Troubleshooting Rstan error “running command…had status 1”

Question

I am getting an Rstan error that I don't understand and this particular error hasn't been addressed in the previous Stack Overflow questions.

The R code to call Rstan is

    # Fit the model
    start = Sys.time()
    pvl_d_fit <- stan(file = 'pvl_d.stan', data = pvl_d_dat, 
              verbose = TRUE,
              warmup = 1000, iter = 4000, thin = 3, chains = 3, 
              init = 'random', pars = c(  
              "A_ind", "mu_A", "la_A",
              "w_ind", "mu_w", "la_w",
              "a_ind", "mu_a", "la_a",
              "c_ind", "mu_c", "la_c"))
    end = Sys.time()                  
    end - start 

And get the following output:

SAMPLING FOR MODEL 'pvl_d' NOW (CHAIN 1).

Informational Message: The current Metropolis proposal is about to be rejected becuase of the following issue:
Error in function stan::prob::normal_log(N4stan5agrad3varE): Random variable is -1.#IND:0, but must not be nan!
If this warning occurs sporadically, such as for highly constrained variable types like covariance matrices, then the sampler is fine,
but if this warning occurs often then your model may be either severely ill-conditioned or misspecified.

Anyone run into this before?

(Note: this message is repeated over and over again).

Error in compileCode(f, code, language = language, verbose = verbose) : 
  Compilation ERROR, function(s)/method(s) not created! cygwin warning:
  MS-DOS style path detected: C:/PROGRA~1/R/R-30~1.2/etc/x64/Makeconf
  Preferred POSIX equivalent is: /cygdrive/c/PROGRA~1/R/R-30~1.2/etc/x64/Makeconf
  CYGWIN environment variable option "nodosfilewarning" turns off this warning.
  Consult the user's guide for more details about POSIX paths:
    http://cygwin.com/cygwin-ug-net/using.html#using-pathnames
C:/Users/User/Documents/R/win-library/3.0/rstan/include//stansrc/stan/agrad/rev/var_stack.hpp:49:17: warning: 'void stan::agrad::free_memory()' defined but not used [-Wunused-function]
C:/Users/User/Documents/R/win-library/3.0/rstan/include//stansrc/stan/agrad/rev/chainable.hpp:87:17: warning: 'void stan::agrad::set_zero_all_adjoints()' defined but not used [-Wunused-function]
c:/rtools/gcc-4.6.3/bin/../lib/gcc/i686-w64-mingw32/4.6.3/../../../../i686-w64-mingw32/bin/as.exe: file9d8452242a.o: too many sections (39807)
C:\Users\User\AppData\Local\Temp\ccOl2E
In addition: Warning message:
running command 'C:/PROGRA~1/R/R-30~1.2/bin/x64/R CMD SHLIB file9d8452242a.cpp 2> file9d8452242a.cpp.err.txt' had status 1 

The stan file is:

    data {
      int<lower=1> n_s;                       // # subjects
      int<lower=1> n_t;                       // # trials
      int<lower=0,upper=4> choice[n_s, n_t];  // # subj. x # trials matrix with choices
      real<lower=-25,upper=2> net[n_s, n_t];  // Net amount of wins + losses   
                                              // (# subj. x # trials matrix) 
    }

    parameters {
      // Group-level mean parameters
      real mu_A_pr;   
      real mu_w_pr;
      real mu_a_pr;  
      real mu_c_pr;

      // Group-level standard deviation  
      real<lower=0> sd_A;
      real<lower=0> sd_w;
      real<lower=0> sd_a;
      real<lower=0> sd_c;    

      // Individual-level paramters    
      real A_ind_pr[n_s]; 
      real w_ind_pr[n_s]; 
      real a_ind_pr[n_s];   
      real c_ind_pr[n_s]; 
    }

    transformed parameters {
      real<lower=0,upper=1> mu_A;   
      real<lower=0,upper=5> mu_w;
      real<lower=0,upper=1> mu_a;  
      real<lower=0,upper=5> mu_c;

      // Individual-level paramters    
      real<lower=0,upper=1> A_ind[n_s]; 
      real<lower=0,upper=5> w_ind[n_s]; 
      real<lower=0,upper=1> a_ind[n_s];   
      real<lower=0,upper=5> c_ind[n_s];   

      // Group-level precision parameters  
      real<lower=0> la_A;
      real<lower=0> la_w;
      real<lower=0> la_a;
      real<lower=0> la_c;     

      mu_A <- Phi(mu_A_pr);
      mu_w <- Phi(mu_w_pr) * 5;
      mu_a <- Phi(mu_a_pr); 
      mu_c <- Phi(mu_c_pr) * 5;  

      for (s in 1:n_s)
      {
         A_ind[s] <- Phi(A_ind_pr[s]);
         w_ind[s] <- Phi(w_ind_pr[s]) * 5;
         a_ind[s] <- Phi(a_ind_pr[s]); 
         c_ind[s] <- Phi(c_ind_pr[s]) * 5;  
      }

       la_A <- pow(sd_A, -2);
       la_w <- pow(sd_w, -2);
       la_a <- pow(sd_a, -2);
       la_c <- pow(sd_c, -2);  
     }

     model {
      vector[4] p;
      real Ev[4];
      real dummy[4];
      real theta;
      real v;

      # Prior on the group-level mean parameters
      # probit scale [-Inf, Inf]  
      mu_A_pr ~ normal(0, 1);
      mu_w_pr ~ normal(0, 1);
      mu_a_pr ~ normal(0, 1);
      mu_c_pr ~ normal(0, 1);

       # Prior on the group-level standard deviation
      sd_A ~ uniform(0,1.5); 
      sd_w ~ uniform(0,1.5); 
      sd_a ~ uniform(0,1.5);    
      sd_c ~ uniform(0,1.5);

       # Individual-level paramters
       for (s in 1:n_s)
      {
        A_ind_pr[s] ~ normal(mu_A_pr, sd_A);      
        w_ind_pr[s] ~ normal(mu_w_pr, sd_w);   
        a_ind_pr[s] ~ normal(mu_a_pr, sd_a);   
        c_ind_pr[s] ~ normal(mu_c_pr, sd_c);    
      }

      for (s in 1 : n_s)  // loop over subjects
      {  
        theta <- pow(3, c_ind[s]) - 1;     

       // Trial 1
        for (d in 1 : 4)  // loop over decks
        {  
          p[d] <- .25;
          Ev[d] <- 0;
        }  
        choice[s,1] ~ categorical(p);

        // Remaining trials
        for (t in 1 : (n_t - 1))  
            { 
            if (net[s,t] >= 0)
            v <- pow(net[s,t], A_ind[s]);
            else
            v <- -1 * w_ind[s] * pow(abs(net[s,t]), A_ind[s]);

          Ev[choice[s,t]] <- (1 - a_ind[s]) * Ev[choice[s,t]] + a_ind[s] * v;      

          for (d in 1 : 4)  // loop over decks
          dummy[d] <- exp(fmax(fmin(Ev[d] * theta, 450), -450));       

          for (d in 1 : 4)  // loop over decks
          p[d] <- dummy[d] / sum(dummy); 

          choice[s, t + 1] ~ categorical(p);
        }
      } 
    }

Answer 1

The "informational message" is probably the most frequently (un)answered question on the Stan mailing list.

• It can mean that your model is ill-defined; to find out if this is the problem, try to use very narrow priors first, until the message disappears or is rare. After that, release the priors into the wild one after the other to find the ill-defined culprit.
• In most cases, however, the message comes from badly defined initial values; the messages often disappear after the burn-in (sorry, Andrew, should be warm-up) phase. If the result is in a reasonable range, possibly obtained with narrower priors, try to use these as initial values first, then add small random variations.
• Sometimes, removing the initial values altogether and let stan to the dirty work helps; but probably you have already tried this.

This said, I am totally confused what your "compile" error means. The model should not run at all with that error.