From: SYMON Paul [s1356194@sms.ed.ac.uk]
Sent: Thursday, March 20, 2014 7:26 AM
To: Andrew Rose
Cc: Schaffer, Mark E; alex.kiermeier@gmail.com; MITCHELL Stephanie;
SCHRÖDER Sarah
Subject: Surprising Similarities: Recent Monetary Regimes of Small
Economies
Attachments: Stata Rose.zip; ATT00001.htm; Edinburgh University charitable
status
Dear
Professor Rose,
We,
i.e. Alex Kiermeier, Stephanie Mitchell, Sarah Schröder and Paul Symon are MSc.
Economics students at the University of Edinburgh in the Scottish Graduate
Program of Economics (SGPE). As part of our programme, we are currently working
on an “Econometrics Project” (overseen by Prof. Mark Schaffer from Heriot-Watt
University) and set out to replicate and challenge the findings of your recent
working paper “Surprising Similarities: Recent Monetary Regimes of Small
Economies.”
First
and foremost we would like to thank you for allowing us access to your dataset
and log files, especially for keeping everything so comprehensive. Having
had a closer look at your dataset and your regressions, we highlighted some
points you may or may not have come discovered since we accessed your files and
wanted to share our results with you.
We
found some duplicates in your dataset that change your results fairly
insignificantly but thought you would still want to know. Attached in the zip
file are copies of our log and do files that compare your original regressions
in ptest1.log to the results when one drops the duplicates. We actually
identified the issue when we tried to xtset your data. The observation that
repeatedly crops up in the regressions is Sudan (2011) although this is not the
only duplicate in your dataset, however we are not sure if the others are in someway
intentional due to the different classifications of monetary regime. The other
log file include does exactly the same thing for the xtivreg regressions.
We
also have reason to believe that the model may be estimated in a fashion other
than the conventional Random Effects GLS. You may have noticed that in several
of your regressions the sigma_u=0. This suggests that Random Effects
degenerates to a Pooled OLS, which can be verified by using the ‘regress’ in
place of ‘xtreg’. Ordinarily one would proceed by performing a Hausman Test,
however as you mention fixed effects is not applicable due to the lack of
between orthogonality conditions. We test the consistency of Random Effects by
using Mundlak fixed effects where the dependent variable is regressed on the
explanatory variables and the mean of the explanatory variables. By using
‘xtoverid’ command (package xtoverid from http://fmwww.bc.edu/RePEc/bocode/x) we test
whether the additional orthogonality conditions of the Random Effects estimator
are overidentifying, thus rendering Random Effects inconsistent. The log file
is also attached for this procedure.
Our
suggestion is that the time-variant regressors are driving the rejection of the
null and that this can be avoided by estimating them using only the within
orthogonality conditions and the time invariant regressors using only the
between orthogonality conditions. This is what we are working towards and would
be more than happy to share our finished project with you if you are
interested?
Your
paper was thoroughly enjoyable to work with and we hope that you find our
insight useful. We would sincerely appreciate any feedback you may possibly
offer.
Alex
Kiermeier, Stephanie Mitchell, Sarah Schröder and Paul Symon
PS-
One last small point, in Table 8 where you present the p-values for the
Hypothesis test there is a typo with regards to the number of significance
stars. There is a p-value in the third column second row (Income=0) that is
equal to 0.14 that has been marked as highly significant (**). We
double-checked and the p-value is correct it is just the stars that are not.