One unusually high leakage reading or low resistance value can change a statistical summary. It may be a transcription error, a fixture problem, or the most important board in the sample. Investigate its origin before removing it from the analysis.
NIST’s outlier discussion distinguishes erroneous data from unusual but potentially meaningful observations and cautions against simply deleting outlying values. A statistical flag identifies a point for review; it does not establish the physical cause.
Trace the observation back to the unit
Retrieve the raw instrument response, unit identity, test attempt, recipe, station, and time. Confirm units and any calculations between acquisition and report generation. Check for a decimal conversion, missing field, duplicated row, or incorrectly linked serial number. Compare the stored reading with the original instrument evidence before assuming that the physical board produced the reported value.
Review the physical test conditions and any operator intervention. Preserve the board’s condition before rework or destructive examination. If the unit is retested, retain the original result and document what changed between attempts.
Separate possible dispositions
| Finding | Data treatment |
|---|---|
| Confirmed transcription or conversion error | Correct through a traceable derived record |
| Measurement invalid under the defined method | Record exclusion reason and retain original evidence |
| Valid unusual product observation | Keep it in the relevant population analysis |
| Cause unresolved | State uncertainty and examine sensitivity of conclusions |
Do not decide exclusion solely because the point prevents a desired capability or yield target. Define the analysis rules before comparing alternatives, and apply them consistently to both favorable and unfavorable observations.
Look for a broader pattern
- Inspect neighboring observations in time and production sequence.
- Compare station, fixture position, lot, and variant.
- Check whether the apparent outlier belongs to a distinct population.
- Investigate a plausible physical mechanism.
- Repeat only the checks needed to resolve that mechanism.
A mixed population can produce several apparently unusual points without any data-entry error. Conversely, a single corrupted value can distort an otherwise coherent group. Use the engineering context to choose the next investigation, then use an appropriate statistical method to summarize the evidence.
When reporting an analysis with exclusions, show how many observations were excluded, why, and where their original records are retained. Include the effect on the conclusion when that sensitivity matters. Avoid presenting cleaned data as though they were the complete original sample.
The retest-history guide explains why a later passing measurement must remain linked to the unusual first result rather than replacing it.
Move from reading to a build-ready project
Use the fabrication and assembly checklists to prepare drawings, production files and test requirements.
Open the DFM checklist