近期论坛无法登录的解决方案


一亩三分地论坛

 找回密码
 获取更多干活,快来注册

一亩三分地官方iOS手机应用下载
查看: 1157|回复: 4
收起左侧

[统计--核心课程] 关于experimental design的困惑

[复制链接] |试试Instant~ |关注本帖
helenaqa 发表于 2016-5-24 06:50:56 | 显示全部楼层 |阅读模式

注册一亩三分地论坛,查看更多干货!

您需要 登录 才可以下载或查看,没有帐号?获取更多干活,快来注册

x
各位统计大神们,请问我们为什么会有各种各样的experimental design呢? 我在学习的过程中一直在问,为什么不单单用regression直接上indicator variable解出来就好了么?
. 鍥磋鎴戜滑@1point 3 acres. Waral 鍗氬鏈夋洿澶氭枃绔,
为什么会有这个design,那个design,这个block,那个factor,感觉linear model只是把内容复杂化了,还不如regression来的直接呢?

望指教!
demonhunter 发表于 2016-5-24 14:38:51 | 显示全部楼层
关注一亩三分地公众号:
Warald_一亩三分地
本帖最后由 demonhunter 于 2016-5-24 14:50 编辑

Sometimes people are only interested in the factors they want to study however the response is influenced by all factors so they need different design strategies to remove the influence of  these nuisance factors.

Sometimes a simple linear regression is not able to fully explain the inner mechanism of your data(for example, to study repeated measurement you need mixed model) .

Performing an experiment and getting data is far more expensive and time consuming than you can expected, so you have to check the collection of designs carefully and adopt one wisely.
. Waral 鍗氬鏈夋洿澶氭枃绔,
____________________________ calalia mentioned Randomization. That means your cannot always have a simple CRD design(assigning treatments to the Experiment Units completely randomly), whose model equals to a linear regression with several indicator predictors. Sometimes you just cannot run you Experiment in this completely random way(like assigning treatment to the plants in some algriculture experiment) so you need a more complicated model(like the model for split-plot deisgn)
. more info on 1point3acres.com
回复 支持 4 反对 0

使用道具 举报

calalia 发表于 2016-5-24 13:17:21 | 显示全部楼层
关注一亩三分地微博:
Warald
为了Random
回复 支持 反对

使用道具 举报

 楼主| helenaqa 发表于 2016-5-25 10:54:36 | 显示全部楼层
demonhunter 发表于 2016-5-24 14:38
Sometimes people are only interested in the factors they want to study however the response is influ ...

Thank you demonhunter and thank you calalia!

Thank you for your thoughtful and detailed discussion. I understand that the design of experiment is both an art and science. And once we have the design and the data, we can analyze the results using an appropriate linear model (CRD, BIBD, split-splot, etc), or with a linear regression model.

I am still puzzled whether there is a difference in the analysis stage, it appears to me the various linear model designs and linear regression do the same with parameters that differ only by a simple linear transformation.

Is that correct? Or is there more to it than that?

回复 支持 反对

使用道具 举报

demonhunter 发表于 2016-5-25 11:25:50 | 显示全部楼层
本帖最后由 demonhunter 于 2016-5-25 11:27 编辑
helenaqa 发表于 2016-5-25 10:54
Thank you demonhunter and thank you calalia!

Thank you for your thoughtful and detailed discuss ...
.鏈枃鍘熷垱鑷1point3acres璁哄潧
"it appears to me the various linear model designs and linear regression do the same with parameters that differ only by a simple linear transformation.". from: 1point3acres.com/bbs
. Waral 鍗氬鏈夋洿澶氭枃绔,
No, many design linear models estimate parameters in a different way. For example, in mixed model:

y = X\beta + Zu +\epsilon, if you want to estimate u, the Random effect you need to use either ML or REML approach, you have to use more complicated Optimization Algorithm than Ordinary Linear model does.
回复 支持 反对

使用道具 举报

本版积分规则

关闭

一亩三分地推荐上一条 /5 下一条

手机版|小黑屋|一亩三分地论坛声明

custom counter

GMT+8, 2017-6-28 08:33

Powered by Discuz! X3

© 2001-2013 Comsenz Inc. Design By HUXTeam

快速回复 返回顶部 返回列表