Fun with chisquares (more fits put on HST paper web page)

From: Robert A. Knop Jr. (robert.a.knop@vanderbilt.edu)
Date: Sat Mar 22 2003 - 09:28:01 PST

  • Next message: Greg Aldering: "Re: Fun with chisquares (more fits put on HST paper web page)"

    Every lightcurve fit I've done uses the new templates. The only things
    that don't are the contours taken from the old paper. (I did refit the
    cosmology there, but used the old gersontable.)

    > 3) In the "Differences" table is the "chirat" column the ratio of the
    > SNminuit lightcurve fit Chi-squares?

    "Differences" is always stroked-filled, unless I screwed up in which
    case it's the opposite. The ratio is stroked/filled, and is the ratio
    of the chisq/dof column from the gersontable. Note that in line A-5,
    it's a werid comparison, since I'm comparing the joint fit chisquare to
    the chisquare from the second stage of a two-stage fit. (The main
    difference, though, is that there are two more degrees of freedom when
    you let the zero float.)

    To compare the effects of only fixing the MW extiction, compare the
    stroked lines of A-2 and A-1. That worsens the chisquare from 57 to
    59.2. The sequence of stroked fits from lines A-1 to A-3 gives you the
    sequence of chisquares. (Then add line A-5 for the simultaneous fit
    method; skip A-4, as that's playing with different K-corrections.)

    > I do note that between the P99 fit of line A-1, and the full treatment
    > you would like to use on A-5, the *change* in Chi-squared is 8.2. Since
    > the *change* in Chi-squared is a useful means of hypothesis testing, we
    > see that the final A-5 fits are almost 3-sigma worse that the original P99
    > fits.

    I've put residuals online.

    I'm not convinced your interpretation of the chisquare change makes
    sense. Yes, adding 6.63 to the minimum chisquare *of a given fit* gives
    you the location of the 3-sigma contour. It's not obvious to me that
    similar sorts of subtractions of chisquares of different fits to
    different sets of data can be similarly interpreted.

    In any event, any such micromanaging of the chisquares of the fits is
    tantamount splitting hares while being faced by a hungry lion. We can
    change the chisquares of the fits by as much or even more more than
    these sorts of differences just by choosing which errors to use on our
    supernovae (e.g. by changing the assumed alpha for erorr calcluations,
    or using redshift errors). This also changes the difference between the
    P99 chisquare and the chisquare from my new fits. See:

      http://brahms.phy.vanderbilt.edu/~rknop/scp/hst/#screwingaround

    *Now* can I move on?

    ...unless we want to sit around and worry about the fact that doing the
    fits with what is probably a better set of errors (i.e. mostly from not
    overestimating the alpha we can use in propogating dstretch into our
    magnitude error bars) has a goodness-of-fit of only about 0.002,
    irrespective of whether you use the P99 gersontable or my new lightcurve
    parameters.

    -Rob

    -- 
    --Prof. Robert Knop
      Department of Physics & Astronomy, Vanderbilt University
      robert.a.knop@vanderbilt.edu
    


    This archive was generated by hypermail 2.1.4 : Sat Mar 22 2003 - 09:28:03 PST