Translating a Research Question into Variables

By the end of this topic

You will be able to read a research question and distinguish the value measured as an outcome, the condition to compare or explain as an exposure, the covariates to record alongside it, and the identifier that distinguishes rows. You will also be able to decide what one row in a table represents before drafting a data contract that fits the question.

Starting point: a question does not yet have column names

Consider the following question.

How do measurements obtained under different treatment conditions differ?

This sentence indicates a research direction, but it is not yet a table that can be analysed. We do not know what the “measurement” is, at which unit the treatment condition was assigned, or how many independent units exist in each condition. The first step in moving from a question to data is not choosing a statistical method; it is separating the roles contained in the sentence.

At this point, do not determine the biological effect of a particular treatment in advance. First define the range of questions that the data design permits, so that numbers calculated later are not made to say more than they can.

Decide what one row means first

Before making a table, answer the question, “What is one row?” A row might be an independent culture unit, or it might be one measurement repeatedly taken from the same unit. Both tables can be valid representations, but if the meaning of a row differs, the analysis unit differs even when the same column names are used.

For example, if three technical measurements were taken from one independent unit, the following representation is possible.

The three rows record three observations, but if the treatment condition was assigned at the culture_id level, they do not become three independent treatment units. The number of observations and the amount of independent information must be recorded separately. The distinction among biological replicate, technical replicate, and experimental unit directly affects later summaries and inference. The next topic examines this issue in detail.

Translate a question into four kinds of columns

outcome: what will be measured to answer the question?

An outcome is the value observed as the result in a research question. It may be a continuous measurement, whether an event occurred, or a count. What matters more than its name is the measurement rule. Record its unit, measurement time, permitted range, and missing-value representation together.

Writing “growth” alone is not enough. The meaning of the outcome changes depending on whether it is the cell count at a particular time, the change from a reference time, or an area measure obtained from an image. Even a column named value must not be mixed in one analysis when its measurement definition differs.

exposure: what will be compared or explained?

An exposure is a condition or explanatory variable used to compare observations. In an experiment it may be expressed as a treatment condition or dose; in an observational study it may be an exposure level or group classification. Also check whether it is categorical or numeric, whether it was determined before the study began, and whether it was calculated after measurement.

However, recording an exposure does not automatically define a causal effect on the outcome. The assignment mechanism, temporal order, potential confounding, and analysis unit must also be reviewed. This topic goes only as far as translating a question into variables; it does not make a causal conclusion.

covariate: context to record alongside the data

A covariate is a variable that may be considered when interpreting an outcome and an exposure. It can represent the context of the study design and measurement process, such as batch, measurement time, or initial state. Not every recorded value automatically becomes a covariate. A value that does not fit the question or temporal order, or that was created after looking at the outcome, can change the interpretation.

Therefore, do not assume that data improve simply because they have more columns. For each column, you should be able to explain why it was recorded, which unit it belongs to, and what role it may play in analysis.

identifier: a name that distinguishes rows and units

An identifier is a value used to distinguish observations or study units. There may be IDs at several levels, such as culture_id, batch_id, and measurement_id. IDs are important for finding duplicates and checking group structure, but they are not themselves interpreted as biological magnitudes or effects.

In particular, take care not to put a numeric ID into a model as if it were a continuous variable. If culture_id=10 does not mean a state five times larger than culture_id=5, the number is a name rather than a measurement with order and magnitude.

Write a small data contract

Once the variables are defined, first write a contract such as the following.

This table is an agreement made before analysis, not an analysis result. The contract is not complete if the measurement unit and time of outcome, the assignment level of treatment, and the existence of repeated measurements are unknown. Leave missing information as a checklist rather than filling it in by assumption.

Return to the research question for interpretation

After translating the question into variables, read the question again.

• Is the condition I want to compare represented as an exposure?
• Are the measurement definition and unit of the result recorded in the outcome?
• Are the unit of one row and the treatment-assignment unit distinguished?
• Can repeated measurements and batch information be identified?
• What question can this table answer, and what question can it not answer?

Passing this check does not establish the effect of a treatment condition. The calculation method, uncertainty, and limitations of the research design have not yet been reviewed. The achievement here is translating a question into language that can be analysed, not obtaining a conclusion in advance.

Common failures and how to check them

First, roles are not distinguished even though every column is called a “variable.” If outcome and identifier are not distinguished, numeric magnitudes can incorrectly be assigned meaning for an ID.

Second, each row is defined as one measurement and then all rows are counted as independent replicates. Because the treatment-assignment unit and measurement unit can differ, preserve the replication structure in separate columns.

Third, the question is changed retrospectively to fit the data. Record the question and variable roles before analysis; if a change is necessary, record what changed and why.

Key takeaways

• A research question can be translated into the roles of outcome, exposure, covariate, and identifier.
• Decide the meaning of one row and the treatment-assignment unit first.
• An ID is a name that distinguishes units; it does not automatically become an analysis variable or effect.
• A variable contract should include data type, unit, permitted values, and replication structure.
• Making a variable table does not establish causation or biological importance.

Next topic

The next article distinguishes observations, experimental units, biological replicate, and technical replicate within the same table. In particular, it addresses pseudoreplication: incorrectly counting repeated measurements from one independent unit as the sample size.

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