More options for field layout about fieldhub HOT 3 OPEN

Followup: So this is basically the design I talked about before:

optim_multi_prep <- multi_location_prep(lines = 27,
l = 3, copies_per_entry = 4, checks = 0, nrows = 3, ncols = 12,
locationNames = c("LOC1", "LOC2", "LOC3", "LOC4"), seed = 1234)

We would like to make kind of a sparse version of this, going down to "copies per entry = l", but to have two replicates at one location, none at another location, and 1 replicate in the other locations. FielDHub at the moment does not support this, as it produces an error message, here is the specification:

optim_multi_prep <- multi_location_prep(lines = 27, l = 3,
copies_per_entry = 3, checks = 0, nrows = 3, ncols = 12,
locationNames = c("LOC1", "LOC2", "LOC3", "LOC4"), seed = 1234)

Error in do_optim(design = "prep", lines = lines, l = l, copies_per_entry = copies_per_entry, :
p-reps option requires that copies_per_entry be greater than the number of locations

Is it feasible to implement this?

Best Gregor

from fieldhub.

There are two things here:

1. More options in the field dimensions: I made a decision to limit the number of experimental layouts in FielDHub to four rows or less. This is because these experiments are usually analyzed using spatial methods like splines and autoregressive models. We use the rows and columns as either fixed or random effects in the models. In some cases, the model can only work successfully if we have enough degree of freedom to estimate the row and column effects. I have seen many spatial models fail when trying to estimate the effect of the row (when there are fewer than four) simply due to a lack of degree of freedom. What do you think about this?

2. Copies per entry are the same as the number of locations: If the number of copies per entry is the same as the number of locations, then the experiment is no longer partially replicated. In such a scenario, it becomes an unreplicated design where each treatment or genotype appears once at each location. However, it's important to note that it won't be possible to estimate the residual error of the treatments/genotypes being tested without allocating checks. I can assist in setting up this feature, but it's crucial for users to add checks; otherwise, no model will work to analyze the data, or the estimates obtained will be meaningless.

from fieldhub.

Dear Didier,

sorry for the late reply and thank you for your time. I know this might sound confusing at first, I'll try to explain better. Our Status quo is a 3x12 design with 4 checks and no replicates at each location, but with 3-4 locations. Our breeders like this because they can compare entries to checks for each experiment and do not spend any plots for "unnecessary" replicates within locations. THe fields, or better said, the whole breeding program contains lots of these 3x12 units. When trying to analyse the whole program (many 3x12 units with checks but no replicates on one location, but multiple locations) we are running into the problem that we cannot estimate errors within experiments (3x12 units) because there are not replicates. Furthermoren checks are placed always on the same spot in 3x12 units on each location, so for some rows, we cannot even estimate row effects, leading to convergence problems with the mixed models. Our team is very reluctant to change since this system is very efficient for planting, but we want to be able to generate well adjusted BLUEs for downstream analysis like GS, therefore we are working our way towards designs that can be analysed better.

The first design I descirbed above is the status quo.
One idea to improve would a p-rep where we split the entries into four quarters, having one location with two reps, two locations with one rep and one location where this quarter is not present, rotation through the quarters and locations in this fashion would give us a p-rep with the same number of plots, but one quarter of the entries replicated at each location. That was what my second example referred to.
Another idea that has come is to systematically distribute the checks a bit better, generating a few different 3x12 experiment arrangement that are stacked together to avoid rows and columns that are not populated by checks.

Do you have any idea how this could be simulated in FieldHub?
Best Gregor

from fieldhub.