Please help with an appropriate type of regression to obtain numerical result (SPSS) with categorical data Please help me to figure out the right approach for fairly straightforward task.
Let say, we asked respondents to state 3 most important attributes (over ten) for the one product. Answers were coded in 3 columns.
Then we asked to generally evaluate product on 5-point Likert scale. (Like/Dislike product).
How to learn which attributes really lead to the higher evaluation and in which direction?
Extra question - what if respondents not just state 3 most important factors but rate them. How to incorporate this information (that I can't figure out at all...)
 A: Your independent variables, the attributes, constitute 10 binary variables with values 1 (selected, important) and 0 (not selected, unimportant). So, recode your 'categorical multiple response set' (3 columns with attributes' codes) into that 'dichotomous multiple response set', and go on.
Regarding the choice of regression. If you dare to take the likert dependent variable an interval one use linear regression. If you take it as ordinal one use ordinal regression. You can also use categorical regression (CATREG) which transforms categorical data into interval nonlinearly (so called "optimal scaling") in order to maximize the prediction by a linear model and then performs that linear regression.
Possible answers to your last paragraph depend on whether the rating of an "important" attribute is the importance again or is a different characteristic.
Update for the last: Since in your comment you indicate that the rating is the importance, then it is simple - because the rating then is the same process as selecting to be important/unimportant. Appoint "not selected" attribute value importance=0, the lowest rating point (of, say, rating scale 1-5): you get rating scale 0-5. This action clearly leaves the scale ordinal and not interval - because you don't know whether the diastema between ratings 0 and 1 is the same as between 1 and 2. This naturally calls you to use categorical regression, as the regression which can use ordinal predictors by "optimally" converting them into interval ones. (Please be aware that CATREG does not allow 0 value; so add 1 to all the values to make the lowest code 1.)
Another strategy to incorporate ordinal predictors in regression is to investigate them as giving polynomial effects (e.g., for example, add a squared copy of every predictor to the regression model).
