0

When I try to fit an Ordered Logit model to a long dataset the functions return an error Error in solve.default(H, g[!fixed]): system is computationally singular: reciprocal condition number = 3.64628e-18.

I've seen this but the data is not available anymore and I couldn't fix my issue. Also, tried this one and this one with no success.

The code that reproduces the dataset I am using is:

# Example of caracteristic dummies
caract <- tibble::tribble(~CPU, ~RAM,
                     1,    1,
                    -1,    1,
                     1,   -1,
                    -1,   -1)


# Weights for choosing the probabilities
weights <- c(0.3954545, 0.2727273, 0.2363636, 0.0954546)

# Simulating the choices block
block <- caract
block$Alternative <- 1:4

# Simulating the dataset
df <- NULL
n_decisors <- 50
set.seed(123456)
i <- 1
for(i in 1:n_decisors){
  block$Decisor <- i
  block$Choice <- sample(1:4, prob = weights)
  df <- rbind(df, block)
}

And the model fitting procedure that returns the error is:

library(mlogit)
# Creating the model data object
df_mlogit <- mlogit.data(data = df, choice = "Choice", 
                         shape = "long", alt.var = "Alternative", 
                         varying = 1:2, ranked = T)

# Fitting the model
m <- mlogit(formula = Choice ~ CPU + RAM, 
            rpar = c(CPU = "n", RAM = "n"), 
            data = df_mlogit)

When I make the block of choices non-exhaustive by sampling rows from caract in each iteration of the df simulation the error doesn't occur but I need to estimate a case where the data is the format simulates in this example. Also, I know nnet::multinom(Choice ~ CPU + RAM, data=df) solve this, but it won't fit a mixed logit model, which is also necessary for me.

Anderson Neisse
  • 118
  • 8
  • Perhaps `nnet::multinom(Choice ~ CPU + RAM, data=df)` might be an alternative for you? – jay.sf Jun 12 '19 at 16:14
  • It actually is @jay.sf but I also need to adjust a mixed logit model, and as far as I know `nnet::multinom(Choice ~ CPU + RAM, data=df)` won't solve this for the mixed logit case. – Anderson Neisse Jun 12 '19 at 16:21
  • I see. You may be interested in this post: https://stats.stackexchange.com/a/365115/163114 – jay.sf Jun 12 '19 at 16:25
  • In fact the frequentist alternative `gmnl` suffers from a similar issue. The data objet must be a `mlogit.data` class and when fitting it returns the error `Error in s + x[[i]] : non-conformable arrays`. Guess I will have to try the bayesian approach. Thanks for the help. – Anderson Neisse Jun 12 '19 at 16:49
  • IMO this is not just a programming issue and may be better asked at [Cross Validated](https://stats.stackexchange.com/help/on-topic) where there are people more familiar with this type of model. Voting to migrate there. – jay.sf Jun 12 '19 at 17:43

0 Answers0