Test after sureg

[Elvira Kasymova]

I am trying to replicate the paper of Schäfer, Schnabel, DiMauro "Bail-in expectations for EU banks: actions speak louder than words" (2016) (lhttps://www.esrb.europa.eu/pub/pdf/...d6db2c5105f789) where they implement SUR on stock returns and CDS spreads for 85 banks to check for abnormal returns at a certain date.

Their system of equations looks like that

R(i,t) = alpha(i)+beta(i)*Rm(t) + D(t)*tau(i,t)+u(i,t)


where Rit is a return on equity, Rm - market return, Dt - dummy equals to 1 at day when event of interest happened. Date of events are similar to all 85 banks, tau(i,t) - estimated coefficient that shows abnormal return at the day of interest.

What puzzles me is a test after SUR regression. The authors report that they test whether average abnormal returns significantly different from zero and report p-value. I didn’t find a command for such a test.

Right now I run sureg on 85 banks:

global eqn1 (Ret_bank_1 Ret_market Dummy_event1 Dummy_event2 … Dummy_event15)
...
global eqn85 (Ret_bank_85 Ret_market Dummy_event1 Dummy_event2 … Dummy_event15)

sureg $eqn1 …$eqn85

I get the coefficients but don’t know how to run a test I need on them.


Thank you very much in advance!

[Clyde Schechter]

The -test- command does this. It offers several different syntaxes depending on whether you are trying to test coefficients being 0, linear expressions being 0, all coefficients in one of the equations being 0, or equality of coefficients across equations. It's not clear to me which of those you want to do, but if you read the help file you will probably recognize what you need. Note that you may need to replay your -sureg- results with the -coeflegend- option to see the correct names to use to refer to particular components of particular equations.

[Elvira Kasymova]

Dear Clyde,

Thank you for your response. I checked the help file, but I didn't find the test I'm looking for. Which is basically to take all the banks as a group and check whether the average coefficient is equal to zero.
H_0: 1/N*[Sum(tau_i)]=0, where N-number of banks, tau_i - coefficient of Dummy_event_day for each bank=abnormal return


Usually in the event studies one would do it by running a regression on abnormal returns as those are saved as a separate variable - the difference btw expected and real return [AR=E(R)-R] as in widely used princeton example (http://dss.princeton.edu/online_help...ntstudy.html):

reg abnormal_return if event_day==1, robust

However in my case, abnormal returns are measured by the coefficients of Dummy_event_day, so I don't have them as variables and can't run a regression.

As this method is used by Schäfer, Schnabel, DiMauro, there must be a way to do it, but I can't find it. Maybe someone knows the solution?

[Clyde Schechter]

H_0: 1/N*[Sum(tau_i)]=0, where N-number of banks, tau_i - coefficient of Dummy_event_day for each bank=abnormal return

Sorry, but I don't know what you're saying here.

I think you need to show the -sureg- output (with the -coeflegend- option) and then give a clearer indication of what coefficients you are referring to.

[Elvira Kasymova]

Sorry, I am new here, I will try to be more precise.

So, the dependent variables are CDSspreads for 85 banks, dependent variables - Dummies for the event that happened for all of the banks at the same day. As it is written in the paper, the coefficient of Dummies eventdaypre_Denmark, eventday_Denmark and eventdaypost_Denmark give me the abnormal returns at the day before, at and after the event of interest respectively.

The Null Hypothesis is that the average abnormal return across all the banks at the day of the event is equal to zero.
Meaning that the average of _b[eventday_Denmark] across all the banks is equal to zero.

Below is an example of sureg output for 3 banks out of 85.

I am very grateful for your help!

[Daniel Abler]

Dear Statalist Community,

I have the same question as the post creator. Basically, a SUR regression is run for 85 banks using the following structure specific to the respective event date.

local subsetvar nHSBCHOLDINGS-nVPBANK
sureg (`subsetvar' = nStoxx50 eventdaypre eventday eventdaypost) if inrange(mdy(Month,Day,Year),mdy(11,28,2012),mdy(3,26,2013))

The goal is then to calculate the average of the coefficient "eventday" across all 85 regressions. I have done this so far. Now the researchers mentioned in #1 apply a test to see if this average value is statistically significantly different from zero. What do you think, how to run such a test in Stata? Personally, I had two ideas: 1) the command

test MeanCoefficient=0

2) Stata's Meta-analysis tools. However, with 1) I got the error message below since the newly generated variable only is a constant and with 2) I am unsure if this is really the same procedure as mentioned by the authors in #1.

MeanCoefficient not found

For completeness, this is how I calculated the average of the coefficient:

mat b=e(b)
local coleqs: coleq(b)local coleqs: list uniq coleqs
foreach eq of local coleqs{
    replace numerator=numerator+_b[`eq':eventday]
}
replace MeanCoefficient=numerator/e(k_eq)

Thanks for the help!