Re: st: stcox: weighted regression changes the # of observations

[jpitblado@stata.com (Jeff Pitblado, StataCorp LP)]

Michael McCulloch <mm@pinest.org>

> I'm conducting cox regression on a dataset measuring survival after 
> treatment. I'm puzzled as to why weighted regression (using   -iweight- 
> after -stcox-) would show a different # of observations in the output. Here 
> are results of the ways I've done the analysis. In each case, I start with 
> only 33 subjects.
>
> (output omitted)

As documented in '[U] 11.1.6 weights'

	iweights, or importance weights, are weights that indicate the
	"importance" of the observation in some vague sense.  iweights have no
	formal statistical definition; any command that supports iweights will
	define exactly how they are treated.  In most cases, they are intended
	for use by programmers who want to produce a certain computation.

However -iweight-s are usually treated computationally as -fweight-s that are
allowed to take on non-integer values.  For commands that have that
interpretation, the sum of the weights is treated as the sample size.  I
believe most of Stata's estimation commands that allow -iweight-s (including
-stcox-) are consistent in this regard.

> *Note: I'm using inverse probability of treatment weighting, with the 
> following command immediately preceding -stcox-:
> stset datedied [iweight=wt_unst_II]

-bootstrap- is not meant for weighted data.   Currently -stcox- with the
-vce(bootstrap)- option does not complain when weights have been -stset, but
it should (we will fix this soon).
 
> Another anomaly is why the bootstrap option yields a narrower ci95.

Standard errors from the bootstrap are not guaranteed to be larger than
the conventionally computed standard errors.

--Jeff
jpitblado@stata.com
*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/