Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.

Re: st: Problems with plots from generalized sensitivity analysis (GSA)

 From To statalist@hsphsun2.harvard.edu Subject Re: st: Problems with plots from generalized sensitivity analysis (GSA) Date Sun, 27 May 2012 12:38:41 -0500 (CDT)

This program -gsa- has not been uploaded on SSC, but I answer
here because the question is posted to statalist, I will
submit the program in a few month and the future users can
refer to this post in the near future. -gsa- is currently
available at my personal website with a few known minor bugs.
--------------------------

Hi Bertel,

I looked at your data and command, and tried to produce a good
contour by myself. In this case, you are right. The small
number of observation is the sole reason why you do not get a
good contour.

I felt that you understand what each option does correctly and
indeed set precision at the small value (1). This does reduce
the variations of the plots due to the error |t-\tilde{t}|,
but does NOT reduce those due to the uncertainty of the
contour, from which your analysis suffer.

As the coefficient has both point estimate and standard
errors, the contour has point estimate (in line) and
variances. This is because of the fact that, with actual data,
the size of omitted variable bias can be different from the
value calculated from the canonical formula of OVB depending
on the control variables. Although the correlation between
pseudo unobservable and control variables is near zero, the
correlation conditional on the treatment variable may not.

First thing you might want to do is to drop control variables
that are only weekly correlated with the treatment and outcome
variables. Indeed, -gsa- produced much better contour with
your data with a few control variables. Another thing is to
use multiple imputation and to use the most conservative
cases.

If you still have the problem, how should you interpret the
widely-dispersed scatter plots and the contour with wide
margin? Because you usually do not know the sign and strength
of the correlation between pseudo unobservable and control
variables conditional on the treatment variable, you might
want to interpret such contour like confidence interval. That
is, first drop 5% of the contour that are closest ("in terms
of what?" would be another question which I have not figured
out), and use the line that connects the plots closest to the
origin as a lower-bound contour. In this way, you can trust
the contour with 95% CI.

--