Jon Heron posted on Tuesday, November 27, 2012 - 1:58 am
I would be interested in your thoughts on how best to utilise the bivariate tech10 output.
In the past I have focussed on the "overall" value and derived a statistically-dubious threshold to compare it to. For instance, if the overall value is a composite from ten 2*2 contingency tables then I'd derive a value from a chi-square with ten degrees of freedom.
This doesn't feel that great, neither does looking for individual large values due to the problem of multiple testing.
I am now wondering about something a little more sophisticated and robust. In my current large LCA I have a couple of hundred bivariate tech10 comparisons. What would you think about a QQ plot of all those individual Pearson values or some other consideration based on their distribution?
It is an important issue, but not sure how to handle it easily. What is really needed is a statistically sound overall summary of TECH10-type tests. Albert Maydeu-Olivares has written about this and we have it on our list to add to Mplus. Perhaps you want to take a look at his writings and/or contact him.
Jon Heron posted on Wednesday, November 28, 2012 - 3:40 am