Nuanced regression question

[Ollie123]

Advertisement - Members don't see this ad

Getting some odd results that means either I misunderstood something fundamental early on in my stats training (and am still misreading it in the book I have in front of me), or SPSS is doing something weird and/or buggy in its background calculations.

My understanding of multiple regression has always been that if you are, say, regressing predictors X and Y on criterion Z, the results for X should be identical to what you would get if you regressed Y alone on criterion Z, saved the residuals, and then regressed X on those residuals.

From running these analyses, they seem similar, but not at all equivalent. Differences are too great to attribute it to rounding. I've checked pretty much every assumption I can think of and nothing weird is going on with the dataset. Any ideas?

 

[defaultmode]

Getting some odd results that means either I misunderstood something fundamental early on in my stats training (and am still misreading it in the book I have in front of me), or SPSS is doing something weird and/or buggy in its background calculations.

My understanding of multiple regression has always been that if you are, say, regressing predictors X and Y on criterion Z, the results for X should be identical to what you would get if you regressed Y alone on criterion Z, saved the residuals, and then regressed X on those residuals.

From running these analyses, they seem similar, but not at all equivalent. Differences are too great to attribute it to rounding. I've checked pretty much every assumption I can think of and nothing weird is going on with the dataset. Any ideas?

long time lurker here (currently in the midst of the admissions process), but I figure I'd give this a shot. While I'm not sure how SPSS works specifically (R user, myself), I believe in general the results you expect would be true if X were treated as orthogonal to Y. Otherwise, some programs will attribute any variance in Z explained by both X and Y to the error term, rather than to either beta. If SPSS has a means of orthoganalizing X with respect to Y you can test this out. i'm sure someone with more knowledge of SPSS and statistics can do a better job here, but hopefully this helps somewhat.