# Does anyone know how to structure multiple percentages per individual for multinomial logistic regression in SPSS?

Per individual I have 6 percentages (summing to 100 per individual) for 6 different behaviors shown by the individual. I want to perform a multinomial logistic regression on this data in SPSS. However, I can only add 1 variable in the ''Dependent" variable/target box in the multinomial logistic regression analysis. In my dataset each row is an individual, and thus each individual has 6 columns of behaviors with the percentage how much that behavior was observed over a certain time.These percentages are the dependent variables. Happy new year!

• You probably need to look for methods for compositional data rather than logistic regression. – mdewey Jan 2 '18 at 13:56
• I checked compositional data and I think this is in the right direction. However, is this possible in SPSS? I cannot seem to find it online – Rob Jan 2 '18 at 19:43
• I've read some more about compositional data. I now irl-transformed my data with CoDaPack software, which is built for compositional data. However, I do not know how to continue my analysis – Rob Jan 2 '18 at 21:00

## 1 Answer

Try to rethink the question you need to answer. Multinomial logistic regression would answer a question such as --

• Does the borrower of a loan end up repaying on time, getting an extension, or defaulting?
• Does a respondent prefer broccoli, anchovies, artichokes, or pineapple as a single pizza topping?
• Does an experimental treatment make the patient turn red, blue, or orange?

In these examples, each case fits into one and only one category along the outcome variable. This is not true of your situation.

• Thank you for your answer. I understand what you mean. Do you perhaps know an alternative? – Rob Jan 2 '18 at 15:51
• An alternative analysis to answer what question? – rolando2 Jan 2 '18 at 16:01
• Sorry, I misread your questions. An alternative analysis to answer the question which percentage (i.e. behavior) has been observed the most. I also have the data in a long form, the observed behavior for each second. So e.g. I have 10 individuals whose behavior was recorded every second for 30 minutes long (thus 18 000 rows). There are 6 different outcomes for the behavior (1-6). I want to test which behavior was shown the most. I tried multinomial logistic regression, but I don't know if the results are correct. For the model fitting information, it gives zero's for almost everything. – Rob Jan 3 '18 at 9:37
• Any thoughts on this? – Rob Jan 3 '18 at 9:42