GOLD STANDARD · SOIL SCIENCE · HOW X WORKS
Find the decision hidden inside the model
Balancing a soil sample on auxiliary means can still miss the shape of the landscape. Distributionally balanced designs ask the sample to resemble the full auxiliary distribution, using energy distance to see nonlinear gaps that a few matched totals can hide.
This guide connects the Gold Standard of Soil Science, Gold Standard of Education, and How X Works hub.
Chapter contents and reading routes
Choose the route closest to your present task. Every chapter remains expanded, so the article also reads naturally from start to finish.
See the design gap
Follow energy distance
Compare designs and estimates
Transfer to fieldwork
Use evidence and next routes
CHAPTER 1 OF 12 · ORIENT THE DESIGN
1. Matching means does not match a landscape
A sample can reproduce the average elevation, clay proxy or vegetation index while placing too many sites in the middle and too few at both extremes. If the soil outcome changes nonlinearly across those auxiliaries, mean balance alone can miss important response regions.
Distributionally balanced design treats the empirical auxiliary distribution of the population as the target. The goal is not a visually even scatter for its own sake; it is a probability sample whose auxiliary shape is close enough to support many smooth, potentially nonlinear outcomes.
CHAPTER 2 OF 12 · ORIENT THE DESIGN
2. Energy distance turns shape mismatch into a criterion
Energy distance compares distributions through distances among sample and population points. It can detect differences beyond selected low-order moments and works with multivariate auxiliary vectors after thoughtful scaling.
A small energy distance means the sampled auxiliary cloud resembles the frame cloud under the chosen metric. It does not mean every soil variable is balanced, because the criterion only sees auxiliaries supplied to it and the geometry imposed by preprocessing.
CHAPTER 3 OF 12 · BUILD DISTRIBUTIONAL BALANCE
3. A worked example where means agree
Imagine a frame with elevation classes coded 0, 1 and 2 in proportions 40%, 20% and 40%. Its mean class is 1. A four-site sample with classes 0, 0, 2, 2 also has mean 1 and represents both extremes.
Another four-site sample with classes 1, 1, 1, 1 has the same mean but none of the low or high terrain. If carbon or moisture responds in a U-shape, the second sample can fail badly. Distributional discrepancy exposes the difference that the mean hides.
CHAPTER 4 OF 12 · BUILD DISTRIBUTIONAL BALANCE
4. The design remains a probability design
The 2026 distributionally balanced design construction uses an optimized circular ordering of the population and then selects a random contiguous block, giving equal inclusion probabilities in its first implementation. The optimization arranges units so possible blocks resemble the population distribution.
Random selection is essential. Simply choosing the subset with the smallest observed discrepancy would be a deterministic balanced set, not automatically a design that supports design-based inference. Keep inclusion probabilities and the randomization mechanism in the record.
| Design target | What it protects | What it can miss |
|---|---|---|
| Selected auxiliary totals | Linear relationships with chosen auxiliaries | Distribution shape and nonlinear response |
| Spatial spread | Local coverage in geographic space | Non-spatial auxiliary imbalance |
| Full auxiliary distribution | Many smooth functions of auxiliaries | Outcomes unrelated to recorded auxiliaries |
CHAPTER 5 OF 12 · BUILD DISTRIBUTIONAL BALANCE
5. Scale and encode auxiliaries before optimization
Distance is sensitive to units. Elevation in metres can dominate a proportion or spectral index unless variables are scaled for their scientific role. Categorical terrain or parent-material classes also need an encoding whose distances make sense.
Inspect missingness and frame error first. An elegant energy-distance calculation on outdated maps can optimize resemblance to the wrong population. Keep raw variables, transformations, scaling constants and excluded units versioned.
CHAPTER 6 OF 12 · COMPARE AND ESTIMATE
6. Nonlinear coverage is a benefit, not a guarantee
If a target varies smoothly with the chosen auxiliaries, closer distributional fit can reduce Horvitz–Thompson variance across a broad family of functions. This is valuable in multipurpose soil inventories where no single outcome dominates.
The word smoothly matters. Abrupt contamination hotspots unrelated to the auxiliaries may still be missed. Add hazard strata, unequal probabilities or a targeted component when the survey has a dominant rare outcome.
CHAPTER 7 OF 12 · COMPARE AND ESTIMATE
7. Compare with cube, local and robust designs
The cube method balances selected auxiliary totals and is powerful when outcomes are approximately linear in them. Local pivotal and related spatial designs promote spread. Distributionally balanced designs target the complete auxiliary distribution and can also promote spatial spread when coordinates are included.
Distributional robustness is different again: it protects performance across an ambiguity set of population or model conditions. A distributionally balanced sample resembles a known frame distribution; a distributionally robust design guards against uncertainty about which distribution matters.
| Method | Primary object | Use when |
|---|---|---|
| Cube balance | Auxiliary totals | Strong linear outcome links |
| Local/spatial balance | Neighbourhood spread | Spatial trend and coverage matter |
| Distributional balance | Empirical auxiliary distribution | Multiple nonlinear outcomes share auxiliaries |
| Distributional robustness | Worst case over an ambiguity set | Frame or model conditions are uncertain |
CHAPTER 8 OF 12 · TRANSFER TO FIELDWORK
8. Changed-context transfer: soil to forest inventory
Move from soil carbon sampling to a forest inventory estimating biomass, habitat and regeneration. The stable mechanism is a probability design that matches the full distribution of remote-sensing and terrain auxiliaries.
The context changes which variables deserve distance. Canopy height, disturbance age and slope replace or supplement soil proxies. Transfer succeeds when the team rebuilds the auxiliary geometry and outcome rationale rather than simply reusing the same scaled matrix.
CHAPTER 9 OF 12 · TRANSFER TO FIELDWORK
9. Field feasibility must survive the ordering
A theoretically representative block is not useful if selected sites are inaccessible, unsafe or outside permission. Feasibility flags should enter before randomization when legitimately part of the target frame, or trigger a pre-specified substitution and weighting protocol afterwards.
Do not silently replace difficult sites with nearby convenient ones. That can erase the very extremes the design selected. Record attempted visits, reasons for nonresponse and the design-consistent adjustment.
CHAPTER 10 OF 12 · TRANSFER TO FIELDWORK
10. Audit distributional fit without tuning to the outcome
Plot or tabulate marginal and joint auxiliary coverage, energy distance, geographic spread and inclusion probabilities for the realized sample. Compare them with alternative probability samples using repeated design simulation when possible.
Do not choose the design after seeing which one produces the preferred soil estimate. That turns design selection into outcome-dependent analysis. Set the criteria before field values are known and preserve simulation code and seeds.
CHAPTER 11 OF 12 · CHECK AND CONTINUE
11. Questions soil teams often ask
Does low energy distance remove the need for weights? No. Use the inclusion probabilities defined by the sampling design. Should every auxiliary enter? No. Redundant or noisy variables can distort distance and increase computation without adding useful coverage.
Is the method always better than unequal-probability sampling? No. When one target and a strong size measure dominate, a targeted unequal-probability design may be more efficient. Distributional balance is attractive for multipurpose inventories and broad nonlinear coverage.
CHAPTER 12 OF 12 · CHECK AND CONTINUE
12. Evidence, specialist routes and the next check
The primary 2026 paper is Distributionally balanced sampling designs in Biometrics; its PubMed record summarises the equal-probability construction, energy-distance target and simulation results. Compare the cube method and the energy-statistics foundation at Székely and Rizzo.
Continue through cube selection and calibration checks, design effects and effective sample size, outcome-specific balance diagnostics, and distributionally robust balance. The next field check is to ask which nonlinear outcome would fail if the sample matched means but missed the auxiliary shape.
Contents · Previous chapter · Return to the Soil Science root