ENEN90028 Chap.5 Controls, Replication and Randomisation
Controls, Replication and Randomisation
A defensible intervention design needs a comparison that separates impact from background variation. Before data, control or gradient locations, independent replication, randomisation, blocking and a suitable after period each address a different threat. The chapter uses BACI logic and pseudoreplication diagnosis to show why repeated observations inside one site cannot replace independent intervention units.
What this chapter covers
- 01
Before baselines and after duration
- 02
Control impact and gradient comparisons
- 03
BACI change in differences logic
- 04
Independent treatment replication
- 05
Random selection and spatial coverage
- 06
Blocking correlation and pseudoreplication
Worked example · free
Repair a one paddock comparison
- 1Grazing differs at paddock level, so each condition has one independent unit.
- 1Treating cores as replicates inflates degrees of freedom and significance.
- 1Sample several independent paddocks per grazing condition.
- 1Randomise cores within paddocks and analyse them as nested observations.
Key terms
- Baseline
- Before-intervention information used to estimate initial condition and natural variation.
- Control site
- A comparable location where the intervention is absent.
- BACI design
- A before-after control-impact design that compares changes between conditions.
- Replication
- Independent repetition of the intervention comparison.
- Randomisation
- Chance based selection or allocation used to reduce bias.
- Blocking
- Grouping units by known heterogeneity before randomised comparison.
Controls, Replication and Randomisation FAQ
Why is BACI stronger than an after-only comparison?
BACI shows whether the impacted condition changed differently from controls while accounting for pre-existing differences and shared temporal change. It still needs independent replication and an appropriate monitoring window; the four cells alone do not guarantee a strong causal design.
Can a gradient replace a true control?
A gradient can be appropriate when no zero-impact location exists. The exposure must be measured clearly, enough independent units must cover the gradient, and alternative factors must be considered. The conclusion then concerns association or dose response unless the broader design supports stronger attribution.
What does randomisation achieve?
Random selection reduces the risk that convenient or unusual locations dominate the sample and supports an unbiased estimate within units. It does not automatically ensure broad coverage, which is why randomisation within blocks or a justified systematic design may be useful.
Assessment move
For each design, draw a table with condition by time and place independent units inside every cell. Count units before observations. Practise explaining the difference between a control site, a control measurement and a laboratory blank. When randomisation appears, state the bias it reduces and whether blocking or coverage rules are also needed.
Start revision by redrawing every scenario as a hierarchy: intervention condition at the top, independent sites or paddocks below it, then cores, quadrats, dates or repeated readings within each unit. Circle the level at which exposure changes. That circled level supplies the replicate count; everything below it improves description of a replicate but does not create another one.
Work through deliberately misleading counts, such as one impacted wetland sampled at ten points for twelve months, until you can say immediately that the treatment still has one independent impacted unit. Then translate the hierarchy into a condition-by-time table.
For a BACI question, check that before and after observations exist for both control and impact conditions, and ask whether several independent units occupy each comparison cell. Explain the target contrast in words: the change at impacted units minus the change at controls. Do not treat a visible after-period difference as sufficient evidence of impact. Next, practise repairing designs.
If selection is convenient, propose a chance-based rule. If a known slope, habitat zone or upstream-downstream position would make unrestricted randomisation uneven, define blocks before viewing outcomes and randomise within them. If no genuine zero-exposure control exists, describe a measured gradient and restrict the conclusion to the association the observational comparison can support.
For every repair, preserve identifiers for unit, block, sub-sample and time so the analysis can retain the nesting. Use three short oral drills. First, distinguish a field control location from a laboratory blank: one supports the environmental comparison, while the other detects procedural background.
Second, distinguish repeated measurement from replication: repeated visits estimate temporal behaviour within a unit, whereas independent units show whether the condition contrast generalises. Third, distinguish random selection from representative coverage: chance reduces observer selection bias, but a small random sample can still miss important zones.
Finish each practice response with a limitation sentence that names the unit of inference, the alternative explanation that remains and the resulting boundary on the claim. This makes the answer useful for quiz discrimination, planning-report justification, presentation questions and the final report.
