1

I am modelling logistic regressions in SPSS, the same model for different countries (well, with slight differences in the independent variables set due to collinearity diagnosis and stepwise results). The model seems to work fine for most countries. In two countries, I am having some issues with the same variable. The variable has three categories (hierarchical). In the output for those two beautiful countries, seems that SPSS take the reference category out, assumes other as reference category and gives exp(B) for the last category.

Is there something wrong going on? What should / can I do about it?

Thanks a lot for your answers! The problem cannot be in the syntax since I am using the same for all countries and works perfectly.

I am using ENTER. The STEPWISE was used as an exploratory method to identify the strongest predictors. According to the -2LL values my variable is not a 'strong' one for this two countries (0.12%). Is not the first time that I read that STEPWISE is a bad choice... I will that in account!

Nevertheless, why do SPSS does that? Is this a real problem, meaning something that I must solve? Or can I use/report the odds for those two countries without that category adding a footnote? Should I run the model without that variable (theorectically the variable is not that important).

Thanks ;)

Hi, All my variables are recoded to the reference category be 1 and most of my variables have three categories. I am using Enter with the following syntax. LOGISTIC REGRESSION VAR= "DV" /METHOD=ENTER "IV's" /CONTRAST ("IV")=Indicator (1) [...] /PRINT=GOODFIT CI(95) /CRITERIA PIN(.05) POUT(.10) ITERATE(20) CUT(.5). Do you have idea what is going wrong?

Joana
  • 11
  • 3
  • This appears technical question, you must be doing something wrong via syntax or menu. I think the question should be moved to StackOverflow. – ttnphns Oct 18 '13 at 19:21
  • 1
    My first suspicion is that this is not a technical problem with SPSS, but a less frequent result of stepwise selection algorithms (see my answer below). I believe this question is on-topic on CV. – gung - Reinstate Monica Oct 18 '13 at 20:34
  • I agree with @gung. This question should stay open. – Peter Flom Oct 18 '13 at 20:47
  • Yes, I am using enter. The reference category is droped the 2nd level becomes the reference category. – Joana Oct 18 '13 at 23:21
  • 1
    I need to know more... did you use the `categorical` button to assign a variable to be categorical or did you make a series of dichotomous indicator by yourself and feed them into the model? If it's the earlier, pay attention after assigning the categorical variable, you can actually set your reference group to either the first or last of the coding scheme (if your desirable ref. group is at the middle then you'll need to recode). If it's the latter, then you should not feed a group of dichotomous indicators into stepwise... bad idea. use `block` button to test the whole categorical variable. – Penguin_Knight Oct 19 '13 at 00:39
  • Hi, All my variables were recode to the reference category be 1 and most of my variables have three categories. I am using Enter with the following syntax. LOGISTIC REGRESSION VAR= "DV" /METHOD=ENTER "IV's" /CONTRAST ("IV")=Indicator (1) [...] /PRINT=GOODFIT CI(95) /CRITERIA PIN(.05) POUT(.10) ITERATE(20) CUT(.5). Do you have idea what is going wrong? – Joana Oct 19 '13 at 10:05
  • The syntax looks fine. One possible reason is that the suspect variable has a very low count in level 1 to begin with. When other predictors were also added, the extra information might have caused the level 1 of the suspect variable to be perfectly predicted (aka perfect collinearity.) – Penguin_Knight Oct 19 '13 at 11:59
  • True, "very low count in level 1". In this analysis I had quasi-complete separation with other variables (very large S.E.). I recoded them and add this "suspect variable" as a proxy of one that I excluded. Can I use the odd ratios as they are (without the same reference category that I have in other countries?) and report that in a footnote? If not, Is it too sloppy to drop the variable just for the two countries here I am having problems?Thanks for your help! – Joana Oct 19 '13 at 15:30

2 Answers2

1

I suspect your problem is you are using stepwise model selection. Automatic model selection algorithms, such as stepwise selection, don't have any 'knowledge' about what the variables mean or how they ought to relate to each other. For example, it is standard advice that if you include a squared term (e.g., $X_j^2$) in a model, you should include the lower level term ($X_j$) as well, even if it's not 'significant'. However, the selection algorithm doesn't know there is any relationship between the two variables, and so you commonly have situations where it will throw out the lower level term but retain the squared term. Likewise, all levels of your categorical variable should be retained or excluded together, but the stepwise algorithm doesn't necessarily 'know' that. I believe this is your problem.

Of course, even if you were to address this issue, stepwise selection algorithms are a truly horrible modeling strategy, and still should not be used. For more on that topic, it may help you to read my answer here: algorithms-for-automatic-model-selection.

As the above suggests, the way to rectify this problem is to stop using stepwise selection. Aside from the problem described in the post, it does not do what people believe it does (i.e., tell them which variables are the 'real' or 'important' ones).

Based on your edit, I'm guessing there may be a different issue. It is possible that you don't have any observations (cases) with those levels of the categorical variable for those countries. You should check to see if that is the case. If so, see if one of the other levels of the categorical variable has instances for every country, and use that one as your reference level instead.

Community
  • 1
gung - Reinstate Monica
  • 132,789
  • 81
  • 357
  • 650
  • That was my first guess, but is not the case… The crosstabs look good, I have missing cases but I have 17% of the cases in the level that is dropped. I also tried to change the reference category, the result is the same. I tried to change the measure (ordinal, nominal, scale), the same result... – Joana Oct 18 '13 at 23:41
0

SPSS Statistics provides ridge, lasso, and elastic net in the CATREG command, but if you need to enter blocks of dummies or other groups together, that can be done in REGRESSION by using blocks of terms via multiple METHOD subcommands.

JKP
  • 1,349
  • 10
  • 7