Removing the blindness

From: Alexander Conley (AJConley@lbl.gov)
Date: Wed Oct 13 2004 - 14:16:51 PDT

  • Next message: Chris Lidman: "Re: Removing the blindness"

    Hello all,

       I have finished (I hope) trying to answer all of the questions that
    you
    have asked me about my analysis. What I need from you is for you all
    to tell me if you are satisfied with my attempt at an answer, or if I
    need
    to look at a particular issue more deeply.

    There is a new version of the paper (1.32) available in the usual place
    ( http://panisse.lbl.gov/~aconley/cmag ) which incorporates many of the
    changes you suggested, plus a few more.

    The biggest change is that probcut is no more. I discovered a subtle
    bias that probcut was introducing in the analysis (it tended to punish
    extremely well observed low redshift SNe with slopes that were a bit
    away from the central value). I ended up excluding 2 low-z SNe by hand
    that probcut had been catching: 97br, 97bp. Their data appears to
    be internally inconsistent, with a much larger degree of scatter than
    can be accounted for by their quoted photometric errors.

    There is also a (blinded) w fit. I haven't combined it with the other
    priors, but there isn't much point in doing so in blind mode anyways.
    Finally, I started putting the systematic errors together, although the
    job isn't done yet.

    Below I have listed all of you in alphabetical order and tried to give
    a brief response to each of the questions you asked of me.

    Ariel:
         Original message:
    http://supernova.lbl.gov/collab/archive/cmag/0006.html

         Q: Have you tried not applying the MW extinction correction and
    seeing
              if that behaves as you expect?
         A: As expected, the m_B residuals (as measured by the RMS) got much
              worse and the B_BV residuals did not.
              More info:
    http://supernova.lbl.gov/collab/archive/cmag/0010.html

         Q: Why do you set s_B = s_V for some fits but not for others?
         A: I think it would be internally inconsistent to force s_B=s_V
    given what
             we know from CMAGIC -- namely, that when there is a bump, s_V <
    s_B.
             Note that I have only allowed s_B ne s_V when there is a bump.
             And it's not like the s_B = s_V fits to SNe with bumps are good
    -- most
             of them are absolutely awful, clearly missing the shape of the
    light curve.
             On a related note, on Saul's urging I did redo all fits so that
    s_B was fit
             to the B data only, and found that using these values for
    stretch, etc. made
             very little difference. See Saul's section below.
             More info:
    http://supernova.lbl.gov/collab/archive/cmag/0007.html

        Q: If the split in the high redshift residuals (between the Knop and
    Barris sample)
              is really due to dust, can you see this in the residuals of
    m_B vs B_BV?
         A: Yes, the matched residuals are consistent with this story. This
    is probably
              my single greatest area of concern with my result -- but I'm
    not sure what
              to do about it. Note that if a 'by-hand' extinction
    correction is applied, the
              sample no longer splits.
              More info:
    http://supernova.lbl.gov/collab/archive/cmag/0007.html

          Q: Have you tried to set limits on intergalactic dust?
          A: No, but we can talk about this after the blindness comes off.

    Chris:
          Original email:
    http://supernova.lbl.gov/collab/archive/cmag/0001.html

          Q: Could you add a stretch vs. scriptM plot?
          A: Done

          Q: You seem less sensitive to the U-enhanced K-correction
    systematic
                than K03. Could you emphasize this more?
           A: Am I? K03 was completely insensitive to changing the U-B
    color of
                the uberspectrum as long as an extinction correction was not
    applied,
                and very sensitive when one was. I don't apply a extinction
    correction,
                and my sensitivity lies somewhere between the two K03 values.
                See the latest version of the paper for actual numbers.

           Q: s_B ne s_V?
           A: Ariel also asked this (although you beat him to the punch).
    See the
                answer in his section.

       Greg:
           Greg's comments:
    http://supernova.lbl.gov/collab/archive/cmag/0013.html
           Greg had many comments. I won't be able to hit them all here. See
             http://supernova.lbl.gov/collab/archive/cmag/0013.html for many
    of my responses.
             Some highlights, and things I didn't get to in that email.

          Q: The first 3 high redshift SNe all seem high. Is their anything
    funny with them?
          A: Nothing I can find. They come from 3 different reductions
    (P99, K03 and Riess 98).
               Applying a by-hand extinction correction does not save them.
    They are also
               all high in a maximum magnitude fit. There was a problem
    with one of the
               lightcurve points for SN1995ba (a photometry error -- I have
    informed Rob),
               but that doesn't really 'save' it.

          Q: Can you break up the residuals by various subsamples and see if
    they
                differ.
           A: By photometry source: The SCP and hi-z team samples are quite
    compatible
                   in terms of mean and scatter. SCP mean: 0.025 avdev:
    0.183 rms: 0.218
                   high-z team mean: 0.006 avdev: 0.183 rms: 0.233. The
    Hawaii (basically
                   Barris) sample has an offset mean and lower rms -- mean:
    -0.13
                   avdev: 0.11 rms: 0.19. The offset is well documented
    (the split in the
                   residuals) -- see Ariel's section.
                By redshift: Split into three chunks: [0,0.45),
    [0.45,0.55),[0.55) (to get roughly
                   equal numbers of SNe), they are consistent. The RMS gets
    marginally
                   smaller at the high redshift end (0.23 -> 0.17), but I
    expect this to some
                   extent because the mean extinction will be lower because
    of selection effects.
                 Summary: I see no evidence for any real problems, except
    the already noted
                   shift in the mean between the Barris sample and everybody
    else (which seems
                   to arise because the Barris sample has unusually little
    host galaxy extinction).

           Q: In the matched residuals (m_B vs B_BV) is the distribution
    narrowing as you
                 move towards positive residuals? Do bump SNe have anything
    to do with this?
           A: Splitting the sample in half, the lower RMS is 0.16, the upper
    is 0.15, so it is
                  slightly narrower, but not by much. Removing all of the
    SNe with bumps
                  decreases the RMS values slightly to lower: 0.16 upper:
    0.14.
                  On a related note, of the 8 SNe with bumps, 6 have
    positive residuals.
                  This is significant at the 1.4 sigma (equiv: 14.4%) level,
    according to
                  the binomial distribution. Bottom line: This doesn't seem
    significant.

           Q: What does the chisquare distribution for the fits to the CMAG
    relations look like?
           A: Too good -- as expected. This is why W03 rescaled their error
    bars -- people
                are clearly underestimating them. I decided that the
    rescaling was far too
                dangerous given the quality of the high-z data. The plot of
    the chisq probability
                (the prob that the chisq should be higher than it is
    observed to be) is
                attached. It is very difficult to say what this
    distribution should look like because
                it is different for every number of points in the CMAG
    region, but it's fair
                to say it should be peaked around 0.5. The big spike at
    around 1 is a bunch
                of SNe that are too well fit.



      Mark:
         Original email:
    http://supernova.lbl.gov/collab/archive/cmag/0011.html
         Q: Could you include the intrinsic error in the stretch vs scriptM
    plot?
         A: Done
         Q: How about propogating the z error into the vertical error bars
    in the
              Hubble residual plots?
          A: Done
          Q: On the resid vs. resid plot, do you take the errors into
    account when
               calculating the pearson's coefficient?
          A: Not yet, but I think this can wait until after the blindness is
    removed
               (since that shouldn't affect these values).

      Saul:

        Q: Could you do a blinded w fit?
        A: Done. See the most recent version of the paper. The blindness
    scheme
              is a bit rough given the almost pathological shape of the
    error contours.
              I have not combined these contours with the other priors yet,
    but there is
              no point in doing so now. Adding the priors and then blinding
              would probably allow the blindness scheme to be violated, so I
    won't do it.
              More info:
    http://supernova.lbl.gov/collab/archive/cmag/0021.html

         Q: Could you redo all of the lightcurve fits using the prescription
    of P99 --
              that is, do a B only fit, then fix the stretch and date of
    maximum to those values
              for the V fit.
          A: Done. Didn't make a difference (see the most recent version of
    paper). As
              detailed at great length in
                 http://supernova.lbl.gov/collab/archive/deepnews/1361.html
              a very long and boring email which nobody probably read, this
    is (a) harder
              than it looks, and (b) wasn't done right in P99. Luckily it
    only affects the V
              magnitudes, so doesn't change fit C. Note that I didn't
    follow this prescription
              for the 8 bump SNe, since the quality of the fit in that case
    is god-awful.

          Q: What effect does applying an extinction correction have?
          A: It shifts the contours almost purely along the long axis inward
    (by 0.4ish).
               Note that this fit can NEVER be published. It's a piece of
    junk found by
               combining lots of extinction values that other people
    calculated, and so
               is horribly internally inconsistent.

          Q: What effect does removing all SNe with bumps have?
          A: A shift of 0.016 along the short axis and -0.14 along the long
    axis, giving
               lower values of om, ol, but a slightly higher value (+0.016)
    for om in a flat
               universe. Not a really large systematic.

    Vitaliy:
        Original email:
    http://supernova.lbl.gov/collab/archive/cmag/0015.html
        Q: Are you going to combine the shifts into a final systematic error?
        A: Yes. I have started doing this in the latest paper draft.

    Okay -- hopefully I didn't miss anything. Please let me know if you are
    happy with my answers, and if not, what I can do to try and answer your
    questions better. Then, the blindness can be removed.

    Alex



    This archive was generated by hypermail 2.1.4 : Wed Oct 13 2004 - 14:17:04 PDT