-------- Original Message --------
Subject: Re: question about abstract
Date: Sat, 5 May 2007 05:39:39 -0400
From: astokes at
camail.harvard.edu
To: Holger Lutz Kern <hlk23 at cornell.edu>
References: <1178323375.463bc9af365ca at
www.camail.harvard.edu>
<463BE28A.1040201 at cornell.edu>
Authors: Tom Laakso and Andy Stokes
Title: Back to the Future: Revisiting the Global Mortality Projections of
Mathers and Loncar Through Bayesian Priors
Abstract: Mathers and Loncar (2006) use the last decade of global vital
registration data to update worldwide projections of deathsthrough 2030.
Their
basic model strives to predict mortality based on just a few demographic
quantities, but is reliant uponextensive by-hand adjustment of regression
outcomes. We attempt to replace these after-the-fact corrections with a
Bayesian
approach to prior beliefs, as implemented in Girosi and King's (2001)
YourCast
package. By specifying a number of priors, we obtain a measure of
specification
uncertainty in the projections. We find that, on average, the 2030
age-specific
mortality rate can only be specified within plus/minus 50% of its measured
value. This far outweighs the plus/minus 10% allowed for in the original
study,
given different global growth scenarios. We then mapped this variance
country-by-country for each unique combination of age and
cause-of-death. While
the specification uncertainty does show strong geographic patterns,
these are
complicated functions of both age and cause-of-death. Therefore, the model
appears to be robust to prior-specification only in very specific subsets of
the global population.
Quoting Holger Lutz Kern <hlk23 at cornell.edu>:
Andy,
just send it to me; I'll forward it to the list...
Holger