Check Which Boards Disappear When Incomplete Test Records Are Removed

A clean analysis table may contain only boards with complete records. If incomplete records occur more often for a particular failure, station, or product state, the resulting table can present an overly favorable picture. Examine what was removed before interpreting the remaining values.

R’s complete-case documentation describes identifying rows without missing values. That operation answers a data-completeness question. It does not establish that the retained rows represent the original board population without distortion.

Account for every eligible unit

Start with the population that should have generated a record, such as the authorized test attempts or boards received by a station. Compare it with the imported dataset. Missing files, missing rows, missing fields, and intentionally unexecuted tests are different conditions.

Retain a reason code for known absence. Examples include early sequence exit, communication error, unavailable optional measurement, and an unparsed file. Do not replace all of them with one generic blank that later software silently drops.

Compare retained and excluded records

Calculate missing counts using the original denominator as well as the cleaned one. Report both board counts and attempt counts when retesting occurs. Review whether one software version exports a field under a different name; a parser mismatch can turn observed measurements into apparent missing data without any station acquisition failure. A dashboard that displays only successfully parsed attempts can hide the data loss entirely.

Choose an analysis policy explicitly

1. State which fields the analysis actually requires.
2. Avoid excluding a row for an unrelated optional field.
3. Preserve exclusion reasons and counts.
4. Evaluate how plausible missing outcomes could affect the conclusion.
5. Document limitations that the available data cannot resolve.

Do not fill an absent measurement with zero, a passing limit, or the previous board’s value simply to keep the row. Any estimation or imputation requires a justified method and must remain distinguishable from observed data. For many production questions, the missingness itself is an actionable finding.

When a logging fault is corrected, retain the period during which it occurred. A later complete dataset does not retroactively supply the missing evidence. The final report should describe the original population, the analyzed subset, and why the difference between them does or does not limit the engineering conclusion.