Almost, Look at Analysis of Covariance (ANCOVA). Best, David Booth

The main reason for using a linear mixed model is when you have repeated measures in your data. In that case you must account for the within "subject" variation as a random effect. If you do not have repeated measure, then there are probably much simpler models you can use such as ANCOVA.

The most accessible tutorials on LME in R that I have found are by Bodo Winter found here: http://bodowinter.com/tutorial/bw_LME_tutorial1.pdf and part 2 here: http://bodowinter.com/tutorial/bw_LME_tutorial.pdf

In R, the best way to test if a fixed effect is significantly different from zero is to create two separate lmer models and compare them via anova. If the model with the term has a significantly lower AIC, then that term is significant. It doesn't give you a p-value for that specific term, only for the model with that term in it. There are some add on packages in R that will convert the F statistic of the lmer model to p values (sjPlot), but it is best to also understand how to use LME first.

-Phil

I don't understood completely your question, so I will respond the main question. "No difference in intercept of two var in mixed model same as their differences = 0?"

At first I am assuming that the "two var" in your question reffers to a two classes of the same variable. If there is no differences between intercepts then we don't can conclude by the existence of differences between the classes when the the value of the independent variable is 0. But we can have different slopes and as we have some distance from the 0, difference can exist.

The rest of the question is confuse. I can't understand if you have, or not, paired data to have a valid diference (1 minus 2). If you need to apply a mixed model then I agree with David Eugene Booth when he writes about ANCOVA model.

Phil J Greer 2

Hello thanks. Question:

what do the lmer models need ot contain when I compare them in ANOVA?

I have only one fixed effect factor variable . therefore I cannot create a reduced model or can I?

In the endeffect I want to test if the mean of a given dataset (that has random "component") is statistically different from 0. thanks

specifically. I have values of carbon content (that are calculated by subtracting paired vegetation1-and veg2 C contents). random would be some spatial factor and I would like to know if the difference of the values isn't the same as 0 which would imply that the veggietypes have different contents of C

I would need more information about your data, but the barest minimum model that you can create with lmer is: lmer(y ~ 1 + (1 | random effect), data= df) this model tells you very little about your data, but it is the barest minimum that you need to have lmer run correctly. In essence, it says Y equals a straight line with a random intercept influenced by the random effect term.

The most important thing about the models that you will be comparing is that the two models MUST have the same set of random effects terms in order to be comparable.

Now let me take a stab at what you model would look like based on my understanding so far. please correct me if I am wrong.

model.1

@Phil I don't agree. consider the Ancova model y=post -pre +(rest of model). This is one way to handle a simple repeated measure. For the whole story, see Prof. Marie Davidian's class notes at the following link: https://www4.stat.ncsu.edu/~davidian/st732/notes.html David Booth

Hello thanks for the answers! My data are in the model

lmer(C~year+ 1|some plots, data =data)

The C is actually difference of measured carbon bound to soil taken between field and flower strip, year is the year the flower strip was turned into field again (its only two levels) and some plots are plots within which C was repeatedly measured.

In the meantime I found some solution through using lsmeans with lme from nlme package. lmer wouldn't work (bug or so)

summary((lsmeans(mod, ~year)) ,..., by=year) it was.

The summary does give u p-value, of lsmeans of yr vs null-hypothesis mean = 0. Whereby, the lsmeans with no other fixed factor is just the mean of each level and the calc. includes the random factor. ;)

David Eugene Booth, I agree with you. Without know fully what the data looks like and with only one explanatory variable, ANCOVA is a far easier model to understand and implement especially when there are only two time points. I was specifically trying to answer the question using a linear mixed model framework.

That being said, I do not think Vera Maria should use a mixed model for this analysis. Especially if the simple model of lmer(C~year+ (1|some.plots), data =data) is failing.

Let's step back and ask how would you present this data graphically?

I, personally, would plot this data using box plots. C would be on the y axis with "some plots" on the x axis and each "some plot" split into the two "year" categories. If you had 4 "some plots", two "year" categories, and multiple samplings for each "year" in each "plot", you would end up with 8 boxes on the entire plot. If you do not have enough multiple samplings for each plot by year category, lmer may not be able to correctly estimate the random effect.

It may be better to just ignore the random effect component of "some plots" for now and run the data with an anova and an interaction between year and plot. You should then add a note that the lack of accounting for plot as a random effect is a limitation, but unavoidable due to lack of data and the time it takes to collect randomized data.

Phil

Phil J Greer

Thank you. The thing is, this is a thesis and I believe that the data,

that were previously used, were already supervised by an expert and it was recommended to use a mixed model with plots as random effects (they have similar land use history)

Then again, when asking a guy on a statstical help board , he said that the design is quite complex i should hire a consultant. Mhh.

I just want to add here the data in a picture... do you really think ANCOVA is appropriate or even lmer is completely WRONG??? it's really quick with this picture. The unfilled circle would denote another usage (year).... thanks

Vera Maria , Tell me if my understanding and assumptions here are correct.

Your "year" variable is your grouping variable and denotes a length of growing time. Thus category 1 was grown for n season(s) and category 2 was for n+1 seasons. From you picture, category 1 is an open circle and category 2 is a closed circle (or the other way around, not that important right now). Also from your picture you have 7 groups of 6, or 14 groups of 3, or some value in-between. Either way, each of the clusters appears to have only one type of growth group within that cluster.

If this is the case, then you have a very serious confound in the model. You may need to group plots together (9 and 8, 2 and 4, 5n and 5s) so that each of your plot areas can have year sampled in each of plots. That may end up inflating the variance in the random effect of "plot", but it will allow the model to run.

The best example I can come up with is: I want to compare fish growth in four lakes (A,B,C, & D). I go there and catch a bunch of fish, but for some odd reason, I catch one species of fish in lakes A and C, and a different species of fish in lakes B and D. What model can I use to analyze this? I was planning on using fish species as a covariate because I know that they grow at different rates, but now fish species is confounded with my random effect of lake. If I have a sampling of both fish species in all four lakes, then I don't have a problem and lmer will work fine.

I hope I am making that clear.

Phil

Yes this makes it clear. So I do not have both species represented in all clusters wherefore I cannot use the lmer. oO if I get that correctly it's now time to tell this to my supervisor soon:D and in the meantime use two lmers and compare the dudes or do the ancova anyway thanks very comprehensive with the fish example!