• between-subject variation in power analysis

    From Norman B. Grover@21:1/5 to All on Sat Jul 9 13:51:45 2016
    Many years ago, I read that when estimating standard deviations in power analysis of the two-sample t-test, one should take a larger value for the treatment group than the control group (which can usually be found in the literature) because in biological experiments intervention often disturbs
    the steady state or, for observational data, the sick group can be expected
    to be less homogeneous.
    I can no longer locate that reference and wonder whether anyone here can
    help me by directing me either to an appropriate source for such a claim or
    to actual published data.
    Any help would be greatly appreciated.

    --

    Norman B. Grover
    Jerusalem, Israel

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Rich Ulrich@21:1/5 to norman@md.huji.ac.il on Sun Jul 10 00:59:24 2016
    On Sat, 9 Jul 2016 13:51:45 +0300, Norman B. Grover
    <norman@md.huji.ac.il> wrote:

    Many years ago, I read that when estimating standard deviations in power
    analysis of the two-sample t-test, one should take a larger value for the >treatment group than the control group (which can usually be found in the >literature) because in biological experiments intervention often disturbs
    the steady state or, for observational data, the sick group can be expected >to be less homogeneous.
    I can no longer locate that reference and wonder whether anyone here can
    help me by directing me either to an appropriate source for such a claim or >to actual published data.
    Any help would be greatly appreciated.


    Well, Jacob Cohen wrote the book on power analysis for the
    social sciences. That would be the first place that I would look.

    Cohen does mention practical considerations, so that might be
    one of them. The consequence of knowing that variances are
    unequal is that you have reason to depart from the usual practice
    of sampling equal Ns in the two groups.

    --
    Rich Ulrich

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Norman B. Grover@21:1/5 to All on Sun Jul 10 12:48:39 2016
    In article <78l3obtmjipp19pg5i5gjmd0riv6ic45gh@4ax.com>, rich.ulrich@comcast.net says...
    On Sat, 9 Jul 2016 13:51:45 +0300, Norman B. Grover
    <norman@md.huji.ac.il> wrote:

    Many years ago, I read that when estimating standard deviations in power
    analysis of the two-sample t-test, one should take a larger value for the >treatment group than the control group (which can usually be found in the >literature) because in biological experiments intervention often disturbs >the steady state or, for observational data, the sick group can be expected >to be less homogeneous.
    I can no longer locate that reference and wonder whether anyone here can
    help me by directing me either to an appropriate source for such a claim or >to actual published data.
    Any help would be greatly appreciated.


    Well, Jacob Cohen wrote the book on power analysis for the
    social sciences. That would be the first place that I would look.

    Cohen does mention practical considerations, so that might be
    one of them. The consequence of knowing that variances are
    unequal is that you have reason to depart from the usual practice
    of sampling equal Ns in the two groups.


    I just replied directly to the poster, sorry. I meant to write here.

    Cohen was the first place I went to, and the second. But he does not
    consider the case common in experimental biology in which treatment implies intervention (nor when a healthy control group is compared with a sample of sick subjects) and there, I believe, the issue arises.

    Yes, with unequal variances one should use unequal Ns.

    Thank you Rich for your help. Sorry about the direct reply.
    --

    Norman B. Grover
    Jerusalem, Israel

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)