multinomial logistic interaction output interpretation

[delliott]

Hi
I'm running an interaction in multinomial logistic regression.  my DV is departure from college and my 2 variables are Strat= college selectivity (16=low; 18=high) and ASE_YE which is a categorical measure of perception of academic self efficacy(1=low; 3=high).  When I ran my model with other covariates my interaction was significant so I asked SPSS to compare groups.  My output is attached.Nominal+Regression.doc

questions:
1) There is no estimate for the final category (18 selectivity and high ASE).  Is this the reference against which the rest of the output should be assessed?  EX: A student with low efficacy perceptions at a highly selective college is 1.23 times more likely to depart than a student with high efficacy perceptions at the same caliber institution.

OR
Are these parameter estimates simply the estimates that correspond to that particular group (i.e. low efficacy in a low selectivity college)?
I'm thinking this is not the case, but if so then what is the estimate for a student in a highly selective college with high efficacy? the intercept?

2) I'm trying to generate odds ratios and probabilities for prototypical cases so I can plot them on a graph.  I know I can generate them from my regression equation but I can't do this b/c when I run my logistic regression without denoting (clicking categorical button in upper right) which of my variables are categorical my constant is outrageous (17.452) because it's picking up the excess varianceand I end up with crazy probabilities like .000000004. However when I run my regressions with my categorical variables denoted then there's always a comparison group and no coefficient for the variable on a whole.
a) is there any way to fix this?
b) if not then how else can I obtain probabilities for prototypical cases so I can create a graph?

Thanks so much in advance!!
~Diane

[Bruce Weaver]

delliott wrote

Hi
I'm running an interaction in multinomial logistic regression.  my DV is departure from college and my 2 variables are Strat= college selectivity (16=low; 18=high) and ASE_YE which is a categorical measure of perception of academic self efficacy(1=low; 3=high).  When I ran my model with other covariates my interaction was significant so I asked SPSS to compare groups.  My output is attached.Nominal+Regression.doc

questions:
1) There is no estimate for the final category (18 selectivity and high ASE).  Is this the reference against which the rest of the output should be assessed?  EX: A student with low efficacy perceptions at a highly selective college is 1.23 times more likely to depart than a student with high efficacy perceptions at the same caliber institution.

OR
Are these parameter estimates simply the estimates that correspond to that particular group (i.e. low efficacy in a low selectivity college)?
I'm thinking this is not the case, but if so then what is the estimate for a student in a highly selective college with high efficacy? the intercept?

2) I'm trying to generate odds ratios and probabilities for prototypical cases so I can plot them on a graph.  I know I can generate them from my regression equation but I can't do this b/c when I run my logistic regression without denoting (clicking categorical button in upper right) which of my variables are categorical my constant is outrageous (17.452) because it's picking up the excess varianceand I end up with crazy probabilities like .000000004. However when I run my regressions with my categorical variables denoted then there's always a comparison group and no coefficient for the variable on a whole.
a) is there any way to fix this?
b) if not then how else can I obtain probabilities for prototypical cases so I can create a graph?

Thanks so much in advance!!
~Diane

Hi Diane.  Here is your syntax:

NOMREG fall_plans_QQ (BASE=LAST ORDER=ASCENDING)
   BY STRAT_Q NEWAGE_Q PARENT_ED_FINAL_Q ASPIRATIONS_Q CAT_ASE_YE
  WITH SEX03_Q RACE_BLACK_Q RACE_ASIAN_Q RACE_HISPANIC_Q RACE_OTHER_Q
          RACE_WHITE_Q ZASE ZSSE ZAI_FINAL_SCALED ZSI_FINAL_SCALED9 ZYEAR_END_SSE
  /CRITERIA CIN(95) DELTA(0) MXITER(100) MXSTEP(5) CHKSEP(20) LCONVERGE(0)  
   PCONVERGE(0.000001)  SINGULAR(0.00000001)
  /MODEL=STRAT_Q*CAT_ASE_YE
  /STEPWISE=PIN(.05) POUT(0.1) MINEFFECT(0) RULE(SINGLE) ENTRYMETHOD(LR) REMOVALMETHOD(LR)
  /INTERCEPT=INCLUDE
  /PRINT=PARAMETER SUMMARY LRT CPS STEP MFI.

You've listed all of the covariates you refer to following BY and WITH in the first few lines; but then on the /MODEL sub-command, you say that you only want the STRAT_Q*CAT_ASE_YE interaction in the model.  And the table of coefficients confirms that this interaction is the only term in the model.  

You've also told SPSS to perform STEPWISE selection, for some reason.  

So, the first thing to do is to get rid of STEPWISE.  Second, make sure that you've actually got the variables you want in your model.  If you want them in the model, they need to be on the /DESIGN line, and they will show up in the table of coefficients.

Finally, given that your DV appears to be binary, is there any particular reason for using NOMREG rather than LOGISTIC REGRESSION?