Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, is already up and running.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

st: Comparing ICC coefficient across different definitions of lvl-2 units in multilevel regression - is bootstrap the way to go?

From   Claus Dalsgaard Hansen <>
To   "" <>
Subject   st: Comparing ICC coefficient across different definitions of lvl-2 units in multilevel regression - is bootstrap the way to go?
Date   Tue, 19 Mar 2013 09:04:16 +0000

Dear Stata-list,

I have a question I hope some of you will be able to help me with.

With a colleague I've been developing small area clusters using GIS for use in multilevel analysis instead of the administratively generated clusters that can normally be used. Now I want to compare the new clusters with the old ones and I've been using Intra Class Correlations as one measure of this.

We have a sample of some 22.000 respondents and these respondents can be grouped into 9 counties, app. 220 parishes and 803 local area clusters that we have generated using GIS by splitting up the parishes in smaller units.

I have run three different multilevel logistic regression models using county, parish and local area as the lvl-2 unit and compared the ICC-coefficient for the three. Is this a valid way of comparing them to each other and to look at the clustering of the answers? As you can imagine, the lvl-2 units do not have the same number of respondents (the counties have between 550 and 4000 respondents, the parishes have between 13 and 638 and the local areas have from 13 to 201 respondents).

In addition we are developing the GIS-generated local areas and have a number of different versions of these clusters that we want to compare to each other in order to see which of these area definitions leads to the greatest clustering of data. We have thought about using ICC for this but are not sure it's the best way?

We've been talking to another colleague who told us that the ICC are problematic when the number of respondents in the lvl-2 units gets too small and he suggested that we compared the local area clusters we have created to a clustering where the individuals have been distributed randomly across 800 artificial units. In that way we could compare and see whether the units we have generated are better than the random one. We have been doing that manually generating a random number and on the basis of that distributing the individuals into new artificial clusters. We want to know whether this can be done automatically e.g. using the bootstrap command?

I'm sorry if the question is a bit difficult to understand but I hope someone will be able to help us.

Bedste hilsner,
Claus D. Hansen
Adjunkt | Institut for Sociologi og Socialt Arbejde
T: 9940 7385 / 2338 1039 | Email: |  
Aalborg Universitet | Kroghstræde 5, 20 | 9220 Aalborg Ø|

*   For searches and help try:

© Copyright 1996–2016 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index