Statistical consistency of results?

1
Statistical consistency of results?

• Consider Prob(c2,n.d.f) for individual runs
  before combining
• For typical cubic function for single run, ndf
  9
• Small fit uncertainties do not mean results are
  reliable
• From Luis summary of last week

2
Alternative combinations
Conclusions (i) A), B), Weighting by 1/dq12
alone for all 6 runs gives unacceptable
chisquared for the average q1, using either slot
0 or same slot reference (ii) C), D),
Weighting by 1/(dq1.sqrt(chisq))2 gives improved
chisq on average but still very small, e-5.
Basic problem is fits for runs 1465, 1471, 1500
are very poor chisq. (iii) Suggest that we
simply reject entire runs for which the
chisquared fit probability is absolutely crazy
(e.g. less than 1.e-4). Also, that we allow to
remove end points at e.g. y1.2mm that do not
look anything like cubic behaviour. AA),
BB), CC), DD) give similar chisquared probability
on the average and similar q1, dq1. (iv) On
assumption that errors from individual runs may
be underestimated, suggest we use CC) for main
result, and DD) for cross-check.