fcr-experimental-design

Experimental Design (fcr-experimental-design)

FCR is demanding about field-experimental rigour. The design must credibly connect the agronomic

question to evidence that generalises across environments. The single most important FCR-specific rule:

field experiments should, unless exceptional circumstances apply, span **at least two seasons and/or

multiple locations/environments**. Design for that from the start.

When to trigger

• Planning a multi-environment trial (MET) or a season×site×treatment layout
• A reviewer questioned randomization, replication, blocking, or G×E inference
• Deciding how to characterise environments (soil, weather, phenology)
• Designing a crop-modelling study (calibration/validation, scenario design)

Field-experiment design essentials

• Multi-environment by design. Plan ≥ 2 seasons and/or multiple sites; define what an

"environment" is (site×year, managed water/N regime). State why the set spans the target population

of environments.

• Randomization & replication. Use a proper randomized design (RCBD, **resolvable

incomplete-block / alpha-lattice**, split-plot for factor hierarchies, strip-plot, augmented for many

genotypes). State replication per environment and the randomization procedure — not "plots were

arranged."

• Blocking & spatial control. Block against known field gradients; plan for spatial analysis

(e.g., row/column or P-spline) where fields are large or heterogeneous.

• Plot management detail. Plot size, borders/guard rows, sowing density, dates, and the management

applied — enough to reproduce the experiment.

• Environment characterisation. Record soil and weather and present weather **in relation to

crop phenology**; this is what lets readers interpret G×E.

Genotype/treatment × environment (G×E)

• Decide up front how G×E will be modelled: factorial structure, which effects are **fixed vs.

random**, and how environments enter (random sample vs. fixed targets).

• Plan stability/adaptation analyses (Finlay–Wilkinson regression, AMMI, GGE biplot) where

ranking across environments is the question.

Crop-modelling design

• State the model and version, the cultivar coefficients, and the **calibration vs.

validation** split (independent data, not the same trials).

• Justify the scenario/factor design and the environments simulated; report what the model adds

beyond the field data (extrapolation, yield-gap decomposition, generalisation).

The generalisation test (FCR-specific)

For your design, write one sentence: *"These environments represent ___, so the result is expected to

hold for ___ (and not for ___)."* If you cannot, the design does not yet support a general,

FCR-worthy claim — add environments or scope the claim.

Design-choice decision table (match layout to the question)

FCR referees expect the layout to follow from the agronomic question and the field's structure, with a

named design and stated randomization. Pick — and justify — before committing plots.

| Situation | Design FCR expects | Note |

|-----------|--------------------|------|

| One factor, field gradient | RCBD, blocks across the gradient | name blocks, give replication |

| Many genotypes, few reps | Resolvable incomplete block / alpha-lattice | recover inter-block information |

| Factor hierarchy (irrigation × N) | Split-plot (water = whole-plot) | report whole-plot + sub-plot error |

| Large/heterogeneous field | RCBD/lattice + spatial model (row–column, P-spline) | pre-plan the spatial term |

| Genotype ranking across environments | MET, environments a random sample | enables AMMI/GGE, stability inference |

Sizing anchors (illustrative, hedged)

No universal minimum exists, but FCR's ≥2-seasons/-environments expectation points to norms worth

calibrating against — as illustrative anchors (confirm against your own variance): MET genotype

trials often run ≥6–8 site-years before stability inference is credible; replication is commonly

3–4 blocks per environment; a response curve wants ≥4–5 levels.

Worked design vignette (illustrative)

Illustrative; the logic is the lesson. A team wants to claim a new wheat cultivar yields more under

reduced N. A weak design — 1 site, 1 season, cultivar unreplicated — cannot separate cultivar from

field position and yields no G×E information. The FCR-grade redesign: **2 seasons × 4 sites (8

environments) on a soil-N gradient, split-plot (N as whole-plot, cultivar as sub-plot), 4 blocks

per environment, 5 N levels for a response curve, and a row–column spatial term** — making the

cultivar × N × environment surface identifiable and testable across environments.

Anti-patterns

• One site, one season presented as sufficient (fails the multi-environment expectation)
• "Randomized" asserted with no design named, no replication count, no layout
• Treating a controlled-environment study as the main evidence (out of scope — see fcr-topic-selection)
• Ignoring spatial heterogeneity in large fields; pseudoreplication (sub-samples treated as reps)
• Calibrating and validating a model on the same data

Output format

【Design】RCBD / alpha-lattice / split-plot / MET / modelling
【Environments】#seasons × #sites; what they represent
【Randomization & replication】procedure + reps per environment
【G×E plan】fixed/random structure; stability analysis if relevant
【Environment characterisation】soil + weather vs. phenology recorded? [Y/N]
【Generalisation sentence】represents ___ → holds for ___
【Next】fcr-data-analysis

Supplementary resources

• [../../resources/external_tools.md](../../resources/external_tools.md) — design packages (agricolae, FielDHub) and crop models (APSIM, DSSAT, STICS)
• [../../resources/official-source-map.md](../../resources/official-source-map.md) — the ≥2-seasons/-environments rule and reproducibility expectations