Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down at the end of May, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

st: question about error in ML maximization


From   Woolton Lee <finished07@gmail.com>
To   statalist <statalist@hsphsun2.harvard.edu>
Subject   st: question about error in ML maximization
Date   Fri, 10 Sep 2010 17:02:06 -0400

Dear Statalist,

I am running a program that estimates a hospital choice model using
patient discharge data then bootstraps the data 100 times using sample
sizes equal to the original dataset.  The first step is used to
compute a mean effect, then the bootstrapping is used to calculate
standard errors for the estimate of the mean.  Whenever I estimate the
model using the unbootstrapped data there is no problem with
convergence.  However, whenever the model is being estimated on the
bootstrapped dataset (usually after about 50-90 bootstraps), I
encounter the error,

"discontinuous region encountered, cannot compute an improvement r(430)".

I would like to better understand what this error means and also what
types of data problems could cause it.  Because this error only occurs
when I am bootstrapping, which by the way is clustered on patient zip
code and stratified by year and market, it seems that one possible
cause is that the bootstrapped samples may contain little variation
causing the model to crash after several iterations.  Could someone
please comment on what this error means and provide some examples of
what could cause it?

When bootstrapping, the mean from each iteration is stored in a vector
(which is later used to compute a standard error), and since the model
typically crashes after completing 50 or more iterations I am thinking
of saving the completed iterates and combining them from another run
with 50 or more iterates to obtain 100 bootstrapped means for
computing the standard errors. Could someone also comment on the
statistical validity of calculating standard errors by pooling
separate runs of the program?  I would like to get some reactions to
what I am thinking of doing.

Thanks for your help,

Wtn
*
*   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/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index