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

From |
Rich Steinberg <rsteinbe@iupui.edu> |

To |
statalist@hsphsun2.harvard.edu |

Subject |
Re: AW: st: RE: wald tests with mfx |

Date |
Tue, 14 Jul 2009 17:23:01 -0400 |

Maarten buis wrote:

--- On Mon, 13/7/09, Rich Steinberg wrote:What I want is the vce of the marginal effects at the mean of various subsamples of the data. I had thought, like you, that if two marginal effects are statistically different in the tobit evaluated at the mean of the full sample, the same would be true when evaluated for any subsample. But when I run mfx on various subsamples, the Z scores of each coefficient are sensitive to the subsample. More is going on than simply applying the McDonald and Moffitt decomposition to a different level of censorship when I run that mfx command. Maybe that is a bug, but my more sophisticated friend thinks that mfx at a subsample mean is computing correctly and I am misunderstanding how to apply it to subsamples. Thus my current query. Martin, and anyone else who is curious, here is the research problem that generated this strategy. We are seeing whether the marginal propensity to make charitable donations differs with the source of income (or annuitized equivalent for wealth variables), among other things, testing the implicit assumption of economists that money is money and consumption behavior of individuals does not depend on where that money came from. So for example, we want to know if the marginal propensity to give out of welfare transfer payments equals that out of earned income, annuitized inheritance income, etc. A single tobit evaluation tells us whether the coefficients on these various income measures are equal around the mean of our full sample. But there are no observations where the respondent received welfare income and had the mean level of other income sources, so the comparison at the mean has no clear interpretation of the sort we want. In our base model, we therefore test for the equality of coefficients within each of 7 subsample means (defined by having positive levels of each of our 7 income variables). (We hope to have more advanced estimates accounting for unmeasured heterogeneity that makes a person receiving welfare different from a person receiving an inheritance, but we are not there yet). Maybe we should estimate the entire tobit separately for each subsample, but as some of the subsamples are small (particularly our main income source of interest, inheritances) we were hoping those regressions could be pooled because the slopes with respect to the latent dependent variable were not significantly different in the subsamples.The marginal effect depends on where you evaluate it. More,specifically it will become less when you evaluate at a pointcloser to the cut-off point. So if you evaluate the marginaleffects of different groups at different means you will add asource of variation in results between these groups that youmay not want, because now you are not comparing like with like.I can think of two suggestions:1) Maybe the overall median lies within the range of allsub-groups. In that case you could avoid the entire problem byestimating the marginal effects at this value.2) You could compare the effects on the latent variable, asthese effects are linear, and thus not depend on where youevaluate them.Hope this helps, Maarten ----------------------------------------- Maarten L. Buis Institut fuer Soziologie Universitaet Tuebingen Wilhelmstrasse 36 72074 Tuebingen Germany http://home.fsw.vu.nl/m.buis/ -----------------------------------------Martin Weiss wrote:<> You are more than welcome! I did not endorse yourresearch strategy as I donot think that what you are attempting is necessary,but on a technicallevel, you can get your hands on the desired returnedresult in thisfashion. Stata provides a full array of postestimation toolsfor every estimationcommand, and this list is usually exhaustive. If youwant the -vce- of theoriginal estimation, you can get it like this: *** sysuse auto, clear generate wgt=weight/1000 tobit mpg wgt gear_ratio, ll(17) estat vce *** HTH Martin -----Ursprüngliche Nachricht----- Von: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu]Im Auftrag von Rich SteinbergGesendet: Montag, 13. Juli 2009 22:51 An: statalist@hsphsun2.harvard.edu Betreff: Re: st: RE: wald tests with mfx Thanks for responding, and so quickly. Goodanswer, but I learned from following your advice that this reproduces only the standard errors, not the covariances I would need for a Wald test. I don't see an e( ) that saves the vce. And I can't use the vce from the original tobit because the sample is changing.Martin Weiss wrote:<> " For the latter, I know from the Help file thatmfx saves what I need as e(Xmfx_se_dydx), but I can't figure out how to see that."*** sysuse auto, clear generate wgt = weight/100 tobit mpg wgt len tu head, /* */ ll(17) ul(24) mfx compute, /* */ predict(e(17,24)) mat l e(Xmfx_se_dydx) mat A= e(Xmfx_se_dydx) matrix list A di A[1,3] *** HTH Martin -----Original Message----- From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu]On Behalf Of Rich SteinbergSent: Montag, 13. Juli 2009 22:10 To: statalist@hsphsun2.harvard.edu Subject: st: wald tests with mfx As a relatively unsophisticated user, I have triedfor a few hours and failed to solve the following problem. After running tobit, I want to test for the equality of marginal effects for the unconditional observable dependent variable. No problem with mfx or dtobit. But I don't want to evaluate these marginal effects test at the mean, median or zero of the full sample prior to testing. Instead, I want to use the estimates from the full sample, but evaluate the marginal effect at means of various subsamples. So I have, for example:tobit totgiv $income $control, ll vce(clusterfid68)mfx if welfare01>0, pred(ystar(0,.)) Now, I want to test for the equality of twoelements of $income in this subsample. But everything I try works on the tobit coefficients, not the mfx output. So how do I retrieve this for a "test" command (ideally) or even display the vce from the mfx to do thetest by hand?For the latter, I know from the Help file that mfxsaves what I need as e(Xmfx_se_dydx), but I can't figure out how to see that. This should be easy, but stumped me.Thanks everyone. * * 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/* * 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/-- Rich Steinberg Department of Economics, 516 Cavanaugh Hall IUPUI Indianapolis, IN 46202-5140 317-278-7221 "Unanswered email since 1955" * * 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/

-- Rich Steinberg Department of Economics, 516 Cavanaugh Hall IUPUI Indianapolis, IN 46202-5140 317-278-7221 "Unanswered email since 1955" * * 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/

**References**:**Re: AW: st: RE: wald tests with mfx***From:*Maarten buis <maartenbuis@yahoo.co.uk>

- Prev by Date:
**Re: st: Treatment for Missing Values - What Options ?** - Next by Date:
**Re: AW: st: RE: wald tests with mfx** - Previous by thread:
**Re: AW: st: RE: wald tests with mfx** - Next by thread:
**Re: AW: st: RE: wald tests with mfx** - Index(es):

© Copyright 1996–2015 StataCorp LP | Terms of use | Privacy | Contact us | What's new | Site index |