w*********8 发帖数: 70 | 1 假设有组binary data,比如说不同年龄人群里面的男女分布。现在我们把性别0,1分
布转化:在25岁的人群里面,n=100, 75female, 那75/100=0.75是我们的第一个值;在
26岁人群里,n=200, 100 female,那100/200是第二个值 and so on....这样的话我们
得到的这组数据就是continuous了。问:这时候可以用linear regression来处理这组
数据吗?首先,做过normality test了,satisfied. 其他几个assumption也满足。
多谢! |
k*******a 发帖数: 772 | 2 不过你怎么解释你的model?
linear regression的y取值范围不可能限制与(0,1) |
a****r 发帖数: 1486 | 3 这跟logistics regression有啥太大区别吗?
【在 w*********8 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 假设有组binary data,比如说不同年龄人群里面的男女分布。现在我们把性别0,1分 : 布转化:在25岁的人群里面,n=100, 75female, 那75/100=0.75是我们的第一个值;在 : 26岁人群里,n=200, 100 female,那100/200是第二个值 and so on....这样的话我们 : 得到的这组数据就是continuous了。问:这时候可以用linear regression来处理这组 : 数据吗?首先,做过normality test了,satisfied. 其他几个assumption也满足。 : 多谢!
|
w*********8 发帖数: 70 | 4 这个应该不是问题。Y取值范围可以限定在(0,1),就在(0,1)之间follow一个
normall distribution啊。
【在 k*******a 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 不过你怎么解释你的model? : linear regression的y取值范围不可能限制与(0,1)
|
M****e 发帖数: 178 | 5 I think this is kind of a typical case for logistic regression with binomial
responses. If you reduce the data to percentage, in some sense you are
losing information. For example, in one population, n=200, n(female)=100,
you have percentage of 0.5; in another population, n=4, n(female)=2, you
still get percentage of 0.5. But apparently you have more certainty that the
first population has a "real" female percentage of 0.5. The data of these
two populations will add different "weights" to logistic regression model.
There seem also other reasons avoiding using percentage in statistical
analysis... (I forget)
BTW, how did you do normality test? For linear regression, response
variables do not need to be normally distributed (residuals after fitting
the model do) |
A*******s 发帖数: 3942 | 6 i am curious what is the reason for difficulty of modeling proportions. is
it because hard to assume its distribution?
binomial
the
these
【在 M****e 的大作中提到】![](/moin_static193/solenoid/img/up.png) : I think this is kind of a typical case for logistic regression with binomial : responses. If you reduce the data to percentage, in some sense you are : losing information. For example, in one population, n=200, n(female)=100, : you have percentage of 0.5; in another population, n=4, n(female)=2, you : still get percentage of 0.5. But apparently you have more certainty that the : first population has a "real" female percentage of 0.5. The data of these : two populations will add different "weights" to logistic regression model. : There seem also other reasons avoiding using percentage in statistical : analysis... (I forget) : BTW, how did you do normality test? For linear regression, response
|
s*r 发帖数: 2757 | 7 你本来有300个observations
现在只有两个了
【在 w*********8 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 假设有组binary data,比如说不同年龄人群里面的男女分布。现在我们把性别0,1分 : 布转化:在25岁的人群里面,n=100, 75female, 那75/100=0.75是我们的第一个值;在 : 26岁人群里,n=200, 100 female,那100/200是第二个值 and so on....这样的话我们 : 得到的这组数据就是continuous了。问:这时候可以用linear regression来处理这组 : 数据吗?首先,做过normality test了,satisfied. 其他几个assumption也满足。 : 多谢!
|
w*********8 发帖数: 70 | 8 多谢!
至于normlity test. 我的理解是,normality assumption是指 response variable
follows normal distribution given each sub-population. 就是说
Y|X~N. 很多test for normality啊,比如S-W test.
binomial
the
these
【在 M****e 的大作中提到】![](/moin_static193/solenoid/img/up.png) : I think this is kind of a typical case for logistic regression with binomial : responses. If you reduce the data to percentage, in some sense you are : losing information. For example, in one population, n=200, n(female)=100, : you have percentage of 0.5; in another population, n=4, n(female)=2, you : still get percentage of 0.5. But apparently you have more certainty that the : first population has a "real" female percentage of 0.5. The data of these : two populations will add different "weights" to logistic regression model. : There seem also other reasons avoiding using percentage in statistical : analysis... (I forget) : BTW, how did you do normality test? For linear regression, response
|
w*********8 发帖数: 70 | 9 所以power下降?
【在 s*r 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 你本来有300个observations : 现在只有两个了
|
D*D 发帖数: 236 | 10 actual proportion is between 0 and 1 but linear regression can give
predictions beyond [0,1]
This is how it was done before logistic regression came into use.
【在 A*******s 的大作中提到】![](/moin_static193/solenoid/img/up.png) : i am curious what is the reason for difficulty of modeling proportions. is : it because hard to assume its distribution? : : binomial : the : these
|
|
|
w*********8 发帖数: 70 | 11 如果我们假设足够多的年龄段那?
恩,不过那还是要少了很多obs.
【在 s*r 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 你本来有300个observations : 现在只有两个了
|
s*r 发帖数: 2757 | 12 应该没啥变化,本来binomial就是多个被努力的和
不过covariate也只能用aggregated value了
【在 w*********8 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 所以power下降?
|
A*******s 发帖数: 3942 | 13 oh, i thought he referred to the difference between binary response {0, 1} and continuous response bounded in (0, 1), not the difference between logit and identity link functions.
【在 D*D 的大作中提到】![](/moin_static193/solenoid/img/up.png) : actual proportion is between 0 and 1 but linear regression can give : predictions beyond [0,1] : This is how it was done before logistic regression came into use.
|
D*********2 发帖数: 535 | 14
re.
likelihood一样的
【在 s*r 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 应该没啥变化,本来binomial就是多个被努力的和 : 不过covariate也只能用aggregated value了
|
w*********8 发帖数: 70 | 15 问了老师,老师说,if you have ungrouped data, you cannot do that. if you
have grouped data, you can do it through percentages. |
p***r 发帖数: 920 | 16 I think it must mean both are okay when you programming, but the sample
size is not negligible.
【在 w*********8 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 问了老师,老师说,if you have ungrouped data, you cannot do that. if you : have grouped data, you can do it through percentages.
|