Announcement

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

  • lclogitml2 yields error message - initial values not feasible

    Hello,

    I am analyzing the results of a Discrete choice experiment and am doing some latent class analysis.
    My conditional logit model has 6 attributes and 4 interactions.

    When estimating the model for certain numbers of classes, the lclogit2 command (presented below) works and yields some results.

    Code:
    lclogit2 decision, group(resp_task) nclasses(5) membership(sig_identity sig_status hp_id hp_status env_id atsci env_norm) rand(att1 des_imp des_top des_none att3 att4 att5 att6 visximp visxtop visxadp visxems visxprc) seed(1234)
    However, when I ask for the goodness of fit measures I get missing values '.'.
    And when I run the lclogitml2 to get the p-values, I get the message 'Initial values not feasible'

    This is the code that generates the error:
    Code:
    lclogitml2 decision, group(resp_task) nclasses(5) membership(sig_identity sig_status hp_id hp_status env_id atsci env_norm) rand(att1 des_imp des_top des_none att3 att4 att5 att6 visximp visxtop visxadp visxems visxprc) from(start, skip) iterate(0)

    This code works for other numbers of classes (such as 2 and 3), so there shouldn't be a typo in the codes.
    It would be really helpful if I could figure out why Stata cannot provide goodness of fit and p-values in these cases.

    Any insight would be highly appreciated!


    Han

    Last edited by Han Kyul Yoo; 12 Sep 2023, 03:57.

  • #2
    Han Kyul Yoo: The missing log-likelihood at convergence indicate that the 5-class model you're interested in may not be empirically identified (i.e., your model is too complex relative to the data you have). To check if this is indeed the case, you may want to play around with different values of -seed()- to initalise the EM algorithms from different sets of starting points, and see if the missing likelihood of -lclogit2- persists. If the problem persists, you'd have to simplify the model specification somehow, e.g., by dropping some variables from your membership function or by considering a smaller number of classes as you've already done.
    Last edited by Hong Il Yoo; 12 Sep 2023, 10:20.

    Comment

    Working...
    X