in some situations, modeling the numerator and denominator separately can be interesting. use 2 separate equations (event count model perhaps? or maybe just ln(y) regressed linearly on the X's) and then simulate to compute the quantities you are interested in, combining the sims together from the two equations. this is more flexible than doing it all together -- but it doesn't necessarily make sense. depends on the substance. more to the point of your question. if you have 0<Y<1, then you could do ln(Y/(1-Y)) and run a linear regression, or you could model Y directly as a beta distribution (which is similarly bounded). Gary --- http://gking.harvard.edu On Mon, Apr 26, 2010 at 8:46 AM, Zuzul, Tiona <tzuzul at hbs.edu> wrote: ratio. It's continuous but bounded at 0,1. The authors right now estimate parameters by running OLS, but we'd like to try running different models to check for dependence. Possible options appear to be gamma and tobit, though we are not sure which would be most appropriate. We'd really appreciate any ideas/comments. Thanks! _______________________________________________ gov2001-l mailing list gov2001-l at lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

Thanks so much for the suggestion and article, Gary and Iain! We are trying to run a beta regression using the betareg package, but keep getting the following error: Error in betareg(sim ~ constraint + stock + direct + indirect + bs + bb +? : ? invalid dependent variable, all observations must be in (0, 1) I have checked, and our dependent variable really DOES range from 0 to 1 so I have absolutely no idea what is going on. Any thoughts? Thank you!! From: gov2001-l-bounces at lists.fas.harvard.edu [mailto:gov2001-l-bounces at lists.fas.harvard.edu] On Behalf Of Iain Osgood Sent: Monday, April 26, 2010 8:59 AM To: Class List for Gov 2001/E-2001 Subject: Re: [gov2001] Continuous, Bounded Dependent Variable For the last idea, here's a quick article I just found discussing using the Beta distribution to model the outcome: http://ideas.repec.org/a/taf/japsta/v31y2004i7p799-815.html Iain On Mon, Apr 26, 2010 at 8:54 AM, Gary King <king at harvard.edu> wrote: in some situations, modeling the numerator and denominator separately can be interesting. ?use 2 separate equations (event count model perhaps? ?or maybe just ln(y) regressed linearly on the X's) and then simulate to compute the quantities you are interested in, combining the sims together from the two equations. ?this is more flexible than doing it all together -- but it doesn't necessarily make sense. depends on the substance. more to the point of your question. ?if you have 0<Y<1, then you could do ln(Y/(1-Y)) and run a linear regression, or you could model Y directly as a beta distribution (which is similarly bounded). Gary --- http://gking.harvard.edu On Mon, Apr 26, 2010 at 8:46 AM, Zuzul, Tiona <tzuzul at hbs.edu> wrote: _______________________________________________ gov2001-l mailing list gov2001-l at lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l _______________________________________________ gov2001-l mailing list gov2001-l at lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l