[R-sig-ME] Repeated measures with a non-linear time effect

[Daniel Rubi]

Thanks a lot 

    On Thursday, July 21, 2016 4:47 PM, Jörg Albrecht <albrechj at staff.uni-marburg.de> wrote:
 

 Hi Dan,
that was a misunderstanding. I meant you should code time as a factor and run a modified version of the model (see code below).
library(lme4)library(lattice)library(multcomp)
df <- data.frame(subject = rep(c("T1", "T2", "T3", "C1", "C2", "C3"), 3),                 group = rep(c(rep("T", 3), rep("C", 3)), 3),                 time = c(rep(1, 6), rep(2, 6), rep(3, 6)),                 measure = c(0, 253, 155, 16, 232, 251, 1035, 1014, 760, 98, 239, 87, 371, 60, 47, 0, 260, 190),                 col = rep(c(rep("red", 3), rep("blue", 3)), 3), stringsAsFactors = FALSE)df$time <- as.factor(df$time)xyplot(measure ~ time|subject, data = df)
# using the lme4 packagefit <- lmer(measure ~ time * group + (1|subject), data = df)summary(fit)cftest(fit)
# using anovafit <- aov(measure ~ time * group + Error(subject), data = df)summary(fit)
Both models are able to detect the interaction between treatment and time.
Hope this helps.
Best,
Jörg
—Jörg Albrecht, PhDPostdoctoral researcherInstitute of Nature Conservation
Polish Academy of Sciences
Mickiewicza 33
31-120 Krakow, Poland
www.carpathianbear.pl
www.globeproject.pl
www.iop.krakow.pl

Am 21.07.2016 um 21:50 schrieb Daniel Rubi <daniel_rubi at ymail.com>:
Hi Jörg,
Unfortunately setting time as integer doesn't change anything whereas setting it as a factor means that the number of random effects will be equal to the number of samples.
Perhaps this is an alternative:fit1 <- gls(measure~time*group,correlation=corSymm(form=~1|subject),weight=varIdent(form=~1|time),data=df)
as it does capture the group and interaction effects but I think I need to contrast it with a null model and I'm not sure what would that null model be.
Any idea?
 

    On Thursday, July 21, 2016 3:27 PM, Jörg Albrecht <albrechj at staff.uni-marburg.de> wrote:
 

 Hi Dan,

most likely the model treats your time covariate as a continuous predictor. Try str(df), then time should appear as integer (int). If you specify time as a factor (with three levels: 1, 2, 3) the model will be able to estimate the time x treatment interaction separately for each time point. However, you still have to decide whether treating time as a factorial variable makes sense for your dataset.

Best,

Jörg


> Am 21.07.2016 um 20:13 schrieb Daniel Rubi via R-sig-mixed-models <r-sig-mixed-models at r-project.org>:
> 
> Hi,
> I have repeated measures from two groups (treatment and control), three subjects in each, over three time points.
> 
> Here's the data in an R data.frame:df <- data.frame(subject=rep(c("T1","T2","T3","C1","C2","C3"),3), group=rep(c(rep("T",3),rep("C",3)),3), time=c(rep(1,6),rep(2,6),rep(3,6)), measure=c(0,253,155,16,232,251,1035,1014,760,98,239,87,371,60,47,0,260,190), col=rep(c(rep("red",3),rep("blue",3)),3), stringsAsFactors=F)
> 
> 
> The plot shows the time x group interaction:
> 
> R code for producing the plot:
> plot(df$time,df$measure,col=df$col,xlab="time",ylab="measure")
> legend("topleft",legend=c("treatment","control"),col=c("red","blue"),pch=1)
> 
> 
> My question is what model to use to capture the time x group interaction.
> I thought:library(lmerTest)fit <- lmer(measure~time+group+time*group+(time|subject),data=df)
> might do it.
> But the summary of this model doesn't really capture that:> summary(fit)Linear mixed model fit by REMLt-tests use Satterthwaite approximations to degrees of freedom ['lmerMod']Formula: measure ~ time + group + time * group + (time | subject) Data: df
> REML criterion at convergence: 210
> 
> Scaled residuals: 
>    Min    1Q Median    3Q    Max 
> -1.228 -0.448 -0.163  0.275  1.923
> 
> Random effects:
>  Groups  Name        Variance Std.Dev. Corr
>  subject  (Intercept) 0.00e+00 0.00e+00    
>          time        3.06e-16 1.75e-08  NaN
>  Residual            1.05e+05 3.25e+02    
> Number of obs: 18, groups:  subject, 6
> 
> Fixed effects:
>            Estimate Std. Error    df t value Pr(>|t|)
> (Intercept)  168.89    286.35  13.78    0.59    0.56
> time          -8.17    132.55  13.78  -0.06    0.95
> groupT        218.33    404.96  13.78    0.54    0.60
> time:groupT    19.83    187.46  13.78    0.11    0.92
> 
> Correlation of Fixed Effects:
>            (Intr) time  groupT
> time        -0.926              
> groupT      -0.707  0.655      
> time:groupT  0.655 -0.707 -0.926
> 
> 
> So my question is what model to use?
> 
> Thanks a lot,Dan
> 
> <Rplot.png>_______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models


   



  
	[[alternative HTML version deleted]]