[Rasch] DIF and multiple comparisons - BH method
S.Kreiner at biostat.ku.dk
Wed Feb 8 01:23:09 EST 2006
One additional advantage of B-H methods compared to Bonferroni is that the
B-H also works when test statistics are correlated. This is very important
since tests for DIF are correlated.
No DIF requires that items and other person covariates are conditionally
independent given the latent trait variable. The only model where items and
covariates also are conditionally independent given the manifest person
score is the Rasch model where the score is a sufficient statistic.Tests of
conditional independence of items and covariates given the total score
using either Partial gamma coefficients or Mantel-Haenszel methods are
therefore Rasch-based methods.
Remember also, that DIF is uniform if and only if the item x covariate x
score table fits a two-factor loglinear model. The Rasch model applies in
the different subpopulations defined by the values of the covariate, but
item parameters of the biased items are different. Comparisons within
subpopulations are objective. Comparisons across subpopulations are not.
At 06:26 07-02-2006 -0600, Fred Wolfe wrote:
>I am not expert on the Benjamina-Hochberg method. I understand that there
>are modifications of BH, but it seems that the original BH is still the
>predominant method now (from what I could see looking at the literature).
>BH would seem to have important advantages over Bonferroni.
>This issue came up when a colleague was examining DIF using non-Rasch
>methods. Using the model form: Item score = Exogenous variable +
>Questionnaire Total, this was examined by 3 methods (with different
>results) that had been used in the literature:
>Partial gamma coefficient (Stata)
>Kendal's tau b coefficient (SAS)
>Variance estimation using bootstrap methods
>Perhaps someone might have comments as to the usefulness/validity of Rasch
>vs. non-Rasch methods for detecting DIF.
>At 04:45 PM 2/6/2006, Mike Linacre (RMT) wrote:
>>Thank you, Randy and Fred,
>>The Benjamini, Y. & Hochberg, Y. (1995) method for multiple comparisons
>>appears to be:
>>(1) Compute individual p-values for each of the N hypothesis tests as
>>though each was the only one.
>>(2) Sort the N p-values by size ascending, n=1,N, so that p(1) is the
>>smallest, and p(N) is the largest
>>(3) Starting from p(N) downwards, look down for the first p(n) which is
>>less than or equal to (n/N) * 0.05 (or your chosen significance level)
>>(4) Tests 1 to n are classified as significant.
>>Is this correct? Have any useful modifications been proposed? This seems
>>easy to implement in software, which I may well do.
>>Editor, Rasch Measurement Transactions
>>rmt at rasch.org www.rasch.org/rmt/ Latest RMT: 19:3 Winter
>National Data Bank for Rheumatic Diseases
>Tel (316) 263-2125 Fax (316) 263-0761
>fwolfe at arthritis-research.org
>Rasch mailing list
>Rasch at acer.edu.au
Department of Biostatistics
University of Copenhagen
Blegdamsvej 3, DK-2200 Copenhagen N, DENMARK
From June 9, 2005:
Øster farimagsgade 5, entr. B
P.O. Box 2099
DK-1014 Copenhagen K, Denmark
Email: S.Kreiner at biostat.ku.dk
Phone: (+45) 35 32 75 97
Fax: (+45) 35 32 79 07
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Rasch