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

Re: st: metandi error: initial values not feasible

From   Roger Harbord <[email protected]>
To   [email protected]
Subject   Re: st: metandi error: initial values not feasible
Date   Wed, 30 Sep 2009 04:10:56 +0100

On Mon, Sep 28, 2009 at 8:30 PM, Jannik Helweg-Larsen <[email protected]> wrote:

> I'm trying to do a diagnostic meta-analysis of  51 studies using the
> metandi routine by Harbord et al- as described in Stata Journal 9.2
> using Stata 10.1.  None of the studies are particularly large( largest
> n~ 400).
> However I keep getting the following error: " initial values not feasible"
> I've tried changing the "nip" values and using the "force" option without luck
> Is there any work-around?

Dear Dr Helweg-Larsen,

This error message originates from -xtmelogit-, which is called
internally by -metandi-. I suggest you try adding the "gllamm" option
to the -metandi- command. This option requests use of the user-written
command -gllamm- for the actual model fitting instead of the built-in
command -xtmelogit-, which is currently the default when it is
available, i.e. in Stata 10 and above. You will need to install
-gllamm- (-ssc install gllamm-) if you do not have it already.

,Since publishing -metandi- in the Stata Journal, I've discovered
that this error does occur occasionally despite the considerable work
that -metandi- puts in to provide reasonably good (and certainly
feasible) initial values. To my surprise, -gllamm- is proving
considerably more reliable at fitting this model than -xtmelogit-,
albeit considerably slower (though not unbearably so -- perhaps a
couple of minutes at most, depending on your CPU).

If that doesn't help, please send me the particular data that is
causing this error and i'll look into it further. (Best to do so
off-list to my work email address, especially as i no longer read
Statalist regularly. My work email address is in the help-file for
-metandi- or via the link below - i only use this googlemail address
for Statalist).

Roger Harbord

*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index