Design and experiments

Model diagnose

image-20200808144119933

image-20200808144131770

Design with both nested and factorial factors

image-20200729191811153

image-20200729191830357

image-20200729192330853

自由度 degree of freedom

道格拉斯试验设计206页

interaction term 的自由度是(a-1)(b-1), 这是因为treatment A 的自由度是a-1,treatment B的自由度是b-1。

test for nonadditivity

  • when to use: one observation per cell, want to test if the interaction term is not significant
  • douglas 203-204, and 205 example

Split plot design

image-20200729225354644

image-20200729225424772

image-20200729225501837

image-20200729225743109

image-20200729225829915

image-20200729225849614

image-20200729230618155

Rules for expected mean squares

image-20200729231013618

你可以死算,不过太烦。

image-20200729231131004

image-20200729231249155

image-20200729231344799

image-20200729231427168

image-20200729231542903

image-20200729231601013

image-20200729231918957

image-20200729231943873

image-20200729232129940

image-20200729232402400

this is because summing over fixed effect equals zero for the interaction terms. (Which means it Is a restricted model)

image-20200730002413853

image-20200730002540717

Qualify exam 2019

image-20200730004046614

image-20200730010521749

image-20200730010541930

  • (a)到(c)问只要用它给定的模型算就可以了

  • (d)问有个小技巧,就是看它的variance组成,比如这里我们观察到这两项的差是在同一个tray里的,所以没有$\sigma_{tray}$ 这一项,但肯定会有$\sigma_{tray:planttype}$和$\sigma_{\epsilon}$。所以我们需要关注哪一个term的ems是含这两项的。很明显是planttype:tray interaction。由于有两个replicate,所以ems(planttype:tray) = $\sigma_{\epsilon}^2+2\sigma_{tray:planttype}^2$

  • (g)问以后都是consider the nested model, 这里tray effect nested在pesticide里面了。所以根据这张表

  • image-20200802110021451

  • image-20200729192330853

  • 这张表里,可以把$\beta$ 看做pesticde effect,把$\tau$看做tray effect,把$\gamma$看做是planttype。

  • 在test pesticide effect的时候,F test分母上除以的是tray effect, 因为nested design没有pesticide:tray这一项。

  • 在test pesticide:planttype:interaction的时候,分母上除以的是planttype:tray interaction。

Qualify exam 2018

image-20200730104214237

image-20200730104237795

image-20200730104258174

image-20200730104315144

image-20200730104334873

image-20200730104351932

image-20200730104409453

image-20200730104429132

image-20200730104443865

image-20200730104458523

image-20200730104514327

  • (a):

    image-20200812111651421

  • (d): if the error variance is now $\sigma^2 V$, the BLUE is

    image-20200812005253291

  • (e) Write the cell mean model into the general form like (a), and show their degree of freedoms are equal. Then based on the R command, write down the design matrix of R, and see what’s actually estimated by R. This can be achieved by rewriting the general form model to make it has the same design matrix as R’s.

Qualify exam 2017

image-20200730162339509

image-20200730162400121

image-20200730162417554

  • 这个问题模型已经给定了。

  • (d)问中mc main effect 的ems(expected mean squared)的计算方法就是看模型中有多少项是含subscript j的。这里我们发现$b_j$, $c_{ij}$, 和$\tau_{jk}$,$\epsilon_{ijk}$ 都是含j的项。而又因为$c_{ij}$ 是fixed and restricted的,所以他不会出现在ems里面。而 $\tau_{jk}$ 是random effect,所以会出现在ems里。在做F-test的时候分子上是MS(mc main effect), 分母上应该含有除$b_j$以外的剩余项,也就是MC:GH interaction 的ems。

  • (e)问中geno main effect 的ems是不geno:GH这一项的,(因为模型里没有,其实就等于假设了他们的interaction为0),所以它的ems只含有$a_i$和$\sigma_{\epsilon}$。在做F-test的时候,分母上是geno:GH的SS加上error的SS,再除以这两项的自由度之和。

  • (g)问中geno:mc interaction只含$c_{ij}$和$\sigma_{\epsilon}$这两项,所以在F test里也除以的是geno:GH和error term的pool error。

  • 这里的error term统统指的是三阶interaction:GH:gend:mc。

  • (h): the design becomes a two-factor block design.

Qualify exam 2016

image-20200730233609224

image-20200730233651803

image-20200730233715470

image-20200730233733471

image-20200730233816188

image-20200730233835846

image-20200730233855107

  • (a) tests the definition of estimable. Refer to:

    image-20200802113512637

  • 这是一道split design的题目,如何做F test取决于你的模型长什么样子。

  • 有两种常见的假设:

  • 第一种是假设whole plot effect:replicate存在,而split plot effect: replicate = 0,也就是被pool进error term里。

  • image-20200731120825865

  • 第二种假设是whole plot effect: replicate存在,split plot effect: replicate 也存在。

  • image-20200731120926765

  • 在选定模型以后按部就班做即可。这里我倾向于选择第二个模型,即split plot error term 包含所有高阶的包含area和depth的项。anova table 长这个样子:

    image-20200822103115955

  • 最后一问的做法是:

    image-20200806171054368

    We can sometimes find a term like MS_whole_plot_error, its expectation exactly has the form that our estimator’s variance has, or has the the shape: a$\sigma_a^2 + b \sigma_b^2$,, in this situation, we can directly use t-test with degree of freedom equal to the term MS_… possess.

QE sample 1

image-20200806220137564

image-20200806220155913

image-20200806220213864

image-20200806220229384

image-20200806220313300

image-20200806220330489

  • Split plot design

  • (d)

  • image-20200816155622465

  • (e) MS_fert / MS_fert:yard

  • (f) MS_prune / ((SS_prune:yard + SS_prune:yard:fert) / 12) 括号里是split error MS 的计算方式.

  • (f)模型有所变化,在确定了fert和pruning method以后,就不用考虑whole plot error了。简而言之,这是一个全新的实验,但我们希望从老的实验中提取一些有效信息去估计新的实验的结果。

    image-20200816155752269

    Since E(MS_sp) = $\sigma_{\delta}^2$ + $2\sigma_{jk}^2$ (this is because l = 1,2) and E(MS_error) = $\sigma_{\delta}^2$. Here, $\hat{\sigma_{SPE}^2}$ is actually $\sigma_{jk}^2$.. So it’s easy to estimate the two terms.

Qualify exam 2014

image-20200805223005808

image-20200805223011663

  • (a): materials

  • (b): sum to zero constraints applied on both lab effect and material effect

  • (c): considered as fixed

  • (d): two way anova with interaction

  • (e) (i): contrast can be $\mu_A - \mu_B = 0$, $\mu_C - \mu_D = 0$, $\mu_A + \mu_B - \mu_C - \mu_D = 0$

  • (e) (ii): the estimated contrasts are $\hat{\mu_A} - \hat{\mu_B}$ , $\sigma^2$ can be estimated by SSE.

  • (e) (iii):

    image-20200805225913760

Qualify exam 2015

image-20200807222647103

image-20200807222705023

image-20200807222723615

  • Radom effect model.

Stat850 assignment

image-20200808141502041

  • image-20200808143732228

image-20200808143818464

  • image-20200808143845902

    image-20200808143923932

image-20200808154101885

  • image-20200808154156391

  • image-20200808154217274

    image-20200808154232320

  • image-20200808154305675

  • image-20200808154342879

image-20200808154535350

image-20200808160825820

image-20200808161110533

image-20200808161132034

image-20200808161150509

image-20200808161203282

image-20200808161227534