Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, is already up and running.

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

Re: st: too good to be true : lr test in mlogit?

From   John Litfiba <>
Subject   Re: st: too good to be true : lr test in mlogit?
Date   Sat, 14 May 2011 11:31:58 +0200

Dear Marten and Joerg,

Thank you for this idea : it can be a interesting solution indeed!
Maybe I shall add 2 remarks here :

1) The log likelihood doesnt converge when I try to fit a random or
fixed effect with xtlogit on my entire dataset..
I have to chose a very "small" (well, compared to the total size of
the sample) of about 10000 observations in order to see the results...
otherwise I get an error message after 3 or 4 iterations

2) The idea of running lets say M regressions over randomly chose
samples could be a solution, but it is statistically valid ? I mean if
I obtain the distribution of the parameters across my M simulation can
I infer something on the parameters of the simulation that should have
been done on the entire dataset ?

Have a good day
Best Regards,

On 13 May 2011 15:11, Joerg Luedicke <> wrote:
> On Fri, May 13, 2011 at 6:56 AM, Maarten Buis <> wrote:
>> On Fri, May 13, 2011 at 11:21 AM, John Litfiba wrote:
>>> But after various tentatives I think that it is not computationally
>>> feasible (at least on my PC) to fit a random or fixed effect logit on
>>> such a large dataset
>> I would not give up on that yet. I would select my model on a sample
>> of your data (I would sample higher level units, rather than
>> individual observations),
> This is what I would do, too. In addition, I would be interested in
> the variation of results from different samples. I would recommend
> setting up a program that is drawing a random sample, say 1%, running
> the model, and saving the results. Then let that program run a 1000 or
> whatever times. (How this can be done is described here:
> Small
> variation in results could then be indicative of the fact that it is
> pointless anyway to have a sample size of > 2mio. In that case, i.e.
> low variation of results, you can also compare your xtlogit model with
> the model you have now etc.  Large variation could be an indication of
> the fact that a very large sample size may be appropriate. You could
> also play with different sample sizes (say, 1%, 3%, 5%) and see how
> the variation of results changes across sample size. At the very
> least, this will provide useful information about your sample and
> might be helpful with respect to the specification of your model.
> J.
> *
> *   For searches and help try:
> *
> *
> *

*   For searches and help try:

© Copyright 1996–2015 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index