Replicates and Independent Samples

Introduction

“We measured six replicates” can describe six independent reactions or six injections of one reaction mixture. Those designs answer different questions. Repeated readings reveal instrument variation; independent preparations reveal variation in synthesis, handling and reaction execution. If dependent readings are counted as independent evidence, uncertainty looks artificially small. Before calculating a standard error or claiming reproducibility, identify what was independently varied or assigned.

Core explanation

The experimental unit is the smallest unit independently assigned to a condition. If a catalyst treatment is assigned to each reaction vessel, vessels are units. Three aliquots drawn from one vessel after reaction share the same chemistry and are subsamples, not three separate catalyst tests. If a coating process is applied to one substrate and ten spots are measured, the substrate may be the unit for the coating comparison, while spots characterize spatial uniformity within it. The unit depends on how the treatment was applied, not simply on how many numbers appear in a file.

Technical repeats include repeated instrument injections, spectral scans or aliquot measurements of one preparation. They help estimate readout repeatability, locate outliers and improve an estimate for that preparation. Independent replicates require new relevant experimental instances: separate syntheses, reaction vessels, electrode coatings or field samples, depending on the claim. Independent does not mean perfectly unrelated—shared reagent lots or apparatus can still create batch effects—but it means the treatment and key preparation steps are repeated separately.

Pseudoreplication occurs when correlated measurements are analyzed as if they were independent. Suppose one catalyst batch is split into ten vials and each is measured. Those vials can test assay repeatability for that batch, but not batch-to-batch synthesis reliability. Treating 100 spectral pixels from one film as 100 independently manufactured films likewise shrinks the calculated standard error without adding evidence that the process reproduces. The remedy is to replicate at the level of the claim and keep the hierarchy in analysis.

Nested designs are common. A study might synthesize three catalyst batches, run two reactions per batch and inject each reaction three times. The data have batch, reaction and injection variation. Report and analyze these levels separately. Averaging injections within each reaction and reactions within each batch may produce three independent batch values for a batch-level comparison. A hierarchical statistical model can use the full structure without pretending all 18 injections are independent. The NIST experimental-statistics handbook identifies replication and randomization as tools for estimating experimental error and reducing bias.

Sampling also matters. Three bottles filled from one mixed tank represent the tank's composition at that time, not three independent river locations. Three independently collected river samples can assess spatial variation if the locations were chosen appropriately. Likewise, repeated battery cycles on one cell form a time series, not independent cells. Multiple cells from separate manufacturing batches are needed for generalizing across production. Repeated cycles are still valuable for degradation trajectories; they answer a different question.

Plan the number of independent units based on expected variation and the effect size worth detecting. Adding more technical repeats cannot substitute for independent units once instrument noise is already small. Balance treatment groups across synthesis batches and instrument days so a batch effect cannot masquerade as treatment. State explicitly which values are shown in plots: individual measurements, per-unit means or group summaries.

Step-by-step reasoning

Identify the claim and treatment assignment. Draw a hierarchy of batches, preparations, specimens, aliquots and readings. Mark which level is independent for the effect being claimed. Collect enough units at that level and randomize their assignment or measurement order. Use technical repeats to estimate lower-level variation, then summarize or model the hierarchy. Report both independent sample count and technical repeat count; do not replace one with the other.

Visual explanation

Draw a branching tree: three separate syntheses at the top, two reaction vessels branching from each synthesis, and three injections branching from each vessel. The tree contains 18 readings, six reactions and three syntheses. If the claim is about reliable synthesis, the relevant independent count is three. If the claim concerns injection repeatability, the lower branches provide useful information. The drawing makes the source of dependence visible.

Real-world analogy

Taking ten photos of one apple gives ten views of that apple, not ten independently grown apples. Photos can reveal surface variation, but they cannot establish how variable an orchard's apples are. Chemical repeats similarly depend on whether the question concerns one sample's measurement or a process across independent preparations.

Real-world example

A group reports a new solar coating with efficiency measured at 20 locations on one film and labels n = 20 . The locations show within-film uniformity, but the coating recipe has been applied only once. To claim reproducible fabrication, make several independent films, ideally on different days or from different precursor batches, and report both between-film and within-film variation. A single film's high-performing spots should not be presented as independent successful fabrication runs.

Why?

Why does pseudoreplication make a confidence interval too narrow? Measurements from one unit share preparation and environment, so their errors are correlated. A standard-error formula that divides by the square root of all readings assumes independence and counts shared information repeatedly. It understates uncertainty about how the method would perform on a new independent unit.

Common misconception

“Every number in a data table is a replicate.” The unit of replication is set by assignment and preparation. “Averaging all readings fixes dependence” may still obscure variation among batches. “Technical repeats are useless” is also false; they estimate instrument noise and within-unit heterogeneity. They simply cannot support a broader claim than their design allows.

Worked example

Two catalyst recipes are each made in three independent batches. Each batch is injected four times, yielding 12 instrument readings per recipe. If within-batch injection standard deviation is 0.2 units but batch means vary by 3 units, the dominant uncertainty for a new batch is batch variation. Reporting n = 12 with a standard error based on all injections would be misleading. For a simple batch-level comparison, average each batch's four injections and compare the three means per recipe. The four readings still help detect a poor injection, but the inferential sample count for recipe reproducibility is three per group.

Quick check

1. A reaction is performed once and its product mixture is injected into a chromatograph five times. How many independent reaction replicates are there? Answer: One. The five injections estimate chromatographic repeatability for the same reaction mixture, not reaction-to-reaction variation.

Exam focus

Name the experimental unit before calculating a sample size or standard error. Distinguish independent synthesis, reaction and instrument-repeat levels. Explain what each level can generalize to. In a plot, label whether error bars represent within-unit measurement noise or between-unit variability. Suggest a revised design that repeats the process at the level of the intended claim.

Advanced insight

Clustered data can be analyzed with mixed-effects or hierarchical models that estimate variation at each level. Such models do not create independent units; they represent dependence more honestly. If every treatment uses a separate synthesis day, treatment and day remain confounded even with many readings. Crossed or randomized assignment across days is needed to identify the treatment effect.

Summary

Replication must match the level at which a claim is made. Technical repeats assess reading variation, while independent preparations assess process reproducibility. Nested data contain several variance components and should not be flattened into one inflated sample count. Identify the experimental unit, randomize where possible and report both independent and technical counts.

Practice questions

1. Ten spots are measured on each of four independently coated films. What counts are relevant to coating reproducibility and spatial uniformity? Answer: Four independent films assess coating reproducibility; ten spots per film assess within-film spatial variation. Forty spot readings are not forty independent coatings.

2. Three reaction vessels are prepared from the same catalyst batch. Can they show that catalyst synthesis is reproducible? Answer: No. They can assess reaction repeatability using that batch. Repeat catalyst synthesis in separate batches to assess synthesis reproducibility.

3. Why should cycles 1 through 100 on one battery not be analyzed as 100 independent battery samples? Answer: The cycles are repeated observations on the same cell and depend on its prior state. They describe one degradation trajectory; independent cells are needed for cell-to-cell variation.