Bio Experimental Design Randomization Blocking

Version Compatibility

Reference examples tested with: designit 0.5+, lme4 1.1-35+, lmerTest 3.1+, pwr 1.3+.

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('<pkg>') then ?function_name to verify parameters

If code throws an error, introspect the installed package and adapt the example to the actual API rather than retrying. designit is an R6 package whose BatchContainer$new(), optimize_design(), and *_score_generator() signatures evolve between releases; confirm against the installed vignette (vignette(package = 'designit')) before relying on argument names.

Randomization and Blocking

"Design the experiment so the statistics will be valid" -> Decide what the experimental unit is, randomize treatments to those units, replicate the unit (not the measurement), and remove known nuisance variation by blocking — so that the analysis model mirrors how the experiment was actually run.

• R: designit::optimize_design() for constrained randomization; lme4::lmer() / lmerTest for the matching mixed model
• The design and the analysis are one decision: "analyze as randomized"

The Single Most Important Modern Insight -- The Experimental Unit, Not the Measurement, Is the n

The most consequential and most violated idea in biological design: the experimental unit (EU) is the smallest entity independently assigned to a treatment — and it, not the number of measurements, is the sample size for inference. Lazic 2018 PLoS Biol 16:e2005282 separates three entities: the biological unit (what conclusions are about), the experimental unit (what randomization acts on, = the n), and the observational unit (what is measured). When observational units are counted as independent replicates, the standard error shrinks illegitimately and p-values become meaningless — pseudoreplication (Hurlbert 1984 Ecol Monogr 54:187). Ten thousand cells from three mice are n = 3, not n = 10,000, for a between-mouse question; mice co-housed in a cage dosed through the chow make the cage the EU, not the mouse. Lazic et al. found ~46% of surveyed animal studies pseudoreplicated. The fix is structural: model the design's hierarchy (random effects) or aggregate to the EU before testing — pseudoreplication is, formally, an omitted random effect.

A second, deeper point (Fisher): randomization is what licenses the p-value. It supplies the physical basis for the error term and converts systematic lurking-variable bias into random error balanced in expectation. Model-based tests are approximations to the randomization distribution. Skip randomization and the causal claim rests entirely on assumptions.

Algorithmic Taxonomy

Decision Tree by Scenario

Choosing and Counting the Experimental Unit

Goal: Identify the EU and therefore the true n before any power or analysis decision.

Approach: Trace the randomization: the EU is the smallest entity to which a treatment level was independently assigned. Anything measured below that level is an observational unit and is summarized (mean/sum) up to the EU, or modeled as a nested random effect — never counted as an independent replicate.

# Between-condition question with multiple cells per donor:
# the donor is the experimental unit, NOT the cell.
# Correct: aggregate observational units to the EU, then test on EU-level values.
library(dplyr)
eu_level <- cells |>
  group_by(donor, condition) |>
  summarise(value = mean(measurement), .groups = 'drop')   # one row per experimental unit
# n for inference = number of donors per condition, not number of cells

Randomization Mechanics

Goal: Assign treatments to units with a documented random mechanism, optionally restricted to guarantee balance on known factors.

Approach: Use a seeded pseudo-random generator (never "haphazard" order, which aliases treatment with processing position/time). For known prognostic factors, restrict the randomization (block/stratify) and then include those factors in the model. When finite-sample imbalance matters, rerandomize against a pre-specified balance criterion (Morgan & Rubin 2012 Ann Stat 40:1263) or use minimization for sequential enrollment (Pocock & Simon 1975 Biometrics 31:103).

set.seed(20260528)                      # record the seed for reproducibility
units <- data.frame(id = sprintf('S%02d', 1:24),
                    block = rep(c('day1','day2','day3'), each = 8))

# Restricted (block) randomization: randomize treatment WITHIN each block
units$treatment <- ave(units$id, units$block,
                       FUN = function(ids) sample(rep(c('ctrl','treat'),
                                                      length.out = length(ids))))
# Also randomize RUN ORDER so processing position is not confounded with treatment
units$run_order <- sample(nrow(units))

Blocking and Local Control

Goal: Remove a known nuisance source from the error term to sharpen the treatment comparison.

Approach: Group units into homogeneous blocks (day, litter, chip, donor), randomize treatments within block, and add the block as a term in the model. The paired t-test is the special case of an RCBD with block size 2. Block only on factors with real between-block variation; blocking on a noise factor spends error df for nothing.

library(designit)                        # constrained assignment; verify API vs installed vignette
bc <- BatchContainer$new(dimensions = list(block = 3, position = 8))
bc <- assign_in_order(bc, samples = units)
bc <- optimize_design(
  bc,
  scoring = osat_score_generator(batch_vars = 'block',
                                 feature_vars = c('treatment')))  # balance treatment across blocks

Split-Plot and Nested Designs -- the Genomics Trap

A split-plot has two experimental-unit sizes and therefore two error strata: a whole-plot factor that is hard to randomize finely (incubator temperature, the sequencing run/lane, the 10x chip, the staining batch) and a sub-plot factor randomized within each whole plot (the individual sample, the genotype). Analyzing a split-plot as a flat factorial uses the wrong, too-small error term for the whole-plot factor and gives anti-conservative tests for exactly the factor that was hardest to replicate. In genomics the lane/run/chip is almost always a whole plot; "batch effects" are frequently a split-plot structure to be modeled, not a nuisance to scrub.

Goal: Match the model's random-effects structure to the design's randomization structure.

Approach: Encode each randomization level as a random effect; fixed effects carry the questions. Crossed vs nested structure determines the denominator for each fixed effect; with few EUs use Satterthwaite or Kenward-Roger degrees of freedom (lmerTest / pbkrtest).

library(lme4); library(lmerTest)
# Whole plot = run (random); sub-plot factor = condition (fixed); cells nested in sample
fit <- lmer(expression ~ condition + (1 | run/sample), data = df)   # run, and sample within run
anova(fit)                               # Satterthwaite df via lmerTest