Statistical Issues for GLAST

1
Statistical Issues for GLAST

• (1) "Automated analysis of the GLAST photon data
  streamusing segmentation techniques (Voronoi
  tessellation, etc.)" 
•                                                   
                        Jeff Scargle
• (2) "Upper limits on sources and spectral
  lines"     Jeff Scargle
•  
• (3) "Maximum Likelihood for LAT Data Upper
  Limits, Significances, and Confidence
  Intervals"                               
   Jim Chiang
•  
• (4)  When the likelihood ratio fails Pilla
  Loader Pat Nolan
• (5) "What does the Statistics Committee at
  the CDF experiment at Fermilab do?
                     Louis
  Lyons
• (6) Discussion

2
CDF Statistics Committee What does it do?

• Louis Lyons
• Oxford SLAC
• 
  
  March 2007

3
CDF Statistics Committee

• Who?
• Luc Demortier, Rockefeller (chair)
• John Conway, UC Davis
• Joel Heinrich, Pennsylvania
• Tom Junk, Illinois
• Louis Lyons, Oxford
• Giovanni Punzi, Pisa
• Where?
• http//www-cdf.fnal.gov/physics/statistics/
• When?
• Once a month since 2000

4
ACTIVITIES

• Frequently asked questions
• Recommendations
• Liasons
• Notes on statistical issues
• Links to Conferences, papers etc
• List of books

5
Frequently asked questions

• Estimating efficiencies near 1
• Pull quantities
• Error on ratio of Poisson counts
• Unbinned maximum likelihood as goodness of fit?
• Combining quantities with unknown correlation
• Parametrising background shapes for fits
• Significance calculation allowing for background
  uncertainties
• Significance from difference in log(L)
• Combining significances from different
  (independent) analyses

6
RECOMMENDATIONS

• Neural networks, support vector machines
• Optimising searches
• Likelihood fits with individual event errors
  (Punzi effect)
• Coverage for Poisson intervals e.g. ?lnL
• Plotting Poisson error bars
• Good and bad random number generators
• Systematics and limits the Manhattan project
• Bayes and Frequentism
• Comparing 2 hypotheses
• Simple facts about p-values
• Blind analyses

7
Conclusions

• Useful for --
• Giving advice
• Spotting some errors
• Aiming for uniform practice
• Answering queries
• Possible improvements --
• More active liasons
• Someone at Fermilab to discuss, rather
  than e-mail
• More on multi-variate methods for
  separating signal from bgd
• More software