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

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

From |
"J. Li" <lij53@univmail.cis.mcmaster.ca> |

To |
statalist@hsphsun2.harvard.edu |

Subject |
Re: st: use -predictnl- to construct confidence intervals |

Date |
Wed, 01 Sep 2010 20:08:14 -0400 |

Stas, Thanks! Your explaination makes sense to me. I think it might be because that -predictnl- is just using delta method to calculate the standard errors numerically for each observation, and when the sample size is too small, the prediction would not be accurate enough based on poor linear approximation at these points. Daisy On Tue, 31 Aug 2010 16:57:27 -0500 Stas Kolenikov <skolenik@gmail.com> wrote: > -predictnl- is not aware of your interpretation of [whatever > complicated nonlinear expression you feed it] as a probability. To > get > more appropriately bounded, although asymmetric, prediction intervals > (technically, those are not confidence intervals, by the way, since > you are not talking about a population parameter), you might want to > run the prediction on the logit scale, and then manually transform > the > prediction and the confidence limits. > > On Tue, Aug 31, 2010 at 4:50 PM, J. Li > <lij53@univmail.cis.mcmaster.ca> wrote: > > Dear all, > > > > I am using a random effects ordered probit model to estimate the > > transition of health status over time. There are four health > states: > > poor, fair, good and excellent. After using -reoprob- estimation, I > > used -predictnl- to construct 95% intervals of the predicted > > probabilities of reporting each health status. I specified the > > expression to -predictnl- for reporting the poor health status as > > following: > > > > predictnl prob1= > > > normprob((_b[_cut1:_cons]/sqrt(1+(_b[rho:_cons]/(1-_b[rho:_cons])))) > - > > (predict(xb)/sqrt(1+(_b[rho:_cons]/(1-_b[rho:_cons]))))), > ci(p1lower > > p1upper) > > > > in which _b[_cut1:_cons] is the estimate of the lowest > cut-point > > and sqrt(1+(_b[rho:_cons]/(1-_b[rho:_cons]))) is a scale > > representing the square root of (1+ σ^2). In Wooldridge 2005 > > ?SIMPLE SOLUTIONS TO THE INITIAL CONDITIONS PROBLEM IN DYNAMIC, > > NONLINEAR PANEL DATA MODELS WITH UNOBSERVED HETEROGENEITY?, it is > > suggests that the original parameters should be divided by this > scale > > to construct the expected probability function in order to > calculate > > the average partical effects. > > > > After the prediction, I got two variables for the upper and lower > > bounds of the 95% confidence interval for each observation. But > > strangely, for some of the observations, the left endpoint is a > > negative value. This is supposed to be a predicted probability, so > I am > > not sure why -predictnl- is generating a negative value here. Did I > do > > something wrong with the expression in -predictnl- ? Many thanks! > > > > Daisy > > * > > * 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/ > > > > > > -- > Stas Kolenikov, also found at http://stas.kolenikov.name > Small print: I use this email account for mailing lists only. > > * > * 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/

- Prev by Date:
**st: re: speed of R graphics vs Stata graphics** - Next by Date:
**st: IN ARMA models** - Previous by thread:
**st: re: speed of R graphics vs Stata graphics** - Next by thread:
**st: IN ARMA models** - Index(es):