Learn how easy it is to sync an existing GitHub or Google Code repo to a SourceForge project! See Demo


mix of cotinuous and discrete distribution

  • RAF B
    RAF B


    I am trying to estimate a two-regime system of implicit equations of the form

    mu(x) + epsilon = 0, if x_{nj} > 0
    mu(x) + epsilon <=0 if x_{nj} = 0

    where mu(x) is highly nonlinear and x is an NxJ data matrix. I managed to estimate the first equation by letting \epsilon ~ N(0, seps) and

    mu(x) + epsilon = y_{nj},

    with y_{nj} = 0. It seems to work, but I have failed to estimate the complete system. I tried by letting y_{nj} be a censored variable such that

    y.ind_{n,j} ~ dinterval(mu_{n,j},-0.00001)
    y_{n,j}~ dnorm(mu_{n,j}, seps)
    peps ~ dchisqr(ppeps)
    seps <- pow(peps,-2)

    but failed to find consistent initial values.

    My main concern is, however, whether censoring is the best way to model this. It may be more appropriate to model the distribution of y as a mix of a truncated continuous distribution and a mass point at y=0. Can this be done in JAGS?

    Any advice would be appreciated. Thanks.

    Last edit: RAF B 2012-10-30