Astronomy and Astrophysics – Astronomy
Scientific paper
Nov 2007
adsabs.harvard.edu/cgi-bin/nph-data_query?bibcode=2007aspc..371..101c&link_type=abstract
Statistical Challenges in Modern Astronomy IV ASP Conference Series, Vol. 371, proceedings of the conference held 12-15 June 200
Astronomy and Astrophysics
Astronomy
1
Scientific paper
We propose a new method to test for goodness-of-fit of a model for low-count Poisson data. Our approach does not resemble the usual methods of approximations to χ^2, but instead explicitly uses the full Poisson distribution. First, we propose to use a simple (Poisson-specific) multiscale model to characterize the ``mismatch'' between a best-fit physical model and the data. Next, we embed this multiscale model into a probabilistic/likelihood framework (via hierarchical Bayes), allowing us to handle statistical uncertainties. We then use MCMC to map out the shape of the joint posterior probability of all of the unknown parameters. Finally, we note that this is a generalization of a problem with a known solution: whether an additional component of known shape is justified by the data. Hence, when the multiscale structure that we use to model the ``mismatch'' between the data and the physical model accounts for a significant number of counts, and/or when the scale-factor for the best-fit physics model is significantly different than one, then the original fit was not ``good-enough''. That is, the fitted multiscale ``mismatch'' isolates the discrepency between model and data. We use models of the Gamma-Ray sky as viewed by CGRO/EGRET as our example. We demonstrate that our method works nicely on several examples, but further work is needed to investigate the method's power.
The authors and CHASC acknowledge support from NSF (DMS 04-06085) and Chandra X-Ray Center, and AISRP NAG5-12082. It is a particular pleasure to acknowledge the SAMSI06 Special Program in Astrostatistics.
Connors Alanna
van Dyk David A.
No associations
LandOfFree
How To Win With Non-Gaussian Data: Poisson Goodness-of-Fit does not yet have a rating. At this time, there are no reviews or comments for this scientific paper.
If you have personal experience with How To Win With Non-Gaussian Data: Poisson Goodness-of-Fit, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and How To Win With Non-Gaussian Data: Poisson Goodness-of-Fit will most certainly appreciate the feedback.
Profile ID: LFWR-SCP-O-735628