Some problems with predictions of Bayesian multinomial model

[Mira Fedotova]

Good afternoon! 

I would like to get the predictions of the model but Stan tells me that something is wrong. Please, help me to understand the problem

predictions_of_bayesian_multi_logit <- "

data {

    int N;                                         // number of individuals 

    int<lower=2> K;                        // number of alternatives

    int<lower=1> D;                        // number of individual predictors

    int<lower=1> C;                        // number of cards for train sample

    int<lower=1> C_tilde;               // number of cards for test sample

    matrix[K, D] X[N, C];                 // individual predictors for train sample

    matrix[K,D] X_tilde[N, C_tilde]; // individual predictors for test sample

    int<lower=1,upper=5> y[N, C];  // outcomes 

}

parameters {

    vector[D] beta;

}

model {

    for (d in 1:D)

        beta[d] ~ normal(0, 5);

    for (n in 1:N)

        for (c in 1:C)

        y[n, c] ~ categorical_logit(X[n, c]*beta);

}

generated quantities {

    int<lower=1,upper=5> y_tilde[N, C_tilde];

    for (n in 1:N)

    for (c in 1:C_tilde)

        y_tilde[n, c] = categorical_logit(X_tilde[n, c]*beta);

}

" 

data_for_aggregate_logit <- list(N = N, y = y_train, X = X_train, X_tilde = X_test, D = D, C = C, C_tilde = C_tilde, K = K)

fit_aggregate_logit <- stan(model_code = predictions_of_bayesian_multi_logit, data = data_for_aggregate_logit, iter = 1000, chains = 4) 

When I`m trying to fit the model, Stan tells me the following:

SYNTAX ERROR, MESSAGE(S) FROM PARSER:

No matches for: 

  categorical_logit(vector)

Function categorical_logit not found.

ERROR at line 26

 24:        for (n in 1:N)

 25:        for (c in 1:C_tilde)

 26:            y_tilde[n, c] = categorical_logit(X_tilde[n, c]*beta);

                                                                     ^

 27:    }

Do you know how the given problem can be solved? Thanks in advance!

[Daniel Lee]

Hi Mira,

Use categorical_logit_rng() in generated quantities. rng stands for random number generator. 

Daniel

--
You received this message because you are subscribed to the Google Groups "Stan users mailing list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to stan-users+...@googlegroups.com.
To post to this group, send email to stan-...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Mira Fedotova]

Hi Daniel! Thanks for your help! 

I`ve changed the model as you said. Now the last row is the following:

y_tilde[n, c] = categorical_logit_rng(X_tilde[n, c]*beta) 

Unfortunately Stan again gives me the error:

SYNTAX ERROR, MESSAGE(S) FROM PARSER:

No matches for: 

  categorical_logit_rng(vector)

Function categorical_logit_rng not found.

ERROR at line 26

 24:        for (n in 1:N)

 25:        for (c in 1:C_tilde)

 26:            y_tilde[n, c] = categorical_logit_rng(X_tilde[n, c]*beta);

                                                                         ^

 27:    }

Please, let me know if you know how to tackle this problem

Kind regards, Mira

To unsubscribe from this group and stop receiving emails from it, send an email to stan-users+unsubscribe@googlegroups.com.

To post to this group, send email to stan-...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to a topic in the Google Groups "Stan users mailing list" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/stan-users/nGoDzNtBBmM/unsubscribe.
To unsubscribe from this group and all its topics, send an email to stan-users+unsubscribe@googlegroups.com.

[Stephen Martin]

According to the manual, categorical_logit_rng doesn't exist, BUT categorical(softmax(logit-scaled predictions)) = categorical_logit; so you could manually softmax the predicted values and use that in categorical_rng

To unsubscribe from this group and stop receiving emails from it, send an email to stan-users+...@googlegroups.com.

To post to this group, send email to stan-...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to a topic in the Google Groups "Stan users mailing list" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/stan-users/nGoDzNtBBmM/unsubscribe.

To unsubscribe from this group and all its topics, send an email to stan-users+...@googlegroups.com.

[Mira Fedotova]

I`ve read the manual and found the answer.

It should be done this way:

y_tilde[n, c] = categorical_rng(softmax(X_tilde[n, c]*beta))

Thank you so much for helping me! :)

Mira

[Mira Fedotova]

Thank you very much! :)

To unsubscribe from this group and all its topics, send an email to stan-users+unsubscribe@googlegroups.com.

[Bob Carpenter]

We really need to add categorical_logit_rng to the language.
I created an issue and this would be a super easy place to
write one's first Stan math library function (no derivatives!):

https://github.com/stan-dev/math/issues/517

The explicit softmax is required for now.

- Bob

> To unsubscribe from this group and all its topics, send an email to stan-users+...@googlegroups.com.