t*****2 发帖数: 94 | 1 电面一个MANAGER:
1) 每个predictor 和response 都是positive correlated, 为什么放在一MODEL里就成
了negative?
是什么原因呀? 怎么解决呀?
谢谢各位了 | S*x 发帖数: 705 | 2 Collinearity
Google that on how to solve it
【在 t*****2 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 电面一个MANAGER: : 1) 每个predictor 和response 都是positive correlated, 为什么放在一MODEL里就成 : 了negative? : 是什么原因呀? 怎么解决呀? : 谢谢各位了
| s****n 发帖数: 119 | 3 假设predictor为x1和x2,response为y。x1和x2之间为negative。x1增加的时候,y增
加;但是x1的增加导致x2的减少;x2减少,y减少。所以x1的变化通过两个途径来影响y
,直接的和间接的。当间接的影响比较大的时候,整体为negative。 | l*********3 发帖数: 22 | | s******h 发帖数: 539 | 5 If anyone is interested, you can show that this won't happen in a linear
model.
【在 t*****2 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 电面一个MANAGER: : 1) 每个predictor 和response 都是positive correlated, 为什么放在一MODEL里就成 : 了negative? : 是什么原因呀? 怎么解决呀? : 谢谢各位了
| k*******a 发帖数: 772 | 6 It is understandable if one becomes negative, but I do not understand if
both become negatively associated.
If only one of them becomes negatively associated, then it is just a
confounding effect. Assume Y = -X1 + 10*X2, and X1 and X2 are positively
correlated, then higher X1 associated with higher X2, which makes Y higher,
so Y and X1 seem to be positively associated. But if both X1 and X2 are put
into model, then X1 becomes negatively associated with Y | a***d 发帖数: 336 | 7 seems to be colinearity. But do both become negative? or just one of them?
May try LASSO, or combine the two variables into a single variable.
【在 t*****2 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 电面一个MANAGER: : 1) 每个predictor 和response 都是positive correlated, 为什么放在一MODEL里就成 : 了negative? : 是什么原因呀? 怎么解决呀? : 谢谢各位了
| p***r 发帖数: 920 | | m*******1 发帖数: 855 | 9 multicollinearity
quasi complete 会出现这个情况吗?
还有confounding 之类的?
请高人出来讲讲 | m*******1 发帖数: 855 | 10 如果是multicollinearity,解决方法就很多了,10个面试8个会问这个.
可以exclude redundant variables
可以redefind variables
可以用Biased regression, 其中包括PCA和ridge regression | s******h 发帖数: 539 | 11 I agree with you. In fact, you can show that in linear model, it can't be
that both variables are negatively associated when put them in the model
while assuming positively associated individually.
,
put
【在 k*******a 的大作中提到】![](/moin_static193/solenoid/img/up.png) : It is understandable if one becomes negative, but I do not understand if : both become negatively associated. : If only one of them becomes negatively associated, then it is just a : confounding effect. Assume Y = -X1 + 10*X2, and X1 and X2 are positively : correlated, then higher X1 associated with higher X2, which makes Y higher, : so Y and X1 seem to be positively associated. But if both X1 and X2 are put : into model, then X1 becomes negatively associated with Y
| p***d 发帖数: 257 | | l******h 发帖数: 855 | |
|