On Wed, 15 Oct 2014, Sven Schreiber wrote: The function package I mentioned in my previous messages covers Heckman's estimator, plus its generalisation by Keane and Sauer (Econometrica 2009). Wooldridge's estimator is already available, since it doesn't require more than a routine for RE probit, which we already have (the probit command with the --random-effects option). ------------------------------------------------------- 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 -------------------------------------------------------

Jan, I've got a script that estimates a Bayesian dynamic panel probit model as described in Koop, Poirier & Tobias. It's not in function package form yet, but if you give me a couple days I could get you a more-or-less user-friendly version. Jack, I'd be glad to do a swap with you too -- it'd certainly be useful for me to compare my results with a frequentist version. Cheers, Peter -----Original Message----- From: gretl-users-bounces(a)lists.wfu.edu [mailto:gretl-users-bounces@lists.wfu.edu] On Behalf Of Jan Tille Sent: Wednesday, October 15, 2014 7:06 AM To: r.lucchetti(a)univpm.it; Gretl list Subject: Re: [Gretl-users] Is there a possibility to estimate a dynamic panel probit model in gretl? Sven: First, let me thank you for your answer as well as providing the source of Wooldridge estimator. Indeed, you are right, such models are not trivial at all. Jack: Great to hear, that you and Ms. Pigini are developing a function package for these kind of models. I am not an expert, but I have read that the Heckman has a higher precision than Wooldrige and is also less subject to bias. Furthermore, if I could be of help testing your package, then I am happy to offer it, but I do not have Stata or any other econometric software at hand, to compare the results ( I am not sure if R provides a package (maybe pglm) for dynamic panel probit, anyway even if it should be possible within R, gretl is much more user friendly and has better documentation, which I am really appreciate). All I can offer is to run your functions on my data set and send the results to you. Furthermore, I can share my data set with you, so that you can benchmark the results. I would come back to you on Friday, because this is the only time of the week, when I have enough time to concentrate on such complex issues. Again, thank you Jack and Sven. Regards, Jan -----Ursprüngliche Nachricht----- Von: Plus.line MailSystem [mailto:cyrus@mailer.plusline.de] Im Auftrag von Riccardo (Jack) Lucchetti Gesendet: Mittwoch, 15. Oktober 2014 09:21 An: Gretl list Betreff: Re: [Gretl-users] Is there a possibility to estimate a dynamic panel probit model in gretl? On Wed, 15 Oct 2014, Sven Schreiber wrote: The function package I mentioned in my previous messages covers Heckman's estimator, plus its generalisation by Keane and Sauer (Econometrica 2009). Wooldridge's estimator is already available, since it doesn't require more than a routine for RE probit, which we already have (the probit command with the --random-effects option). ------------------------------------------------------- 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 ------------------------------------------------------- _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users

On Thu, 16 Oct 2014, Jan Tille wrote: The functions Claudia and I are working on should work ok with unbalanced panels too. We should be able to send you something tomorrow so you can try things out. Of course, if others are interested, just drop us a line: the more people help debugging the code, the better. ------------------------------------------------------- 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 -------------------------------------------------------

I've also got a "balanced" flag in my script that allows for unbalanced data. PS -----Original Message----- From: gretl-users-bounces(a)lists.wfu.edu [mailto:gretl-users-bounces@lists.wfu.edu] On Behalf Of Riccardo (Jack) Lucchetti Sent: Thursday, October 16, 2014 12:00 PM To: Gretl list Subject: Re: [Gretl-users] Is there a possibility to estimate a dynamic panel probit model in gretl? On Thu, 16 Oct 2014, Jan Tille wrote: The functions Claudia and I are working on should work ok with unbalanced panels too. We should be able to send you something tomorrow so you can try things out. Of course, if others are interested, just drop us a line: the more people help debugging the code, the better. ------------------------------------------------------- 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 -------------------------------------------------------

Dear Peter, dear Jack, please forgive me to cause you so much inconvenience while trying to help me. Unfortunately I think, that your functions, might probably not work for my model/problem. The reason why I think this might be is that I have seen in the gretl documentation, that "dpanel" requires the definition of the lag p. Well basically, my lag is p=1, so far so good. However, the problem lies in my data. Please let me explain (though it is not directly related to gretl, but more general) my problem. I am trying to estimate whether an individual whose variable of interest is above the sample median (y>=median => y*=1, else 0) in period t has a significantly positive probability to repeat being above the sample median in t+1 (p=1), "conditional" upon the individuals that were already in my sample in period t. Going forward in time, more and more individuals enter the sample (i.e. an unbalanced panel). However, this means, while an individual might be above the median in period t+1, "conditional" on upon the individuals exisiting in period t, it might as well be below the median in the new period t, when the new individuals entered the sample. I don't know if that is clear, maybe an stylized example brings more clarity. Lets assume that there are three 12-Months periods for a sample of mutual funds. In t=1 I calculate that fund i's compound return equals 5% and that this is above the median for all funds with a full 12-months series of returns. In t=2 I calculate the fund i's return again and it equals -2%, but is still above the median for period t=2 and all funds that were available in period 1. Meanwhile new funds entered the market that also have a 12-month history by the end of period 2, which is then relevant, to assess whether funds above the median in period 2 are also above the median in period 3. Now, ranking all funds available at the end of period 2, it might be that the -2% are no longer above the median of all funds. You see, I can not simply take the lag, as is suggested in "dpanel" I hope, it could make it clear, that it is not a gretl problem at this stage, but a data problem. Again, sorry to cause so much inconvenience. Kind regrads, Jan t=1 t=2 t=3 1 5% -2% 1% 2 1% 1% 2% 3 2% -3% 3% 4 3% -4% 4% 5 4% -5% 5% 6 2% 6% 7 3% 7% 8 4% 8% 9 5% 9% -----Ursprüngliche Nachricht----- Von: Plus.line MailSystem [mailto:cyrus@mailer.plusline.de] Im Auftrag von Summers, Peter Gesendet: Donnerstag, 16. Oktober 2014 22:22 An: r.lucchetti(a)univpm.it; Gretl list Betreff: Re: [Gretl-users] Is there a possibility to estimate a dynamic panel probit model in gretl? I've also got a "balanced" flag in my script that allows for unbalanced data. PS -----Original Message----- From: gretl-users-bounces(a)lists.wfu.edu [mailto:gretl-users-bounces@lists.wfu.edu] On Behalf Of Riccardo (Jack) Lucchetti Sent: Thursday, October 16, 2014 12:00 PM To: Gretl list Subject: Re: [Gretl-users] Is there a possibility to estimate a dynamic panel probit model in gretl? On Thu, 16 Oct 2014, Jan Tille wrote: The functions Claudia and I are working on should work ok with unbalanced panels too. We should be able to send you something tomorrow so you can try things out. Of course, if others are interested, just drop us a line: the more people help debugging the code, the better. ------------------------------------------------------- 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 ------------------------------------------------------- _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users

Well, sorry, I thought I am, because I want to know whether being above the median is followed by being above the median and that is binary (i.e. an LDV), isn't it? Jan -----Ursprüngliche Nachricht----- Von: Plus.line MailSystem [mailto:cyrus@mailer.plusline.de] Im Auftrag von Sven Schreiber Gesendet: Freitag, 17. Oktober 2014 12:34 An: gretl-users(a)lists.wfu.edu Betreff: Re: [Gretl-users] Is there a possibility to estimate a dynamic panel probit model in gretl? Am 17.10.2014 um 12:28 schrieb Jan Tille: I haven't read this thoroughly (sorry), but at first glance it sounds as if you aren't really in the probit model framework. If that's true it's not a data problem but a conceptual one, and no dynamic panel probit package (existing or forthcoming) would help you. good luck anyway, sven _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users

Jack, sorry. Let's start with the clearer things first. The unbalanced nature appears both due to entries as well as exits. If I am conceptually wrong and the assumptions are violated, than I would agree, that it does not make much sense using an inadequate method. Well I guess the problem is, that a fund might have two different ranks in one time period, depending on the peer group that is used. I know, that in the literature to assess this kind of problem, multiple cross-sectional regressions are used and then you either look at each year (for example: Brown, Goetzman and Ibbotson (1999) - http://www.jstor.org/stable/10.1086/209603). I was hoping that a panel approach would yield better results. I could of course use the annual returns, that do not depend on other funds' performance, however, it is very likely that returns are not persistent due to moving markets. Therefore, the hypothesis was that relative performance is persistent. Sven: I see your point, yes, being above the median is clearly endogenous, nonetheless it is of interest, whether being above the median is repeatable . Maybe it would indeed be a better approach to run several cross-sectional regressions on a yearly basis. Jan -----Ursprüngliche Nachricht----- Von: Plus.line MailSystem [mailto:cyrus@mailer.plusline.de] Im Auftrag von Riccardo (Jack) Lucchetti Gesendet: Freitag, 17. Oktober 2014 13:20 An: Gretl list Betreff: Re: [Gretl-users] Is there a possibility to estimate a dynamic panel probit model in gretl? On Fri, 17 Oct 2014, Jan Tille wrote: I must confess your problem is not exactly clear to me; I read your previous message several times and I can't make up my mind on it, but, like Sven said, I am doubtful as to whether applying a dynamic panel probit to your data is the right thing to do from a statistical point of view. The fact is, it sounds like the unbalancedness of your panel is endogenous: from what I gather, you'd have a peculiar form of panel attrition (on entry instead of exit, as is more common), so a few assumptions for the dynamic panel probit model would be violated and you'd get inconsistent estimates. That said: (1) I may be wrong (I usually am) (2) nobody is stopping you from using a dynamic panel probit model anyway, if only to see what happens if you do. Claudia and I will send you our code as soon as we manage to put togther some decent documentation to it. ------------------------------------------------------- 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, 15 Oct 2014, Jan Tille wrote: Yes. Claudia Pigini and I are working on a (rather ambitious) function package for doing just that. It's not ready yet for public consumption, mainly because it's pretty much undocumented as of now, but it already works quite well. We can send you the gfn file so you can help us test it if you want. ------------------------------------------------------- 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 -------------------------------------------------------