SCM

[#1278] Discuss Maddala/Powers/stata MoveStay differences

View Trackers | Bugs | Download .csv | Monitor

Date:
2011-02-10 16:31
Priority:
2
State:
Open
Submitted by:
Ott Toomet (otoomet)
Assigned to:
Ott Toomet (otoomet)
Hardware:
All
Product:
None
Operating System:
All
Component:
None
Version:
None
Severity:
minor
Resolution:
None
 
URL:
Summary:
Discuss Maddala/Powers/stata MoveStay differences

Detailed description
(by Nigel Melville)

Movestay is widely used in the Stata community.
Here is a document from Dan Powers that provides clear explanation of switc=
hing models in stata:
http://umich.com.cn/teaching/download/20092/CDAcoursematerials/handouts/Sel=
ectRegQ.pdf

If you visit the document by Powers, you will see a simulation example on p=
age 12-13 of switching regression. The data generating process is given on =
page 9.

When I run the same model as that on page 12/13,

I get the following:

[cid:3379615321_13702523]

As you can see, the rho1 (selection 1 FALSE) is .3143.
In Stata movestay (widely used program) the FALSE selection is rho2 and is =
given by -.3143.

There is a different sign here.

My conclusions, based on the discussion in Powers on pages 8-9 on "Interpre=
ting the Covariance Parameters" is that

1. Powers starts with the positive-signed disturbance term in selection =
equation on page 6 (deviating from Maddala), so the rho2 must be changed in=
sign for correct interpretation.
2. Your sampleSelection package starts with the negative signed disturban=
ce (consistent with Maddala) and so the rho1 can be immediately interprete=
d without any sign change.

Anyhow, once again I appreciate this discussion and I think we are getting =
closer to enlightenment :-)

Best, Nigel


Hi Nigel

On 3 February 2011 23:15, Melville, Nigel <npmelv@umich.edu> wrote:
When you say "all error terms are added (rather than subtracted)" this does=
not seem to be consistent with Maddala's (1983) formulation which I have copied b=
elow. Note the negative sign on selection equation 9.62.

Followup

Message
Date: 2011-02-12 21:25
Sender: Arne Henningsen

Hi Ott! I would be very happy if you clarify the selection issue of the Tobit-5 model in the vignette -- particularly a description of how rho1 and rho2 can be interpreted will be very helpful for many users who estimate a Tobit-5 model. I don't think that it is a disadvantage if the vignette more and more deviates from the published version.
Date: 2011-02-12 05:22
Sender: Ott Toomet

The correct form of the link is:

http://umich.com.cn/teaching/download/20092/CDAcoursematerials/handouts/SelectRegQ.pdf
Date: 2011-02-12 04:16
Sender: Ott Toomet

Should we add these comments to the vignette? I am inclined to do so, but then we will deviate more and more from the published version.

Attached Files:

Changes:

Field Old Value Date By
HardwareNone2011-02-12 21:25arne
Operating SystemNone2011-02-12 21:25arne
SeverityNone2011-02-12 21:25arne
assigned_tonone2011-02-12 21:25arne
Thanks to:
Vienna University of Economics and Business University of Wisconsin - Madison Powered By FusionForge