Choose your reading route
eduKateSG · Why Mathematics? · Split-Plot Designs: Hard-to-Change Factors, Error Strata and Mixed Models
Start with the question, audit the worked example, or jump to the assumptions and practice. Every chapter stays open for uninterrupted reading.
Understand the question
Define the quantity and follow the mechanism.Route 2 · Chapters 9–16
Reproduce the example
Audit the numbers one step at a time.Route 3 · Chapters 17–24
Check assumptions
See where design and calibration matter.Route 4 · Chapters 25–32
Explore applications
Connect the method to a concrete investigation.Route 5 · Chapters 33–42
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why is mathematics important when one factor is expensive or slow to change while another can vary quickly? A split-plot design randomises hard-to-change factors to large whole plots and easy-to-change factors within them. The resulting hierarchy has more than one experimental-unit size, so effects need the correct error stratum and usually a mixed model.
Did You Know? Eight measured items do not provide eight independent assignments of a whole-plot factor when that factor changed only four times. The number of genuine randomisations, not the number of rows in a spreadsheet, determines the evidence for that factor.
A useful first reading therefore asks three questions: which physical units received the hard factor, which smaller units received the easy factor, and how many independent randomisations occurred at each level? Once those are clear, the model, degrees of freedom and uncertainty follow the design. The worked example in this article keeps the arithmetic simple while showing why a large subplot table can still contain only a small amount of whole-plot information. This distinction is the safeguard against pseudoreplication and the reason design must be read before a p-value. It also helps students see why independent assignment, physical units and uncertainty belong to one connected explanation for real experimental work.
Full chapter index
Chapters 1–8 · Understand the question
Chapters 9–16 · Reproduce the example
Chapters 17–24 · Check assumptions
Chapters 25–32 · Explore applications
Chapters 33–42 · Study and verify
Some settings cannot be reset for every observation. An oven temperature may apply to a full batch, a school timetable to an entire class, or an agricultural irrigation level to a large field strip. Forcing complete randomisation can be costly, unsafe or impossible. Split-plot design records the operational restriction and randomises at each feasible level instead of pretending every measurement received every factor independently.
The whole plot is the larger experimental unit receiving the hard-to-change factor. Subplots are smaller units inside each whole plot receiving easy-to-change factors. These are roles, not merely areas on a diagram. A batch can be a whole plot and the items within it subplots. Identify which physical unit changes when each factor changes; that answer determines the randomisation and error structure.
A split-plot can still use a factorial treatment structure. Let A have two hard-to-change levels and B two easy levels. Every whole plot receives one A level, then its subplots receive both B levels in random order. A main effect is learned from differences among whole plots. B and A-by-B use within-whole-plot information as well as the hierarchy specified by the model.
First randomise A among whole plots within any replicate or block. Then independently randomise B among subplots inside each whole plot. If order matters, record it at both levels. This two-stage procedure generates two sources of random variation. Analysing the data as if all A labels were independently shuffled across subplots breaks the connection between the design and the standard errors.
Use four whole plots, two assigned A-low and two A-high. Each contains B-low and B-high subplots. Responses are 48 and 56 in whole plot 1, 61 and 73 in whole plot 2, 63 and 75 in whole plot 3, and 50 and 58 in whole plot 4. The table makes the hierarchy visible: paired B observations share the same whole-plot environment.
Average the two B responses within each whole plot. The means are 52, 67, 69 and 54. The two A-low whole plots average 53; the two A-high whole plots average 68. The illustrative A effect is 15 units. Its uncertainty comes from variation among whole plots receiving the same A level, not from treating the eight subplot measurements as eight independent A assignments.
| Whole plot | A | B low | B high | Whole-plot mean |
|---|---|---|---|---|
| WP1 | Low | 48 | 56 | 52 |
| WP2 | High | 61 | 73 | 67 |
| WP3 | High | 63 | 75 | 69 |
| WP4 | Low | 50 | 58 | 54 |
Within each whole plot, subtract B-low from B-high. Differences are 8, 12, 12 and 8, averaging 10. This B effect benefits from local comparison inside whole plots. The shared environment often makes it precise, though a mixed model still accounts for dependence. Reporting only the average difference without the number and diversity of whole plots would hide the real replication.
The B effect averages 8 when A is low and 12 when A is high, so the difference-in-differences is 4. The interaction asks whether the easy-factor effect changes with the hard-factor setting. It is not an error term or nuisance. Its standard error depends on the design and covariance structure. A visible interaction can make isolated main-effect summaries scientifically incomplete.
Whole-plot variability reflects differences among large units beyond their assigned A levels. Subplot variability reflects smaller-unit variation within whole plots. Classical balanced ANOVA uses separate mean squares for these strata. Modern mixed models represent the hierarchy directly, often with random whole-plot intercepts. NIST stresses that split plots have more than one experimental-unit size and therefore more than one error term.
A single ordinary regression residual pools variation from both levels and may assign too much apparent information to A. If subplot noise is small but whole plots vary strongly, the error can be severe. Conversely, a crude analysis can also waste information for B. Correct analysis follows the randomisation: each effect is compared against variation from units to which that effect could actually have been independently assigned.
| Quantity | Estimate |
|---|---|
| A main effect | 15 |
| B main effect | 10 |
| A×B interaction | 4 |
One model writes response=overall mean+A+B+A×B+random whole-plot effect+subplot error. Blocks or replicates can add further random or fixed terms according to sampling. The random whole-plot intercept induces correlation between subplots sharing a whole plot. Estimate variance components with enough independent whole plots; a complex covariance model cannot manufacture replication that the design never provided.
The NIST Engineering Statistics Handbook describes split plots as restricted-randomisation designs arising from hard-to-vary factors, batches or consecutive processing. It defines whole-plot and subplot experimental units and warns that using one error term can lead to invalid conclusions. Read the NIST guidance.
To improve evidence for A, add independent whole plots or complete replicates that randomise A again. Adding more subplots per whole plot may improve B estimates and whole-plot means, but it does not create new A randomisations. This distinction prevents pseudoreplication. Power planning should vary whole-plot count, subplot count and both variance components separately rather than using only total observations.
A replicate may contain a fresh set of whole plots receiving all A levels. Day, production lot or site can serve as a block if each contains comparable randomisation. Do not call repeated measurements within the same physical batch independent replicates. Report whether blocks are fixed conditions of interest or sampled levels used to generalise. The design diagram should show which factors cross and which units nest.
A factor is hard to change relative to the run schedule, not an immutable property. With different equipment or planning, it might change more frequently. The design should state why restriction is necessary. Convenience alone can reduce precision for A. Compare the cost of extra resets with the value of stronger whole-plot replication before committing to the hierarchy.
If all A-low whole plots run before A-high, A is confounded with time despite the split-plot label. Randomise whole-plot order within feasible blocks. Randomise subplot order within each whole plot to protect B from warm-up or fatigue. When restrictions force a fixed sequence, measure relevant drift and narrow causal claims. Restricted randomisation is still randomisation, not permission for arbitrary order.
Lost or extra subplots make the design unbalanced. Mixed models can use observed data under stated missingness assumptions, but precision and weighting change. A whole plot with many subplots should not automatically dominate an estimand meant to average whole plots equally. Define the target population and weighting. Report planned and realised counts at every level so the analysis can be audited.
Losing a whole plot is more damaging than losing one subplot because it removes an independent A assignment and all nested observations. Recalculate replication, balance and degrees of freedom. If the loss is related to treatment or outcome, missing-at-random assumptions may be implausible. Do not replace the entire plot with predictions and then count those predictions as observed replication.
Subplots are nested within whole plots because each belongs to one larger unit. Treatment factors A and B may be crossed when every A level is combined with every B level. Confusing nesting and crossing changes the model matrix and denominator. Draw a diagram with units and factor assignments before writing formulas. The physical experiment usually reveals the correct structure more clearly than software menus.
Suppose a timetable policy is assigned to schools or classes, while two worksheet prompts are assigned to students or lessons within each class. The policy is a whole-plot factor and the prompt a subplot factor. Student outcomes share teacher and class conditions. Enough independent classes are needed for policy inference. Consent, spillover, baseline attainment and appropriate school clustering remain important beyond the design arithmetic.
Irrigation may be applied to large field strips, with fertiliser varieties randomised to smaller plots inside them. Soil gradients can motivate blocks across the field. Irrigation effects use strip-level replication; fertiliser effects use subplot contrasts. Machinery paths or water flow can create interference across boundaries, so maps, buffer zones and execution records matter. A neat ANOVA cannot correct a factor applied to the wrong physical unit.
Solution concentration may require preparing a batch, while current and temperature can vary for individual items. Concentration becomes a whole-plot factor. NIST uses this type of process to illustrate how execution creates a split plot even if the original treatment list looked like a fully randomised factorial. The analysis must describe what operators actually did, including resets, batches and run order.
A model architecture may require a costly training run, while evaluation threshold or decoding settings can change cheaply within the trained model. Evaluations nested in one training run are not independent architecture replicates. Random seeds or fresh training runs can supply whole-plot replication. Dataset slices and hardware batches may create additional levels. Report the hierarchy instead of presenting every evaluation metric as an independent experiment.
Power for A depends heavily on whole-plot count and whole-plot variance. Power for B depends on within-plot contrasts, subplot variance and the number of whole plots. Interaction power needs variation across both structures. Simulate the planned mixed model with plausible variance components. A plan based on total n alone can look large yet have almost no degrees of freedom for the expensive factor.
Plot whole-plot means by A, within-plot B differences, residuals at both levels and outcomes in run order. Inspect influential whole plots because a single large unit can drive A. Check whether variance changes with settings and whether interaction remains after sensitivity analysis. Verify the random-effect structure against the assignment record rather than choosing it solely by fit statistics.
Reproduce the worked A effect of 15, B effect of 10 and interaction of 4. Fit a model with a random whole-plot intercept and compare its A standard error with naive ordinary regression. Then delete one subplot and one whole plot in separate simulations. The comparison demonstrates which losses affect which stratum and why replication cannot be counted from rows alone.
Split-plot designs trade convenience for unequal precision across effects. They do not remove confounding from nonrandom whole-plot order, solve interference, justify untested covariance structures or generalise beyond sampled whole plots. Very few whole plots make variance components unstable. Complex nesting can also make an effect non-estimable. Simplify the design when the operational hierarchy cannot be executed reliably.
Report each factor and level, which is hard to change, whole-plot and subplot experimental units, randomisation at every stage, block and replicate structure, observations per unit, missing data, model formula, variance components, effect estimates and stratum-aware uncertainty. Include a design diagram or table. Readers should see immediately how many independent assignments support each conclusion.
Use a split plot when restricted randomisation is genuinely necessary and enough whole plots can be replicated. If every factor can change freely, a completely randomised factorial may be simpler and more precise. If the hard factor cannot be randomised at all, treat the study as observational for that contrast. Design around the real assignment mechanism rather than the desired p-value.
Use one row per subplot with columns for replicate, whole-plot ID, whole-plot order, A, subplot ID, subplot order, B, response, units and missingness. Create a second table summarising whole-plot means and counts. Formulas should reproduce A, B and interaction contrasts while keeping whole-plot identifiers visible. Avoid copying down a whole-plot label without recognising its shared assignment.
Generate the two-stage randomisation, simulate whole-plot and subplot errors, fit a mixed model and compare with naive ordinary regression. Unit tests should check that each subplot belongs to exactly one whole plot, every planned B level appears within each plot and randomisation seeds are recorded. Run sensitivity analyses for imbalance, alternative covariance assumptions and influential whole plots.
Agricultural scientists, process engineers, food scientists, educators and machine-learning researchers use split plots when factor changes occur at different physical or computational scales. The key career skill is recognising experimental units and matching standard errors to assignment. Statistical software helps, but practitioners also need operational knowledge, safety, subject expertise and clear communication.
Use paper airplanes: room fan setting is hard to change and assigned to whole sessions; wing adjustment is changed for individual planes within each session. Run several independently randomised sessions at each fan setting, randomise plane order inside sessions and measure flight distance. Compare naive and mixed-model uncertainty, then explain why extra flights inside one session do not replace additional sessions.
More rows do not mean more independent whole plots. A hard-to-change factor is not automatically random. The subplot error is not the denominator for every effect. A statistically significant B effect does not validate A. Random effects do not mean “unimportant”; they represent sampled variation. Balance on paper does not excuse running all settings in a fixed time order.
The factorial-experiment guide explains main effects and interactions. The randomised-block guide explains local control of nuisance variation. This article adds the hierarchy created when factors are randomised to different unit sizes.
Students should begin with physical units: what receives A, what receives B and what is measured? Parents can ask whether a classroom study counts classes or students as independent evidence for a class-wide policy. Draw the hierarchy, mark each randomisation and calculate one within-plot contrast. That habit protects against pseudoreplication in many fields.
- Identify whole plots and subplots.
- Reproduce A=15, B=10 and A×B=4.
- Explain the two error strata.
- Add one whole-plot replicate.
- Delete one subplot and refit.
- Compare a mixed model with naive regression.
### Why split the plot? Some factors can change only at a larger unit or slower schedule.
### Which factor has less precision? Often the whole-plot factor because it has fewer independent assignments.
### Can ordinary regression work? Only if it represents the hierarchy and covariance correctly.
### Do more subplots improve A indefinitely? No; independent whole plots are the main replication for A.
Report the estimand, treatment and randomisation structures, experimental units, blocks, factor-change restrictions, whole-plot and subplot counts, run order, missingness, model, variance components, contrasts, uncertainty, diagnostics and scope. State every deviation from the planned hierarchy and preserve code that reconstructs the assignment.
A split plot is often discovered by asking how settings were physically applied. The mathematics then turns that execution history into covariance, degrees of freedom and honest uncertainty rather than erasing it.
Count independent assignments at each level, verify randomisation and run order, match each effect to the appropriate stratum, inspect whole-plot influence and reproduce the mixed-model contrasts. If one factor was never randomised, describe that limitation directly.
Trace each reported contrast through the assignment process. For A, count the whole plots that independently received each level and show the between-plot comparison. For B, show the within-plot changes and how they are combined across plots. For A-by-B, show the difference between the B contrasts at the two A levels. This trace makes the denominator and degrees of freedom understandable without relying on a software label such as “Type III test.”
Run a pseudoreplication stress test. Duplicate every subplot row without creating new physical observations and refit both naive and mixed models. A sound workflow should not treat those copies as new evidence. Then add genuine whole plots and observe how A precision improves. Add genuine subplots within existing plots and observe the different effect on B. The contrast teaches why sample size must be counted at the level of randomisation and why total rows can be a misleading headline.
Audit operational execution with timestamps, batch logs and setting-reset records. Confirm that the hard-to-change factor actually changed where the protocol says it did, that subplot settings were reset between units, and that order was not perfectly aligned with treatment. If measurement devices were recalibrated or operators changed, add those events to the design diagram. Unrecorded process changes can masquerade as whole-plot effects or interactions.
Check model sensitivity. Compare a balanced ANOVA calculation, a random-intercept mixed model and a cluster-robust sensitivity analysis when appropriate. Agreement increases confidence; disagreement deserves investigation of small-sample degrees of freedom, variance assumptions and influential plots. Do not choose the method that produces the smallest p-value. Choose the method whose assumptions best match assignment and sampling, then report alternatives as sensitivity evidence.
End with the practical decision: how much does resetting A cost, how valuable is extra whole-plot replication, and what uncertainty remains acceptable? A slightly larger experiment with fewer subplots per plot can be far more informative for A than a huge table nested inside two batches. Design resources around the effect that matters rather than around the cheapest measurements.
Continue through the Mathematics Learning Hub. Rebuild the central design, verify its identities, inspect one failure case and explain which conclusion the evidence can and cannot support. The most useful next step is a small reproducible calculation, followed by a plain-language account of the assumptions.
Extended practice laboratory
Practice case 1: Identify experimental units
Setup: Use a process with oven temperature and coating type.
Task: Name the units receiving each factor.
Checking target: Temperature belongs to ovens or batches; coating belongs to items within them.
Practice case 2: Randomise at two levels
Setup: Use four whole plots and two subplots per whole plot.
Task: Randomise temperature among whole plots and coating within each.
Checking target: Do not pretend that eight item labels are independent temperature assignments.
Practice case 3: Reproduce the effects
Setup: Use the eight responses in the worked table.
Task: Calculate A=15, B=10 and the A-by-B interaction=4.
Checking target: Separate arithmetic effect size from its correct error stratum.
Practice case 4: Use the whole-plot error
Setup: Compare whole-plot means among plots receiving the same A level.
Task: Estimate variation at the larger experimental-unit level.
Checking target: Do not use subplot residuals to test the hard-to-change factor.
Practice case 5: Use the subplot error
Setup: Compare B levels within each whole plot.
Task: Estimate the repeated within-plot contrast and its variation.
Checking target: Account for the shared whole-plot environment.
Practice case 6: Add an imbalance
Setup: Lose one subplot observation.
Task: Fit a mixed model to the observed data.
Checking target: Report the changed precision and avoid hand-balanced ANOVA formulas.
Practice case 7: Change the randomisation
Setup: Run every A-low plot before every A-high plot.
Task: Explain the new time confounding.
Checking target: Restricted randomisation still needs random order among whole plots.
Practice case 8: Plan replication
Setup: Add blocks or replicates containing both A levels.
Task: Show which replication supplies whole-plot degrees of freedom.
Checking target: More subplots alone do not repair weak whole-plot replication.
Practice case 9: Build an education study
Setup: Treat school-wide timetable as A and worksheet prompt as B.
Task: Locate the school, class and student units.
Checking target: Protect against pseudoreplication and school-level confounding.
Practice case 10: Write the conclusion
Setup: Report factor effects and interaction with stratum-aware uncertainty.
Task: State the actual randomisation and unit sizes.
Checking target: Name any drift, imbalance and model sensitivity.