使用multcomp设置三路交互作用的对比

Question

我有一个BACI试验，涉及三个站点在2007年和2011年之前和之后的测量。我在设置对比时遇到了问题，以便我能够看到每个站点的治疗效果（与对照组相比）的年份效应是否不同。

我尝试过扩展vignette中的示例，但没有成功。我也尝试过lsmeans包。

下面是示例数据框和代码，以及我想测试的明确对比。

library(nlme)
library(multcomp)
require(lsmeans)

# I have these versions:
# R version 3.2.4 Revised (2016-03-16 r70336)
# Platform: x86_64-w64-mingw32/x64 (64-bit)
# Running under: Windows 7 x64 (build 7601) Service Pack 1
# nlme_3.1-126
# multcomp_1.4-4 



exdf <- expand.grid(Site = c("Z",  "Y", "X"), 
                   Plot = 1:30,
                   Year = c("2007", "2011"), 
                   Treatment = c("Control", "A", "B", "A+B"))
exdf$Site <- factor(exdf$Site)
exdf$Plot <- factor(exdf$Plot)
exdf$Year <- factor(exdf$Year)
exdf$Treatment <- factor(exdf$Treatment)
exdf$Treatment <- relevel(exdf$Treatment, ref = "Control")

exdf$Response <- rnorm(mean = 1150, sd = 130, n = nrow(exdf))

mod1 <- lme(data = exdf,
            fixed =  Response ~ Year * Treatment * Site,
            random = ~1|Plot)


## would like to test (obviously that's not actual code)

## How do the treatments A, B, A+B differ in their interaction with 
## year, compared to the control treatment (ie do the slopes differ)

SiteX:Year:TreatmentControl - SiteX:Year:TreatmentA == 0
SiteX:Year:TreatmentControl - SiteX:Year:TreatmentB == 0
SiteX:Year:TreatmentControl - SiteX:Year:TreatmentA+B == 0

SiteY:Year:TreatmentControl - SiteY:Year:TreatmentA == 0
SiteY:Year:TreatmentControl - SiteY:Year:TreatmentB == 0
SiteY:Year:TreatmentControl - SiteY:Year:TreatmentA+B == 0

SiteZ:Year:TreatmentControl - SiteZ:Year:TreatmentA == 0
SiteZ:Year:TreatmentControl - SiteZ:Year:TreatmentB == 0
SiteZ:Year:TreatmentControl - SiteZ:Year:TreatmentA+B == 0

## Also potentially - does the interaction between site and year
## differ across sites? 
SiteX:Year:TreatmentA - SiteY:Year:TreatmentA == 0
SiteZ:Year:TreatmentA - SiteY:Year:TreatmentA == 0
SiteZ:Year:TreatmentA - SiteX:Year:TreatmentA == 0

SiteX:Year:TreatmentB - SiteY:Year:TreatmentB == 0
SiteZ:Year:TreatmentB - SiteY:Year:TreatmentB == 0
SiteZ:Year:TreatmentB - SiteX:Year:TreatmentB == 0

SiteX:Year:TreatmentA+B - SiteY:Year:TreatmentA+B == 0
SiteZ:Year:TreatmentA+B - SiteY:Year:TreatmentA+B == 0
SiteZ:Year:TreatmentA+B - SiteX:Year:TreatmentA+B == 0


## Started doing the example from 
#https://cran.r-project.org/web/packages/multcomp/vignettes/multcomp-examples.pdf
# Got a bit stuck

tmp <- expand.grid(Year = levels(exdf$Year),
                   Treatment = levels(exdf$Treatment),
                   Site = levels(exdf$Site))
X <- model.matrix(~ Year * Treatment * Site, data = tmp)
glht(mod1, linfct = X)


## Gives all the interactions not in the form I want..
tt <- lsmeans(mod1, specs = ~Year:Treatment|Site)
comps <- pairs(tt, interaction = TRUE)
summary(comps,  adjust = "holm")

- Walden

---

关于lsmeans中的“not n the form I wasn't” - 你想要什么形式？似乎你想要类似于~Year|Treatment*Site的东西，也许是该结果的pairs()吗？ - Russ Lenth

谢谢。通过“form”，我是指测试代码中列出的比较。我尝试了您的建议（与我的代码末尾类似），但没有测试上面的比较。一个稍微不同的公式，ls <- lsmeans(mod1, specs = ~Year*Treatment|Site); pairs(tt, interaction = FALSE) 提供了正确的结构，但会测试每个年份和治疗方法的组合。 - Walden

Answer 1

让我们分步骤构建它。首先，您想要获取每个其他因素组合的两年差异。在lsmeans中很容易实现。

tt = lsmeans(mod1, specs = ~ Year | Treatment:Site)
dd = pairs(tt, reverse = TRUE)   # (2011) - (2008) for each T*S comb
summary(dd, by = NULL)

你现在可以根据自己的喜好比较这些差异：

pairs(dd, by = "Site")
pairs(dd, by = "Treatment")

... 或者略微比较它们：

pairs(lsmeans(dd, "Treatment"))
pairs(lsmeans(dd, "Site"))

- Russ Lenth