Mad Men to Maths Men: Concentration and Buyer
Power in Online Advertising” (2021) by Francesco Decarolis and Gabriele Rovigatti published in the
American Economic Review. The authors motivate their analysis as follows:
Online advertising sales are the main fuel of all of the major digital platforms. In the in-
ternet era, advertising means capturing the attention of consumers who are browsing the
web and this requires both detailed data to effectively target the ad to the right customers
and algorithms to bid in the online auctions where ad space is sold. These needs have
led to a major, but understudied, shift in the industry: rather than bidding individually,
advertisers increasingly delegate their bidding to highly specialized intermediaries. This
concentration of demand within a few large intermediaries raises the question of coun-
tervailing buyer power. Can the emergence of intermediaries counterbalance the highly
concentrated supply of online ads? (pages 3299-3300)
The authors establish causality by leveraging a novel data set, natural language processing, and
instrumental variables.
1By “run”, we mean that the TA can click “do” in the do-file editor and the whole do-file runs through and produces
the desired results and produces all the requested output.
1
The third ingredient is an instrumental variable (IV) strategy. Instruments are needed
for two reasons: measurement error in the proxy for demand concentration and potential
omitted variable bias. For instance, there might be unobservable shocks to the popularity
of some keywords that drive changes in both revenue and demand concentration. Similar
to Dafny, Duggan, and Ramanarayanan (2012), we address this problem by exploiting
the variation in intermediary concentration driven by changes in network ownership of
MAs. In our sample period, there were 21 acquisitions and 2 divestments, affecting 6 out
of the 7 agency networks. These merger and acquisition (MA) operations, especially the
larger ones involving a multiplicity of markets, are a useful source of variation in demand
concentration as the revenue dynamics in each local market are too small by themselves
to cause the MA operations. We extensively discuss this empirical strategy and evaluate
its robustness. (page 3301)
Now your task is to replicate some simple aspects of the author’s study and try to understand how
IV helps to resolve the measurement error and omitted variable bias problems.
Questions
1. The dataset analysis1.dta2 contains the main data used in the paper. The outcome of interest
is Google’s estimated log revenue logr hat (measured in log dollars). This variable is an estimate
for Google’s revenue from advertising on searches with a specific cluster of keywords. Think of
a cluster of keywords as a “market” or industry.
The covariate of interest is a Herfindahl-Hirschman index (HHI) HHI hat. This variable measures
concentration on the demand side of the market. Firms want to advertise their products beside
Google search results. If these firms are competitive, then HHI is low. If firms are oligopolistic
or not competitive, then HHI is high. HHI ranges from 0 to 1.
Merger and acquisition operations mean that firms combine or break apart. When this happens,
the market concentration changes. The instrument variable is sim, the simulated change in
market concentration (HHI) that occurs from merger and acquisition operations.
Load the dataset into STATA, get an overview over the dataset (STATA: describe and sum).
2. Create a scatter plot showing the correlation between HHI hat and sim. Include a linear predic-
tion in your graph. Explain the intuition of this figure.
3. Now consider the model:
logr hatmt = α+ βHHI hatmt + ϵmt, (1)
where m indexes markets and t indexes years.
2Decarolis, Francesco, and Rovigatti, Gabriele. Data and Code for: From Mad Men to Maths Men: Con-
centration and Buyer Power in Online Advertising. Nashville, TN: American Economic Association [publisher],
2021. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2021-09-29.
https://doi.org/10.3886/E130502V1
2
(a) Run an OLS regression of revenue in the market-year on market concentration (HHI) using
robust standard errors.
(b) Interpret the coefficient β. Is it statistically significant? Is it economically meaningful?
(Answer in max. 2-3 sentences)
(c) According to the paper, what is the potential problem that could cause OLS to be incon-
sistent? (Hint: Read the Identification Strategy section on page 3316.) (Answer in max.
2-3 sentences)
4. As explained before, the authors use mergers and acquisitions as an instrument for changes in
market concentration.
(a) Run TSLS manually (i.e., run the first stage, and then run the second stage, without using
the automated IV command) and compare the TSLS estimate to the OLS estimate.
(b) Run TSLS using the automated STATA command ivregress 2sls.
(c) According to the authors, what is the reasoning for the validity of this instrument? (Hint:
Read the Identification Strategy section on page 3317.)
(d) Should we worry about weak instruments in this application? Conduct a formal test.
(e) The authors consider the more general model:
logr hatmt =α+ βHHI hatmt (2)
+ βnnumberofresultsmt + βbbrandedmt + βllong tailmt
+
∑
m
γm1{market = m}
+
∑
t
γt1{year = t}+ ϵmt,
where numberofresults is the number of search results (in millions), branded and long tail
represent keywords that may be endogenously used by advertisers, 1{market = m} is
a dummy variable that takes the value of one for market m and zero otherwise, and
1{year = t} is a dummy variable that takes the value of one for year t and zero other-
wise. Estimate (2) (use the automated TSLS STATA command). (Hint: You need to
create the dummy variables for each of the markets identified by the variable numind us-
ing tabulate numind, generate(dummies numind) and each of the years identified by
the variable year using tabulate year, generate(dummies year). Then include them
as covariates in your estimation using dummies*.) Interpret your results (use a causal
interpretation) and explain why the authors include these additional covariates?