s*******r 发帖数: 769 | 1 longitudinal的data analysis的最大问题好像是missing data
好像有很多讨论,最常用的是不是last observation carry forward?
用(一段时期的)平均值来代替missing values,怎么样? |
a****m 发帖数: 693 | 2 longitudinal date难道非得是balanced between group? 如果可以unbalance,
missing value 是可以分析的吧 |
P****D 发帖数: 11146 | 3 很少真见谁这么干。
贱妾平时做的都是unbalanced,缺个observation啥的就直接忽略。
【在 s*******r 的大作中提到】 : longitudinal的data analysis的最大问题好像是missing data : 好像有很多讨论,最常用的是不是last observation carry forward? : 用(一段时期的)平均值来代替missing values,怎么样?
|
g**r 发帖数: 425 | 4 如果你用MIXED MODEL,你的假设是MAR,忽略就可以了。
如果你不能假设MAR,去找一下SELECTION MODEL 和PATTERN MIXTURE MODEL的资料。
LOCF一般是在用ANOVA的时候才用的,虽然简单,这个现在不时髦了:被人证明存在明
显误差。
另外,还有一个保守的做法是BASELINE CARRY FORWARD,这个就狠了, PENALTY很大。 |
d******g 发帖数: 130 | 5 You may want to try last-value-carried-forward or perform multiple
imputation using SAS Proc MI
【在 s*******r 的大作中提到】 : longitudinal的data analysis的最大问题好像是missing data : 好像有很多讨论,最常用的是不是last observation carry forward? : 用(一段时期的)平均值来代替missing values,怎么样?
|
s*******r 发帖数: 769 | 6 I read a book and it said Mixed effects models can take better care of
missing data than gee. but how?
can anyone explain this to me?
thanks
【在 d******g 的大作中提到】 : You may want to try last-value-carried-forward or perform multiple : imputation using SAS Proc MI
|
s*****n 发帖数: 2174 | 7 关键要看missing mechanism是什么,
是 MCAR, MAR, 还是 NIM? NIM 里面又分2种, outcome-based missing 和 random-
effect-based missing
简单的方法(比如mean-substitution, LOCF)最多只对MAR适用, 对于 NIM, 恐怕得用
MCMC, EM.
【在 s*******r 的大作中提到】 : longitudinal的data analysis的最大问题好像是missing data : 好像有很多讨论,最常用的是不是last observation carry forward? : 用(一段时期的)平均值来代替missing values,怎么样?
|
c*******n 发帖数: 300 | 8 If MAR is assumed, the MLE is valid when using mixed effects model.
You can see some missing value books for details.
【在 s*******r 的大作中提到】 : I read a book and it said Mixed effects models can take better care of : missing data than gee. but how? : can anyone explain this to me? : thanks
|
l****a 发帖数: 352 | 9 as i understood, gee is based on moment, can't do well for missing value
mixed effects model is based on likelihood, can handle missing value
【在 s*******r 的大作中提到】 : I read a book and it said Mixed effects models can take better care of : missing data than gee. but how? : can anyone explain this to me? : thanks
|
s*******r 发帖数: 769 | 10 it is true that MRM (mixed effect regression model) can take better care of
missing data than GEE
MRM uses full likelihook but GEE uses quasi likelihood
【在 l****a 的大作中提到】 : as i understood, gee is based on moment, can't do well for missing value : mixed effects model is based on likelihood, can handle missing value
|
l******g 发帖数: 2429 | 11 Nebraska-Lincoln 的 Seth Spain 正在做一个longitudinal的methods研究,他的成果
已经可以说是能彻底颠覆longitudial的传统方法了,organizational research
methods上明年大概能发出来 |