[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

From |
"Timothy Dang" <[email protected]> |

To |
stata <[email protected]>, [email protected] |

Subject |
Re: st: marginal effects from clogit with lots of control dummies |

Date |
Wed, 19 Sep 2007 11:43:53 -0700 |

Replying to myself... I suspect I've identified something I was previously doing wrong, although problems continue. In the regression: -------------- xi: clogit choiceflag NotOO Price OtherThings i.Group*NotOO i.Period*NotOO i.Subject*NotOO, group(id) -------------- ...I was correctly treating NotOO, Price, OtherThings as features of the alternative chosen, but the control variables are features of the overall decision. So, I'm trying to shift things to - asclogit - as follows: -------------- xi: asclogit y NotOO Price OtherThings, case(id) alternatives(n) basealternative(0) casevars(i.Group i.Period i.Subject) -------------- and that leads to more-immediately-evident problems than the previous approach, even if the previous approach was wrong. The output goes: -------------- note: _Iuniquesub_48 dropped because of collinearity note: _Iuniquesub_40 dropped because of collinearity note: _Iuniquesub_32 dropped because of collinearity note: _Iuniquesub_24 dropped because of collinearity note: _Iuniquesub_16 dropped because of collinearity note: variable r_minprice has 72 cases that are not alternative-specific: there is no within-case variability note: model has collinear variables; convergence may not be achieved Iteration 0: log likelihood = -361.60637 (not concave) -------------- (uniquesub is the actual variable name I've noted as "Subject" above, and r_minprice is one of the "OtherThings".) ...and so on, (not concave) at every step, and failing to converge. using the - difficult - option doesn't seem to help. Question: does the warning "note: model has collinear variables; convergence may not be achieved" mean that some variables are perfectly collinear? I assume since it's dropping some collinear variables that the warning means there are variables which are closely but not exactly collinear? In any case, I suspect that a problem like this is a matter of me thinking more about my data, but if anyone has a clever idea, I'd love to hear it. Thanks -Timothy On 9/18/07, Timothy Dang <[email protected]> wrote: > Hello- > > I'm estimating the rules for buyer purchase decisions in an experiment > with what amounts to panel data. Buyers can purchase any one of 3 > goods or purchase nothing, and this occurs repeatedly for 20 periods. > The basic decision model is: > > Pr(choice i) = e(b0+b1*Price_i+b2*OtherThings_i) > > ----------------------------------------------------------------------------- > 1+ Sum(for j=1,2,3)[e(b0+b1*Price_j+b2*OtherThings_j)] > > I'm taking each buyer decision and breaking it up into 4 records--one > for each possible decision, in order to use clogit for the estimation. > To control for dependencies I'm actually running the regression as: > > xi: clogit choiceflag NotOO Price OtherThings i.Group*NotOO > i.Period*NotOO i.Subject*NotOO, group(id) > > In the above "id" is a unique number for the original record before I > breaks it up into 4 records. So "id" is the number of a specific > decision. "NotOO" is a flag which is 1 when the decision is actually > to purchase something and 0 when the decision is to purchase > nothing--its coefficient will be the b0 in the model. Group, Period, > and Subject are all things which could be reasonably expected to have > interdependencies. I interact them with NotOO to maintain the > distinction between purchasing something and nothing. > > The regression results seem reasonable. Individually, none of the > controls are significant, but jointly they are. > > But I also want marginal effects, and when I ask for those the > marginal effects of the variables I'm actually concerned with are very > nearly zero and insignificant. I'm pretty sure this is wrong, and an > artefact of the way I'm doing the regression. When I leave off the > dummy terms the coefficients are very nearly the same, and marginal > effects are significant in both size and p-value. Actually, the one > exception to this is the constant term which goes from around 2 to > around 25, andfrom statistically significant to not, when the control > dummies are added. > > So, I'm hoping for advice on either the mechanics or theory of what's going on. > > As an aside, I'm using clogit rather than mlogit because of what > appears to be a bug with mlogit. When I run the same model with mlogit > (using a single record per decision instead of 4 records, and > constraints to fit the model), one of the coefficients for one of the > choices winds up getting dropped for no sensible reason. > > Thanks > > -Timothy > > ------------------------------ > Timothy O'Neill Dang / Cretog8 > 623-587-0532 > One monkey don't stop no show. > * > * For searches and help try: > * http://www.stata.com/support/faqs/res/findit.html > * http://www.stata.com/support/statalist/faq > * http://www.ats.ucla.edu/stat/stata/ > -- ------------------------------ Timothy O'Neill Dang / Cretog8 623-587-0532 One monkey don't stop no show. * * For searches and help try: * http://www.stata.com/support/faqs/res/findit.html * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/

**References**:**st: marginal effects from clogit with lots of control dummies***From:*"Timothy Dang" <[email protected]>

- Prev by Date:
**Re: st: boostrapping in quantile regression** - Next by Date:
**Re: st: Chi-square test for Categorical Data Analysis** - Previous by thread:
**st: marginal effects from clogit with lots of control dummies** - Next by thread:
**st: mkdir with local** - Index(es):

© Copyright 1996–2024 StataCorp LLC | Terms of use | Privacy | Contact us | What's new | Site index |