Blocking and Paired Designs

A block is a group of experimental units that are more similar to one another than to units in other blocks, with respect to some nuisance variable that is not the variable of interest. In a randomized block design (RBD), each block contains one replicate of every treatment, and treatments are randomly assigned to units within each block. The key contrast with a completely randomized design (CRD) is where randomization happens: in a CRD, treatments are assigned completely at random across all units; in an RBD, randomization is restricted to occur separately inside each block. This restriction is deliberate — it is what allows the analysis to remove block effects from the comparison of treatments. A paired design is the special case where each block has exactly two units — often the two members of a matched pair (littermates, twins, adjacent plots) or, most powerfully, the same experimental unit measured under two different conditions (before/after, drug/placebo crossover, left arm/right arm). Pairing controls for every stable characteristic of the unit — its genotype, baseline physiology, location — because that unit serves as its own control. The relevant statistic is then the paired difference, $d_i = x_{i,1} - x_{i,2}$, and the analysis reduces to testing whether the mean difference $\bar{d}$ differs from zero, using a one-sample $t$-test on the differences rather than a two-sample comparison of raw values. The essential logic in both cases is variance partitioning. In any experiment, the total variability among all observations can conceptually be split into: variation due to the treatment (what we want), variation due to the blocking variable (nuisance, but predictable), and residual variation (everything else, assumed random noise). A completely randomized design lumps the blocking variation into the error term, inflating it and making real treatment effects harder to detect. Blocking subtracts the block-to-block variation out of the error term explicitly, so the denominator of the test statistic shrinks — and, all else equal, the test becomes more powerful.

Step 1: Identify the nuisance variable. Before running the experiment, the biologist identifies a source of variability among experimental units that is expected to affect the outcome but is not the variable under study — soil fertility gradient, litter of origin, day of experiment, sex, initial body mass, batch of reagent. This requires biological knowledge: a good blocking variable correlates with the outcome; a useless one does not.

Step 2: Form blocks of similar units. Units are grouped so that within-block variability in the nuisance variable is small, while between-block variability captures most of the nuisance variation. For the hillside fertilizer example, this might mean grouping fields into horizontal bands at similar elevation (similar moisture), rather than leaving moisture to vary freely within a comparison.

Step 3: Randomize treatment assignment within each block. Every block receives every treatment exactly once (in the simplest RBD), and which unit within the block gets which treatment is decided by randomization. This step is not optional — it is what justifies treating the residual variation as unbiased "noise" rather than a systematic artifact, exactly as randomization does in an unblocked design (see Randomization and Replication).

Step 4: Measure the outcome for every unit.

Step 5: Partition the variance in the analysis. For a randomized complete block design with $t$ treatments and $b$ blocks, the total sum of squares is decomposed as:

The corresponding two-way ANOVA table has treatment and block as two factors (block is typically treated as a nuisance factor, not tested for a treatment-like effect of scientific interest, though its mean square is still computed). The F-test for the treatment effect is:

Because $SS_{block}$ has been subtracted out of $SS_{error}$, $MS_{error}$ is smaller than it would be in an unblocked analysis of the same data — provided blocks really do differ in the outcome. This is the entire mechanism: blocking does not change the treatment effect estimate on average, but it shrinks the error variance used to judge whether that effect is real. For a paired design, the same logic collapses to one dimension. Instead of running a two-sample $t$-test on $x_1$ values versus $x_2$ values (which would include all the unit-to-unit variability in its pooled variance), the paired $t$-test computes:

where $\bar{d}$ is the mean of the within-pair differences and $s_d$ is the standard deviation of the differences. Because subtracting $x_{i,2}$ from $x_{i,1}$ removes any additive unit-specific effect (e.g., a mouse's baseline metabolic rate), $s_d$ reflects only the variability in the treatment effect itself across units, which is almost always far smaller than the variability in raw measurements.

The randomized block design was not a hypothetical convenience — it was engineered to solve a real, expensive problem in agricultural science. Working at Rothamsted Experimental Station in England during the 1920s, Ronald A. Fisher inherited decades of fertilizer and crop-variety trial data collected under haphazard, unrandomized field layouts. Fields at Rothamsted, like nearly all agricultural land, have systematic fertility gradients — drainage, historical manuring, soil depth all vary continuously across a field — so plots near each other tend to yield similarly regardless of treatment. Before Fisher, the standard practice was to arrange treatments in a fixed, non-random pattern (often alternating strips) across a field and simply compare average yields. This confounds treatment with position: if plot fertility declines from north to south, and treatment A always sits north of treatment B, an apparent "treatment A is better" result cannot be distinguished from "north plots happen to be better soil." There is no statistically valid way to compute a standard error for such comparisons, because the systematic spatial pattern violates the independence assumption needed for classical significance tests. Fisher's solution, published in Statistical Methods for Research Workers (1925) and elaborated in The Design of Experiments (1935), was twofold: (1) divide the field into blocks — compact regions, each small enough that soil conditions within a block are relatively uniform — and place one plot of every treatment in each block; (2) randomize the assignment of treatments to plots within each block, using an actual randomization device (originally, random number tables or drawn lots), rather than any systematic or "haphazard-by-eye" ordering. The classic demonstration ran across Rothamsted's long-term wheat and other crop plots. When yields were analyzed with the block structure explicitly modeled (two-way ANOVA with blocks and treatments), the fertility gradient's substantial contribution to yield variance was mathematically subtracted from the error term. Fisher and his successors showed, both by reanalysis of historical trials and by new randomized-block field experiments, that this could reduce the estimated error variance by a large fraction — commonly citing gains equivalent to needing only half to two-thirds as many replicate plots to detect the same treatment difference, compared with a completely randomized layout on the same heterogeneous field. Crucially, blocking also made it valid to compute a standard error and a p-value at all, since randomization within blocks guarantees that any remaining variation is due to chance, not a systematic confound with position. The broader evidentiary point, repeated across a century of agricultural, animal-breeding, and clinical trials since Fisher, is this: whenever nuisance variation is large and identifiable (soil position, litter of origin, batch, day, surgeon, subject), blocked or paired analyses of the same raw data recover statistically significant, biologically real effects that unblocked analyses of that data miss or estimate with much wider confidence intervals. This has been confirmed repeatedly by direct comparison — reanalyzing the same experimental dataset both ways and comparing the resulting standard errors and p-values — a demonstration many statistics courses replicate directly on Rothamsted's own century-long Broadbalk wheat plots.

Example 1 — A genetics-adjacent field trial (randomized complete block). A plant physiologist tests three fertilizer treatments (Control, Nitrogen-only, Nitrogen+Phosphorus) on wheat yield (kg/plot), using 4 blocks arranged along a known slope (Block 1 = wettest, Block 4 = driest). Each block contains all three treatments, randomly assigned within the block.

Notice the moisture gradient is large: block means range from 14.3 to 26.0 — nearly a two-fold difference driven purely by position, nothing to do with fertilizer. In a completely randomized analysis that ignored blocks, this 11.7 kg spread would be folded entirely into "unexplained error," potentially swamping the ~8 kg spread among treatment means. Blocking removes it. Computing (approximately, using the standard RBD ANOVA formulas):

This large block sum of squares (234.3) is subtracted out of what would otherwise be residual error, leaving a much smaller $SS_{error}$ computed from the leftover deviations (treatment effect roughly parallel across blocks, since N+P is consistently ~8–4 kg above control in every block). The resulting $F$-ratio for treatment is large, and the fertilizer effect — real and consistent across every block — is detected with high confidence, exactly because the huge moisture-driven noise never enters the denominator.

Example 2 — A paired physiological design (drug crossover). A clinical researcher tests whether a new bronchodilator increases forced expiratory volume (FEV₁, in liters) in asthma patients, compared to placebo. Because baseline lung function varies enormously between people (age, height, disease severity), the researcher uses a paired crossover design: each of 8 patients receives both placebo and drug (in randomized order, with washout between), and FEV₁ is measured on each occasion.

Raw FEV₁ values range from 1.5 L to 3.8 L across patients — a huge between-subject spread that would dominate an (incorrect) unpaired two-sample $t$-test and could easily mask a real drug effect. But the differences are far more consistent:

Standard deviation of differences, $s_d \approx 0.208$ L. The paired $t$-statistic:

With 7 degrees of freedom, $t \approx 3.91$ exceeds the critical value for $\alpha = 0.05$ (two-tailed, $t_{0.025,7} \approx 2.365$), giving a clearly significant result ($p < 0.01$). Had the analysis instead (incorrectly) treated the 16 measurements as two independent groups of 8, the enormous between-patient variability in baseline FEV₁ (1.5 to 3.8 L) would inflate the pooled standard deviation so much that the same 0.29 L average improvement might fail to reach significance — the drug's real effect would be obscured by exactly the nuisance variation that pairing eliminated.

Clinical trials routinely use paired (crossover) designs when a condition is chronic and stable enough that each patient can serve as their own control, dramatically reducing the sample size needed to detect a treatment effect — critical when patients are scarce, as in rare-disease trials. Matched-pairs case-control studies in epidemiology pair each disease case with a control of similar age, sex, and background risk, isolating the exposure of interest from these known confounders, directly building on the logic covered in Confounding Variables.

Agricultural field trials, the method's birthplace, still use randomized blocks (and their refinement, Latin squares and incomplete block designs) to compare crop varieties, fertilizers, and pesticides across fields with unavoidable soil and microclimate gradients — this is the backbone of modern plant breeding programs run by agricultural universities and seed companies worldwide. Laboratory animal science blocks by litter (controlling shared genetic and maternal effects among littermates) and by cage or batch (controlling shared environmental exposure), a design choice increasingly mandated by animal ethics boards precisely because it reduces the number of animals needed to reach a valid conclusion. Batch-effect correction in genomics (e.g., blocking RNA-sequencing samples by processing batch or sequencing run) is a direct descendant of the same logic, now essential in high-throughput biology where technical variation between processing days can otherwise be mistaken for biological signal. Understanding exactly how blocking achieves its power gain is also the necessary foundation for later distinguishing genuinely controlled, blocked experiments from Observational vs Experimental Studies where such deliberate grouping and randomization are not possible.

Ask the tutor: In my paired experiment, some subjects show carryover effects from the first treatment into the second measurement — does this invalidate the paired analysis, and how should I redesign the study?