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

  • Implementing a mixture model of Gaussians in Stata

    Dear Statalisters,

    for my thesis I need to compare and evaluate statistical models provided by the literature concerning my subject. It is about predicting the dependent variable based on a set of independent variables (categorical as well as continuous).

    One model I want to recreate is a mixture of Gaussian distributions. The paper I am referring to is Altman, Kalotay - Ultimate recovery mixtures in Journal of Banking & Finance 40 (2014), p.116-129. The model is described in section 4 (I am not sure If I am allowed to provide a link, but it is easy to find). Basically, the dependent variable is modelled by using a probability weighted mixture of Gaussian likelihoods. To estimate the joint posteriors of the parameters (mean, standard deviation of mixture component and probability) the Marcov Chain Monte Carlo technique of Gibbs sampling is used.

    I am new to more complex models in Stata and I do not know how to obtain the information I am interested in. In the end, I want to compare the model predictions to other models and need to estimate the RMSE and MAE and some summary statistics and being able to contrast different distributions of subgroups depending on their characteristics. I have problems figuring out which command to use in Stata, therefore I am more than happy if someone could elaborate on my problem and where to find information how to recreate this model.

    I thought about a finite mixture model as it gives probabilities for each class as well as mean and standard deviation for the Gaussians, but in the paper it is not clearly mentioned that they use such an approach. Also, with a FMM it is not clear to which latent class an observation belongs. In the paper the “mixture assignments depend on an ordered probit model”.

    Kind regards