[R-sig-ME] Need advice on model specification for repeated measures on split-plot design
Taylor, Jason
Jason_Taylor1 at baylor.edu
Wed Nov 24 19:50:14 CET 2010
Hello,
I sent this earlier this week but realized my data table had been unfolded in the email. Here is my post again with the data table formatted correctly.
Thanks,
Jason
I am new to the mixed-model world. I posted a couple of weeks ago but received no feedback so I have plowed on ahead and tried to figure things out on my own. Any feedback on my current perspective would be very useful.
I essentially have a split plot design with repeated measurements.
I have 12 experimental sampling units.
Each sampling unit is assigned to one of 3 nutrient treatments (4 replicates per treatment).
Within each sampling unit I have a grazed and ungrazed treatment.
Each sampling unit was sampled repeatedly on day 0 (before nutrients and grazer manipulations began), day 14 and day 28.
I am interested in the fixed effects of nutrient, grazing, time and their interactions but also want to account for potential random differences in sampling units.
First few lines of data are listed below for reference.
obs Sample unit Date Nut Graz Response
1 S1 DAY0 Low G 2.014
2 S5 DAY0 Low G 2.487
3 S8 DAY0 Low G 2.144
4 S11 DAY0 Low G 1.946
5 S1 DAY0 Low UG 2.199
6 S5 DAY0 Low UG 1.666
7 S8 DAY0 Low UG 1.642
8 S11 DAY0 Low UG 2.288
9 S3 DAY0 Med G 2.786
10 S6 DAY0 Med G 1.766
11 S7 DAY0 Med G 1.756
12 S12 DAY0 Med G 0.943
13 S3 DAY0 Med UG 2.124
14 S6 DAY0 Med UG 1.33
15 S7 DAY0 Med UG 1.847
16 S12 DAY0 Med UG 1.424
17 S2 DAY0 High G 1.775
18 S4 DAY0 High G 1.838
19 S9 DAY0 High G 2.971
20 S10 DAY0 High G 2.14
21 S1 DAY14 Low G 2.014
22 .......
I believe the correct code for a starting model in lme4 that accounts for the fixed effects and their interactions + the random effect of sample unit would be:
model.1<-lmer(response~Nut*Graz*Date+(1|Sample unit))
I realize now that if I wanted to do any nested random effects my data is currently coded poorly (implicit nesting of the variables). However, if I understand correctly, I don't need to incorporate (1|Stream:Graz) or (1|Stream:Date) as nested random effects because:
1) autocorrelation between Graz or Date treatments within a sample unit will be accounted for by treating the sample unit as a random effect,
2) they are not pseudo-replicates but actual fixed factors that I am interested in, and
3) Graz only has 2 groups and I may only use day 14 and 28 for date (also resulting in 2 groups) since Day 0 was not influenced by any of the treatment effects. Estimating variance from only 2 groups would be pointless.
Thanks in advance for any comments,
Cheers,
Jason
More information about the R-sig-mixed-models
mailing list