Re: st: How low can the percentage of uncensored cases be in heckprob?

[Steven Samuels <sjhsamuels@earthlink.net>]

I am not familiar with -heckprob-, but I doubt if the *percent* of  
uncensored observations matters much.

 -heckprob- fits two probit models. I know of results related to  
Margaret's question only for logit models.  For a single logistic  
regression model, the relevant sample size is the smaller of the  
number of events or non-events. Peduzzi et al. (1996) showed that the  
ratio of this number to the number of predictors should be at least  
15:1 to avoid bias from over-fitting.

-Steve

Refs:
Peduzzi PN, Concato J, Holford TR, Feinstein AR. (1995) The  
importance of events per independent variable in multivariable  
analysis, II: accuracy and precision of regression estimates. J Clin  
Epidemiol; 48: 1503–10.

Peduzzi PN, Concato J, Kemper E, Holford TR, Feinstein AR. (1996) A  
simulation study of the number of events per variable in logistic  
regression analysis. J Clin Epidemiol; 49: 1373–9.

M Babyak. (2004) What You See May Not Be What You Get: A Brief,  
Nontechnical Introduction to Overfitting in Regression-Type Models.  
Psychosomatic Medicine 66:411-421.  Full text:
http://www.psychosomaticmedicine.org/cgi/content-nw/full/66/3/411/




On Nov 11, 2008, at 12:15 PM, Maarten buis wrote:

> --- "Tyler, Margaret C D" <margaret-tyler@uiowa.edu> wrote:
> > In the example in the Stata reference -H heckprob, there are 95 total
> > and 59 uncensored observations, so 62% are uncensored. In my own
> > situation I have only about 19% uncensored. Is it still appropriate
> > to use heckprob for my analysis? I have run the equations and gotten
> > what seem to be valid results. rho is non-significant.
>
> You are obviously pushing your luck with that many censored cases. It
> is no longer very popular to make statements like you need at least N
> observation or p% uncensored cases for technique t to be appropriate
> (whatever appropriate may mean). So I don't think you will get the
> answer you are looking for. However, what you can do is run some
> simulations and see how well (or bad) your estimator behaves with a
> small number of uncensored cases. At the last Summer North American
> Stata Users' Group meeting I gave a talk on using Stata for doing this
> type of simulations, you can get the materials from:
> http://ideas.repec.org/p/boc/nsug08/14.html
>
> Hope this helps,
> Maarten
>
>
> -----------------------------------------
> Maarten L. Buis
> Department of Social Research Methodology
> Vrije Universiteit Amsterdam
> Boelelaan 1081
> 1081 HV Amsterdam
> The Netherlands
>
> visiting address:
> Buitenveldertselaan 3 (Metropolitan), room N515
>
> +31 20 5986715
>
> http://home.fsw.vu.nl/m.buis/
> -----------------------------------------
>
>
>
> *
> *   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/


*
*   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/