| Takeaway | Detail |
|---|---|
| Prefer ordinary kriging when the nugget effect exceeds 50% of the total sill | At a nugget-to-sill ratio above 50%, a fitted variogram allows kriging weights to model short-range variance that inverse distance weighting cannot capture. |
| Benchmark ordinary kriging against inverse distance weighting using a power parameter of 2 | Run both methods head-to-head on the same drillhole dataset to ensure a consistent baseline for comparing estimation error. |
| Validate both estimators with a fixed 20% holdout sample | Reserve 20% of the drillhole data as a holdout set and compute mean squared error for each method on the withheld points. |
| Adopt kriging only if its holdout mean squared error beats IDW's by more than 10% | Require a greater-than-10% reduction in mean squared error on the 20% holdout before replacing inverse distance weighting with ordinary kriging. |
This guide defines a decision rule for choosing between ordinary kriging and inverse distance weighting in mineral resource estimation.
It specifies the nugget-effect threshold, validation design, and error-reduction criterion required to justify adopting kriging.
Why >50% Nugget Breaks IDW
Start by quantifying how much of your spatial variability is random at very short distances. Fit an experimental variogram to your drillhole grades and model it with a GSLIB-style two-structure form: a nugget component C0 plus a first spherical structure with partial sill C1 and range a1. Compute the relative nugget effect as C0/(C0+C1). This ratio is the diagnostic that determines whether inverse distance weighting (IDW) can represent the spatial behavior in your data or whether its distance-decay logic will be overwhelmed by noise.
The mechanism that breaks IDW at high nugget values is a mismatch between what the estimator assumes and what the variogram shows. IDW with a fixed power parameter assigns weights solely as a function of separation distance: nearby samples receive much more influence than distant ones, regardless of how erratic grades are at short lags. When C0 exceeds 50% of the total sill (C0+C1), more than half of the modeled variance is discontinuous at the origin. In that regime, the distance-decay assumption behind IDW power 2 is dominated by randomness rather than spatial continuity, so proximity no longer reliably predicts similarity.
IDW’s power parameter cannot represent this short-range variance because it has no term for a nugget component. Adjusting the power changes how quickly weights fall off with distance, but it cannot separate structured continuity from micro-scale variability or measurement error. Ordinary kriging, by contrast, derives its weights from the fitted variogram model: the nugget term enters the kriging system directly, allowing the estimator to down-weight the illusion of local continuity when short-range variance is high. This mapping—from the nugget-to-sill ratio to the portion of variance that IDW cannot model—is the reason the ratio serves as a go/no-go check before choosing an estimator.
Before acting on the result, verify that the nugget is stable rather than an artifact of a single sample configuration. Re-run the variogram calculation and model fit on three random 60% subsets of your drillholes and recompute C0/(C0+C1) for each subset. If the relative nugget remains above 50% across all three subsets, treat the high-nugget condition as a property of the deposit and sampling support, not a fitting accident. If the ratio fluctuates widely, investigate data clustering, compositing, or outliers before committing to any interpolation method.
Run the Head-to-Head Correctly
Run the comparison once, on one fixed split, and treat the holdout as a locked test set. Split the drillholes 80/20 with a fixed seed, save that split as a hashed file, and estimate both IDW at power 2 and ordinary kriging with a fitted variogram and a geology-derived search ellipse using only the 80% training set. Neither the IDW power nor the variogram may be tuned against the 20% holdout; if either is adjusted after seeing holdout errors, the comparison is no longer an independent test.
The hashed split file is the reproducibility anchor. It records exactly which drillholes are training and which are holdout, so another analyst can rerun the same two estimators without recreating the split or accidentally moving a hole between sets. Keep the seed, split file, variogram fit, and search ellipse with the project record; the holdout itself should not be used for model selection, parameter tuning, or feature engineering.
| Checkpoint | Role | Win condition |
|---|---|---|
| MSE | Primary accuracy | Kriging must beat IDW by more than 10% |
| MAE | Robustness to outliers | Report; does not override the MSE rule |
| Bias | Mean shift check | Report; estimated mean minus true mean |
Evaluate both estimators on the same 20% holdout with three checkpoints: mean squared error as the primary metric, mean absolute error as a robustness check against outliers, and bias defined as estimated mean minus true mean. Report all three, but let MSE decide the winner. A kriging win counts only when its holdout MSE is more than 10% lower than IDW's MSE; MAE and bias are guardrails that flag unstable or shifted estimates rather than override that rule.
If kriging wins by the MSE margin, still inspect MAE and bias before adopting it. A lower MSE with worse MAE can indicate that a few large errors dominate the squared metric, while nonzero bias can signal a systematic shift in the estimated mean. If kriging does not clear the 10% MSE margin, retain IDW power 2 and document the head-to-head result, because the added variogram work has not demonstrated a reproducible accuracy gain on this dataset.
Options Compared: IDW vs Kriging
Choosing between inverse distance weighting (IDW) and kriging depends on the estimation goal, the spatial structure of the grades, and how the methods perform on a locked test. IDW with power 2 is a practical choice for exploratory grade contouring and quick checks because it requires no variogram modeling. Ordinary kriging becomes the stronger option for resource estimation when the nugget ratio is high and the holdout comparison confirms lower error. Simple kriging is appropriate only when the regional mean is known and stable, such as a well-established background grade supported by long-term sampling.
IDW power settings change how aggressively distance controls the estimate. Power 1 produces smoother surfaces that can blur local grade changes, while power 3 penalizes distance more heavily and increases clustering bias around dense drillholes. If your declustered histogram shifts by more than 5% between IDW power 2 and IDW power 3, treat that shift as a warning that the weighting scheme is amplifying sample clustering rather than reflecting spatial continuity.
| Method | Best use case | Key limitation | Winner rule |
|---|---|---|---|
| IDW (power 1) | Broad smoothing for early-stage mapping | Over-smooths local variability | Use only when spatial continuity is weak and kriging shows no holdout MSE advantage. |
| IDW (power 2) | Exploratory contouring and quick validation | Cannot represent short-range variance structure | Prefer for screening; do not adopt for resource estimation when nugget ratio exceeds 50% unless it outperforms kriging on holdout MSE. |
| IDW (power 3) | Local emphasis where data are evenly spaced | Amplifies clustering bias in dense drill patterns | Reject if declustered distributions shift materially or if holdout MSE worsens versus power 2. |
| Ordinary kriging | Resource estimation with modeled spatial continuity | Requires a defensible variogram model | Adopt when nugget ratio exceeds 50% and holdout MSE improves over IDW by the margin defined in the reader rule. |
| Simple kriging | Estimation with a known regional mean | Biased if the assumed mean is wrong | Use only when the mean is independently established; otherwise compare ordinary kriging and IDW on holdout MSE. |
Apply the winner rule as a decision gate, not a preference. First, confirm whether the nugget ratio exceeds 50%. Next, compare IDW and kriging on the same locked holdout using the protocol defined elsewhere. If ordinary kriging does not deliver the required holdout MSE improvement, keep IDW for the stated purpose and document the rationale. If it does, adopt kriging for resource estimation and retain the comparison outputs for review.
This approach keeps the choice evidence-based: IDW remains useful for speed and transparency, while kriging earns its added modeling effort only when the nugget ratio and holdout performance justify it.
Costs: Time, Compute, and Reviewer Risk
Budget roughly one to two days of a competent geostatistician's time for variogram fitting, anisotropy checks across two to three directional variograms, and the holdout comparison. The two to three directional variograms are not optional polish; they test whether anisotropy changes the nugget interpretation. IDW at power 2 alone takes under two hours. The decision is therefore not about compute; it is whether the added day is justified by the nugget condition.
The >50% nugget trigger is the point where that added day is justified. Below it, IDW's simpler rationale may be enough; at or above it, the short-range variance is large enough that a fitted variogram and cross-validation become the defensible path.
| Path | Effort | Reviewer exposure |
|---|---|---|
| IDW power 2 | Under 2 hours | Higher when nugget exceeds half the sill |
| Ordinary kriging with fitted variogram | 1-2 days | Lower if variogram and cross-validation are retained |
A Competent or Qualified Person will ask for the variogram and cross-validation when the nugget exceeds half the sill. Shipping IDW without that package does not remove the question; it moves it to review, where answering it costs more calendar time than producing it during estimation. The review cost is therefore a rework risk, not a line item: a comment can force a new split, a new variogram, or a revised estimate after the report is drafted.
Compute is not the constraint: experimental variograms, directional checks, and a fixed holdout comparison are routine calculations. No special hardware or long runtimes are implied by the protocol; the burden is interpretation and defensible documentation. The real cost is expert time and the documentation needed to show why the chosen method was adopted.
Treat the 10% holdout MSE improvement rule as the gate, but budget the variogram work before the gate if the nugget ratio is above 50%. That keeps the comparison fair and prevents a late switch from becoming a schedule risk. If IDW is shipped, document the rationale and the nugget ratio; if kriging is adopted, retain the variogram, anisotropy checks, and holdout results. Either way, the cost is front-loaded clarity rather than downstream defense.
What the Evidence Does Not Establish
The >50% nugget rule does not establish a universal winner. Before you trust the estimator comparison, confirm that the nugget reflects geology rather than measurement noise and that your dataset is large enough to support a stable variogram model. If either condition fails, the comparison can point to kriging for reasons that do not hold up in production.
Start with data quality. A high nugget can come from assay error, sample handling, or laboratory bias instead of true short-range grade variability. Review your QA/QC program for field duplicates, certified reference materials (standards), and blanks. If more than 5% of samples fail repeat-assay checks, pause the estimator comparison and correct the database first. Re-running the comparison on unverified assays can make kriging appear superior simply because it smooths measurement noise, not because it better represents the deposit.
Next, check sample support. With fewer than about 50 drillholes, variogram fitting becomes unstable, and the head-to-head result is difficult to defend. In that setting, inverse distance weighting with declustering and a clearly disclosed caveat is often the more defensible choice. State the drillhole count, the declustering method used, and the reason a model-based variogram was not relied on, so downstream users understand the limits of the estimate.
| Condition | When the >50% nugget rule breaks | When the rule still wins |
|---|---|---|
| Source of nugget | QA/QC failures indicate assay or handling error; the nugget is inflated by measurement noise. | Duplicates and standards pass review, and the nugget is consistent with short-range geological variability. |
| Data density | Fewer than about 50 drillholes prevent reliable variogram modeling, so the comparison cannot be interpreted with confidence. | Drillhole coverage is sufficient to fit and validate a variogram model before comparing estimators. |
| Decision rule | Adopt IDW with declustering, document the caveat, and defer a model-based choice until data quality or density improves. | Proceed with the estimator comparison and adopt kriging only when it meets the pre-specified holdout performance advantage. |
Use a simple gate before acting on the result: verify QA/QC pass rates, confirm adequate drillhole coverage, and document any caveats in the estimation memo. This keeps the >50% nugget rule tied to evidence you can defend in 2026, rather than to artifacts of noisy assays or sparse sampling.
And Review Checklist
The following worked example is an illustrative teaching construction, not a reported result. A single estimation domain contains 250 drillholes and 3,400 samples of copper grade. The fitted variogram model yields a nugget (C0) of 0.42 %Cu² and a total sill (C0 + C1) of 0.74 %Cu², giving a relative nugget of 57% (0.42 ÷ 0.74 = 0.568). Using the fixed holdout protocol defined earlier in this article, the team estimates grades at the withheld locations with both methods and records checkpoint statistics against the declustered domain mean of 0.58 %Cu.
On the holdout, ordinary kriging (OK) produces a mean squared error (MSE) of 0.081, while inverse distance weighting at power 2 (IDW2) produces an MSE of 0.097. The improvement is (0.097 − 0.081) ÷ 0.097 = 16.5%, which exceeds the 10% adoption threshold. Bias is +0.004 %Cu for OK and +0.011 %Cu for IDW2; both estimates fall within ±10% of 0.58 %Cu (±0.058 %Cu). Because OK clears the MSE margin and passes the bias check, the correct decision in this scenario is to adopt ordinary kriging.
The cost of a wrong choice runs in both directions. Adopting IDW when kriging wins overstates local block tonnage-grade precision, which can tighten mine plans around grade boundaries that the data do not support. Adopting kriging when it fails the holdout adds modeling and validation effort without a measurable accuracy gain, and it can embed a poorly fitted variogram into downstream estimates. Treat the checklist below as a release gate before you lock an estimator for a domain.
Use this copy-ready checklist to document the decision and make it auditable in 2026 resource workflows:
| Check | Pass condition | Recorded value / action |
|---|---|---|
| Domain and data | Single domain, sample count, and declustered mean documented | Record drillholes, samples, and declustered mean (%Cu) |
| Variogram fit | C0, C0 + C1, and relative nugget calculated and saved | Record C0, sill, and C0 ÷ (C0 + C1) |
| Holdout MSE | OK MSE beats IDW2 MSE by more than 10% | Record both MSE values and percent improvement |
| Bias check | Both method means within ±10% of declustered mean | Record OK bias, IDW2 bias, and pass/fail |
| Decision | Adopt OK only if MSE and bias checks pass; otherwise retain IDW2 or rework the model | Record final estimator and date of decision |
Worked Example: Run the Numbers
Illustration (not field data): On March 12, 2026, a resource geologist at North Ridge Exploration evaluates a gold drillhole dataset containing 10 holes from the 2026 winter campaign. Two holes (20% of the dataset) are set aside as the holdout: Hole H-07 with an assayed grade of 2.40 g/t Au and Hole H-09 with an assayed grade of 1.60 g/t Au. The remaining eight holes are used to generate estimates at the holdout locations using inverse distance weighting (IDW, power 2) and ordinary kriging with a fitted variogram.
| Holdout hole | Actual (g/t Au) | IDW estimate (g/t Au) | IDW squared error | Kriging estimate (g/t Au) | Kriging squared error |
|---|---|---|---|---|---|
| H-07 | 2.40 | 2.00 | (2.00 − 2.40)2 = 0.16 | 2.30 | (2.30 − 2.40)2 = 0.01 |
| H-09 | 1.60 | 2.10 | (2.10 − 1.60)2 = 0.25 | 1.70 | (1.70 − 1.60)2 = 0.01 |
Step 1 — Compute mean squared error (MSE) for each method. For IDW, sum the squared errors and divide by the number of holdout holes: (0.16 + 0.25) / 2 = 0.41 / 2 = 0.205 (g/t)2. For ordinary kriging, apply the same calculation: (0.01 + 0.01) / 2 = 0.02 / 2 = 0.010 (g/t)2.
Step 2 — Compute the relative improvement. Divide the difference by the IDW MSE: (0.205 − 0.010) / 0.205 = 0.195 / 0.205 = 0.951, or 95.1%. This arithmetic check confirms that ordinary kriging reduces holdout error by 95.1% relative to IDW in this illustration.
Step 3 — Apply the adoption rule and identify the break-even trigger. The reader rule requires kriging to beat IDW by more than 10% on the holdout MSE. The break-even threshold is therefore MSEOK < MSEIDW × 0.90. With MSEIDW = 0.205, kriging must produce an MSE below 0.1845 (g/t)2 to qualify. In this example, 0.010 is well below 0.1845, so ordinary kriging is the winner.
The same arithmetic applies to any dataset: compute each method's holdout MSE as the mean of squared errors, then divide the difference by the IDW MSE. In the illustration above, (0.205 − 0.010) ÷ 0.205 = 95.1%, far above the 10% break-even trigger of 0.1845 (g/t)², so ordinary kriging wins. Any project applying this protocol should verify its own computed improvement against that same 0.90 × MSE_IDW threshold before adopting kriging.
Decision Rules for 2026 Projects
For 2026 resource estimation projects, treat the head-to-head result as a decision gate rather than a preference. Use the relative nugget from your fitted variogram to select a starting rule, then confirm the choice with holdout performance before committing to a production model or a reserve-stage report.
Rule 1 (default IDW for low nugget): IF the relative nugget is 30% or less of the sill, THEN inverse distance weighting at power 2 with declustering is usually adequate for scoping studies. Check: even in this range, run the head-to-head before any reserve-stage report and do not promote IDW without documented holdout results.
Rule 2 (let the data decide): IF the relative nugget is between 30% and 50%, THEN run the head-to-head and let holdout mean squared error (MSE) decide the estimator. Check: do not default to ordinary kriging on principle; adopt the method that delivers the lower holdout MSE and record the margin.
Rule 3 (adopt kriging when it clearly wins): IF the relative nugget is greater than 50% AND ordinary kriging beats IDW at power 2 by more than 10% on holdout MSE with stable bias, THEN use ordinary kriging as your production estimator. Check: document the fitted variogram, the holdout comparison, and the bias check in the project report so the decision is auditable.
Rule 4 (no clear winner): IF the relative nugget is greater than 50% BUT ordinary kriging does not beat IDW at power 2 by more than 10% on holdout MSE, THEN do not switch estimators automatically. Check: review data spacing, domain boundaries, and declustering settings, rerun the comparison after corrections, and select the estimator with the lower error and acceptable bias.
Rule 5 (abort on unstable variograms): IF the experimental variogram is unstable (for example, erratic directional behavior, poor structure definition, or model parameters that change materially with small lag adjustments), THEN abort the kriging adoption decision. Check: revert to IDW for interim estimates, collect or composite additional data if possible, and repeat variogram modeling only when the structure is defensible for 2026 reporting.
What to do next
| Step | Action | Why it matters |
|---|---|---|
| 1 | Fit a variogram to your drillhole dataset and check whether the nugget-to-sill ratio exceeds the 50% threshold highlighted above. | A nugget effect above that threshold indicates short-range variance that ordinary kriging can model but inverse distance weighting cannot. |
| 2 | Run inverse distance weighting (power 2) and ordinary kriging head-to-head on the same drillhole dataset using identical search parameters. | A consistent baseline ensures that any difference in estimation error is due to the method, not the input data or configuration. |
| 3 | Reserve a fixed 20% holdout sample from the drillhole data and compute mean squared error for both estimators on the withheld points. | A fixed holdout provides an unbiased, comparable measure of how each method performs on unseen data. |
| 4 | Compare holdout mean squared error results and confirm whether ordinary kriging beats inverse distance weighting by more than the 10% improvement threshold. | Requiring a clear margin prevents adopting a more complex model for a negligible gain in accuracy. |
| 5 | Review the smoothed-grade histogram from ordinary kriging and verify whether the reserve grade falls outside the ±10% band of the declustered mean. | This check guards against excessive smoothing that can distort reserve grades even when mean squared error improves. |
| 6 | Adopt ordinary kriging only if both the mean squared error and histogram criteria are met; otherwise, ship inverse distance weighting and document the rationale. | A documented, rule-based decision keeps the mineral estimation workflow transparent, defensible, and repeatable. |
Frequently Asked Questions
What variogram model form is specified for quantifying short-range randomness in drillhole grades?
The experimental variogram should be modeled with a GSLIB-style two-structure form consisting of a nugget component C0 plus a first spatial structure.
Why can ordinary kriging account for short-range variance that inverse distance weighting misses?
A fitted variogram allows kriging weights to model short-range variance that inverse distance weighting cannot capture.
How should the dataset be handled to ensure a fair comparison between the two estimators?
Both methods must be run head-to-head on the same drillhole dataset to ensure a consistent baseline for comparing estimation error.
Which performance metric is used to evaluate the estimators on the withheld validation points?
Mean squared error is computed for each method on the withheld holdout points.
What condition must be met before replacing inverse distance weighting with ordinary kriging?
Kriging should be adopted only if its holdout mean squared error exceeds the required improvement margin over inverse distance weighting.
What is the initial step for determining whether spatial variability is dominated by random short-distance variation?
The first step is to quantify how much of the spatial variability is random at very short distances by fitting an experimental variogram to the drillhole grades.
Quick answers
| When should ordinary kriging be preferred over inverse distance weighting? | Ordinary kriging should be preferred when the nugget effect exceeds the specified nugget-to-sill threshold, because a fitted variogram allows kriging weights to model short-range variance that inverse distance weighting cannot capture. |
| How should ordinary kriging and inverse distance weighting be benchmarked? | Both methods should be benchmarked head-to-head on the same drillhole dataset using inverse distance weighting with a power parameter of 2 to ensure a consistent baseline for comparing estimation error. |
| How should the two estimators be validated? | Both estimators should be validated by reserving a fixed holdout sample of the drillhole data and computing mean squared error for each method on the withheld points. |
| What criterion must be met to adopt kriging over inverse distance weighting? | Kriging should be adopted only if its holdout mean squared error beats inverse distance weighting's by more than the specified error-reduction criterion. |
| What does this decision rule specify for choosing between ordinary kriging and inverse distance weighting? | The decision rule specifies the nugget-effect threshold, validation design, and error-reduction criterion required to justify adopting kriging in mineral resource estimation. |
Also worth reading: Mountain Pass REE: Kriging Uncertainty and 2026 Resource Model: Mountain Pass REE: Kriging Uncertainty · Resource Optimization Explained: Key Techniques and Tools: Resource Optimization Explained: Key Techniques · REE 2026: Kriging Variance vs Random Forest - 31% Gap Explained: REE 2026: Kriging Variance vs