Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Heckprobit: Endless iterations for some specifications of selection model

    Hello everyone,

    I'm working on my first empirical project ever and got stuck running heckprobit.

    My dependent variable is House, indicating whether s.o. owns a house. So it's binary coded (1 if s.o. owns a house, 0 if not).
    Unfortunately I have this information from only 129 individuals, while my whole dataset consists of 523 persons. As I doubt the representativity of the sample of 129, I want to test (and possibly remediate) selection bias via heckprobit.

    Now I have the following two questions/ problems:
    First I had a selection model with only three variables. All with a p-value of less than 0.1, but likelihood-ratio test reported at the bottom of the output gave a p–value equal to 0.7197, which would imply, if I understand correctly, that the Heckman selection equation with my data is not a good idea and not necessary.
    Then, with some ttests to detect differences in variables between the large (523) and small (129) sample, I tried to find further variables for the selection equation. As a result, I added six variables to the selection equation, so now I'm working with a selection model of 9 variables. All but one variable are dummies.
    What happened was that on the one hand the likelihood-ratio test reported at the bottom of the output has a p–value of 0.0183, which would imply that the Heckman selection equation with my data is useful and better than standard probit. On the other hand, all variables in the selection model except for one have p-values between 0.12 and 0.88.
    The only significant variable is part of a set of dummies, so that I cannot simply include this (significant) one but leave out the others (with quite high p-values) of this dummy variable set, right?

    Trying to find a more parsimonious selection model, I again excluded some of the variables with high p-values. But I came across the following problem:
    For some specifications of the selection model Stata is endlessly "Fitting full model" with thousands of iterations, where log likelihood doesn't change any more. When I then stop the process via "Break" Stata tells me "parameter named athrho not found". So far, I didn't find out what the problem with the respective selection models is. To me it seems random whether a specification of the selection model can be calculated within a second or two or it leads to endless iterations.

    Does someone know what to do?
    Your help is greatly appreciated!

    Thanks a lot!
    Simon Kuhn

  • #2
    Hello Statalist,

    Sorry for my (too) long first post, but unfortunately I still have the problem with heckprobit.

    For some specifications of the selection model, Stata is endlessly "Fitting full model" with thousands of iterations "(not concave)", where log likelihood doesn't change any more.
    When I then stop the process via "Break", Stata tells me "parameter named athrho not found".

    I'm unable to find out what the problem with the respective selection models is. Sligth changes to the selection model make it incomputable.
    To me it seemsrandom whether a specification of the selection model can be calculated or it leads to endless iterations.

    Does anyone have an idea?
    Thank you very much in advance!
    Simon

    Comment


    • #3
      You didn't get a quick answer. It is hard to help you with your problem since it seems to derive from an interaction between the specifics of your model and your data. There are maximum likelihood models that simply won't estimate. You seem to have found which variants estimate and which don't. You might try some of the maximization options to see if they help (but often they don't help).

      Comment


      • #4
        Thanks for making time for my question, Phil!
        I was hoping for a more or less tangible reason that I then could consider, but it looks like I just need to choose one of the selection models that Stata is giving results for.
        Thanks anyway!
        Simon

        Comment

        Working...
        X