tolerance level for zero frequencies in multinomial logistic regression

[Greg]

Hi everyone,

I checked the forum and couldn't find anything specific regarding my question. I'm running MLR in which all the predictors are categorical. I ran series of crosstabs to see if there are zero/missing cases between the IVs and DV. Everything is fine (no zero cells), except for combining some categories on some IVs.

However, I'm still getting a certain percentage of zero frequencies, particulary over the 50% range. (The parameters' standard errors are fine).

I would greatly appreciate it if anyone suggest any ideas as to what else I can do?

Thanks,
Grigoris

[Neda Faregh]

Could you provide more information about your data?

What percentage of hits (1, or yes) do you have in the data? What is your overall N? How many predictor variables? How many leves of the DV? Do you get output warnings regarding a problem with the Hessian Matrix?

Crosstabs certainly show the major cells, but it is always possible to have zero cells, especialy if your DV category of interest is a small percentage of your total N. I don't know your data so I'm guessing at this...but remember that it is always possible that there are no cases for the cell that represents, say, employed females within a certain age and certain income category, in a certain geography, with a particular disease profile, who are also endorsing your DV category of interest. You wold have to run several crosstabs and use various filters to get exactly the number of people that would fit this example category I just gave.

possible ways around the problem are to either increase your overall N, decrease your IV's, or collapse some or all of your IV categories, and even collapse some of the DVcategories if possible. Any of these options would effectively increase the number of cases you get per cell.

Hope this helps,

 

Neda

---

From: grigoris <[hidden email]>
To: [hidden email]
Sent: Tuesday, March 3, 2009 12:19:53 PM
Subject: tolerance level for zero frequencies in multinomial logistic regression

Hi everyone,

I checked the forum and couldn't find anything specific regarding my
question. I'm running MLR in which all the predictors are categorical. I ran
series of crosstabs to see if there are zero/missing cases between the IVs
and DV. Everything is fine (no zero cells), except for combining some
categories on some IVs.

However, I'm still getting a certain percentage of zero frequencies,
particulary over the 50% range. (The parameters' standard errors are fine).

I would greatly appreciate it if anyone suggest any ideas as to what else I
can do?

Thanks,
Grigoris
--
View this message in context: http://www.nabble.com/tolerance-level-for-zero-frequencies-in-multinomial-logistic-regression-tp22313360p22313360.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Henrik Lolle]

Facing the problem of empty cells for combinations of independent
variables it can be impossible to estimate the usual goodness of fit
test, but the parameters of the model should be okay.

Best,
Henrik

Quoting NEDA FAREGH <[hidden email]>:

> Could you provide more information about your data?
> What percentage of hits (1, or yes) do you have in the data? What is
> your overall N? How many predictor variables? How many leves of the
> DV? Do you get output warnings regarding a problem with the Hessian
> Matrix?
> Crosstabs certainly show the major cells, but it is always possible
> to have zero cells, especialy if your DV category of interest is a
> small percentage of your total N. I don't know your data so I'm
> guessing at this...but remember that it is always possible that there
> are no cases for the cell that represents, say, employed females
> within a certain age and certain income category, in a certain
> geography, with a particular disease profile, who are also endorsing
> your DV category of interest. You wold have to run several crosstabs
> and use various filters to get exactly the number of people that
> would fit this example category I just gave.
> possible ways around the problem are to either increase your overall
> N, decrease your IV's, or collapse some or all of your IV
> categories, and even collapse some of the DVcategories if possible.
> Any of these options would effectively increase the number of cases
> you get per cell.
> Hope this helps,
>
> Neda
>
>
>
>
> ________________________________
> From: grigoris <[hidden email]>
> To: [hidden email]
> Sent: Tuesday, March 3, 2009 12:19:53 PM
> Subject: tolerance level for zero frequencies in multinomial logistic
> regression
>
> Hi everyone,
>
> I checked the forum and couldn't find anything specific regarding my
> question. I'm running MLR in which all the predictors are categorical. I ran
> series of crosstabs to see if there are zero/missing cases between the IVs
> and DV. Everything is fine (no zero cells), except for combining some
> categories on some IVs.
>
> However, I'm still getting a certain percentage of zero frequencies,
> particulary over the 50% range. (The parameters' standard errors are fine).
>
> I would greatly appreciate it if anyone suggest any ideas as to what else I
> can do?
>
> Thanks,
> Grigoris
> --
> View this message in context:
> http://www.nabble.com/tolerance-level-for-zero-frequencies-in-multinomial-logistic-regression-tp22313360p22313360.html
> Sent from the SPSSX Discussion mailing list archive at Nabble.com.
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>

************************************************************
Henrik Lolle
Department of Economics, Politics and Public Administration
Aalborg University
Fibigerstraede 1
9200 Aalborg
Phone: [phone]
************************************************************

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD