Timeline for How to best model interaction effect of two continuous predictor variables?
Current License: CC BY-SA 3.0
7 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Feb 1, 2016 at 22:48 | comment | added | Frank Harrell | That was a teaching example but you are right it involved some cheating. I wanted to show that you could get more evidence for some kind of interaction if you concentrated the interaction effect into fewer parameters. | |
| Feb 1, 2016 at 22:44 | vote | accept | Clark Chong | ||
| Feb 1, 2016 at 22:44 | comment | added | Clark Chong | Thanks! I read the narrative in the book (Chap 10) carefully. I am wondering why the decision was made to take the simpler model containing only $X1 \cdot X2$ as interaction term? Also, since the process is not blind to Y, will we run the risk of multiple comparison when we test out different non-linearity and interaction specification? | |
| Feb 1, 2016 at 12:31 | comment | added | Frank Harrell | See expanded answer | |
| Feb 1, 2016 at 12:30 | history | edited | Frank Harrell | CC BY-SA 3.0 |
Expanded answer to include answers to OP's follow-up quesitons
|
| Feb 1, 2016 at 5:09 | comment | added | Clark Chong | Thank you Dr Harrell for the reply! I found some description of the tensor spline in 2.7.2 of your 2015 book and 2.9.2 of your lecture note. Is there a worked example of the same in either of the sources above? Y = 1 is about 1.5% of the total sample (2 million observations), I suspect that frequency should be sufficiently large (is rarity of the event a problem?)? Lastly, is your last sentence meant to say that the percentile-cut is arbitrary because it is not individual characteristic? Thanks! | |
| Jan 31, 2016 at 13:21 | history | answered | Frank Harrell | CC BY-SA 3.0 |