[R] odds ratios in multiway tables (stratified)

[Wim Bertels]

On Thu, 2008-07-31 at 10:31 -0500, Marc Schwartz wrote:
> on 07/31/2008 07:40 AM Wim Bertels wrote:
> > On Wed, 2008-07-30 at 12:13 -0500, Marc Schwartz wrote:
> >> on 07/30/2008 08:48 AM Wim Bertels wrote:
> >>> Hi,
> >>>
> >>> does anyone know of a function to calculate odds ratios in multiway
> >>> tables (stratified) (+ the other usual statistics involved)
> >>>
> >>> i mean:
> >>> say we have a table r*c*d,
> >>> For every d (depth) we have a r*c table,
> >>> and in this table the odds ratio's are calculated for every 2*2 subtable
> >>> in it.
> >>>
> >>> logically this function would look like):
> >>> ORs(multiwaytable)
> >>> or
> >>> ORs(data$var1r,data$var2c,data$var3d)
> >>>
> >>> (eg. not taking the lot together, keeping the paradox of simpson in
> >>> mind)
> >>>
> >>> mvg,
> >>> Wim
> >> In ?mantelhaen.test, there is some code in the examples using the
> >> UCBAdmissions data set. There is also code for the Woolf test in the 
> >> same example.
> > 
> > thanks Marc,
> > but CMH testing supposes a common odds ratio, (hence the woolf test)
> 
> Strictly speaking that is not correct, whether one is using the Woolf 
> test or the Breslow-Day test. In fact, since you reference SAS below, my 
> copy of "Categorical Data Analysis Using the SAS System" by Stokes et al 
> from 1995 (back in the days when I was using SAS) notes this in the 
> chapter on stratified 2x2 tables (second para on page 53).
> 
> This is also noted in CDA 2nd Edition, Agresti (2002), on page 235.
> 
> That being said, a significant Woolf or BD test should give you pause 
> relative to the validity of the _common_ odds ratio and to consider 
> alternatives that enable the analysis of more interesting relationships 
> in your data.

i'm not interest in a common odds ratio as such,
more in the "fixed" effects between the different categories

> 
> > i am looking for a way to just get all the odds ratios calculated, with
> > a family p-values (for each one) and/or family confidence intervals
> > (since i am doing multiple testing then,.. data snooping..)
> > [i know this is easily done in SAS, but i prefer R..]
> > 
> >> In addition, there is similar code in the 'vcd' and 'rmeta' CRAN packages.
> > 
> > tnx,
> > structplot looks nice as an extra
> 
> I presume that is strucplot in the vcd package?
> 
> You might also want to look at the mh() function in the Epi package, 
> which I noted doing a quick search this morning.
> 
> Also, if the assumption of the homogeneity of odds ratio is not valid in 
> this situation, you may want to consider that there is an interaction 
> going on and that an alternative analytic approach, such as logistic 
> regression with an appropriate interaction term might make sense here.

i know, i completely agree,
helas, i tried this route with lots of variaties,
but had no succes in finding a good model 
(btw: i am still looking for a goodness of fit statistic for multinomial
regression, also called baseline regression, like eg the goeman and
lecessie statistic, but it seems not be implemented in R)

> 
> This came up in a prior thread, which you might find helpful:
> 
> https://stat.ethz.ch/pipermail/r-help/2007-February/126254.html

nicely explained,
also the function to recode to a 3D table is very usefull,
tnx,
Wim

> 
> Regards,
> 
> Marc
> 
>