When it comes to multiple predictors, I've seen conflicting interpretations of MARS models and hoping for some clarification.
Some fake results, let's say it's predicting household income via years of age and education...
estimate
(intercept) 50000.00
h(age-40) 50.00
h(40-age) -100.00
h(16-edu_yrs) -200.00
How I'd interpret these findings is the following. For Age, as you approach 40 years, your coefficient is 100 per year and, after 40, it's 50 per year. As for Education, as you approach 16 years, your coefficient is 200 per year. After 16, it has no effect.
So, a few examples:
- 25 year old with 14 years education: 50000-100*(40-25)-200*(16-14) = 48100.
- 60 year old with 20 years of education: 50000+50*(60-40) = 51000.
However, I've seen someone in the field much longer than I interpret the results in a different way such that the coefficients for a predictor are in relation to the median value of the other predictor(s). For example, the finding for education listed above occurs at the median value of age.
Which interpretation is correct?
