w******y
发帖数: 45 | 1 一个很初级的问题。multicollinearity的后果是不是主要就在于本来可能是
significant的variable变成insignificant了？
如果V1和V2是highly correlated, say r=.95，但是它们的regression coefficients
都是significant，那是不是说multicollinearity没造成什么影响呢？ | a**n
发帖数: 3801 | 2 the coef estimates cannot be taken seriously due to
big st dev.
coefficients
【在 w******y 的大作中提到】
: 一个很初级的问题。multicollinearity的后果是不是主要就在于本来可能是
: significant的variable变成insignificant了？
: 如果V1和V2是highly correlated, say r=.95，但是它们的regression coefficients
: 都是significant，那是不是说multicollinearity没造成什么影响呢？
| t*****i
发帖数: 68 | 3 My understanding is that you can say they are significant, but you can barel
y say anything about their magnitude
coefficients
【在 w******y 的大作中提到】
: 一个很初级的问题。multicollinearity的后果是不是主要就在于本来可能是
: significant的variable变成insignificant了？
: 如果V1和V2是highly correlated, say r=.95，但是它们的regression coefficients
: 都是significant，那是不是说multicollinearity没造成什么影响呢？
| f*******r
发帖数: 257 | 4 When you have significant results, even though you have highly correlated
regressors, most likely you have a large data set. Although you have
significant results, the question you should ask yourself is: given x1 and
x2 have a correlation of 0.95, do they really both need to be in the model?
If your theory says yes, then it's fine. But if x1 and x2 have a
correlation of .95, they are very likely to be measuring the same construct. | w******y
发帖数: 45 | 5 If all I care is the sign of the coefficient, positive or negative, then the
magnitude doesn't matter, given it's significant, right?
barel
【在 t*****i 的大作中提到】
: My understanding is that you can say they are significant, but you can barel
: y say anything about their magnitude
:
: coefficients
| w******y
发帖数: 45 | 6 Yes, I am also confused about this. I tried to measure two different
concepts, but the variables are so highly correlated. I prefer not to remove
any of them because they are the two main predictive variables in the model
. Or, can I use two models with x1 in one and x2 in the other, instead of
having both in one model?
Another question is, if x1 and x2 are put separately in two models, their
coefficients are both positive. But if they are put together in one model,
x1 is positive but x2 is nega
【在 f*******r 的大作中提到】
: When you have significant results, even though you have highly correlated
: regressors, most likely you have a large data set. Although you have
: significant results, the question you should ask yourself is: given x1 and
: x2 have a correlation of 0.95, do they really both need to be in the model?
: If your theory says yes, then it's fine. But if x1 and x2 have a
: correlation of .95, they are very likely to be measuring the same construct.
| a**n
发帖数: 3801 | 7 that's expected
you can think x1=a, x2=a+b
the impact of a is (b1+b2) and the impact of b is b2
your b2 of x2 actually measures the additional impact of the b factor.
remove
model
【在 w******y 的大作中提到】
: Yes, I am also confused about this. I tried to measure two different
: concepts, but the variables are so highly correlated. I prefer not to remove
: any of them because they are the two main predictive variables in the model
: . Or, can I use two models with x1 in one and x2 in the other, instead of
: having both in one model?
: Another question is, if x1 and x2 are put separately in two models, their
: coefficients are both positive. But if they are put together in one model,
: x1 is positive but x2 is nega
| w******y
发帖数: 45 | 8 Thanks! Then I have a silly question. When people hypothesize x2 has a
certain impact on DV, does it refer to the overall effect (a+b) or the
additional impact (b)?
【在 a**n 的大作中提到】
: that's expected
: you can think x1=a, x2=a+b
: the impact of a is (b1+b2) and the impact of b is b2
: your b2 of x2 actually measures the additional impact of the b factor.
:
: remove
: model
| s*****w
发帖数: 2065 | 9 If you use x1 as a control variable, then should be the additional impact.
depends on your economic interpretation.
【在 w******y 的大作中提到】
: Thanks! Then I have a silly question. When people hypothesize x2 has a
: certain impact on DV, does it refer to the overall effect (a+b) or the
: additional impact (b)?
| w******y
发帖数: 45 | 10 好像我对这方面的知识太缺乏了，再接着请教一下。
打个比方，如果我的研究是想证明身高体重和寿命的关系。我用了一堆理论来假设个子
高的人寿命长，又用了一堆理论来假设胖的人寿命长。分别regress身高和体重on寿命
，两个变量都是significantly positive。但是因为这两个变量highly correlated，
把它们同时regress寿命的话，身高是significantly positive,但是体重是
significantly negative。
我的问题是：
1.关于体重和寿命的假设被confirmed没有？when the model only contains weight,
its coefficient is positive. but if the model contains both height and
weight, the coefficient of weight is negative. which one shall i use to
discuss the hypothesis?
2. 这个结果可信吗？Can i trust the
【在 s*****w 的大作中提到】
: If you use x1 as a control variable, then should be the additional impact.
: depends on your economic interpretation.
| t*****i
发帖数: 68 | 11 In this case, you may want to consider the ridge regression method.
,
【在 w******y 的大作中提到】
: 好像我对这方面的知识太缺乏了，再接着请教一下。
: 打个比方，如果我的研究是想证明身高体重和寿命的关系。我用了一堆理论来假设个子
: 高的人寿命长，又用了一堆理论来假设胖的人寿命长。分别regress身高和体重on寿命
: ，两个变量都是significantly positive。但是因为这两个变量highly correlated，
: 把它们同时regress寿命的话，身高是significantly positive,但是体重是
: significantly negative。
: 我的问题是：
: 1.关于体重和寿命的假设被confirmed没有？when the model only contains weight,
: its coefficient is positive. but if the model contains both height and
: weight, the coefficient of weight is negative. which one shall i use to
| f*******r
发帖数: 257 | 12 As long as you believe they both should be in the model, the results are
still valid, even though they are highly correlated. When you have large
data set, the unstableness brought by multicollinearity will be alleviated
by large sample. That is, multicollinearity brings impreciseness; but large
sample overcomes it. Every thing is the same in interpretation. b1 is the
effect of x1 controlling x2. Weight has a negative effect on longevity
after controlling height.
What I am uncomfortable abo | w******y
发帖数: 45 | 13 I am dealing with panel data using the tobit model and don't know how to use
ridge regression in this context. :(
【在 t*****i 的大作中提到】
: In this case, you may want to consider the ridge regression method.
:
: ,
| w******y
发帖数: 45 | 14 Thank you! This helps a lot. Indeed I am using a large data set with around
1500 observations.Even after I randomly divided the data set into several
samples, the result remains the same. So I guess multicollinearity is not a
big problem for the model per se, right?
This is actually the critique I got from a reviewer for my manuscript, which
says the two main variables are highly correlated and the result is
problematic. I want to respond that it is not a serious issue given the fact
that the co
【在 f*******r 的大作中提到】
: As long as you believe they both should be in the model, the results are
: still valid, even though they are highly correlated. When you have large
: data set, the unstableness brought by multicollinearity will be alleviated
: by large sample. That is, multicollinearity brings impreciseness; but large
: sample overcomes it. Every thing is the same in interpretation. b1 is the
: effect of x1 controlling x2. Weight has a negative effect on longevity
: after controlling height.
: What I am uncomfortable abo
|
|
