Blocked Full Factorial Experiment
- Introduced by
- R. A. Fisher (blocking principle); full factorial DOE tradition
- Year
- 1935 (Fisher); systematized through 20th-century DOE literature
- Type
- Experimental design
- Family
- Experimental design
What does it do?
A blocked full factorial experiment tests every combination of all factor levels while grouping experimental runs into homogeneous blocks to isolate a known nuisance variable.
This design preserves the power to detect all main effects and interactions of the factors of interest while preventing batch-to-batch, day-to-day, or machine-to-machine variability from inflating experimental error.
The intuition behind the method
Imagine testing three fertilizer types at two doses across four different fields. The fields vary in soil quality, which you cannot control but can account for. A blocked full factorial design runs all six fertilizer-dose combinations within each field (block), so field-to-field differences are separated from the treatment effects. Within each block the runs are randomized. The result: you see the full picture of how fertilizer type and dose interact without that picture being blurred by soil heterogeneity. Blocking buys precision; full factorial coverage buys completeness.
History
Blocking as a method to control nuisance variation was formalized by R. A. Fisher at Rothamsted Experimental Station and described in The Design of Experiments (1935), motivated by agricultural field trials. The full factorial principle was developed alongside analysis of variance in the same period. Integration of blocking with full factorial structure became standard DOE practice through the influential work of Box, Hunter, and Hunter in Statistics for Experimenters (1978) and Douglas Montgomery's Design and Analysis of Experiments, first published in 1976 and now in its ninth edition.
When to use it
Use a blocked full factorial experiment when you need to estimate all main effects and every interaction among a manageable set of factors AND a known nuisance variable (batch, day, operator, equipment unit) threatens to inflate error if ignored. It is appropriate when factor count is small enough (typically k = 2-4 factors at 2-3 levels) that running all combinations is feasible and the nuisance can be grouped into discrete homogeneous blocks. Do NOT use it when the total runs required exceed practical capacity — prefer a fractional factorial or response surface design instead. Do NOT block when no meaningful nuisance variable can be identified, as unnecessary blocking wastes degrees of freedom.
Limitations and considerations
- Run count equals (number of blocks) times (all factor-level combinations), which becomes very large for more than three or four factors or three levels.
- Requires a known, groupable nuisance variable; if no such variable exists, blocking provides no benefit and wastes degrees of freedom.
- Assumes block-by-treatment interactions are negligible; if treatments respond differently across blocks the additive model is misspecified.
- Not applicable when block size cannot equal one complete replicate of all treatment combinations — an incomplete block design is needed instead.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Provides unconfounded estimates of all main effects and every interaction among the factors studied.
- Blocking removes a known nuisance source, reducing experimental error and increasing statistical power without requiring additional treatment replicates.
- Yields the most complete information about factor effects within the chosen factor set — no aliasing of main effects or interactions.
- Results are directly interpretable via standard ANOVA tables and interaction plots once block effects are partitioned out.
- Supported by standard statistical software (R, SAS, Minitab, JMP) with well-established analysis routines.
Applications
- Manufacturing and process engineering: studying how temperature, pressure, and catalyst concentration affect product yield across different production batches as blocks.
- Agricultural research: evaluating all combinations of irrigation level and fertilizer type across fields with differing soil characteristics used as blocks.
- Pharmaceutical development: testing formulation factors such as pH and excipient type with production day or laboratory analyst as the blocking variable.
- Food science: assessing ingredient-combination effects on flavor profiles across sensory-panel sessions serving as blocks.
- Behavioral and educational research: investigating teaching-method combinations across school cohorts or time periods while blocking on cohort.
- Quality improvement: identifying which machine settings control defect rates when experiments must be spread across multiple work shifts or operators.
How it works
How it works
Identify all experimental factors and their levels. In a full factorial design every factor-level combination is included: for k factors each at two levels this yields 2^k treatment combinations. With more levels or factors the run count grows rapidly, so resource feasibility must be confirmed before committing to a full factorial structure.
Identify one or more nuisance sources — time batches, operators, machines, days, or raw-material lots — that introduce systematic but uncontrollable variation. These become the blocking variable. Each block should be internally homogeneous, and the number of runs per block must accommodate one complete replicate of all treatment combinations.
Within each block, randomly assign the full set of treatment combinations to experimental units. Randomization within blocks protects against unknown within-block trends; blocking itself removes known between-block differences. If the block size cannot accommodate all treatment combinations, an incomplete-block extension is needed instead.
Execute the experiment strictly according to the blocked randomization plan. Record all responses as planned and document any deviations, equipment changes, or environmental anomalies. Maintaining protocol fidelity within each block is critical to preserving the intended separation between treatment and block effects.
Fit a linear model that includes the blocking factor as an additive term alongside all factorial main effects and interactions: Y = mu + block + A + B + AB + ... + error. The block term absorbs between-block variance, reducing residual error and increasing the precision of factorial effect estimates. ANOVA tables are then examined for significant main effects and interactions.
Report estimated factorial effects and interaction plots with blocking effects removed. Assess practical and statistical significance of main effects and interactions. Verify model assumptions — normality of residuals and constant variance within blocks. Report the block structure transparently so readers understand what nuisance was controlled.
Common pitfalls
- Omitting the block term from the analysis model: analyzing a blocked design as completely randomized discards the variance reduction blocking achieved and inflates error, reducing power.
- Treating the blocking variable as a factor of scientific interest: blocks are nuisance controls, not treatment factors; misclassifying them distorts the study objectives and degrees-of-freedom allocation.
- Failing to randomize runs within blocks: without within-block randomization, time trends or operator learning effects contaminate treatment comparisons.
- Block size mismatch: if each block cannot accommodate one full replicate of all treatment combinations, standard analysis assumptions break down — use a balanced incomplete block design instead.
- Factor proliferation: adding more factors rapidly multiplies required runs; beyond four or five factors a fractional factorial or screening design is almost always the practical choice.
Frequently asked questions
What is the difference between a blocked full factorial design and a randomized complete block design (RCBD)?
An RCBD typically tests a single set of treatments (one factor) within each block. A blocked full factorial design tests all combinations of two or more factors within each block, enabling estimation of interactions. The RCBD is a special single-factor case; the blocked full factorial generalizes it to multiple factors with full interaction estimation.
Can I have more than one blocking variable?
Yes. Designs with two blocking variables (e.g., rows and columns) are called row-column or Latin square extensions. However, adding blocking factors constrains assignment further and the assumption of no block-by-block-by-treatment interaction becomes more demanding. Consult a DOE specialist when using multiple blocking variables.
What happens if my block size is smaller than the number of treatment combinations?
Standard blocked full factorial analysis requires each block to contain exactly one complete replicate of all treatment combinations. If blocks are smaller, you need an incomplete block design — such as a balanced incomplete block design (BIBD) — where not every treatment appears in every block but overall balance is maintained.
Should I include a block-by-treatment interaction in my model?
By default, no. The standard assumption is that block effects are purely additive. Including block-by-treatment interactions consumes degrees of freedom and is usually unjustified. If you genuinely suspect such interactions, consider a mixed-effects model treating blocks as random rather than fixed.
My factor count is five or more — should I still use a full factorial with blocking?
Almost certainly not. A 2^5 full factorial requires 32 runs per block, multiplied by the number of blocks. This is usually prohibitively expensive. A fractional factorial or Plackett-Burman screening design followed by a focused full factorial on the key factors is a far more practical strategy.
Related methods
Sources
Montgomery, D. C. (2017). Design and Analysis of Experiments (9th ed.). Wiley.
ISBN: 978-1119113478Copied.Fisher, R. A. (1935). The Design of Experiments. Oliver and Boyd.
Open publication referenceCopied.