A split-plot design is an experimental setup where two factors are tested at different scales, with one factor applied to larger units (whole plots) and a second factor applied to smaller units (subplots) nested inside them. This two-layer structure creates two separate levels of randomization and two separate error terms for statistical analysis, which is what makes split-plot designs distinct from simpler factorial experiments.
The design originated in agricultural research but is now used across manufacturing, clinical studies, and any field where changing the level of one factor is harder or more expensive than changing another.
How Whole Plots and Subplots Work
The core idea is straightforward. You have two factors you want to test, but you can’t (or don’t want to) randomize both of them at the same small scale. So you randomize them in two stages.
In the first stage, you divide your experimental space into large sections called whole plots, and you randomly assign levels of one factor (Factor A) to those large sections. In the second stage, you divide each whole plot into smaller sections called subplots, and you randomly assign levels of a second factor (Factor B) to those smaller sections within each whole plot.
A classic agricultural example makes this concrete. Suppose you’re testing three crop varieties and four fertilizer rates. The varieties need large plots because you’re planting with machinery that can’t easily switch between seeds on a small scale. So you assign each variety to a whole plot. Then you split each whole plot into four subplots and randomly assign the four fertilizer rates within each one. Every whole plot contains all four fertilizer levels, but only one crop variety.
Why Not Just Use a Standard Factorial Design?
In a fully randomized factorial design, every combination of factors would be assigned completely at random to individual experimental units. That sounds ideal, but it’s often impractical. Some factors are inherently harder to change than others.
In industrial settings, for instance, a factory might need to keep oven temperature constant for an entire day’s production runs while varying ingredient mixtures from run to run. Changing the oven temperature between every single run would be costly and time-consuming. A split-plot design handles this naturally: the day becomes a whole plot (with one temperature), and each individual run becomes a subplot (with a different ingredient mixture).
A paper manufacturing study illustrates the same logic. Researchers testing three pulp preparation methods and four cooking temperatures would first produce a batch of pulp using one method, then divide that batch into four samples and cook each at a different temperature. The pulp method is the whole-plot factor because once you’ve prepared a batch, you can’t re-randomize it at the individual sample level.
The Two Error Terms
The most important statistical consequence of a split-plot design is that it produces two separate error terms, not one. This is where many researchers make mistakes, and it’s the feature that most distinguishes split-plot analysis from a standard two-factor analysis.
The whole-plot error measures the variability among whole plots that received the same treatment. It’s used to test the significance of Factor A (the whole-plot factor). The subplot error measures the variability among subplots within whole plots, and it’s used to test Factor B (the subplot factor) and the interaction between A and B.
This matters because the subplot factor is generally tested with greater precision than the whole-plot factor. There are more subplots than whole plots, so the subplot error term is based on more observations and tends to be smaller. In practical terms, if you care more about detecting differences in one factor than the other, you should assign the higher-priority factor to the subplot level.
How the ANOVA Table Looks
A split-plot ANOVA table has a distinctive layered structure. Instead of one pooled error row at the bottom, it contains two error rows that partition the total variability into whole-plot and subplot layers.
- Factor A (whole-plot factor): tested against the whole-plot error
- Whole-plot error: the variability among whole plots receiving the same treatment
- Factor B (subplot factor): tested against the subplot error
- A × B interaction: also tested against the subplot error
- Subplot error: the residual variability at the subplot level
If you have “a” levels of the whole-plot factor, “N” total whole plots, and “t” levels of the subplot factor, the whole-plot error has N minus a degrees of freedom, while the subplot error has (N minus a) times (t minus 1) degrees of freedom. The subplot error almost always has more degrees of freedom, which is part of why subplot-level comparisons are more powerful.
Setting Up Randomization
Randomization always happens in two stages, and the arrangement of the whole plots can follow different standard designs.
If whole plots are arranged as a completely randomized design, you simply assign the levels of Factor A at random across all available whole plots, with no blocking. Then within each whole plot, you randomly assign levels of Factor B to the subplots.
If whole plots are arranged as a randomized complete block design (which is more common in field experiments), you first group your whole plots into blocks, then randomly assign levels of Factor A within each block. Then within each whole plot, you randomly assign levels of Factor B to the subplots. This adds a blocking structure on top of the split-plot structure, which helps control for environmental variability across the field or facility.
The key point is that both stages require their own independent randomization. You don’t just arrange things systematically. Each level of Factor A is randomly assigned to whole plots, and each level of Factor B is independently randomly assigned within every whole plot.
Common Analysis Mistakes
The most frequent error in split-plot analysis is ignoring the nested structure and analyzing the data as if it came from a simple factorial experiment. When you do this, you use a single pooled error term for all F-tests, which is wrong. Specifically, it makes the test for the whole-plot factor too liberal (you’ll find “significant” effects that aren’t real) because you’re using the smaller subplot error where you should be using the larger whole-plot error.
A related mistake is using the wrong error term for contrasts. If you’re comparing specific levels of Factor A, you need to use the whole-plot error. If you’re comparing levels of Factor B or looking at interaction contrasts, you use the subplot error. Mixing these up inflates or deflates your test statistics.
The nested blocking structure (subplots nested within whole plots, which may be nested within blocks) is the defining feature of the design, and the analysis must respect it. Most modern statistical software can handle split-plot models correctly, but you have to specify the model structure explicitly rather than relying on default settings for factorial experiments.
Split-Plot vs. Strip-Plot Designs
A close relative of the split-plot is the strip-plot (also called a split-block design). In a split-plot, Factor B is nested within the levels of Factor A: each whole plot is subdivided and Factor B is randomized inside it. In a strip-plot, both factors are applied in strips that cross each other. Factor A is applied in rows and Factor B in columns, so their intersection forms the experimental unit for the interaction.
Strip-plot designs create three error terms instead of two: one for rows (Factor A), one for columns (Factor B), and one for row-by-column intersections (the interaction). They’re used when both factors are hard to apply at a small scale, not just one of them. In agriculture, this happens when both planting method and irrigation scheme require large contiguous areas.
When a Split-Plot Design Makes Sense
You should consider a split-plot design when one factor is genuinely harder, slower, or more expensive to change than the other. If both factors are equally easy to randomize at the individual unit level, a standard factorial design is simpler and gives you equal precision for both factors.
Split-plot designs also arise naturally in repeated-measures studies, where the same subjects are measured at multiple time points. The subject is the whole plot (assigned to a treatment group), and the time points are the subplots. This framing makes it clear why you need two error terms: between-subject variability (whole-plot error) is different from within-subject variability over time (subplot error).
The tradeoff is straightforward. You get better precision for the subplot factor and worse precision for the whole-plot factor, compared to a fully randomized design with the same total number of observations. If you’re primarily interested in the subplot factor or the interaction, that tradeoff works in your favor.