Multinomial Logit Hessian Matrix Error

[starraf]

Hi all,
 
I’m running into a problem running a multinomial logistic regression in SPSS 17. After running the model, I receive the error message “Unexpected singularities in the Hessian matrix are encountered. This indicates that either some predictor variables should be excluded or some categories should be merged”.
 
After a search on the internet and through the listserv for this error message, I’ve reached the conclusion that this is likely a result of one of my predictor variables.
 
A little background on my model- I’m using 4 categorical variables in the model, one of which has 15 categories. The other three have 3 categories each. My dependent variable has 8  categories. When I run the model without the 15 category variable (we’ll call this x1 for ease of typing), it runs fine and converges. When I include x1 or run the model with x1 alone, I receive the error message.
I believe that this is due to empty cells when I cross tab x1 with my dependent variable. There are about 6 empty cells in this crosstabulation. Thinking about the theory behind my model and what each of the variables is representing, it makes sense to me that these cells are empty.
 
I’ve used the ‘delta’ option when running the regression to add .5 to each empty cell and the model runs without error.
 
My question is- does this sound like the appropriate way to deal with this issue? I plan to check the output of the model out in more detail before moving forward with it- but I wanted to get an initial check from the group first to see if there are any drawbacks to this approach.
 
I realize that another alternative is to collapse some categories within x1, however I would like to avoid that if possible.
 
A final note is that my total number of cases is over 60,000- so while I realize x1 has many categories, I don’t *think* that should be an issue in this case- especially given that the cases fall fairly evenly within these categories (but please correct me if I’m wrong!).
 
Thanks!
Sarah

[MaxJasper]

Hi Sara,

 

You might be able to use a modified model as follows:

 

ln[(m_11k*m_22k) / (m_12k*m_21k)] = lamda + k * delta.

 

delta= a new variable you add to your dataset based on some criteria like:

 

delta = var1 * (var2 = 1) * (var3 = 1).

 

Then PASTE your commands into syntax and edit syntax and add delta to:

 

/DESIGN = var1, ............, delta. [delete unnecessary cross producted variable).

 

You might need delta1, delta2,....new variables.

 

Max.

 

Hi all,
 
I’m running into a problem running a multinomial logistic regression in SPSS 17. After running the model, I receive the error message “Unexpected singularities in the Hessian matrix are encountered. This indicates that either some predictor variables should be excluded or some categories should be merged”.
 
After a search on the internet and through the listserv for this error message, I’ve reached the conclusion that this is likely a result of one of my predictor variables.
 
A little background on my model- I’m using 4 categorical variables in the model, one of which has 15 categories. The other three have 3 categories each. My dependent variable has 8  categories. When I run the model without the 15 category variable (we’ll call this x1 for ease of typing), it runs fine and converges. When I include x1 or run the model with x1 alone, I receive the error message.
I believe that this is due to empty cells when I cross tab x1 with my dependent variable. There are about 6 empty cells in this crosstabulation. Thinking about the theory behind my model and what each of the variables is representing, it makes sense to me that these cells are empty.
 
I’ve used the ‘delta’ option when running the regression to add .5 to each empty cell and the model runs without error.
 
My question is- does this sound like the appropriate way to deal with this issue? I plan to check the output of the model out in more detail before moving forward with it- but I wanted to get an initial check from the group first to see if there are any drawbacks to this approach.
 
I realize that another alternative is to collapse some categories within x1, however I would like to avoid that if possible.
 
A final note is that my total number of cases is over 60,000- so while I realize x1 has many categories, I don’t *think* that should be an issue in this case- especially given that the cases fall fairly evenly within these categories (but please correct me if I’m wrong!).
 
Thanks!
Sarah