Discussion: View Thread

  • 1.  Failure Discovery Rate and Bonferroni , etc (SAS)

    Posted 04-05-2012 21:09
    This message has been cross posted to the following eGroups: Statistics in Epidemiology Section and Statistical Consulting Section .
    -------------------------------------------

    Hi-

    I have microarray data to analyze. I have several paired observations which I want to 
    analyze among different groups.

    I am a SAS user/programmer and was wondering if anyone recommends using PROC TTEST
    over PROC GLM OR PROC MIXED.

    Also I plan to use PROC MULTEST to generate q-values and was wondering how to choose
    a cut off or threshold for statistical significance versus randomly choosing P<0.05 as significant.

    Also, what should I do with the q-values? Should I display those values in graphs instead of 
    the raw P-values from the paired analyses?

    Thanks.
    -------------------------------------------
    Margaret Linan
    Epidemiology and Biostatistics, MPH (c)
    Temple University
    -------------------------------------------


  • 2.  RE:Failure Discovery Rate and Bonferroni , etc (SAS)

    Posted 04-05-2012 21:38

    I haven't done much with microarrays so I may not be the best person to answer.  I do think that more complicated models are required for microarrays. So probably a mixed model would be best.  Also I often use proc multtest. I like the results presented the way they are in the SAS output.  Raw p-values followed by adjusted p-values in columns with the specific multiple testing procedures as headers. 
    -------------------------------------------
    Michael Chernick
    Director of Biostatistical Services
    Lankenau Institute for Medical Research
    -------------------------------------------








  • 3.  RE:Failure Discovery Rate and Bonferroni , etc (SAS)

    Posted 04-06-2012 09:38
      |   view attached
    Kerby Shedden of the University of Michigan recently gave a talk on this to the UM R users' group.

    I have the slides.  I *think* they're attached (the interface is, uh, bad).

    -------------------------------------------
    Barry DeCicco
    Statistician,
    University of Michigan Health System
    -------------------------------------------






    Attachment(s)

    pdf
    R-screening.pdf   414 KB 1 version
    Original Message


  • 4.  RE:Failure Discovery Rate and Bonferroni , etc (SAS)

    Posted 04-06-2012 09:53

    To add to my comments on large scale inference, there was a one day conference in DC last year that I attended.  it was organized in honor of Brad Efron.  There may be a conference website where you can go to see the presentation slides.
    -------------------------------------------
    Michael Chernick
    Director of Biostatistical Services
    Lankenau Institute for Medical Research
    -------------------------------------------






    Original Message


  • 5.  RE:Failure Discovery Rate and Bonferroni , etc (SAS)

    Posted 04-06-2012 08:05
    Susan clearly has more experience and provides better advice.  I comment here only because I do have a little aquiantance with micro array problems and do have an interest in them.  i know that as she says the key issue is usually the large p small n problem where the number of tests being performed can be huge but the number of cases small.  That is what motivated Benjamini and others to create the FDR multiple testing rule. To my knowledge this has required a lot of innovation from statisticians.  The use of empirical Bayes methods and Bayesian Hierarchical model is commonly used.  There is now even an IMS monograph on this by Brad Efron  titled "Large Scale Inference".
    I have no experience with software tools that are specifically designed to address this.  It does seem that Susan has good advice in that regard.

    -------------------------------------------
    Michael Chernick
    Director of Biostatistical Services
    Lankenau Institute for Medical Research
    -------------------------------------------








  • 6.  RE:Failure Discovery Rate and Bonferroni , etc (SAS)

    Posted 04-06-2012 12:14
    Margaret,

    Below is a useful link from NIH with many good tips regarding statistical analysis and visualization of microarray data such as the "shrinkage" SE estimation method of Tibshirani, false discovery rates (FDR) and the quantile and volcano plots for visualization. I hope it's helpful:

    http://discover.nci.nih.gov/microarrayAnalysis/Statistical.Tests.jsp

    Regards,
    -------------------------------------------
    Edit Kurali
    GlaxoSmithKline
    -------------------------------------------