Progeny trials (or progeny testing or genetic tests or whatever you call them) are a real money pit. They are super useful, with many functions(*) but they are expensive as hell. Their establishment, maintenance and assessment are a constant money sink.

Progeny trials follow an experimental design, through which we try to isolate signal from environmental noise. They also follow a mating design that we keep track of via the pedigree (either through a list of crosses or using marker information). Putting those two designs together starts producing a more complex analysis, which becomes even more complicated as we also include multiple environments, multiple traits, etc.

So, Why am I complicating the analysis? Because I want to squeeze as much value of those bloody expensive trials. Over 25 years ago, 1997 to be precise, I read a very cool article: “Accounting for Natural and Extraneous Variation in the Analysis of Field Experiments” by Arthur Gilmour, Brian Cullis and Arūnas Verbyla (available for free here). It is a beautiful example of model building AND value extraction from a single trial. What is the point of leaving money on the table (or genetic signal in the trial)? None.

These days there are multiple options for statistical software for running those “complex” analyses. I use ASReml-R, you may use something else. There are diminishing returns, there are simplifications that are a good idea, but please, keep on polishing those analyses.

(*) That’s another post, of course.

I was having a chat with analysts that just had a project dumped on their lap. They were questioning previous analyses as complex and were thinking of doing something much simpler and effective to make some progress on the project. There had been many delays due to personnel issues, so it was the time to move faster. They were chatting with me/asking my opinion because I was familiar with one of the data sets.

I was struggling with some of the assumptions they were making until I asked “Have you visited the trials? At least one of them…”. The answer was “no, we haven’t”.

It may sound like a silly question, as an analysis of an experiment is just applying a recipe, isn’t it? Nevertheless, this is not your typical textbook analysis, where everything is nice and square and tidy. These are forestry experiments, growing on the crappy end of land use classes, on hilly terrain, where different parts of the trial have different aspect, slope, fertility, drainage, etc… That’s why we care about the experimental design and anything else that helps us reduce environmental noise (x/y positions, spatial modelling of residuals, etc).

Becoming familiar with the site, the material being tested and the history of the experiments is not optional.