Divergent transitions after warmup to be sloved

Modeling
, , , ,
9

Ah, looks like that was the wrong way to non-center the lognormal, according to this post from Bob.

Try with the following model:

data{
  int<lower=0>  Ntotal;
  int<lower=0>  K;
  vector[Ntotal] y;
  int s[K];
}
parameters {
  real<lower=-10,upper=10> mutheta;
  real<lower=-10,upper=10>  musigma;
  real<lower=0,upper=10> tausigma;
  real<lower=0,upper=50> tautheta;
  vector<offset=mutheta,multiplier=tautheta>[K] theta; 
  vector[K] sigma2_raw;
}

transformed parameters {
  vector[K] sigma2_jacob = musigma + sigma2_raw * tausigma;
  vector<lower=0>[K] sigma2 = exp(sigma2_jacob); 
}

model {
  int pos;
  pos = 1;

  sigma2_raw  ~ std_normal();

  for (k in 1:K) {
    segment(y, pos,s[k])  ~ normal(theta[k],sqrt(sigma2[k])); 
    pos = pos + s[k];
    theta[k]  ~ normal(mutheta, tautheta) ;
  } 
  mutheta  ~ uniform(-10,10);   
  musigma  ~ uniform(-10,10);   
  tausigma ~ uniform(0,10);
  tautheta ~ uniform (0,50);

  target += sum(sigma2_jacob);
}

generated quantities {
    real<lower=0,upper=1> Ppop ;
    real<lower=0,upper=1> Ppopdiff;
    real<lower=0> meanvar;

    Ppop = mutheta>0;
    Ppopdiff = mutheta>0.5;
    meanvar = exp(-musigma+tausigma/2);
}
2 Likes