Getting an ill-typed argument error

Modeling
1

Please share your Stan program and accompanying data if possible.

Hello,

I’m trying to slowly build a GLM model with a target variable that has a quasi-poisson distribution. I’m not sure if this distribution is possible in STAN but I’m trying it with the poisson distribution. I’m getting the below error.

Semantic error in 'C:/Users/JORDAN.HOWELL.GITDIR/PycharmProjects/pythonProject/marine_bayes/model/marine_bayes.stan', line 20, column 4 to column 26:
   -------------------------------------------------
    18:              a[n] = mu + credit_model_normalized[n]*credit_beta;
    19:      }
    20:      h_pp ~ poisson_log(a);
             ^
    21:  }
   -------------------------------------------------

Ill-typed arguments to '~' statement. No distribution 'poisson_log' was found with the correct signature.

My model is as follows:

data {
    int<lower=0> N; // number policy
    real<lower=0> h_pp[N]; // hull pure premium
    vector[N] credit_model_normalized;// normalized credit
}

parameters {
    real<lower=0> mu;
    real credit_beta ; //credit coefficient
}

model {
       mu ~ normal(0,3);
       credit_beta ~ normal(0,5);

       vector[N] a;
    for (n in 1:N) {
            a[n] = mu + credit_model_normalized[n]*credit_beta;
    }
    h_pp ~ poisson_log(a);
}

instead of

model{
vector[N] mu = alpha+beta*x;
y~normal(mu,sigma);
}

Is this because I’m using the wrong distribution or another issue? If it’s a distribution issue, is it possible to run a STAN model with a quasi-poisson distribution?

2

The issue is that poisson distributions only accept integer outcomes, and so aren’t compatible with the real-type outcome h_pp

3

Thank you for that confirmation. This is an insurance loss data set where losses follow a quasi-poisson distribution. Does STAN have a quasi-poisson option?

4

Unfortunately not, your closest option for overdispersed count data in Stan would be the negative-binomial

5

I tried the below from the docs…

data {
    int<lower=0> N; // number policy
    real<lower=0> h_pp[N]; // hull pure premium
    vector[N] credit_model_normalized;// normalized credit
}

parameters {
    real<lower=0> mu;
    real credit_beta ; //credit coefficient
}

model {
       mu ~ normal(0,3);
       credit_beta ~ normal(0,5);

       vector[N] a;
    for (n in 1:N) {
            a[n] = mu + credit_model_normalized[n]*credit_beta;
    }
    h_pp ~ neg_binomial(a);
}

I get the following;

Ill-typed arguments to ‘~’ statement. No distribution ‘neg_binomial’ was found with the correct signature.

I’ve also tried neg_binomial_2 a and neg_binomial_2_lpmf and get the same error. What is the correct command to use in the model block?

EDIT: @maxbiostat has edited this post for syntax highlighting.

6

As I mentioned earlier, methods for count data are not compatible with outcomes defined as real. In your data block, change the definition of h_pp to int:

int<lower=0> h_pp[N]; // hull pure premium
7

Ok. I see now that I miscommunicated. When modeling insurance losses, the losses can be modeled via a poisson distribution but are real numbers (i.e. the dollar amount of the loss) so the actuarial literature states to use a quasi-poisson to model the data. (frequentist models).

There is also a large part that states to use a tweedie distribution since there is a large amount of zeros (people who have made no claims).

Is there anything like a tweedie or other zero-inflated distribution that I can use that has the signature of a poisson but excepts real numbers?

8

Ah I see what you mean, unfortunately not. A user has previously implemented their own tweedie distribution in Stan, which might be applicable in this case: Tweedie Likelihood (compound Poisson-gamma) in Stan