Notice: On March 31, it was **announced** that Statalist is moving from an email list to a **forum**. The old list will shut down on April 23, and its replacement, **statalist.org** is already up and running.

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

From |
Steve Samuels <sjsamuels@gmail.com> |

To |
statalist@hsphsun2.harvard.edu |

Subject |
Re: st: Testing equality of odds ratios outside of logistic regression |

Date |
Mon, 10 May 2010 08:50:00 -0400 |

I agree with Timothy that some kind of validation is needed to assess and compare predictions. -roccomp- will do pairwise comparisons of ROC curves, and predictions from each model can be graphed against the predictions from the others. To factor in differential costs of incorrect decisions, as Timothy suggests, use -senspec- , by Roger Newson. Steve On Mon, May 10, 2010 at 7:52 AM, Mak, Timothy <timothy.mak07@imperial.ac.uk> wrote: > Dear Michael, > > I'm not sure the proposed tests answer your question. You want to know (or 'test') whether several predictors predict the outcome equally well. But the tests offered test whether the odds ratios are the same. To me: Odds ratio are the same neither implies nor is implied by prediction being equally well. > > Prediction is usually assessed by validation: > > This requires a validation dataset and a training dataset. > Suppose you use your data to estimate the following: > Odds when predictor A + > Odds when predictor A - > Odds when predictor B + > Odds when predictor B - > etc. > > Now you need to apply a decision rule to these odds: > e.g. Predict outcome + if odds > 1; Predict outcome - otherwise > > Now say you have a validation dataset obtained elsewhere. One possible assessment strategy is: Score 1 for each correct prediction; Score 0 for each incorrect one. Comparing the scores for the different predictors: then you have a comparison of prediction performance. > > The point for this rather pedantic example is that: To compare prediction, you need a few more things than an odds ratio: (1) You need a decision rule, and (2) you need a scoring system, and (3) you need a validation dataset. > > Cross-validation is often used to overcome (3). For example, leave-one-out cross-validation means that you repeat the above leaving out one observation at a time from the training dataset and use that remaining data-point as your validation data. > > Frequently, (1) and (2) can be assessed together by a scoring rule that says, e.g.: Score = -(Actual outcome - predicted probability)^2 > > Another thing is: you usually want a scoring system that reflects the cost of making a wrong decision. Sometimes, a false positive may be much worse than a false negative, or vice versa. Usually decision rules are designed to hold either the false negative or false positive probability at a fixed level, while minimizing the other. > > I hope this helps towards your question. > > Yours, > Tim > > > > -----Original Message----- > From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Steve Samuels > Sent: 10 May 2010 01:18 > To: statalist@hsphsun2.harvard.edu > Subject: Re: st: Testing equality of odds ratios outside of logistic regression > > Actually Michael M.'s response answers your question, because you did > not ask for a simultaneous test. > > Steve > > On Sun, May 9, 2010 at 8:16 PM, Steve Samuels <sjsamuels@gmail.com> wrote: >> You can use -logistic- and test the interactions or -cc- with a test >> for homogeneity >> >> ***************** >> use http://www.stata-press.com/data/r11/bdesop, clear >> des >> gen tobac20 = tobacco <=2 >> xi: logistic case i.tobac20*i.alcohol [fw=freq] >> testparm _ItobXa* >> cc case tobac20 [fw=freq], by(alcohol) bd >> ********************** >> or >> On Sun, May 9, 2010 at 7:49 PM, Michael I. Lichter <MLichter@buffalo.edu> wrote: >>> I'm assisting on a paper where we examine the relationship between each of >>> four dichotomous predictors variables and one dichotomous outcome variable. >>> Prediction is our primary objective. The predictors are all measures of more >>> or less the same thing, and we want to know whether, *without controlling >>> for any of the others*, they predict the outcome equally well. We want to be >>> able to say, "If you could only pick one of these variables as a predictor >>> of the outcome, it wouldn't make any difference which one you selected." >>> >>> For each of the predictors we calculate a odds ratio and a corresponding >>> confidence interval. The odds ratios are very similar in magnitude and have >>> confidence intervals that overlap almost entirely. We did not do any formal >>> tests, not knowing of any offhand, and, because this isn't a central point, >>> we didn't think it was very important. When we reported that the odds ratios >>> were essentially equal, a reviewer objected that we had not tested for >>> equality. Any suggestions? >>> >>> In logistic regression, by the way, two of the four variables emerged as >>> significant predictors and two did not, controlling for the others. That is >>> of interest, but it doesn't answer my initial question. At least, I don't >>> think it does. >>> * >>> * 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/ >>> >> >> >> >> -- >> Steven Samuels >> sjsamuels@gmail.com >> 18 Cantine's Island >> Saugerties NY 12477 >> USA >> Voice: 845-246-0774 >> Fax: 206-202-4783 >> > > > > -- > Steven Samuels > sjsamuels@gmail.com > 18 Cantine's Island > Saugerties NY 12477 > USA > Voice: 845-246-0774 > Fax: 206-202-4783 > > * > * 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/ > -- Steven Samuels sjsamuels@gmail.com 18 Cantine's Island Saugerties NY 12477 USA Voice: 845-246-0774 Fax: 206-202-4783 * * 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**:**st: Testing equality of odds ratios outside of logistic regression***From:*"Michael I. Lichter" <MLichter@Buffalo.EDU>

**Re: st: Testing equality of odds ratios outside of logistic regression***From:*Steve Samuels <sjsamuels@gmail.com>

**Re: st: Testing equality of odds ratios outside of logistic regression***From:*Steve Samuels <sjsamuels@gmail.com>

**RE: st: Testing equality of odds ratios outside of logistic regression***From:*"Mak, Timothy" <timothy.mak07@imperial.ac.uk>

- Prev by Date:
**st: RE: line plot** - Next by Date:
**st: RE: AW: RE: AW: RE: AW: AW: RE: -ciplot- -eclplot- combine with box-options?** - Previous by thread:
**RE: st: Testing equality of odds ratios outside of logistic regression** - Next by thread:
**Re: st: Testing equality of odds ratios outside of logistic regression** - Index(es):