[R] difference of the multinomial logistic regression results between multinom() function in R and SPSS

[peter dalgaard]

On Jan 5, 2012, at 02:10 , Yoo Jinho wrote:

> Dear all,
> 
> I have found some difference of the results between multinom() function in
> R and multinomial logistic regression in SPSS software.
> 
> The input data, model and parameters are below:
> 
> choles <- c(94, 158, 133, 164, 162, 182, 140, 157, 146, 182);
> sbp <- c(105, 121, 128, 149, 132, 103, 97, 128, 114, 129);
> case <- c(1, 3, 3, 2, 1, 2, 3, 1, 2, 2);
> 
> result <- multinom(case ~ choles + sbp + choles:sbp, abstol=1.0e-20,
> reltol=1.0e-20, MaxNWts=10000);
> 
> However, the estimated coeffcients and standard errors of the coefficeints
> are different from the SPSS.
> 
> For instance,
> 
> the estimated coefficients of the variable "choles" are 0.1946555 and
> 0.6244513 from the above result, but the SPSS result are  0.213120 and
> 0.662575.
> 
> Standard errors are much more different.
> 
> Why these kind of discrepancies occur?

Usually because the parametrizations differ and/or one of the programs (not always R) has convergence problems.

In the present case, I suspect that you missed the point in ?multinom about scaling variables on the rhs. 

> colSums(fitted(result))
       1        2        3 
2.999702 4.001649 2.998649 

suggests that convergence accuracy is not the greatest (the likelihood equations should set these equal to the observed counts: 3-4-3). 

Also notice that the interaction means that the interpretation of coefficients for choles are that they are effects for sbp==0, way outside of your data, so smaller discrepancies may be getting multiplied.

Fuller output from SPSS is needed to say anything more. Is the deviance smaller, e.g.?



-- 
Peter Dalgaard, Professor,
Center for Statistics, Copenhagen Business School
Solbjerg Plads 3, 2000 Frederiksberg, Denmark
Phone: (+45)38153501
Email: pd.mes at cbs.dk  Priv: PDalgd at gmail.com