ASA Connect

 View Only

  • 1.  Nonparametric tests

    Posted 03-30-2019 13:13
    One of the frequently made assumptions for example for the sign test is continuity of the underlying distribution(s). What are the consequences of violating this assumption, if any, if ties with the hypothesized median are discarded? How about other nonparametric tests? Many examples of using the sign test are discrete (e.g. Zar's Handbook, pg. 35). Thanks for reading and commenting.

    ------------------------------
    Leszek Gawarecki
    ------------------------------


  • 2.  RE: Nonparametric tests

    Posted 04-01-2019 07:28
    I think of the Sign Test as a close relative of the McNemar's Test, in which ties are discarded.  I think it's fine to discard them.  The only effect is on the power of the test if there are a lot of ties.

    We also discard ties in a Wilcoxon Signed Ranks Test.

    ------------------------------
    Edward Gracely
    Drexel University
    ------------------------------

    Original Message


  • 3.  RE: Nonparametric tests

    Posted 04-03-2019 11:53
    I published a paper on this subject last summer. It does not address the sign test specifically - it is really about two-sample tests. References within should be helpful even if the actual paper is not.

    https://journals.plos.org/plosone/article/authors?id=10.1371/journal.pone.0200837

    Monnie

    ------------------------------
    Monnie McGee
    Associate Professor
    Southern Methodist University
    Dallas,TX United States
    ------------------------------

    Original Message


  • 4.  RE: Nonparametric tests

    Posted 04-03-2019 15:09
    Tank you for sharing, I fund your article very interesting. I can see now the complexity of the question of ties and discrete distributions. Truly appreciate your response.

    ------------------------------
    Leszek Gawarecki
    ------------------------------

    Original Message


  • 5.  RE: Nonparametric tests

    Posted 04-04-2019 06:52
    ​McNemar's Exact Conditional Test is the way to go here.  Also, a more powerful exact unconditional test is available in STATXACT. This is a well researched concept.  Of course, there is a push toward not having P-values.  But McNemar also offers a conditional relative risk ratio with confidence limits (P[X>Y]/P[Y>X]

    Jon Shuster

    ------------------------------
    Jon Shuster
    ------------------------------

    Original Message


Related Content