st: another missing data problem

["Lachenbruch, Peter" <Peter.Lachenbruch@oregonstate.edu>]

In an application, two measures are made on an individual.  One may be missing in some of them, either the first or the second.  At issue is the difference between them.  The number of observations is limited and the chance of finding others is zero.  I have suggested using the non-missing observation so the difference is zero - this will reduce the variance (I think) and so imply somewhat greater effects.  I've been thinking about imputing, but i'm not sure the effort is merited (no publication is contemplated).  The client is not really aware of correlation of observations.  To complicate matters, the client wants a percentage change which means dividing by the first observation - if it's zero, we have a big prolbem.
I have some ad hoc solutions:  make the rule symmetric and have tghe per cent change the sme whethere the first or second is 0.  I.e. (final - start)/start is the same as (final-start)/final if either is 0.  Not a perfect solution. 
Has anyone had this problem?


Peter A. Lachenbruch,
Professor (retired)
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/