[R] Different LLRs on multinomial logit models in R and SPSS

Ben Bolker bbolker at gmail.com
Thu Jan 6 17:17:13 CET 2011

Sören Vogel <sovo0815 <at> gmail.com> writes:

> Hello, after calculating a multinomial logit regression on my data, I
> compared the output to an output retrieved with SPSS 18 (Mac). The
> coefficients appear to be the same, but the logLik (and therefore fit)
> values differ widely. Why?

  Since constants that are independent of the model
parameters can be left out of a log-likelihood calculation
without affecting inference among models, it is quite common
for likelihoods to be calculated differently in different
software packages (for example, the (n/2 log(2*pi)) constant in the
Gaussian log-likelihood) -- this is true even among different
computations within the R ecosystem.  What's important is the difference
among log-likelihoods between models, which is the same (up to
the numeric precision of what you've shown us) for both
software packages.

> 135.02-129.8
[1] 5.22
> 204.2904-199.0659
[1] 5.2245

  Ben Bolker

More information about the R-help mailing list