SPSSX Discussion

Multinomial Logistic Regression

Classic

List

Threaded

3 messages Options

rct

Multinomial Logistic Regression

I was running Multinomial LR using SPSS 16 for data that has two groups-
Hindus (N=212) and Muslims (N=159). Outcome variable (Reaction to norm
violation) is nominal and has 3 categories (Retribution, Retaliation and
Reconciliation). Reconciliation is the reference category. The obtained
frequencies for Hindus for these 3 categories are 68, 17 and 127,
respectively. The six predictors include a categorical variable- ideological
preference ( MC=38; CC=174). For this predictor (ideology) B coefficient
returned is 20.319 Standard error shows up as 0 for retaliation category and
Exp (B) as 1.499E-9. With 3 more predictors B coefficient for Retaliation
reported is -19.959 , SE again is equal to zero and Exp(B) =2.147E-9.
Can you help me understand these findings in the light of the following
message
"For Hindu Either the maximum likelihood estimates do not exist or some
parameter
estimates are infinite.
The NOMREG procedure continues despite the above warning(s). Subsequent
results shown are based on the last iteration. Validity of the model fit is
uncertain."
Thanks

--
Sent from: http://spssx-discussion.1045642.n5.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

Rich Ulrich

Re: Multinomial Logistic Regression

That's probably not a useful solution. Get rid of predictors

that are true by definition or tautology.

The category of "Hindu either" rather sounds like it would

be nested under Hindu, never Muslim. That's probably

not useful for a prediction that wants to use other variables.

I expect that if you look at two-way and 3-way cross-tabulations,

you will see zeros in places that prevent completed computations.

You can get infinite coefficients when you have complete

separation of groups, unless a computer program makes some

ad-hoc adjustment.

Most likely (but not always), the other coefficients in the results

will not be meaningful, either

Rich Ulrich

From: SPSSX(r) Discussion <[hidden email]> on behalf of rct <[hidden email]>
Sent: Friday, March 27, 2020 7:26 AM
To: [hidden email] <[hidden email]>
Subject: Multinomial Logistic Regression

===================== 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

Rich Ulrich

Re: Multinomial Logistic Regression

In reply to this post by rct

Here are more comments on the problem.

Do any of the variables show strong prediction by

univariate testing? Is this an attempt tease out new

associations, or is the multi-variable analysis needed

in order to "control for" some effect that you know is

strong and present?

In your stated problem: Back up and look at the logic of

the variables. Look at your hypotheses. Retribution is

barely distinguishable from Retaliation, in my vocabulary.

I would test (1) Retribution vs Retaliation (dropping the

rest of the sample); and (2) Reconciliation vs the other two.

That's probably what the two roots of the multi-variable

analysis would reflect.

I'd use logistic regressions for those two particular contrasts,

which seem to convey the hypotheses that are interesting.

As an alternative analysis, you /could/ run multiple-group

discriminant function (dummy-coding your categorical

predictor as two contrasts) -- Its advantage is that the

procedure comes with some nice ancillary information,

like, plotting group centroids against the two functions.

The test of (1) will have a smaller N and (thus) less power

of analysis than (2). You might reduce the list of predictors

to those that are most likely on logical grounds, in order

to help against the lack of power.

Rich Ulrich