I have confronted with a problem in the analysis of categorical variables having more than two response categories during binary logistics regression in SPSS. I have 16 independent variables and a binary dependent variable. For instance,
For instance, some of the independent variables are as:
Distance (kilometers)
This error might occur because of insufficient data and too many independent predictors. The category to be selected as reference is based on your question and aims of your model. I attached some links.
http://www.ats.ucla.edu/stat/spss/output/logistic.htm
http://www.ats.ucla.edu/stat/spss/webbooks/reg/chapter3/spssreg3.htm
This error might occur because of insufficient data and too many independent predictors. The category to be selected as reference is based on your question and aims of your model. I attached some links.
http://www.ats.ucla.edu/stat/spss/output/logistic.htm
http://www.ats.ucla.edu/stat/spss/webbooks/reg/chapter3/spssreg3.htm
I agree with Mehmet; you probably don't have enough data to fit the number of parameters you're trying.
The low confidence intervals aren't actually an error, it looks like they're a necessary consequence of the other problem in your model. Notice that the standard errors are huge (implausibly huge, in fact). That also explains why the z-statistics approach 0 (the Wald z statistic is just B/SE) and why the p-values approach 1. Most of the betas are also implausibly high; for example, the first is 128, which is so high that it's indistinguishable from 100% probability. So the weird CIs you see are not an isolated error, but a consequence of a more basic error that has affected everything else about the model as well.
I'm not sure why you are getting 0's for confidence limits in your output, and seeing the full command being executed might help that. Also, seeing your dependent variable and how it's coded could assist, as well. Your statement that you "...took the first category as reference category ..." isn't altogether clear either and may be the source of your problem. Logistic regression needs to see only "0's or 1's" in the dependent variable being regressed, so unless you have your dependent variable coded like that you will get this kind of error. If the dependent variables are coded like those you present in your question, they may need to be recoded into binary variables with one value coded as "0" and all the others coded as "1". This means creating a new dependent variable, using the recode command or create variable command and applying regrouping logic for combining the non-reference values as the second binary coded value. Best of luck.
The dependent variable is "high use and low use" and coded as " 1 and 0". My main problem is how to get odds for the predictors which have more than two response categories (categorical variables with many levels) and which code may be given to the reference category of a predictor - whether 0 or 1 ? Thanks
OK. Perhaps rescaling the independent categorical variables could help... Instead of having the independent variables spread across multiple categorical assignments, I would explore regrouping them into one single independent variable, scaled as either real numbers, with each category coded with the value associated with the category value, or midpoint if they represent a range of values; or scaled as an ordinal variable , low to high, with coded values 1 - 5 if there are five separate values. Recode one of the independent variables across all the separate categories of that variable, into one variable with either real values, or ordinal values and rerun your regression with just that independent variable one of your dependent variables. If that works, recode the rest of your independent variables in similar fashion to how you did it for the first. Best of luck...
Hi Sanaullah,
As far as I know, even for your independent variables (or predictors), you should use dummy coding. I think you may solve your problem if you create new variables using this approach. The number of variables that you would have to create can be calculated by k-1, where k is the number of levels within your categorical variable.
Best
Thank you, Dear Rodrigo, Py for adding your answer to the question.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS