Hello,
I have several observations for every participant and want to make a logistic regression to ascertain the influence of continuous variables (behaviors) on categorical ones (3 types of performance). Is there a way to do this controling for the different subjects in SPSS? tia
Hello Rita,
Your question is not entirely clear in what you are attempting to do. It sounds like you have 3 different categorical variables that would you like to predict from other continuous variables. If the categorical variables are binary/dichotomous (e.g., yes/no), then logistic regression would likely work. In this case, it would simply be setting the binary variable as the dependent variable and the continuous variable as the independent variable. If your situation is more complicated (e.g., you have the same binary variable measured repeatedly for an individual), then you would have to switch to a more complicated analysis such as hierarchical linear modeling (aka multilevel modelling).
If you clarify your design a bit more, I'm sure more folks would be able to help.
#Objection!
What are you measuring? Is there any order in your outcome or they are purely categorial? Do you want to measure the change in each person? what is your hypothesis based on previous knowledge?
That's very important to receive help. The logistic regression MAKES no sense IF you don't have a binary (yes/no) outcome and the study design is NOT a case control or prevalence study. You can try an ORDINAL logistic regression, but it is very difficult to interpret.
I personally don't get the point of logistic regression in other study designs.
Hello Rita,
Your question is not entirely clear in what you are attempting to do. It sounds like you have 3 different categorical variables that would you like to predict from other continuous variables. If the categorical variables are binary/dichotomous (e.g., yes/no), then logistic regression would likely work. In this case, it would simply be setting the binary variable as the dependent variable and the continuous variable as the independent variable. If your situation is more complicated (e.g., you have the same binary variable measured repeatedly for an individual), then you would have to switch to a more complicated analysis such as hierarchical linear modeling (aka multilevel modelling).
If you clarify your design a bit more, I'm sure more folks would be able to help.
If you mean your categorical variable have three states (e.g. green, yellow and red for safe, caution and danger), then, as advised by Hedley, logistic regression is not suitable.
However, if you mean you have three different categorical variables, as advised by Andrew, logistic regression would likely work.
Rita, to measure continuous behaviors and their association with characteristics, I suggest you consider multinominal logistic regression. it is similar to ordinal in that the outcome has more than 2 categories, but it does not assume the characteristics are equally spaced in degree. If you were looking at over-eating as a characteristic and you considered anxious, stressed, jolly as your characteristics, then they are not graded equally. Another words, you cannot quantify the distance along a scale between stress and anxiety. With that said, yes, you can add many adjustment variables such as socio-demographics - age, gender, income, education, marital status as well as biomarkers such as body mass index, chronic diseases, blood pressure...etc. The problem is interpretation. I think you should run simple cross-tabulations concerning having the characteristic (yes, no) and having the behavior (yes, no). It may be more informative if you consider each level of behavior (eating disorder - on a scale or severity) with whether or not they have a characteristic - (anxiety (yes, no). This could be conducted with logistic regression and several adjustment variables could be added with respect for the sample size and distribution. Make sure you have enough participants in all categories to achieve power. You should work with a statistician who is familiar with your field.
As others have pointed out, you may need to consider a multinomial or ordinal logistic regression model.
In response to "controlling for participants" I'm guessing you have a repeated measures design where each person responds 2 or more times. In this case a conventional logistic regression (including ordinal and multinomial variants) is inappropriate. The two main options in SPSS are probably:
- linear mixed models (also called multilevel models)
- GEEs (generalized estimating equations)
Personally I would run the model in R using one of the linear/mixed multilevel packages. However, for SPSS I have found the multilevel commands a bit limited and would probably suggest looking at GEEs. The GEE interface used to be pretty horrible so I would suggest running the model via syntax.
Just to re-iterate Thom's point, this is one of the procedures that many people find simpler in R than SPSS (and there aren't many). Suppose RESP is a 0 or 1 for the response, PART is the participant number, TASK is a categorical variable for which of the three tasks it is, and BEHAV is a continuous behavior, the following code (depending on your design and research question) could be appropriate in R. This assumes each participant has three lines of data, one for each task.
library(lme4)
summary(glmer(RESP~as.factor(TASK)*BEHAV + (1|PART), family=binomial))
For ordinal models you can use lmer-style syntax with the ordinal package. I'm not sure if SPSS can run multilevel ordinal models at all.
Thanks everyone. After reading all your kind directions I believe what is more suitable to me is performing multilevel in R. Again, thank you so much for your time.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS