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]

Re: st: clustered SE smaller than

From   Austin Nichols <>
Subject   Re: st: clustered SE smaller than
Date   Thu, 21 Mar 2013 11:37:03 -0400

Robert G. LaChausse <>:
Cluster-Robust SEs are biased downward, and you need a large number of
clusters to get the asymptotic properties working for you, as
discussed in e.g.

If the clusters were balanced, one might expect 22 to be enough for
asymptotic properties to prevail.
With cluster sizes ranging from 13 to 321, you might need 50 or more.

That said, it is possible that errors are negatively correlated within
site and pretest scores are positively correlated, leading to a
negative impact of correction for clustering on SEs.

On Thu, Mar 21, 2013 at 11:22 AM, Robert G. LaChausse
<> wrote:
> Hi- I'm very new to using Stata and certainly no statistician.
> I have a clustered RCT with treatment and control conditions where subjects
> are clustered in 22 school sites. Subjects provided pretest and posttest
> scores. The level of random assignment was at the school (site) level.  I'm
> using  Stata to obtain robust standard errors for an ITT analysis. I am
> attempting to predict a posttest score (continuous measure) from the group
> variable (treatment or control; binary) controlling for the pretest score.
> I ran the analysis without considering  the clusters (.regress posttestscore
> pretestscore Group) and then again using  .regress posttestscore pretest
> score Group, vce(cluster site).  I wanted to see for myself the differences
> in SE's.
> My understanding is that the robust SE should be larger than the regular OLS
> SE. I found that the robust SE were actually a little smaller (ie .17
> regular SE and .14 robust SE). The cluster sizes range from 13 to 321. There
> are 22 clusters. The only thing I can think of is that the ICC is negative
> but don't know why. There aren't any outliers making one site different from
> another.  Is it possible to run the .regress posttestscore pretest score
> Group, vce(cluster site) using a random effects approach so that the ICC are
> all positive? Would that change the interpretation of the ITT analysis? What
> would be the command?
> Thanks, RGL
*   For searches and help try:

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