Known Limitations

Reference for important cases the product cannot detect or validate automatically from data and metadata alone.

Licklider can automate many useful checks, but some risks cannot be inferred reliably from the available data structure alone.

Pseudoreplication without usable subject structure

Pseudoreplication detection depends on the data exposing a subject-like or repeat-structure signal that the system can interpret.

If no usable subject identifier or repeated-observation structure is present, the product cannot reliably detect that apparently separate rows came from the same biological source.

Researcher responsibility:

  • confirm the independence structure of the design before interpreting the analysis as independent observations

Study-design facts not encoded in the table

The product can inspect the table and associated metadata. It cannot directly inspect the real-world conduct of the study.

That means it cannot verify facts such as:

  • whether randomization was actually performed correctly
  • whether blinding was actually preserved
  • whether inclusion and exclusion rules were applied consistently in practice

Causal claims

The system can support association-oriented analysis and can flag some risky language patterns, but it cannot determine causation from wording or data shape alone.

Even when a result is statistically strong, causal interpretation still depends on design, intervention structure, confounding control, and domain knowledge.

Pre-registration compliance

The product does not independently verify external registry records or guarantee that a completed analysis matches a pre-registered plan in every detail.

Any status that depends on pre-registration should be read as requiring external human verification unless explicitly documented otherwise.

Zero-cell tables and confidence interval interpretation

When a 2x2 contingency table contains one or more cells with a count of zero, Fisher's exact test can return a confidence interval lower bound of 0.0. This is a mathematically correct result under the conditional exact method used by the underlying engine (SciPy contingency.odds_ratio with kind="conditional").

Licklider does not flag this as an error or suppress the result. The output is typically:

  • odds_ratio: the computed value, which may be 0.0 or very large depending on which cell is zero
  • ci_low: 0.0
  • ci_high: a finite upper bound

What this means for interpretation: A lower bound of 0.0 indicates that the data remain compatible with no association at the lower end of the plausible range. It does not mean the true odds ratio is zero. In zero-cell settings, the raw table and the upper bound usually carry more interpretive weight than the lower bound alone.

Researcher responsibility: Zero-cell tables have limited inferential value regardless of the method used. Before interpreting a Fisher result from a zero-cell table, consider whether the experimental design produced enough observations to support a meaningful conclusion. Adding observations to the sparse cell, or treating the result as descriptive rather than confirmatory, are both reasonable responses.

Practical reading rule

If a risk depends on hidden study-design facts rather than on the structure of the available data, do not assume the product can validate it automatically.