[18655] in s-news-athena
Re: [S] E & M Routines or Software
daemon@ATHENA.MIT.EDU (Simon Jackman)
Sat Jul 17 06:17:14 1999
Date: Sat, 17 Jul 1999 03:10:43 -0700 (PDT)
From: Simon Jackman <simon@tamarama.stanford.edu>
Reply-To: jackman@stanford.edu
To: BILL COX <COX.BILL@epamail.epa.gov>
Cc: s-news@wubios.wustl.edu
In-Reply-To: <s78f44ac.024@RT-MAIL2.RTP.EPA.GOV>
Message-Id: <Pine.LNX.4.10.9907170243290.23994-100000@tamarama.Stanford.EDU>
Mime-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
The question "does anyone have code to do EM?" crops up every now and then
on the list, so I thought I would chime in with a reply...
The question is ill-posed. The EM algorithm is incredibly general, in the
sense that the underlying ideas can be used in a wide variety of settings
-- EM is often brought up in the context of making imputations for missing
data, but the method can be used to estimate parameters in many models ---
anything with latent variables lurking in the model (i.e., "missing data")
is open to estimation by EM (e.g., many GLMs, the Kalman filter, models
with censoring, mixture models, latent class models, Markov regime
switching time series, certain kinds of hierarchical or random
coefficient models).
Because of what EM does and the way it does it, the implementations are
almost by definition problem-specific. So ask instead, for example, "does
anyone have code to estimate <model x> by EM?" or "does anyone have code
for using the EM code to make imputations for multivariate Normal data
that is missing at random?", etc.
Note also that EM is often a stepping stone to Markov chain Monte Carlo
methods, where instead of simply plugging in the expected value of the
"missing data" you plug in a draw from the conditional distribution of the
"missing data", but thats another story...
Regards
-- Simon Jackman
Simon Jackman, Asst Professor, ph: +1 (650) 723-4760
Dept of Political Science, fax: +1 (650) 723-1808
455 Serra Mall, Bldg 160, jackman@stanford.edu
Stanford University, http://tamarama.stanford.edu/simon
Stanford CA, 94305-2044, USA.
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news
