ASA Connect

 View Only

  • 1.  1:N matching

    Posted 07-07-2016 11:23

    Hi,

    I have paired data, 1 case were matched with multiple controls.  The outcomes include continuous and binary variables.  I wonder what method I can use to deal with the correlation within pairs.  Mix model might be a choice. Are there methods, like "modified" paired t test, or mcnemar tests?

    Thank you!

    ------------------------------
    Jun Liu
    Sr. Biostatistician
    ------------------------------


  • 2.  RE: 1:N matching

    Posted 07-08-2016 13:51
    Jun,

    I am confused by your statement " 1 case were matched with multiple controls".  Do you mean out of N samples only 1 was paired and the others were not?  I do not think there is any statistical test that can address a situation were some of the samples are matched (paired) and the others are not.  You may have to check that sample and possibly exclude from analysis.  Supposing you mean there is 1 pairing variable (or possibly more) for all the samples and you have collected both nominal (dichotomous) and continuous data on all the N subjects.  Then you have two scenario:

    1. For the continuous data, you can do a paired t-test or its nonparametric counterpart Wilcoxon's signed rank test.
    2. For the nominal (dichotomous) paired data, you can use McNemar's test for symmetry. Cannot use this test for the former and vice versa.

    You may have to use some other techniques if you have multiple matching (pairing). Some clinical or behavioral statisticians may suggest something better depending on the exact nature of your design.  Good luck.

    Ajit K. Thakur, Ph.D.
    Retired Statistician



    Original Message


  • 3.  RE: 1:N matching

    Posted 07-08-2016 14:50

    Sorry for misleading statements.  I meant every case was matched by more than 1 controls.  For example, every case was matched by 2 controls.  In this case, how to apply paired t-test (how to define the difference between pairs) or McNemar test(how to define the concordance or discordance)?   Thank you!

    ------------------------------
    Jun Liu
    Sr. Biostatistician

    Original Message


  • 4.  RE: 1:N matching

    Posted 07-11-2016 13:53

    Hi Jun,

    I think the book "Statistical Methods in Cancer Research Volume I - The Analysis of Case-Control Studies" by N. E. Breslow and N. E. Day (1980) could be helpful for your case. Specifically, Chapter 5.3, titled with "1:M matching: dichotomous exposures", talks about estimation, hypothesis test, confidence limits, and homogeneity of the relative risk. It may fit your need. 

    Best,

    Zhulin He

    ------------------------------
    Zhulin He

    Original Message


  • 5.  RE: 1:N matching

    Posted 07-08-2016 19:59

    For the continuous outcomes, mixed models is a good choice. For the binary outcomes, I think what you want is called Conditional Logistic Regression. And if you have SAS, the Proc PHReg documentation has an example of how to use Proc PHReg to do Conditional Logistic Regression for M:N matching. If you read all the way to the end of the page, they discuss specifically how to handle ties in the case of 1:N matching. Additionally, Proc Logistic now has a Strata statement (in SAS 9.4), and the SAS documentation for this indicates that this addition allows one to use Proc Logistic for doing Conditional Logistic Regression on 1:N matched sets and and M:N matched sets.  

    ------------------------------
    Eric Siegel, MS
    Research Associate
    Department of Biostatistics
    Univ. Arkansas Medical Sciences

    Original Message