I have a data set with 5 different kinds of nutrient statuses and I want to see whether they are associated with categorical / ordinal grades at school. I have multiple covariates which I will included in the analyses. Due to missing values I have used the multiple imputation strategy (5 times).

So now I have this databases and I want to do a optimal scaling regression (CATREG). However when I do this SPSS says:

The following variables have values less than or equal to zero, which are considered as missing in this procedure: Smoking, etc.

Why is this and what can I do about it?

  • 5
    $\begingroup$ Optimal scaling procedures require categorical data coded as positive integers. There is Discretize button in the procedure, to help you recode continuous variables automatically in a useful way, or you can recode manually. $\endgroup$
    – ttnphns
    May 21, 2014 at 15:34
  • $\begingroup$ Positive or non-negative? $\endgroup$ Jul 29, 2015 at 16:14

1 Answer 1


1) Why is this? This is simply a peculiarity of this analysis. No substantive reason for it.

2) What can you do about it?

First, what are the normal ranges of the (non missing) values on these variables? Do these include values of zero and below? If yes, then recode them.

If no, then inspect the imputed data files and see how many values it concerns. If you imputed lots of missing values, chances are that by mere chance a handful of them end up at/below zero. In this case you could consider recoding the non imputed data files to values well above zero and then rerun the multiple imputation procedure. Imputed values at/below zero should now be much less likely.

  • $\begingroup$ Some variables are coded as either being there are not there (0 and 1) so if I would recode those to say 10 for no and 20 for yes that would solve the problem? I am also not quite sure what to do with this dicretize button, could anyone explain? $\endgroup$
    – user45954
    May 21, 2014 at 19:50
  • $\begingroup$ Well, given that CATREG considers values at/below zero as missing, you can't use codes 0-1 for dichotomies, right? Note that this has nothing to do with the imputation process, the problem is already there in your raw, unedited data. Forget about the discretize button. Even if you want to discretize anything, you don't want SPSS to decide for you where to draw the category boundaries. $\endgroup$
    – RubenGeert
    May 22, 2014 at 4:49
  • $\begingroup$ Ok great :) I will tranform these variables. However if I do not use the discretize button, SPSS will use the default setting... is that ok then? $\endgroup$
    – Inge
    May 22, 2014 at 6:35
  • $\begingroup$ And another question is it better to use ordinal or spline ordinal? $\endgroup$
    – Inge
    May 22, 2014 at 7:04
  • $\begingroup$ 1) What do you want to discretize and why? In this discussion you haven't explicitly mentioned any continuous variables so far. 2) I'm not sure but I believe spline ordinal is a more restricted model than just ordinal. I don't know about the exact details. P.s. if you find answers/comments helpful you can upvote them. $\endgroup$
    – RubenGeert
    May 22, 2014 at 8:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.