s*******e
发帖数: 226 | 1 I'm working on a random intercept model. I wonder why the model gives all
significant covariates if I exclude a constant intercept, but the results
become insignificant when I include the constant intercept.
I know in most cases an intercept term should be included in a regression
even it's not significant. But is there a theory about why to include
intercept or not?
Thanks in advance for any input. |
P****D
发帖数: 11146 | 2 This is pretty weird. Did you plot your data? How does the plot look like?
【在 s*******e 的大作中提到】
: I'm working on a random intercept model. I wonder why the model gives all
: significant covariates if I exclude a constant intercept, but the results
: become insignificant when I include the constant intercept.
: I know in most cases an intercept term should be included in a regression
: even it's not significant. But is there a theory about why to include
: intercept or not?
: Thanks in advance for any input.
|
s*r
发帖数: 2757 | 3 i guess you misread the sas output. |
s*******e
发帖数: 226 | 4 Could you please tell more about your thoughts.
The -2RLL and AIC are about the same,but the covariates are only
significant in the model without intercept term.
Thank you.
【在 s*r 的大作中提到】
: i guess you misread the sas output.
|
s*******e
发帖数: 226 | 5 up up
【在 s*******e 的大作中提到】
: Could you please tell more about your thoughts.
: The -2RLL and AIC are about the same,but the covariates are only
: significant in the model without intercept term.
: Thank you.
|
s*r
发帖数: 2757 | 6 paste sas output
【在 s*******e 的大作中提到】
: up up
|
s*******e
发帖数: 226 | 7 Thanks in advance.
The first set ist with intercept. The second set is without intercept.
The Mixed Procedure
Model Information
Data Set WORK.TEMP
Dependent Variable loggev
Covariance Structures Variance Components,
Autoregressive
Subject Effects
【在 s*r 的大作中提到】
: paste sas output
|
p***r
发帖数: 920 | 8 I think it just mean that the covariates is insignificant, and you should
accept the reality |
l*********s
发帖数: 5409 | 9 re
【在 p***r 的大作中提到】
: I think it just mean that the covariates is insignificant, and you should
: accept the reality
|
s*******e
发帖数: 226 | 10 Why we have to include an intercept term in the model, given that in my
model when all covariates are zero the dependent variable should also be
zero?
should
【在 p***r 的大作中提到】
: I think it just mean that the covariates is insignificant, and you should
: accept the reality
|
|
|
s*r
发帖数: 2757 | 11 it is not easy to see
but i have difficulty to understand why you had the intercept both as a
random effect and a fixed effect in the model.
【在 s*******e 的大作中提到】
: Thanks in advance.
: The first set ist with intercept. The second set is without intercept.
: The Mixed Procedure
: Model Information
: Data Set WORK.TEMP
: Dependent Variable loggev
: Covariance Structures Variance Components,
: Autoregressive
: Subject Effects
|
s*******e
发帖数: 226 | 12 In a random intercept model, we add a random intercept as an error
component and a constant intercept. It's like
y_it= a + bx_it + u_i + error_it
【在 s*r 的大作中提到】
: it is not easy to see
: but i have difficulty to understand why you had the intercept both as a
: random effect and a fixed effect in the model.
|
D******n
发帖数: 2836 | 13 ur terminology confuses people here. Intercept is intercept, it can only be
one.
(a+u_i) is ur intercept here, it is random and it is different across differ
ent groups. ( it is like in the anova model yi=mu+a_i+eij, u dont say there
is a constant effect mu and another effect a_i)....
you should list the model statement u wrote in ur sas script, then we might
know what u really did...
【在 s*******e 的大作中提到】
: In a random intercept model, we add a random intercept as an error
: component and a constant intercept. It's like
: y_it= a + bx_it + u_i + error_it
|
s*******e
发帖数: 226 | 14 Yes, my question is whether the constant intercept (a) is necessary to
be included given that a random intercept u_i (often called error
component) is specified.
The results indicates that the key covariates are significant when the
constant intercept is excluded, and become insignificant when the
constant intercept is present. Why it makes such difference.Which model
should I believe?
only be
differ
there
might
【在 D******n 的大作中提到】
: ur terminology confuses people here. Intercept is intercept, it can only be
: one.
: (a+u_i) is ur intercept here, it is random and it is different across differ
: ent groups. ( it is like in the anova model yi=mu+a_i+eij, u dont say there
: is a constant effect mu and another effect a_i)....
: you should list the model statement u wrote in ur sas script, then we might
: know what u really did...
|
D******n
发帖数: 2836 | 15 As i said it really depends how u write your sas code, what u write down is
not necessary what u think it should be.......
again (a+u_i） is the intercept, not a, in this model u_i~N(0,sigma)
u can just write u_i too, in this model u_i~N(a,simga)
duh....
【在 s*******e 的大作中提到】
: Yes, my question is whether the constant intercept (a) is necessary to
: be included given that a random intercept u_i (often called error
: component) is specified.
: The results indicates that the key covariates are significant when the
: constant intercept is excluded, and become insignificant when the
: constant intercept is present. Why it makes such difference.Which model
: should I believe?
:
: only be
: differ
|
p***r
发帖数: 920 | 16 usually intercept is included, i think your problem is like
the true model is
y=b+(a.x); where (a ~= 0) is insignificant, but (b) is significant non
zero, and now you are using another model
y= a'.x; of course this makes your estimated a' become significant non-
zero. But I guess you might find in this case the RSS become bigger
too.
model
【在 s*******e 的大作中提到】
: Yes, my question is whether the constant intercept (a) is necessary to
: be included given that a random intercept u_i (often called error
: component) is specified.
: The results indicates that the key covariates are significant when the
: constant intercept is excluded, and become insignificant when the
: constant intercept is present. Why it makes such difference.Which model
: should I believe?
:
: only be
: differ
|
s*******e
发帖数: 226 | 17 Below is my SAS code for the problem. As you said, u_i would be
estimated as u_i~N(a,simga)if the intercept is specified as u_i (random)
only, then the coefficient of covariates should not change no matter a
is included or not. But it is not the case in my results.
"
proc mixed data=temp method=reml covtest ;
class permno year;
model loggev= lagq2 compiii softyr hardyr compiii*lagq2 softyr*lagq2
hardyr*lagq2 log_market_value diver pcsale_ratio
stockindexrt/ solution noint;
*noint is specified
【在 D******n 的大作中提到】
: As i said it really depends how u write your sas code, what u write down is
: not necessary what u think it should be.......
: again (a+u_i） is the intercept, not a, in this model u_i~N(0,sigma)
: u can just write u_i too, in this model u_i~N(a,simga)
: duh....
|
s*******e
发帖数: 226 | 18 Which model is right then?
BTW, how to calculate pseudo R square and RSS based on the SAS output
given that I use restricted maximum likelihood estimation?
【在 p***r 的大作中提到】
: usually intercept is included, i think your problem is like
: the true model is
: y=b+(a.x); where (a ~= 0) is insignificant, but (b) is significant non
: zero, and now you are using another model
: y= a'.x; of course this makes your estimated a' become significant non-
: zero. But I guess you might find in this case the RSS become bigger
: too.
:
: model
|
p***r
发帖数: 920 | 19 I think your final model should be
y= intercept,
that's it
【在 s*******e 的大作中提到】
: Which model is right then?
: BTW, how to calculate pseudo R square and RSS based on the SAS output
: given that I use restricted maximum likelihood estimation?
|
s*r
发帖数: 2757 | 20 1)after a second thought i think you are right in that a continuous variable
, including intercept, can be both fixed and random in a mixed effect model
2)people below have explained your question about why some effects become
insignificant after including the intercept. let me clarify
the fixed intercept has an estimated value -5 if it is included, in this
case, the estimates for both lagq2 and compiii are larger than 1.5
when the fixed intercept is excluded, estimates for both lagq2 and compii
【在 s*r 的大作中提到】
: it is not easy to see
: but i have difficulty to understand why you had the intercept both as a
: random effect and a fixed effect in the model.
|
|
|
s*******e
发帖数: 226 | 21 I got your point, which helps indeed. But in your example, you assume
y=a+bx is the correct form. By the same token, if we assume y=bx is the
correct model, the results will be different.
If you try the following;
data two;
DO x=0 TO 100 BY 1;
y=UNIFORM (0)*10*x;
OUTPUT ;
END;
run;
title "with intercept";
proc mixed data=two;
model y=x /s;
run;
title "without intercept";
proc mixed data=two;
model y=x /s noint;
run;
You will find the model without intercept has a bit lower BIC and x is
much more
【在 s*r 的大作中提到】
: 1)after a second thought i think you are right in that a continuous variable
: , including intercept, can be both fixed and random in a mixed effect model
: 2)people below have explained your question about why some effects become
: insignificant after including the intercept. let me clarify
: the fixed intercept has an estimated value -5 if it is included, in this
: case, the estimates for both lagq2 and compiii are larger than 1.5
: when the fixed intercept is excluded, estimates for both lagq2 and compii
|
s*******u
发帖数: 2249 | 22 根据你数据做出的模型INTERCEPT不是有统计意义？你模拟的数据INTERCEPT是有意义还是无意义？
为什么INTERCEPT有统计意义，还要去假定一个Y=A＊X是正确的模型？而且你模型的INTERCEPT=-5.7038,绝对值对比其他系数也挺大的，怎么就可以忽略？
【在 s*******e 的大作中提到】
: I got your point, which helps indeed. But in your example, you assume
: y=a+bx is the correct form. By the same token, if we assume y=bx is the
: correct model, the results will be different.
: If you try the following;
: data two;
: DO x=0 TO 100 BY 1;
: y=UNIFORM (0)*10*x;
: OUTPUT ;
: END;
: run;
|
s*r
发帖数: 2757 | 23 i just want to show you how the coefficient on the intercept can be divided
and shared by the coefficient of other factors when you force the intercept
out.
【在 s*******e 的大作中提到】
: I got your point, which helps indeed. But in your example, you assume
: y=a+bx is the correct form. By the same token, if we assume y=bx is the
: correct model, the results will be different.
: If you try the following;
: data two;
: DO x=0 TO 100 BY 1;
: y=UNIFORM (0)*10*x;
: OUTPUT ;
: END;
: run;
|
s*******e
发帖数: 226 | 24 This is a good point. In my model, the dependent variable is of firm
risk, and the covariates are a series of firm characteristics and market
conditions. Strictly speaking, when all those covariates are zero, which
means the firm is out of market or no longer exists, the firm's risk
should also be zero. In this sense, the intercept does not have any
meaningful effect.
I'm not sure about the intercept issue only because it's very common in
finance literature to include an intercept in regression
【在 s*******u 的大作中提到】
: 根据你数据做出的模型INTERCEPT不是有统计意义？你模拟的数据INTERCEPT是有意义还是无意义？
: 为什么INTERCEPT有统计意义，还要去假定一个Y=A＊X是正确的模型？而且你模型的INTERCEPT=-5.7038,绝对值对比其他系数也挺大的，怎么就可以忽略？
|
s*******e
发帖数: 226 | 25 So back to the questions, given the different results between including an
intercept or not, which model is the right one?
divided
intercept
【在 s*r 的大作中提到】
: i just want to show you how the coefficient on the intercept can be divided
: and shared by the coefficient of other factors when you force the intercept
: out.
|
D******n
发帖数: 2836 | 26 i think u still misunderstand the meaning of intercept.
if u think it is zeros, then why would u fit a random intercept model?
【在 s*******e 的大作中提到】
: This is a good point. In my model, the dependent variable is of firm
: risk, and the covariates are a series of firm characteristics and market
: conditions. Strictly speaking, when all those covariates are zero, which
: means the firm is out of market or no longer exists, the firm's risk
: should also be zero. In this sense, the intercept does not have any
: meaningful effect.
: I'm not sure about the intercept issue only because it's very common in
: finance literature to include an intercept in regression
|
s*******e
发帖数: 226 | 27 Allow the intercept of a firm randomly around zero?
【在 D******n 的大作中提到】
: i think u still misunderstand the meaning of intercept.
: if u think it is zeros, then why would u fit a random intercept model?
|
D******n
发帖数: 2836 | 28 then it is not zero.....
and the model is odd...
if u believe intercept is zero just make it zero, u can make other factor ra
ndom, which u already did....thats fine, dont make the intercept random agai
n.
【在 s*******e 的大作中提到】
: Allow the intercept of a firm randomly around zero?
|
s*******e
发帖数: 226 | 29 Doesn't a random intercept account for the heterogeneity across
subjects?
factor ra
agai
【在 D******n 的大作中提到】
: then it is not zero.....
: and the model is odd...
: if u believe intercept is zero just make it zero, u can make other factor ra
: ndom, which u already did....thats fine, dont make the intercept random agai
: n.
|
