controlling for age in ANCOVA in age-matched groups

Discussion in 'Scientific Statistics Math' started by scale.symmetry, Dec 26, 2006.

  1. Dear sci.stat.math,
    I have two groups (a clinical population and an age and gender matched
    control population).
    The two groups have a large age-range.
    The two groups are significantly different on a number of cognitive
    These cognitive measures also vary substantially over the age-range.

    Should I co-vary for age in the ANCOVA?
    It seems to me the answer is 'yes', because age is a known source of
    'error' variance, potentially overwhelming the effect of interest (the
    cognitive differences between the groups).
    However, I have also been told that one shouldn't control for age if
    one has groups matched for age.
    Scale Symmetry.
    scale.symmetry, Dec 26, 2006
    1. Advertisements

  2. wrote in
    Which, presumably, was why you matched on that feature. You (and the
    human subjects review board) really should have asked and answered this
    question before doing the study.
    You have not told us your dependent variables or your research questions
    yet. You have only suggested that cognitive measures are different.

    Adding a matching variable may not result in problematic collinearity in
    an simple linear model. but it will affect the interpretation. Adding a
    matching variable to a logistic model should have unfortunate
    consequences, but adding interactions between matching variable and other
    covariates of interest should be examined in either simple or generalized
    regression models. Your modeled age <effect> in a simple model will be
    "right in the middle" of the two groups <effect>, what ever those
    <effect>s are. The other covariate effects will then be referenced to the
    adjusted (and averaged) age effect. Is that what you want? It sounds as
    though you don't really know.

    Sometimes you do need to work closely with a statistician rather than
    just winging it. This might be a good time for that. The effort that went
    into the study and that possible benefits of a proper analysis will
    justify getting a proper consultation. A newsgroup answer is unlikely to
    be sufficient.
    David Winsemius, Dec 27, 2006
    1. Advertisements

  3. Dear David,
    Thanks for the helpful part:
    The rest of your comments are based on incorrect assumptions not
    relevent to the content of my question and I'll just let them slide by.
    Scale Symmetry.
    scale.symmetry, Dec 27, 2006
  4. Thanks Rich. That was my original statistical advice, and rounds out my
    understanding of the issues more.

    We were not using a matched design.

    The age-range is very wide (6-65 years) and the cognitive variables are
    known to vary in a nonlinear fashion over this range (depending on the
    variable, usually an inverted 'U' if you put age on a log scale, but
    not always). So we used a linear and quadratic control term for age in
    the main ANCOVA.

    Scale Symmetry
    scale.symmetry, Dec 28, 2006
    1. Advertisements

Ask a Question

Want to reply to this thread or ask your own question?

You'll need to choose a username for the site, which only take a couple of moments (here). After that, you can post your question and our members will help you out.