Hi Dennis, The first thing I did with my data was to explore it with 6 graphs (wet-high, med, and solo-; dry-high, med, and solo-) and gave me very interesting patterns: seed size in wet treatments is either negatively correlated (high and medium densities) or flat (solo). But dry treatments are all positively correlated! There is a very interesting switch there.
I also figured out why I can't do three way interactions. I explored the structure of my data with str(mydata) and it shows that water treatment has three levels when it should have just two. Then I went back to the excel sheet, tried to sort the data by water treatment and I discover a single data point from the wet treatment sticking out by itself. That is why R reads three levels and since it is only one point, there cannot be any stats of course. thanks E On Thu, Oct 14, 2010 at 9:27 PM, Dennis Murphy <djmu...@gmail.com> wrote: > Hi: > > On Thu, Oct 14, 2010 at 7:50 PM, Eugenio Larios < > elari...@email.arizona.edu> wrote: > >> Hi Dennis, >> >> thank you very much for your help, I really appreciate it. >> >> I forgot to say about the imbalance, yes. I only explained the original >> set up, sorry. Let me explain. >> >> It is because in the process of the experiment which lasted 3 months I >> lost individuals within the plots and I actually ended up losing 2 whole >> plots (one dry and one wet) and some other individuals in other plots. >> > > That still leaves you balanced at the plot level :) Fortunately, you have > enough replication. If you have missing subplots within the remaining plots, > that would be another source of imbalance at the subplot level, but you > should have enough subplots to be able to estimate all of the interactions > unless an entire treatment in one set of plots was missing. > > It's worth graphing your data to anticipate which effects/interactions > should be significant; graphs involving the spatial configuration of the > plots and subplots would also be worthwhile. > >> >> My study system has this special feature that allows me to track parental >> seed sizes in plants germinated in the field, a persistent ring that stays >> attached to the root even when the plant has germinated, so some of the >> plants I lost did not have this ring anymore. It happens sometimes but most >> of the time they have it. Also, some plants disappeared probably due to >> predation, etc That made my experiment imbalanced. >> > > That's common. No big deal. > >> >> Do you think that will change the analysis? Also, do you think I should >> use least squares ANOVA (perhaps type III due to the imbalance?) instead of >> LMM? What about the random effects that my blocking has created? >> > > Actually, with unbalanced data it's to your advantage to use lme() over > ANOVA. Just don't place too much importance on the p-values of tests; even > the degrees of freedom are debatable. With unbalanced data, it's hard to > predict what the sampling distribution of a given statistic will actually > be, so the p-values aren't as trustworthy. > > You mentioned that you couldn't fit a three-way interaction; given your > data configuration, that shouldn't happen. > > (1) Get two-way tables of water * density, one for the counts and one for > the averages, something like > > with(mydata, table(water, density)) > aggregate(log(fitness) ~ water + density, data = mydata, FUN = mean, na.rm > = TRUE) > > In the first table, unless you have very low frequencies in some category, > your data 'density' should be enough to estimate all the main effects and > interactions of interest. The second table is to check that you don't have > NaNs or missing cells, etc. > >> >> I am new to R-help website so I wrote you this message to your email but I >> would like to post it on the R website, do you know how? >> > > Wag answer: I hope so, since I managed to view and respond to your message > :) > > More seriously, in gmail, the window that opens to produce replies has an > option 'Reply to all'. I don't know if your e-mail client at UofA has that > feature, but if not, you could always cc R-help and put the e-mail address > in by hand if necessary. Most mailers are smart enough to auto-complete an > address as you type in the name, so you could see if that applies on your > system. > > I keep a separate account for R-help because of the traffic volume - if you > intend to subscribe to the list, you might want to do the same. It's not > unusual for 75-100 e-mails a weekday to enter your inbox... > >> >> Thanks again! >> >> Eugenio >> >> >> On Thu, Oct 14, 2010 at 5:34 PM, Dennis Murphy <djmu...@gmail.com> wrote: >> >>> Hi: >>> >>> On Thu, Oct 14, 2010 at 3:58 PM, Eugenio Larios < >>> elari...@email.arizona.edu> wrote: >>> >>>> Hi Everyone, >>>> >>>> I am trying to analyze a split plot experiment in the field that was >>>> arranged like this: >>>> I am trying to measure the fitness consequences of seed size. >>>> >>>> Factors (X): >>>> *Seed size*: a continuous variable, normally distributed. >>>> *Water*: Categorical Levels- wet and dry. >>>> *Density*: Categorical Levels- high, medium and solo >>>> *Plot*: Counts from 1 to 20 >>>> The *response variable *(Y) was the number of seeds produced at the end >>>> of >>>> the season. >>>> >>>> The experiment started 15 days after plants germinated in the field. >>>> 20 plots were chosen where there was high enough density so I could >>>> manipulate it. In an area where artificial irrigation was possible for >>>> the >>>> wet treatment, dry treatment was natural precip. >>>> Water was blocked so 10 plots were wet and the other 10 were dry. >>>> Randomly >>>> assigned. >>>> Within those 20 plots 6 focal plants were chosen and randomly assigned >>>> the >>>> three densities. (split plot design) >>>> I did not control for seed size since it is continuous and normally >>>> distributed, hoping that with 120 plants total (6 in each 20 blocks) I >>>> could >>>> get all kind of sizes for every treatment. It worked ok. >>>> >>> >>> From the description, it appears you have the following: >>> * water is a whole-plot treatment, each level assigned to 10 plots >>> * seed size is a plot-level covariate >>> * whole plot units are the plots >>> >>> At this level, the ANOVA table is >>> >>> Water 1 >>> Seed size 1 >>> Water x seed size 1 >>> Whole plot error 16 [plots] >>> >>> The split plot treatment is density, and after its main effect is >>> accounted for, it is crossed with every term in the whole-plot ANOVA: >>> >>> Density 2 >>> Density * Water 2 >>> Density * seed size 2 >>> Density * Water * seed size 2 >>> Residual 92 [subplots] >>> >>> Total df = 119 >>> >>> The ANOVA exercise is useful for understanding the structure of the >>> split-plot design - it is not exactly what lme() will fit (especially the >>> df), since lme() is fitting the model via likelihood rather than least >>> squares. >>> >>> Your full lme model, including the test of unequal slopes in the two >>> water levels, should be >>> >>> m <- lme(log(fitness) ~ seedsize * water * density, random = ~1|plot, >>> data=dataset) >>> >>> Without the unequal slopes term (i.e., a parallel slopes model), it >>> should be >>> >>> m2 <- lme(log(fitness) ~ (seedsize + water) * density, random = ~1 | >>> plot, data = dataset) >>> >>> The specification of the first two terms on the RHS of the model formula >>> is associated with the whole-plot structure of your design. >>> >>> I have been trying to analyze this with lme (library NLME). I am not >>>> quiet >>>> sure which are my random variables. models I have used are: >>>> >>>> m<-lme(log(fitness)~seedsize*density,random=~1|plot,data=dataset) >>>> m<-lme(log(fitness)~seedsize+density+water,random=~1|plot,data=dataset) >>>> >>>> I have also tried to include plot and water as random effects: >>>> >>>> >>>> m<-lme(log(fitness)~seedsize+density+water,random=~1|plot/water,data=dataset) >>>> >>>> I am actually not sure if I am using the right random variables here. >>>> Also >>>> for some reason, it won't let me include seedsize*density*water triple >>>> interaction >>>> >>> >>> You mentioned imbalance in your mail header - how imbalanced are you >>> talking about? The structure of the imbalance could have some impact on >>> which effects are or are not estimable, depending on its severity. >>> >>> >>> HTH, >>> Dennis >>> >>> >>>> help! >>>> thanks >>>> >>>> -- >>>> Eugenio Larios >>>> PhD Student >>>> University of Arizona. >>>> Ecology & Evolutionary Biology. >>>> (520) 481-2263 >>>> elari...@email.arizona.edu >>>> >>>> [[alternative HTML version deleted]] >>>> >>>> ______________________________________________ >>>> R-help@r-project.org mailing list >>>> https://stat.ethz.ch/mailman/listinfo/r-help >>>> PLEASE do read the posting guide >>>> http://www.R-project.org/posting-guide.html >>>> and provide commented, minimal, self-contained, reproducible code. >>>> >>> >>> >> >> >> -- >> Eugenio Larios >> PhD Student >> University of Arizona. >> Ecology & Evolutionary Biology. >> (520) 481-2263 >> elari...@email.arizona.edu >> > > -- Eugenio Larios PhD Student University of Arizona. Ecology & Evolutionary Biology. (520) 481-2263 elari...@email.arizona.edu [[alternative HTML version deleted]] ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
