二分类重复测量资料的GEE模型拟合

是 否 +2 论坛币

k人 参与回答

经管之家送您一份

应届毕业生专属福利!

求职就业群

赵安豆老师微信：zhaoandou666

经管之家联合CDA

送您一个全额奖学金名额~ !

立即领取

感谢您参与论坛问题回答

经管之家送您两个论坛币！

+2 论坛币

大家好！
我在用SAS做二分类重复测量资料的GEE模型时，发现：当选择作业相关矩阵为可交换相关矩阵时，自变量较少可拟合得出估计结果，但自变量较多时，会出现迭代收敛错误而停止估计。请问有哪位高手知道，这是为什么呢？

扫码加我 拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

分享0 收藏2 回帖

关键词：重复测量资料 模型拟合 重复测量 GEE 二分类 自变量 测量 模型 资料

I fit GEE quit a lot. Indeed, convergence is of problem sometimes. As you noticed, it can be resulted from many reasons: small sample size, too many covariates, in particular too many levels in class variables, too complicated covariance structures and so on. In your case, I would guess something happened because of your covariates. For example, zero count cells due to further thinning your data or many records removed from the model since missing values in covariates set. Sometimes those problems may show in the form of semi-/complete separation of the data. You may also want to look at some discriptions by Paul Allison etc.
JingJu

非常感谢您的回复，对我帮助很大。我会按照您的提示，在协变量的选择上更加慎重一些。
另外，我还想向您请教的是，在协变量的初筛过程中，也就是做粗OR分析时，应该用简单Logistic回归分析（不考虑应变量同一个体的相关）呢，还是用双变量GEE模型分析（考虑作业相关矩阵）？我发现：简单Logistic回归分析与不同作业相关矩阵的GEE模型分析，自变量的估计系数完全相同或偶有小差别，但自变量P值却相差很大，结论总是相反，常常为简单Logistic回归分析有统计学意义，而GEE模型无。我应该以哪个模型来初筛协变量呢？
还有，我理解作业相关矩阵为independent时，GEE模型应该与简单Logistic回归分析一致，而实际情况并非如此，同样出现了二者估计系数一致，而P值统计结论相反的情况，这又如何解释呢？
期待您的答复，我刚开始接触GEE，好迷茫，谢谢您百忙之中帮助我。

zzq98370 发表于 2010-9-16 16:44
还有，我理解作业相关矩阵为independent时，GEE模型应该与简单Logistic回归分析一致，而实际情况并非如此，同样出现了二者估计系数一致，而P值统计结论相反的情况，这又如何解释呢？

Where do you refer to this conclusion? GEE is even not a strict MLE, is it?

哦，真抱歉，我连这点最基本的区别都没想到。不过，您觉得自变量的初筛应该用哪个方法呢？

zzq98370 发表于 2010-9-17 08:06
哦，真抱歉，我连这点最基本的区别都没想到。不过，您觉得自变量的初筛应该用哪个方法呢？

Very hard to answer. In terms of SAS, it doesn't provide model selection in GENMOD so far. People think, when fitting models, you should weigh the balance bewteen its fitting and practical meaning. Some people say, if you discard those repeated measures but using cross sectional data instead(in Logistic), you still will get at least 80% efficiency.
when comparing covariance structure or included covariates in GEE, you may look at QIC , which came to GENMOD after 9.1.3.