Intraclass correlation coefficient to determine use of cluster-robust standard errors

[frederick lim]

Hi all,

I'd like to find out what is the threshold commonly used to determine if a calculated intraclass correlation (ICC) coefficient indicate cluster-robust standard errors should be used instead of the usual White's heteroskedasticity robust standard errors. Any pointers on the associated reference papers is also greatly appreciated. Currently, I used the following command to calculate the ICC but would like some indicator on whether the observed ICC value is considered high or low that indicate whether cluster-robust standard errors should be used.

loneway Z Y

where Y is the dependent variable and Z is the clustering variable.

Thanks.

Fred

[Joseph Coveney]

Wouldn't the decision rest on whether there is a natural clustering present in the data rather than on some arbitrary ICC threshold? If you're trying to do a frequentist inference, wouldn't you want the model to be settled prior to examining the data?

[frederick lim]

Hi Joseph,

Thanks for the response. The dataset I'm working on is based on a consumer panel data in four cities. While in this case, it is noted that there is an inherent natural clustering present, I wanted to find out whether consumer behavior is homogenous across cities or indeed, correlated at the city level. I felt the need to check the presence of clustering effects rather than applying cluster-robust standard errors directly. What do you think? And whether by looking at the ICC, can we make any sort of inference regarding the significance of the clustering effects, if present.

[Carlo Lazzaro]

Frederick:
Your query is about clustering, but I think it similar to what happens when a decision between an OLS and a -mixed- model should be made.
I'm afraid there's no hard and fast rule about that topic.
Despite some authors considered amn ICC around .10 enough to develop a hierarchical linear model (http://www.tandfonline.com/doi/pdf/1...326985EP3502_6), others (Nezlek JB. Social and Personality Psychology Compass 2/2 (2008): 842–860, 10.1111/j.1751-9004.2007.00059.x), in line with Joseph's reply, underscore the imporatnce of the data structure (ie, the existence of nesting, or Clustering, in your case).

[Clyde Schechter]

While I agree with both Joseph Coveney and Carlo Lazzaro, I would also raise a different question? Are the clusters you are referring to the four cities? If so, that is too few to get valid cluster-robust variance estimation. While there is no consensus on just how many clusters you need, I think everybody would agree that four is too few.

[frederick lim]

Hi Clyde, yes. Indeed, the cities are the possible clustering level I could consider in my panel data. In this case, could I use the few cluster problem resulting in invalid cluster-robust variance estimation as a rationale for not using cluster-robust SE?

[Clyde Schechter]

could I use the few cluster problem resulting in invalid cluster-robust variance estimation as a rationale for not using cluster-robust SE

Definitely!

[frederick lim]

That's great. This helps a lot! Do you have any good reference I could use to cite this? I came this article " a practitioner's guide to cluster-robust inference" which seems appropriate. Would be useful to have another one or two more.

[Clyde Schechter]

See http://cameron.econ.ucdavis.edu/rese...5_February.pdf, especially section VI.

[frederick lim]

Yes, this is the same as the one I posted. Thanks again Clyde.

[frederick lim]

Thanks Joseph and Carlo as well for the thoughts.