st: Unbalanced Panel analysis with systemaic attrition

["Patrick Fahrun" <pfahrun@uni-koblenz.de>]

Dear Statalister, 

I am currently investigating startup performance in terms of whether they
reached a next round of funding (binary coded dependent variable) as well as
how long it takes them to reach the next round of funding (continuous
dependent variable). The number of observations per startup ranges from only
one to five rounds of funding (see below). 

     Freq.  Percent    Cum. |  Pattern
---------------------------+---------
     2055     61.53   61.53 |  1....
      896     26.83   88.35 |  11...
      302      9.04   97.40 |  111..
       75      2.25   99.64 |  1111.
       12      0.36  100.00 |  11111
---------------------------+---------
     3340    100.00         |  XXXXX
AVG: ~ 1.5 observations per startup

In addition, the period between founding rounds varies widely (e.g. from a
couple of days to several years). I adjusted the time variable by coding the
first observed round with t =1, the second observed round of funding with
t=2 and so on to use it as my time id.  However, I am not sure whether this
accompanies problems for a panel analysis?

I read anywhere that unbalanced panels are usually not a problem in terms of
biases, as long as the attrition is random. However, in my specific case the
attrition is systematically as companies, which not survive (do not reach a
new round of funding) drop out of my panel. How to take this into account in
an analysis? Or can I just ignore it?

What I have done so far`?
For my binary dependent variable (reached next funding round?) I used
xtlogit (RE), however I couldn?t test whether a FE model is more appropriate
through the Hausman test as the degrees of freedom are not sufficient  for
the estimation. For my continuous dependent variable (period to next round)
the Hausman test indicates a FE-model. However the FE model shows a negative
adjusted R-square as long as I do not restrict my panel to startups with > 2
observations. 

In a nutshell, I am pretty unsure whether it is advisable to use a panel in
my specific case. I was already wondering if it might be better just to
select one observation per company (randomly vs. first observed round) and
estimate a normal logit and regression. 

What type of analysis would you recommend? Is there a solution for this kind
of unbalanced (panel) dataset with systematic attrition? I didn?t find any
literature by now which solves my issues?

Thank you very much in advance. 

All the best
Patrick Fahrun

Master Student Entrepreneurship & New Business Venturing
Rotterdam School of Management


*
*   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/