Re: st: Subgroup analysis

[David Bai <db555@mail.com>]

Thank you, Clnde. I have looked at the coefficients in the white and 
African American subgroups, and found that they are not very close. So 
there is no evidence of a lack of statistical power, based on what 
you've suggested. Regarding the second suggestion, I have good reasons 
to believe that some interactions could be significant based on the 
subgroup analysis.

Maybe I didn't make it clear in my last inquiry, so let me clarify 
here:  My concern is not about whether the interaction effect is 
significant, but how to interpret the UNEXPECTED differential 
predictors' effects on the outcome across race/ethnicity if the 
interaction effect is significant. Assuming my analysis is correct, and 
found differential effects across race/ethnicity, but there is NO good 
theory to explain this difference, how can we explain it from 
statistical perspective? Lack of statistical power, with 28 predictors 
for a sample size of 600? Or African Americans are very homogeneous in 
the distributions of these predictors, and therefore it is hard for the 
analysis to distinguish any variations in the effects and therefore 
find non-significance in the results? or other interpretations?

Looking forward to more wisdom/insight to be shared with me. Thank you,

David



-----Original Message-----
From: Clyde Schechter <clyde.schechter@einstein.yu.edu>
To: statalist@hsphsun2.harvard.edu
Sent: Thu, Jul 8, 2010 9:33 am
Subject: Re: Re: st: Subgroup analysis


Comparing the statistical significance of effects in two sub-populations
is rather perilous.

I have two suggestions.  First, since you have already done the
race-specific analyses, just look at the coefficients in the White and
African American subgroups, disregarding standard errors and p-values.
Are the coefficients similar?  If so, you may well be simply finding a
lack of statistical power to detect in a subgroup of 600 subtle effects
that achieve statistical significance in your larger combined sample.

Second, and more formally, before even running the subgroup analyses, I
would have added race X predictor interaction terms to the model and 
then
tested the significance of those interaction terms.  If _they_ are not
significant, then the conclusion would be that your data do not provide
evidence of difference across races (which is not the same as evidence 
of
no difference across races).  If the interaction terms _are_ 
significant,
then the coefficients of those interaction terms give you estimates of 
the
cross-race differences in effects.

Hope this helps.

Clyde Schechter, MA MD
Associate Professor of Family & Social Medicine

Please note new e-mail address: clyde.schechter@einstein.yu.edu

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/

 

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/