Split-Plot vs Strip-Plot vs Split-Split-Plot: which error structure fits your field

Split-plot, strip-plot and split-split-plot all handle two or more factors in one field experiment. They are not interchangeable. What separates them is the randomisation restriction, and the randomisation restriction is what dictates which error term tests which factor. Choose the wrong layout in the analysis and you test the main-plot factor against the wrong residual, which usually understates its real precision.

The decision in one sentence

Use a split-plot when one factor must sit on large plots for practical reasons (irrigation, tillage, sowing date) and the other can be randomised within them. Use a strip-plot when both factors need large strips and you care most about their interaction. Use a split-split-plot when a third factor is nested inside the sub-plots.

Why the error terms differ

In a split-plot the main-plot factor is randomised over whole plots, so it is tested against the main-plot error (Error a). The sub-plot factor and the interaction are randomised within whole plots, so they are tested against the sub-plot error (Error b), which is smaller. That asymmetry is deliberate: the factor you can randomise more finely is estimated with more precision.

Design             Error strata    What each error tests
Split-Plot         Error a         Main-plot factor (A)
                   Error b         Sub-plot factor (B), A x B
Strip-Plot         Error a         Horizontal factor (A)
                   Error b         Vertical factor (B)
                   Error c         A x B interaction
Split-Split-Plot   Error a         Main-plot factor (A)
                   Error b         Sub-plot factor (B), A x B
                   Error c         Sub-sub factor (C) and its interactions

Worked example: irrigation and variety

Two irrigation levels (I1, I2) on whole plots, three varieties (V1, V2, V3) on sub-plots, three replications. Grain yield t/ha, in the long form StatVeda accepts:

rep,irrigation,variety,yield_t_ha
1,I1,V1,3.81
1,I1,V2,4.22
1,I1,V3,4.05
1,I2,V1,4.95
1,I2,V2,5.41
1,I2,V3,5.18
2,I1,V1,3.92
2,I1,V2,4.31
2,I1,V3,4.12
2,I2,V1,5.05
2,I2,V2,5.52
2,I2,V3,5.27
3,I1,V1,3.75
3,I1,V2,4.18
3,I1,V3,3.98
3,I2,V1,4.88
3,I2,V2,5.36
3,I2,V3,5.11

Split-plot reading

Irrigation (the whole-plot factor) is tested against Error a, which has df = (r - 1)(a - 1) = 2. Variety and the irrigation by variety interaction are tested against Error b. Because Error b is built from many more df, the sub-plot comparisons are sharper than the whole-plot comparison.

If it had been a strip-plot instead

If irrigation ran in horizontal strips and variety in vertical strips, neither factor sits on a clean whole plot. The interaction gets its own error (Error c) and is the term estimated with most precision. Strip-plot is the right call when the interaction is the scientific question and both factors are operationally large.

What changes in the output

On the same numbers, the F statistic for irrigation, its standard error of the difference, and its critical difference all change depending on which design you declare, because the denominator mean square changes. A split-plot run as if it were a factorial RBD pools all the residual into one error, which inflates the whole-plot precision and understates the sub-plot precision. The pattern, illustrative:

Term              Tested against   Typical relative SE
Irrigation (A)    Error a          larger (fewer df, whole plots)
Variety (B)       Error b          smaller
A x B             Error b          smaller
(strip-plot)
A x B             Error c          smallest (interaction stratum)

How to pick before you plant

Ask which factor you cannot apply on small plots. Irrigation, deep tillage, sowing date and fumigation usually have to go on large plots, so they become the main-plot factor in a split-plot. If both factors are operationally large and the interaction is what you are after, strip-plot. If a third, smaller treatment (a spray, a foliar nutrient) is nested inside, split-split-plot.

Common mistakes

Analysing a split-plot as a two-factor RBD, which throws away the two-stratum error structure and reports a single wrong CD. Putting the operationally large factor on the sub-plot (impossible to randomise that finely). Using a split-split-plot when there is no genuine third level of nesting, which wastes df on an error stratum that carries no information.

When you are unsure between split-plot and strip-plot

If you can physically randomise one factor within the large plots of the other, it is a split-plot and the sub-plot factor gains precision. If you cannot, and both run as crossed strips, it is a strip-plot and the interaction gains precision. The deciding question is never which is more advanced, it is which randomisation the field operations actually allowed.