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: RE: Frontier code- r(1400) error

Subject   Re: st: RE: Frontier code- r(1400) error
Date   Sun, 10 Apr 2011 22:19:43 +0300 (EEST)

Hello all,

As I mentioned before, I am using a translog model, there are lots of
interaction terms in the model. I want to estimate efficiency scores year
by year. For the first year I started with a simple cobb douglas model and
then add the interaction terms one by one, finally I found a significant
model by the way I omit some interaction terms since they were creating
the error. For the second year I want to apply the same model, but the
first year's model doesn't work with the second year data. I got the same
error like "initial values not feasible". For every year the model will be
different. The original variables are included in the model for every
year. Only the interaction terms are different. The efficiency scores are
not so much different when one or more interaction terms are omitted. My
question is: Can I have a different model for every year? Will it create a
very serious bias in the efficiency scores?


> I have no specialist knowledge here, and others may have much better
> answers. But I think your question answers itself. The evidence is
> that with your model and data Stata is straining very hard to use a
> normal, and an exponential is a much better idea. How else can anyone
> interpret this information? Have you tried independent diagnostics,
> e.g. plotting observed vs predicted or residual vs predicted, if that
> makes sense?
> 2011/4/6  <>:
>> Dear Nick,
>> I tried the Cobb Douglas function at first then I include the other
>> variables to the model one by one. The results are much more better.
>> Another thing I want you to ask is about the distribution assumption.
>> When
>> I use the exponential distribution, the results are significant even if
>> the iteration number is small (for instance: 1000). However, when I use
>> the half normal distribution with 1000 iteration, some of the parameters
>> appear as dots in the results. Sometimes increasing the iteration number
>> works in such cases. Is it logical to increase the iteration number? Or
>> if
>> the results appear as dots even with the 1000 iteration, then it is not
>> needed to be continued?
>> Thanks
>> Hande
>>> Your original posting was
>>> Evidently, that guess of 29 observations was wrong. You are still left
>>> with the suggestion that your model appears too complicated and that
>>> one strategy is to simplify it radically. When you get a model that
>>> does not produce estimates, you can complicate it step by step to see
>>> where the problem lies.
>>> Nick
>>> 2011/4/2  <>:
>>>> Dear Nick and Gordon
>>>> Thank you for your advices. But Gordon says that "you have only 29
>>>> observations". I might have expressed myself in an incorrect way. I
>>>> have
>>>> 4
>>>> output variables, 3 input price variables for the 29 firms. Actually I
>>>> have 208 observations. It seems enough to estimate the frontier, isn't
>>>> it
>>>> ?
>>>> Thank you
>>>> Hande
>>>>> Nick's answer is correct.  You have 27 parameters plus the additional
>>>>> parameters for the efficiency error distribution and only 29
>>>>> observations.  This will never produce a satisfactory result.
>>>>> Translog frontier models can be difficult to estimate under the best
>>>>> of circumstances without trying to over-determine the frontier.  You
>>>>> should start by estimating the basic log-linear Cobb-Douglas form
>>>>> (dropping all of the interaction terms) and then introduce
>>>>> interactions individually and very carefully.  Even then it is
>>>>> unlikely that you will get any convincing results with such a small
>>>>> sample.
> *
> *   For searches and help try:
> *
> *
> *

*   For searches and help try:

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