ASA Connect

 View Only
  • 1.  Fishers r to z transform for complex survey data

    Posted 11-01-2018 09:15
    Dear ASA Colleagues:

    Does anyone have suggestions for references on computing Fishers r to z transformation for data from a complex survey sample with weights, strata, and clusters?

    ------------------------------
    Brandy Sinco, BS, MA, MS
    Statistician and Programmer/Analyst
    ------------------------------


  • 2.  RE: Fishers r to z transform for complex survey data

    Posted 11-03-2018 11:33
    ​Brandi, I think this may be what you are looking for.

    • Fisher, R. A. (1921), "On the 'Probable Error' of a Coefficient of Correlation Deduced from a Small Sample," Metron, 1, 3–32.



    ------------------------------
    Robert Lucas
    Principal
    Robert M. Lucas Consulting
    ------------------------------



  • 3.  RE: Fishers r to z transform for complex survey data

    Posted 11-05-2018 14:24
    Robert,

    Thanks for the article link.  I'm trying to find an article that adapts Fisher's formula for a complex sample design with weighting, stratification, and clustering.

    ------------------------------
    Brandy Sinco, BS, MA, MS
    Statistician and Programmer/Analyst
    ------------------------------



  • 4.  RE: Fishers r to z transform for complex survey data

    Posted 11-05-2018 15:44
    Brandi,

    This topic is covered in Appendix C of this book:
    Wolter, K.M. (2007). Introduction to Variance Estimation, 2nd edition. Springer-Verlag.

    Compute Fisher's z in the standard way:
         z = (1/2) ln[ (1 + \hat{r}) / (1 - \hat{r})  ]
    where \hat{r} is the estimate of the correlation that uses the survey weights. 
    Then, put a confidence interval on z and back-transform the endpoints to the correlation scale.
    If the confidence level is 0.95 and the first-stage sample is large, the normal approximation interval based on z would be

        z +/- 1.96 sqrt{ v(z) }

    where v(z) is a variance estimate for z that accounts for the complex sample design. The easiest way to estimate v(z) is probably some kind of replication, like the jackknife or bootstrap.  Using a multiplier from a t-distribution (instead of 1.96) would give a more conservative (wider) interval.

    The back-transform, which you apply to the endpoints, is

       r = [ (exp(2z) - 1 ] / [ (exp(2z) + 1 ] 

    rv

    ------------------------------
    Richard Valliant
    University of Michigan
    ------------------------------



  • 5.  RE: Fishers r to z transform for complex survey data

    Posted 11-05-2018 21:02
    All.

    Fisher's Metron article can be found here:  http://hdl.handle.net/2440/15169

    The link is to a collection of papers of R.A. Fisher housed by the University of Adelaide.

    ------------------------------
    Jeffrey Smith
    Mathematician
    U.S. Army Research Laboratory
    ------------------------------