13th October 2011 at 1:02 pm #3010MCMember
Argh, I’m trying to a multinomial logistic regression in SPSS.
Qs 1) I have managed the basics of running:
DV: 4 unordered ctageories
IV: 3 unordered categories
Where the output is largely ok and makes sense. Ok with the exception of my goodness of fit box doesn’t contain any statistics. Not as in not significant, but pearson chi square is .000, df is 0 and sig. is a full stop. I do have a chi square statistics in my model fitting information box following the -2 log likelihood and I do have pseudo r-square. The answer may be blindingly simple but basically, why do I not have a goodness of fit statistic, is it only relevant when more/certain variables are included and therefore I shouldn’t worry, or does it mean I’ve done something wrong in input? I can’t seem to find a pattern of it happening/not happening, i.e.
running IV->DV= no goodness of fit
running IV->DV when another IV is used to define the subpopulation but isn’t in the model= yes goodness of fit statistic
running model with 3 IVs-> DV=no goodness of fit.
The last example refutes my original idea that maybe I just needed several IVs so I’m back to having no clue.
Next qs: As well as running the above IV (3) vs. DV (4) from qs 1, I want to test the effect that another IV (is binary) may have on the relationship. I want to sort of simulate a hierarchical linear regression but in multinomial form due to my unordered categorical DV. I believe this can be achieved by creating interaction effect and using custom stepwise in the model and from reading around, I think I have to do something like:
Run multinomial: IV(3) -> DV (4)
Run multinomial: IV(2) -> Dv (4)
Run multinomial: IV(3) +IV(2) + IV(3)xIV(2) -> DV (4)
And I am under the impression I compared the loglikelihoods or coefficients or something for each model. However, am very much not sure and I haven’t yet found a clear resource on this for SPSS (exists for Stata) so does any know how to would go about testing the effect of IV(2) on the relationship of IV(3) and DV (4) in SPSS?
Thanks very much in advance for taking the time to help a confused idiot.14th October 2011 at 10:39 am #3015MCMember
Hi, my supervisor advised me to run multinomial logistic regressions on the classes from a latent class analysis, as I think it’s conceptually tidy (working in probailities throughout etc.) and appropriate for my research question (I’m looking at likelihood/risk factors). I’ve tried to break them down into binary LR, but sadly it reduces my sample size for some of the groups to the extent that I don’t have sufficient cases to run the binary. Thanks for the resource and for taking the time to respond, i’m giving it a read now.17th October 2011 at 9:01 pm #3014Jeremy MilesParticipant
Have you tried running a crosstab and getting a chi-square of the table? This should give the same chi-square as your multinomial, and might reveal problems (like zero cells).
The multinomial helps you to see what was significant.11th December 2011 at 8:12 am #3013KELWYN D’SOUZAMemberHi
I am using multinomial logistic regression to study driver distraction. My dependent variable has four outcome categories (not distracted, slightly distracted, distracted, very distracted). The first outcome category (not distracted) is set as the baseline. The SPSS 17.0 analysis consists of three independent binary comparisons. Now, when I calculate the Predicted Probability P(Y) of an outcome category Y occurring, I am using the binary logistic regression equation suggested by Field (2009) page 300 for each of the three pairs of MLR output as follows:
P(Y1) = 1/(1+e-Y1),
Y1 = β0 + β1X1 + β2X2 + β3X3 + ——— + βkXk
In my discussion of the results, I am mentioning the probabilities are in comparison (versus) to the baseline category. For example, the likelihood of outcome B occurring VERSUS A (baseline) occurring will increase by ……
Please let me know if this is a plausible approach, or do I need to use the following more complex MLR equation for predicting each of the three outcomes:
P(Y1) = eY1/( eY1+ eY2+ eY3)
Please respond at your earliest convenience. If you are interested, I can send you my paper.
Field, A. (2009). Discovering Statistics Using SPSS. SAGE Publications Ltd., 3rd Edition.11th December 2011 at 10:17 am #3012AnonymousInactive
if your reponse is in terms of probabilities and not in terms of odds (which is what you end up with after using a logit link in any type of logistic regression-like problem) then yes, you need to use a forumla to transform your SPSS output (which is in log-odds) to probabilities…29th December 2011 at 12:42 am #3011KELWYN D’SOUZAMember
Thanks, Oscar. My response is in probabilityvalues. So, I am on the right track and getting positive results. Your message has confirmed that my methodology is correct. Kelwyn.
- You must be logged in to reply to this topic.