Hello Prof. Lucchetti and Prof. Pigini, After estimating a ivprobit model with Hessian parameter covariance matrix, I got this message on the top of the results "Brrr! Halving params". The results are exactly the same in Stata. Is this something I need to be concerned with? What does it mean? Thanks. Best regards,

Thanks very much, Prof. Lucchetti for the explanation. The results have been produced without problem, but as you have noticed, I do have two coefficient estimate around 0.001 level (e.g., 0.00148632, -0.00270550). I will try to adjust the scale and re-estimate the model as you have suggested. Best regards, Juehui Shi (Richard) ------------------------------------------------------------ ---------------- RA, TA, and Lecturer PhD Management Candidate University at Buffalo, SUNY School of Management Department of Operations Management and Strategy Jacobs Management Center Buffalo, NY 14260-4000 Contact information Tel.: (716) 334-9757 [preferred] Email1: juehuish(a)buffalo.edu Email2: rchrdshi(a)gmail.com ------------------------------------------------------------ ----------------- https://www.linkedin.com/in/richardrjs http://ssrn.com/author=2339807 On Sun, Oct 16, 2016 at 8:42 AM, Riccardo (Jack) Lucchetti < r.lucchetti(a)univpm.it&gt; wrote: > On Sun, 16 Oct 2016, Juehui Shi wrote: > > Hello Prof. Lucchetti and Prof. Pigini, >> >> After estimating a ivprobit model with Hessian parameter covariance >> matrix, >> I got this message on the top of the results "Brrr! Halving params". The >> results are exactly the same in Stata. Is this something I need to be >> concerned with? What does it mean? Thanks. >> >> Best regards, >> > > That is a debugging statement that is triggered in a relatively rare > case, when the automatic initialisation procedure gives parameter values > which result in some observation having an estimated probability less than > 6.3051168e-16 (not particularly healthy from a computational point of > view). Therefore, the algorithm starts form a more conservative set of > numbers. If you don't experience convergence problems in the end, there is > nothing to worry about. > > It's difficult to say anything more specific without looking at your > data. However, you may want to check the scale of your variables: if > anything has a coefficient larger than 1000 or smaller than 0.001 in > absolute value, it may be a good idea to change the unit of measurement for > that variable; this usually helps quite a bit with the numerical procedures. > > Hope this helps, > > > ------------------------------------------------------- > Riccardo (Jack) Lucchetti > Dipartimento di Scienze Economiche e Sociali (DiSES) > > Università Politecnica delle Marche > (formerly known as Università di Ancona) > > r.lucchetti(a)univpm.it > http://www2.econ.univpm.it/servizi/hpp/lucchetti > -------------------------------------------------------

On Wed, 19 Oct 2016, Juehui Shi wrote: > Hello Prof. Lucchetti and Prof. Pigini, > > I'd like to cite the HIP package in my paper. Can you write me the proper > citation? Thanks. I really wouldn't know. But thanks anyway! :)

Thanks, Prof. Cottrell. Professors, I have one additional question. Although the coefficient estimates of HIP match those of Stata IV Probit exactly, I have found that the endogeneity test (Wald) result is a little bit different. For example, the HIP give me Wald (t = 4.06864, p = 0.0437), but the Stata gives me Wald (t = 3.68, p = 0.0551). My sample size N = 2204. Cragg-Donald shows 107.096 and both AR and Wald tests reject the null of weak instruments. As shown, this is a delicate situation. My question is that should I reject or not to reject the null of exogeneity in this case? and What might attribute to the difference in the test? Thanks.