Dear Rachel,
In the case of discrete time event history analysis, it is better to use
robust standard error as you suspect. Though I don't know SAS command,
"clustering on country" sounds fine.
On the other hand, inclusion of duration measure is another issue. It
addresses not violation of independence assumption but time dependence
of the probability of event. Usually, you'd better include it as well.
Best,
Kentaro
****************************************************************
Kentaro FUKUMOTO, Ph. D.
Professor, Department of Political Science, Gakushuin University
tel:+81-3-3986-0221 ext.4913 fax:+81-3-5992-1006
http://www-cc.gakushuin.ac.jp/~e982440/index_e.htm
****************************************************************
(2010/07/06 0:49), Rachel Sullivan Robinson wrote:
> Dear Colleagues,
>
> I am in search of a little advice and hope that you might be able to help. I have a discrete-time event history analysis and want to make sure that I am calculating the standard errors for the regression coefficients correctly. The details:
>
> - Discrete-time event history analysis based on logistic regression, using SAS
> - Units of observation are five-year periods for sub-Saharan African countries
> - Dependent variable is binary indicator showing whether a country adopted a population policy during a given period
> - Independent variables are both fixed and time-varying
> - 41 countries are observed for a minimum of two, and a maximum of four, periods, producing 115 observations.
>
> Because each country is included more than once, the data fail the independence assumption, so regular estimates of standard errors will be biased. I have identified several possible solutions:
>
> - Use the PROC SURVEYLOGISTIC command in SAS, clustering on country (this actually produces smaller standard errors, which is a happy outcome)
> - Use the PROC GENMOD command in SAS, with a logit link function and country as the "repeated subject"
> - Use some other program (like Stata) to run the regression that is more amenable to including robust variance estimates (e.g., Huber-White)
>
> My question is whether any one of these three options is better than the others?
>
> Also, I have not included any measure of duration in the model. Some of the literature I've read seems to suggest that doing so helps address the impact of the lack of independence between observations on calculations of standard errors. Should a measure of duration (either dummy variables for each period or a variable for the number of prior events) be included to facilitate calculation of errors?
>
> If you have any suggestions for the best way to estimate the standard errors in the situation described above, and whether doing so requires including a duration variable, I would greatly appreciate it.
>
> Thank you very much,
> Rachel Robinson
>
> ~~~~~~~~~~~~~~~~~~~~~~~~~~
> Rachel Sullivan Robinson
> Assistant Professor
> School of International Service
> American University
>
> 4400 Massachusetts Avenue, NW
> Washington, DC 20016
> Tel: 202-885-1636
> Fax: 202-885-2494
> Email: [log in to unmask]
> **********************************************************
> Political Methodology E-Mail List
> Editors: Xun Pang<[log in to unmask]>
> Jon C. Rogowski<[log in to unmask]>
> **********************************************************
> Send messages to [log in to unmask]
> To join the list, cancel your subscription, or modify
> your subscription settings visit:
>
> http://polmeth.wustl.edu/polmeth.php
>
> **********************************************************
>
>
**********************************************************
Political Methodology E-Mail List
Editors: Xun Pang <[log in to unmask]>
Jon C. Rogowski <[log in to unmask]>
**********************************************************
Send messages to [log in to unmask]
To join the list, cancel your subscription, or modify
your subscription settings visit:
http://polmeth.wustl.edu/polmeth.php
**********************************************************
|
