Box-Tidwell problems in SPSS (Logit Regression)

Hello everyone

I am currently performing a logit regression in SPSS. I am faced with problems when testing the linearity assumption using the Box-Tidwell approach.

I have 10 continuous variables and for each of them I have computed their Natural Logarithm. For testing the assumption, I run a logit regression with all continuous variables multiplied with their natural logarithm as independent variables + their normal form.

My problem is that SPSS cannot run this regression, and gives me the error: "Estimation terminated at iteration 20 because of perfect fit is detected - this solution is not unique"

When I run the regression with less variables, I am able to get an output.

How do I perform a Box-Tidwell regression on my 10 continuous variables in SPSS?

Best regards
Thomas :)


Less is more. Stay pure. Stay poor.
Not an SPSS user, but wondering if your dependent variable can be created by an independent variable or a combination of them?

Does the model run when you just have the 10 variables included, so no NL transformed variables?


Super Moderator

1) What is your sample size? Perfect fit might arise because you have as many predictors as you have cases.

2) What is your motivation for taking the logarithm of your independent variables?

3) Given that you have log-transformed your IVs, why are you also including the untransformed IVs in your equation?

PS. If you cross-post on other forums, please alert us so as to avoid duplicate effort - it seems like you've also posted at


Less is more. Stay pure. Stay poor.
BG, the linear relationship with logit is tested for continuous variable by including the main term times NL version.
My sample size is 415 and actually I have 9 continuous independent variables.

I am able to run the regression with just the variables included and no NL variables.

I have checked the correlations between my dependent variable and all independent, and there is no correlations >0.7 --> I cannot create the dependent variable from the IV.

As HLSmith explains, the Box-Tidwell approach is conducted by included the non-NL continuous IVs together with the

Sorry for cross-posting on other forums.


Super Moderator
Interesting. I hadn't actually come across the Box-Tidwell test. Um, this might be asking the obvious, but you definitely haven't included the DV as one of the predictors, right?


Less is more. Stay pure. Stay poor.
Are you creating the variable and using it as a variable in the modelor trying to create it in the model? In other programs you create the variables in a data step.

Can you get a reduce version of the model to run or just test one at a time?