[R] Permutation tests in {coin}

[Henric Winell]

Lars Bishop skrev 2013-10-05 22:17:
> Hello,
>
> I'm trying to get familiar with the coin package for doing
> permutation tests. I'm not sure I understand the documentation
> regarding the difference between distribution = "asymptotic" and
> "approximate" in the function independence_test.

The use of "asymptotic" or "approximate" leads to two different 
approximations of the exact conditional distribution under the null 
hypothesis.

The basis of 'coin' is a multivariate linear test statistic, and it can 
be shown (Strasser and Weber, 1999) that its conditional distribution is 
asymptotically normally distributed under the null hypothesis 
("asymptotic").  Alternatively, another approximative conditional null 
distribution can be obtained using conditional Monte Carlo procedures 
("approximate").

> Are permutations of the test statistic actually computed in the
> asymptotic case, or only when the distribution is specified as
> approximate?

In the asymptotic case, the univariate normal distribution is used when 
the test statistic is a scalar and the chi-squared distribution is used 
when the test statistics is a quadratic form.  For a multivariate 
maximum-type test, permutations may be used in the asymptotic case but 
only in the sense that probabilities from the corresponding multivariate 
normal distribution are obtained by numerical integration using Monte 
Carlo procedures.  In the latter case, the 'mvtnorm' package is used and 
further details can be found in its documentation.

In the approximate case, permutations are always used irrespective of 
the type of test statistic.

> When should I use each option?

You should only use "asymptotic" in situations where you trust the
asymptotic approximation. ;-)  In all other cases, use "approximate" or,
if possible, "exact".


HTH,
Henric



>
> Thanks Lars.
>
> [[alternative HTML version deleted]]
>
> ______________________________________________ R-help at r-project.org
> mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do
> read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>