N Disclosure and Attrition Trail
How Licklider tracks sample size from input to analysis, where attrition is recorded, and what the disclosure requirements are for figures with reduced N.
The number of observations that enter an analysis is not always the same as the number in the original dataset. Rows may be dropped due to missing values, outlier exclusions, or preprocessing steps. Licklider tracks this reduction from input to analysis and records it as part of every figure's audit trail.
What gets tracked
For each figure, Licklider records four core counts:
| Measure | Description |
|---|---|
| Input N | The number of rows in the dataset before any analysis-time filtering |
| Analysis N | The number of rows actually used in the analysis |
| Dropped rows | The difference between input N and analysis N |
| Dropped fraction | Dropped rows as a proportion of input N |
When a group column is present, these counts are also broken down by group. When a timepoint column is present, the N at each time point is recorded separately.
These counts are shown separately because a single final N is often not enough to interpret what happened. Input N, Analysis N, and Dropped fraction answer different questions: how much data started the process, how much was actually analyzed, and how much was lost on the way. Group and timepoint breakdowns make it easier to see whether attrition is concentrated in one condition or accumulates over time.
Where to find it
Figure Inspector — Audit Trail tab Select a figure and open the Audit Trail tab in the Inspector. The Sample attrition section shows the N counts recorded at the time the figure was generated.
The attrition information is captured from the figure's generation metadata and reflects the state at figure generation time, not the current dataset state.
Severity levels
Licklider evaluates the severity of sample attrition and shows it alongside the N counts:
| Severity | Meaning |
|---|---|
| No warning | Attrition is within acceptable bounds |
| Low | Some rows were dropped; the figure is still valid |
| Medium | A notable fraction of rows were dropped |
| High | A large fraction of rows were dropped |
The severity label does not block the figure — it informs your judgment about whether the analysis N is sufficient for the conclusions drawn.
Licklider cannot determine automatically whether attrition is random, whether the remaining sample is still representative, or whether a given severity label implies acceptable bias for your scientific question. A figure with "No warning" can still be biased if the missing observations are systematically different, and a figure with "High" attrition is not automatically invalid if the loss is well understood and appropriately disclosed.
Group-level attrition
When group information is available, the attrition section shows the N for each group. If the number of dropped rows is substantially different across groups — for example, if one group lost a large proportion of observations while others did not — this asymmetry is recorded.
Differential attrition between groups is a potential source of bias: if the observations that were dropped are not missing at random, the remaining data may not represent the intended population equally across conditions.
Timepoint-level N
When a timepoint column is present, the attrition section shows the N at each time point. A reduction in N across time points indicates that some observations were lost over the course of the study.
This information is especially relevant for repeated measures and longitudinal analyses, where the N at each time point affects the interpretation of trends.
Disclosure requirements
When attrition is detected, Licklider requires that it is acknowledged before the figure can be used in a claim-bearing export. The disclosure must confirm that the sample size reduction has been considered and reported.
The disclosure status is visible in the Inspector. If attrition is present but the disclosure is unresolved, the Inspector will indicate that confirmation is required.
LOD exclusions When observations are excluded because their values fall below the limit of detection, this is tracked separately from general sample attrition. LOD exclusions have their own disclosure requirement and appear in a separate section of the Inspector.
Design Rationale
This page follows a simple rule: any reduction in sample size that affects a figure should remain visible at the level of that figure, not buried in an earlier preprocessing step. That is why Licklider records input-to-analysis attrition at generation time and keeps the counts attached to the figure's audit trail.
The count breakdown is also intentional. A single summary N can hide whether rows were lost evenly, concentrated in one group, or accumulated at later time points. Showing input N, analysis N, dropped rows, and dropped fraction separately makes the path from dataset to analyzed subset easier to audit.
Severity labels are meant to support review, not replace it. They summarize how much data was lost, but they do not claim to diagnose whether the attrition is ignorable or biased. That judgment still depends on study design, missingness mechanism, and domain context.
LOD exclusions are separated from general attrition because they represent a specific measurement-limit problem rather than a generic drop in row count. Keeping them distinct helps the disclosure record show not just how many observations were lost, but why.
What this page does not cover
- How outlier exclusions contribute to attrition → see Outlier Exclusion Log
- How missing data is handled during preprocessing → see Preprocessing Audit Log
- How replicate structure affects N interpretation → see Replicate Structure
- How missing data patterns are analyzed → see Missing Data and Attrition