Indicator Function Parameterization Help!

Question

From what I can tell the Indicator Function Parameterization output is for when you have categorical predictors in your model.  I actually never even knew about this output window until recently, and in my regression analyses have always simply used the p values and coefficients from the main parameter estimate window. I recently conducted a model that contained 1 continuous IV, and 2 binary IVs as well as all combinations of terms for their interactions. When I look at the main parameter estimates vs the indicator function parameterization estimates, the interaction I am interested in goes from significant in the main output to NS in the IFP output.  Which one is technically the best to use? Have I been reporting the wrong results this whole time?

Mark_Bailey · Answer

I hope that you learned that categorical factors may use either parameterization.

How do you think about the effects of the factors? The default Effect parameterization interprets the effects as a difference from the average response. So if Factor = level 1 has an effect of -10, it means that the response is 10 units below average. The hypothesis tests assume a null hypothesis of parameter equal to zero, or no effect or change from the average response. It is the default parameterization because that is how most people think about effects.

The GLM parameterization is a historical standard used by statisticians. It is the default in SAS PROC GLM. So it is included in JMP for comparison sake.

plfazeli · Answer

Thanks for your reply.  Do you have any insight on which would be most appropriate to interpret? Given the details I provided about my variables?

Mark_Bailey · Answer

The linear statistical model may be parameterized in different ways to suit different purposes or interpretations. The differences change the meaning of the hypothesis tests of the estimates. The differences change the null hypothesis, so it is not expected that the significance will be the same for all parameterizations.

Please see Indicator Parameterization Estimates.