Rob,
The data structure you describe is discrete-measured continuous duration
data. Spells that are measured in this way are properly treated as
interval-censored duration data. There is a large literature in
biostatistics on estimation of survival models given follow-up measurement
schemes. You know the maximum and minimum of the underlying continuous
duration, but don't know the exact time. If an opinion spell was measured
to begin in the kth wave and end in the jth wave, the duration is at most
j-(k-1) (i.e. started just after the measurement prior to k and changed
exactly at j) and at minimum (j-1)-k (i.e. started right at k and changed
right after the wave before j). Your dependent variable is an interval
[j-(k-1), (j-1)-k], within which you know the true duration lies. If the
waves aren't evenly spaced, the interval calculation should be adjusted for
the time between the relevant waves. There is an R package 'intcox', which
has an implementation of an algorithm for the CPHM with interval-censored
data. The references in the package documentation provide good background
reading on this topic. If the average duration - in discrete waves - is
less than 20 or so, parameter estimates will be inconsistent and p-values
will be biased downwards by as much as 30%. This is true if you simply use
the discrete measurements in a continuous-time model, or in a discrete-time
event history model. I have a working paper on this subject, so feel free
to contact me off-list if you'd like to see an aspiring political
scientist's take on the matter.
Regards,
Bruce
--
Bruce A. Desmarais
Ph.D. Candidate
UNC Chapel Hill
Department of Political Science
www.unc.edu/~bdesmara
On Sun, 17 Jan 2010 11:14:21 +0000, Robert Johns
<[log in to unmask]>
wrote:
> Dear all,
>
> I've got a query about the Cox hazard models used by Gartner (APSR, 102:1
> (2008), pp. 95-106). He runs a panel experiment in which monthly
casualty
> rates during a conflict are manipulated, and the hazard models gauge the
> impact of these manipulations on the duration of public support for
> military action. Gartner's panel has ten waves, i.e. ten opportunities
to
> choose between leaving the troops in or bringing them home. In planning
a
> similar design, I was wondering about the amount of flexibility available
> concerning this number of waves. I guess that, under the model's
> assumptions, duration is a continuous variable, but obviously it is
> necessarily discrete in this kind of design. So can anyone suggest a
rough
> rule on how many time periods would be needed to estimate such hazard
> models accurately?
>
> Many thanks,
>
> Rob Johns.
>
> **********************************************************
> 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
**********************************************************
|