Econometrics 2

305-0007-SYW

Q.1. “Market power and banking regulations: Evidence from RDD application to the Brazilian banking
market?”

In this first paper De Genaro et all propose to study how increase in banking regulations lead to cost
transfers of said regulations to consumers, by banks. More specifically, the authors propose to study said
effect through a Regression discontinuity model, arguing that the Total Assets/GDP = 10% ratio set-up by
an international committee to monitor banks acts as a random threshold, creating a discontinuity
between different sized banks. As soon as a bank surpasses the 10% threshold it needs to maintain
additional capital in the range o 1-2.5% of their risked weighted assets, which the authors claim it’s an
additional cost, which banks will try to pass to consumers.

The authors use a Sharp RD, since they assume that as soon as a bank TA/GDP >10% it is immediately
subject to monitoring, and as such it incurs the cost.

The main assumptions of a RD are: Smooth outcomes in the absence of treatment: should there be no cut
off, the outcome would be unchanged; No manipulation: we need to assure that there is no manipulation
around the threshold and in case we cannot feasibly prove the population is the same we need to control
for specific individual/group characteristics.

Starting by analyzing the assumption of smooth outcomes some doubts arise about the population itself.
For an RD to be valid the population we are studying should be similar, or have the same characteristics.
However, we can see, per example in Fig.1, how the population of “banks” is markedly different. The small
banks are nowhere near the cut-off, and most “large” banks are well above 10% (more around the 15%
TA/GDP level), with such an asymmetric distribution the reasoning that small and large banks are similar
in cost and revenue structure might not hold. Furthermore, looking at table 2 we can see how even the
structure/characteristics of the 2 “large” banks are quite different, with income fee share, spread, ROE
and ROA varying a lot from bank to bank. It would have been interesting to see a table regarding banking
characteristics and structure to the left and right of the cut-off in order to prove that the population is the
same, and that characteristics of the banks are similar both sides and do not define the TA/GDP ratio and
treatment.

In regards to Manipulation, a strong argument can be done regarding one bank: Santander. The authors
do not state clearly, but by omission they imply that banks do not control their TA/GDP ratio. It is a
reasonable assumption if we assume they do not know what GDP of a given year will be, and that they
don’t have full control of their Assets. However, this thesis has some holes. GDP estimates are constantly
published, and it is clear banks have some power on defining their Assets since they are the deciders, of
per example, the amounts of loans granted in a given year. So the assumption that a Bank TA/GDP ratio
is random is flimsy. Particularly, if we look at Fig.1 we can see how Santander crosses the threshold around
3 times, only to go below the Ratio again. Reasonable arguments could be made that Santander was
working in a way as to avoid being above the threshold, and thus not incurring in additional costs.

De Genaro et all seem to recognize that their banking “population” has different characteristics that vary
dependent on the cut off, which they should control for, and as such control for Total Assets, HHI, financial
margin, non-performing loans, GDP, percentage change of GDP and Equity. I believe some controls in

1
Econometrics 2
305-0007-SYW
regards to banking structure could also be reasonably added, like Spread and Fees income share, as well
as Costs with personnel.

Still regarding controls, it would have been interesting, if possible, for the authors to control for regional
activity. It seems plausible that small banks have a very niche market there are serving, they might be
rural banks, or expatriate banks, so controlling for the region of operation/revenues, could have brought
some interesting results.

Also, regarding methodology, the authors leave what seems like an important topic unaddressed: serial
correlation between observations and the error term. A bank’s previous Total assets is correlated with
the previous period total assets. A country’s GDP is also correlated with previous year/quarter GDP. What
impact this could have on inference is not addressed by the authors.

The data the authors collect seem to be valid, assuming no manipulation of it by the banks themselves.
They have a large dataset, with 192 individuals, with 4 observations each per year, from 2009 to 2010.
Their data, however, as we have argued before lacks a very important component: observations around
the cut-off. Santander is always close to the threshold, surpassing it sometimes (which as we have
discussed might mean manipulation happening); we also have Bradesco with some observations close to
the threshold, and for some initial years Caixa Economica Federal.

The lack of reliable data around the cut-off becomes apparent when we look at Fig. 2 and Fig.3. In Fig. 2,
regarding the observations to the left of the threshold, we can distinguish two different clusters. One far
left, and one tightly packed around -0.0025 and 0 of the cut-off. The far-left observations seem to pertain
to the small banks, and the cluster close to the threshold to Santander (and probably Caixa Economica
Federal for one period). The “regression” line presented for the observations to the left of the cut-off is
dubious, at the very least. If we were to reduce the bandwidth around the cut-off to -0.05 and 0.05,
including only the “Santander” observations (the ones clustered below the cut-off) it is clear that no RD is
present. The trend to the right and left of the cut-off seems continuous. The same can be said of Fig.3,
even more markedly. If we were to remove the outliers from Fig.3 we would have a continuous regression.
Or, if we decreased the bandwidth to around -0.05 and 0.05 of the cut-off, no discontinuity seems to be
present.

Furthermore, some other effects might be happening that would lead to increase in fees by large-banks.
The authors analyze 2009 to 2019, a period which still faced some global instability, with per example the
European Debt crisis happening in 2010 to 2012, depending on the way to measure it. It is clear that large
banks would have more exposure to international Assets, obligations, treasury certificates, and so on. So,
the increase in fees or spreads by large banks could be a compensation for profits lost elsewhere, rather
than costs incurred by additional Capital.

So, to conclude, the lack of observations around the cut-off makes it hard to prove that an actual
discontinuity occurs. Furthermore, the differences between populations to the left and right of the cut-
off make inference dubious and the internal validity of the paper can be contested on those terms.

2
Econometrics 2
305-0007-SYW
Q.2.

Educated to be trusting? Evidence from Europe

In this study the authors aim is to prove the relationship between education and trust. Kan and Lai argue
that education, quality and/or quantity increases individuals social capital, in this case measured by trust.
The Authors use a survey response from the European Values study to measure trust and individuals’
socio-economic variables, and then use education as the explanatory variable to trust. A 2SLS model is
presented, where mandatory education is regressed as the independent variable in years of schooling in
the first stage, and then years of schooling is used to predict trust in the 2SLS. So, an IV method is utilized.

The variable of interest is trust, which is takes the value of 1 or 0, depending on the answer of the
individual to the question: “generally speaking would you say most people can be trusted or that you can’t
be too careful in dealing with people?”. This measure can already rouse suspicion: the terms of the
question are dubious, and not very specific, individual understanding of this statement can vary greatly.
Furthermore, given a dependent variable that can be between 0 and 1 a question should be asked has to
why not using probit, tobit, or logit models would not be more appropriate rather than a Linear probability
model (which this study ends up being, in a way).

The authors argue the use of mandatory schooling as an instrument on the basis of it being an exogenous
variable, affecting years of schooling, but not trust. This is a way to quell their fears that schooling and
trust are endogenous. We will discuss further ahead whether this instrument fully destroy’s endogeneity.
The authors also use a treatment strategy in which an individual is treated if it was born within 7 years of
after the first-affected cohort (in their respective country).

In an IV methodology, the key assumptions are: Exclusion Restriction, the instrument only affects the
dependent variable outcome through it’s effect on the endogenous variable; Relevance: The instrument
affects the endogenous regressor; Exogeneity: The instrument is as good as randomly assigned.

Regarding the first assumption some doubt on it’s validity can be casted. Trust of an individual in society
motivates him (or not) to act socially and improve society, mainly through public, democratic and
government activism. Higher trust leads to higher requirements for socially provided goods, so higher
trust would lead to higher requirements on education, in per example, demanding that mandatory
education increases in years. So, the relation between the instrument and the dependent variable can be
endogenous.

Relevance of the instrument seems valid, it is clear that higher mandatory schooling leads to more years
of schooling, a point proven by table 2.

As for the assumption of exogeneity, and the randomness of the instrument, the argument presented
against exclusion restriction also applies here. Changes in mandatory schooling are not arbitrary dictated,
they are the conclusion of long public debates and demands, and such public debates only happen with
social interaction and social trust.

Also regarding the methodology, the authors state they control for country and year specific variables,
but we are left wondering which. The presence of GDP per capita, institutional quality, crime rate and so
on, are incredibly valuable in understanding trust differences between countries. The authors also do not

3
Econometrics 2
305-0007-SYW
mention what individual characteristics are controlled for. Socio-economic factor could play a greater role
in trust, rather than education, and if such is not controlled some doubts on the estimates is possible.

The study variables of interest and dependent variable are measured in individual terms, however there
is a strong argument that in-between group correlation is present. People on the same country will
probably have correlated opinions on trust. As such some care should have been made in terms of
standard errors, such that we could eliminate the variation coming from intra-group correlation between
observations.

As for the data, the authors use the EVS survey result, a cross-sectional data for 4 different periods, as
well as reported data of “educational reform”. It should be noted the authors only mention one
educational reform over the whole period (1981-2008), which is rather strange. It is clear that the
European countries presented, by 2008, had a mandatory education of around 12 years. That no other
“educational reforms” are mentioned, neither their impact sounds misplaced.

It should also be noted, as for the data, that two of the countries with lower mandatory education,
Portugal and Spain, and some of the lower trust levels, were, at the time of this “educational reform”,
under a dictatorship. Furthermore theses “treated” cohorts lived under a dictatprship for around 10-15
years. As is apparent living and growing up in a dictatorship could have higher impact on “trust” than
education years. Given that Portugal has a particular low level of trust, this could be pushing the results
to “significance”, but rather by the political environment ono growing up than by education itself.

In table 3 we can see the paper results presented. The lack of controls other than age and gender is
worrying. It is not clear whether controls were omitted or simply not inserted. Country GDP per capita,
crime rate and so on could be highly explanatory of trust, as well as individual characteristics like income.
Not presenting the results or mentioning them can make one wonder on the validity of the results, and
whether some p-hacking might be present.

In the paper the authors also discuss horizontal vs vertical education. They find evidence only for one of
these types of education producing results on trust. The full strategy for defining type of education is not
mentioned, and again, the controls are missing.

In conclusion, the results of the study can be questioned due to the dubious validity of the IV key
assumptions present. The exclusion restriction and exogeneity assumptions are not given a compelling
case on why they hold up, and the authors themselves recognize such in the Conclusions, pointing that
the results are “suggestive”, since they assume exogeneity in teaching. Some more data concerns we
have raised also muddle inference.