Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

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

Re: st: ML nonconvergence (sample) reasons

From   Nick Cox <[email protected]>
To   [email protected]
Subject   Re: st: ML nonconvergence (sample) reasons
Date   Wed, 28 Mar 2012 09:44:16 +0100

Non-convergence can be thought of as a matter of program, model and
data failing to find common ground. Where the blame lies can not be
determined abstractly, and blame might in principle be shared, for

1. The program might not be able to find a good fit even though it
exists. Sometimes, indeed quite often, you need to tweak how the
search for a good fit is conducted.

2. The model just might be a lousy idea for the data. What I think I
see on this list often are many attempts to fit highly complicated
models to data that may not be strong or numerous enough to take the

3. Or that might be twisted round to point an accusatory finger at the
data. It's the data's fault.

If only the data were to hand you could plot statistical people in a
DMP space according to the fractions of the time they blame Data,
Model or Program. The graphics could be done by -triplot- (SSC).


On Wed, Mar 28, 2012 at 9:06 AM, Benjamin Niug
<[email protected]> wrote:

> I have a rather statistical question: Having posted the
> non-convergence of the ML/Probit (-ivprobit-) command before (ML
> estimation not converging after the -ivprobit- where I had two
> endogenous variables and for IVs).
> I would like to know what the reason for this is. Especially, I was
> wondering whether non-convergence was potentially someow related to
> sample characteristics. Does anybody know whether this could be the
> case?
*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index