3

I am familiar with the dummy variable trap in normal OLS, in which we should include one less dummy variable than the total of categories to avoid the problem of multicollinearity.

However, I was wondering if it is also the case in duration models. Specifically, I am running a Weibull survival model with six dummy variables, exhaustive and mutually exclusive, and there is no sign of multicollinearity (Even if all the coefficients of the dummy variables are negative, which makes it a bit harder to interpret as I'm not sure what the dummies are measuring themselves against).

So the question is: does the dummy variable trap also occur in survival/duration models?

Stephan Kolassa
  • 95,027
  • 13
  • 197
  • 357
Not Fisher
  • 43
  • 3
  • Are you also fitting an intercept? – Scortchi - Reinstate Monica Apr 13 '15 at 15:03
  • Yes, there is also an intercept. Also the AIC is (slightly) higher when I exclude one of the dummy variables, which implies that the model fit is better with all the dummies. – Not Fisher Apr 13 '15 at 15:58
  • How are you determining that you don't have perfect collinearity in the design matrix? – Scortchi - Reinstate Monica Apr 13 '15 at 16:06
  • No variables are excluded because of collinearity. I'm using STATA, so in the case of OLS the program warns when I include all dummy variables. – Not Fisher Apr 13 '15 at 16:12
  • 1
    Very trusting of you! Check the condition number of $X^\mathrm{T}X$ (where $X$ is the design matrix). Or try changing the order in which coefficients are input to the model (I've just experimented with the `survfit` function in R - fitting a Weibull model when the design matrix isn't full rank also gives one of the infinity of possible solutions without warning; it gives different solutions when you vary the order of the predictors). – Scortchi - Reinstate Monica Apr 13 '15 at 16:39
  • You're right, I tested for it and the variables are collinear. Thanks for the help! – Not Fisher Apr 13 '15 at 17:49
  • 1
    Should be the same as in OLS... –  Apr 13 '15 at 12:47
  • @Not Fisher: You're welcome! Think about it: the dummy variable trap only involves the specification of predictors, so is relevant to any linear model. I'll add an answer with more detail if no-one else does, or if those who've already answered don't expand theirs. – Scortchi - Reinstate Monica Apr 13 '15 at 20:59

1 Answers1

0

Dummy Variable trap usually occurs in proportional hazards models/semi-parametric models very well. and yes! you should check significant improvement of each dummy variable you add.

Hemant Rupani
  • 1,193
  • 9
  • 19