Distinguishing Observation Units, Experimental Units, and Replicates

By the end of this topic

You will be able to distinguish the unit of a row that stores a measurement from the experimental unit to which a treatment condition was assigned. You will separate biological replicate from technical replicate and formulate questions for deciding whether repeated measurements increase the number of independent samples.

Starting point: do 12 measurements mean 12 samples?

Suppose a study prepares four independent culture units and measures each unit three times. The final table contains 12 values. Is the sample size therefore 12?

This question cannot be answered from the numbers alone. We must know the unit to which the treatment condition was assigned, whether the three measurements came from different biological units, and whether they are repeated measurements of the same unit. The number of observations is the number of records; the amount of independent information depends on the study design and assignment structure.

Separate four kinds of units

Observation unit

An observation unit is the object or measurement level at which a value is recorded once. It can be expressed as one image, one well, or a measurement at one time point. It is closely related to a row in a data table, but is not always identical to one because several measurements can be stored in one row in a wide format.

Experimental unit

An experimental unit is the smallest unit to which an intervention or condition is assigned. If a condition is assigned to an independent culture unit, that culture unit is the experimental unit. Even when the same condition is placed in several wells, do not count wells alone as independent treatment units if the actual assignment occurred at a higher level.

This is why the analysis record must state “Where was the treatment assigned?” Before the name of an analysis method, examine the level of assignment.

Biological replicate

A biological replicate is a repeat obtained from biologically independent units. What counts as an independent unit must be determined from the research question and design; it is not determined merely because IDs differ. Design records must also be checked to decide whether subsidiary measurements divided from the same source are independent biological replicates.

Technical replicate

A technical replicate is a repeated measurement from the same biological or experimental unit, made to examine variation in the measurement process. It is useful for assessing measurement stability, but it does not erase the fact that it came from the same independent unit. Do not therefore convert the number of technical replicates directly into the number of biological replicates.

Read the hierarchy in a table

The following is a long-format example in which one row records one measurement.

The IDs in the table indicate different levels. measurement_id distinguishes records, well_id distinguishes the subordinate unit where measurement occurred, and culture_id indicates a candidate higher-level independent unit. These columns alone cannot establish independence automatically. Treatment assignment and sample generation must be considered together.

For example, if the three rows in condition A all have culture_id=C01, the three observations are not three biological replicates. Conversely, if C01, C02, C03 were independently assigned to conditions, the design record can be used to examine whether each culture is an independent unit.

Where pseudoreplication arises

Pseudoreplication is the problem of treating non-independent repeats as independent samples and thereby overstating the amount of information. A typical starting point is counting several technical measurements as different biological replicates.

This is not simply a “row-count error.” It confuses the level at which sample size is counted, the level at which variation is estimated, and the unit at which a condition difference was observed. Repeated measurements from the same independent unit can show measurement variation, but they do not create new independent units.

This does not mean that technical replicates must always be discarded. Their treatment differs according to whether their purpose is to check measurement error and whether each subordinate unit has a separate intervention and an independent sample. Whether to summarise by an average, reflect hierarchy in a model, or revisit the design is decided from the research question and data structure together.

Order for deciding how to handle repeats

Keep the following sequence in the record.

1. At which unit was the condition or treatment assigned?
2. At which unit was one outcome value measured?
3. How many values came from the same higher-level unit?
4. Does the repetition represent biological variation or variation in the measurement process?
5. How will the independent unit of analysis and the sample size to report be defined?

When answering these questions, do not first count the rows of the data table. Research notes, assignment records, sample lineage, and the data schema may all be needed. If the information is unknown, do not assume independence; recording design_unknown is safer.

Check with a small schema

The repetition structure can be represented in a data contract as follows.

analysis_unit: culture_id
observation_unit: measurement_id
condition_assigned_at: culture_id
technical_repeat_key: well_id
outcome: value

This schema does not require a particular analysis method. Instead, it shows the level of ID at which independence was reviewed. If condition_assigned_at is empty, the design record must be completed before comparing treatment conditions.

Return to the research question for interpretation

When answering “Do outcomes differ between conditions A and B?”, the key is the structure of independent assignment units, not the number of rows belonging to A and B. More technical replicates can reveal the measurement stability of each unit in greater detail, but they do not by themselves add support for an independent comparison between conditions.

Moreover, the label biological replicate does not by itself guarantee all aspects of research independence. Record the unit at which samples were independently generated, assigned, and measured, then choose summaries and models that match that level. Average comparisons or significance tests are not selected yet because the analysis unit must be decided first.

Common failures and how to check them

Writing “there are 12 rows, so n=12”

Rows may be the number of observation records. Unless an analysis document defines what n counts, the number of observations and the number of independent replicates will be confused.

Automatically deleting technical replicates

Repeated measurements can be used to understand quality and variation in the measurement process. The problem is not deleting repeats; it is failing to distinguish their role from biological replicate.

Choosing a test before checking the ID level

Whether to compare group means or model a hierarchy depends on the question and design. Record the assignment unit and repetition structure first, then choose a method.

Key takeaways

• The observation unit is the level at which a value is recorded; the experimental unit is the level at which a condition is assigned.
• Biological replicate and technical replicate have different meanings for independence.
• Repeated measurements from the same independent unit increase the row count, but do not automatically increase the number of independent samples.
• Pseudoreplication occurs when data rows, independent information, and the analysis unit are confused.
• Before selecting an analysis method, record the assignment level and repetition structure in a schema.

Next topic

The next article organises variable data types, schema, and provenance. Then stat-004 summarises data with a defined analysis unit through centre, spread, and distribution.

References

• Bioconductor OSTA, experimental design: https://bioconductor.org/books/release/OSTA/pages/bkg-exp-design.html
• ASA GAISE Reports: https://www.amstat.org/education/guidelines-for-assessment-and-instruction-in-statistics-education-%28gaise%29-reports

The explanations, table, schema, and scenario in this article are independently developed educational material by BioStatPy. They do not reproduce wording, expression, or example arrangement from external sources.