[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

RE: st: correcting standard errors for model with repeatedobservations in same time period

From   Adrian de la Garza <>
To   <>
Subject   RE: st: correcting standard errors for model with repeatedobservations in same time period
Date   Wed, 7 May 2008 12:15:46 -0400

Dear Austin,

Yes, I have about 40-something countries so I do agree that it makes sense to cluster by country and/or time.

I do believe that I also need to account for selection (missing obs when no bonds are issued) given that we only observe the spread of a bond when a bond is indeed issued... otherwise we don't observe anything... and at all times, countries have the option to issue or not issue a bond. Thiis is the typical selection problem where we observe a wage for a working person but not for someone who chooses not to participate in the labor market, so when you regress wages on various individual characteristics of workers, you need to account for selection to correct for this bias.

Anyway, that is not the problem I'm referring to here anyway. My question is what I do about repeated observations in a same time period (year, quarter) when I do observe spreads? For example, the fact that for any given time period I may see 3 Mexico issues or 4 Brazil issues?

Thank you once again.


> Date: Wed, 7 May 2008 08:11:37 -0400
> From:
> To:
> Subject: Re: st: correcting standard errors for model with repeated observations in same time period
> Adrian de la Garza :
> I can certainly see an argument for clustering on country (assuming
> you have 50 or more in your data) to account for serial correlation
> within country, or clustering on time (
> g time=yq(year,quarter)
> format time %tq
> ) to account for the correlation of errors across countries at a point
> in time, but I can't see why you need to account for missing obs when
> no bonds were issued.
> Maybe you want two-way clustering along the lines of
> On Wed, May 7, 2008 at 1:14 AM, Adrian de la Garza
>  wrote:
>> Dear all,
>> I am running a linear regression that looks like this:
>> bondspread = a + macro*b + covariates*c + public/private_dummies + country_FE + e
>> where bondspread is the spread of a bond issued by a public or a private entity in a given country in period (year, quarter); macro are a bunch of macroeconomic variables such as the growth of GDP, total external debt/GDP ratio, etc.; covariates are a bunch of individual characteristics; and I have a dummy that tells me whether the institution issuing the bond was public or private, and country fixed effects (and maybe I will add year FE later and other things... but this is what I have now).
>> The problem is that I think that I should correct the standard errors of a simple OLS because one same country may issue multiple bonds in a given (year,quarter)... so, what I am doing is to cluster by country or at least by region (like "Latin America" or "East Asia" and such)... but I think I should do something else. Does anybody have an idea of how to make such correction or whether it is necessary at all?
>> ...So, basically, Mexico can issue one or multiple bonds in a given period (because one of multiple different companies and local governments decide to issue debt) or can simply NOT issue at all (when spread is missing).
>> I know I probably need some sort of selection model (a la Heckman) to account for the fact that some countries may or may not issue at any given time... but do I need to account for the clumping of issues at any given time? And if so, how?
> *
> * For searches and help try:
> *
> *
> *

Stay in touch when you're away with Windows Live Messenger.
*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index