Compositional Data Warning

How Licklider detects compositional data structure and what confirmation is required before compositional results can be used as claims.

Compositional data is data where multiple variables represent parts of a whole - for example, the relative abundance of species in a sample, the proportions of different cell types in a tissue, or the percentage breakdown of a portfolio across categories. By definition, the values in each row sum to a constant (typically 1 or 100%).

Standard statistical methods assume independence between variables. In compositional data, this assumption is violated: increasing one component necessarily decreases the others. Applying standard regression or comparison methods to compositional data can produce misleading results.

How Licklider detects compositional structure

Licklider looks for the following signals:

  • Multiple columns whose values are bounded between 0 and 1 (or 0 and 100 for percentages)
  • Rows where the values across those columns sum to a constant
  • Column names suggesting compositional variables (for example: fraction, proportion, abundance, pct, share, composition)
  • Request text that suggests compositional context (for example: "composition", "relative abundance", "breakdown")

When these signals are present together, the data is flagged as potentially compositional. The strength of the detection depends on how many signals are present.

What you are asked to confirm

When compositional structure is detected, Licklider presents the following options:

Acknowledge and continue with disclosure

You confirm that the data is compositional, accept the associated limitations, and will include a disclosure in the output. This path is appropriate when compositional methods are not available for the current analysis but the result is still informative with the limitation noted.

Not compositional - proceed normally

The variables are not compositional. They happen to be bounded or sum to a constant for other reasons, and standard analysis is appropriate. Selecting this resolves the check.

Descriptive only

The compositional structure is present and the analysis cannot be treated as inferential under these constraints. The result is treated as descriptive only.

Severity and blocking behavior

When the detection confidence is high, an unresolved compositional warning blocks claim-bearing export. The Inspector will indicate that confirmation is required.

When the detection confidence is low - for example, when only one or two signals are present - the warning appears as a note in the figure's disclosure rather than as a blocking requirement.

What this page does not cover