Outliers and repeat testing

Why it matters

Sooner or later a result looks wrong: one replicate far from the others, a patient value that does not fit their history, a batch result just outside specification. The temptation is to repeat the test, get a "better" number and report that. In regulated and clinical labs this is called testing into compliance, and it is one of the most serious data integrity failures there is. Labs need a disciplined way to decide when a result may be rejected, when a repeat is justified, and which result gets reported.

Rule one: investigate before you reject

A result may be invalidated only for a documented, assignable cause: something you can show went wrong. Examples:

• A recorded instrument error or failed system suitability.
• A documented pipetting or dilution mistake noticed at the time.
• A sample problem: clotted, haemolysed, wrong container, short volume.
• A transcription error, proven by the raw data.

"It looks too high" is not an assignable cause. Without one, the original result stands and goes into the investigation.

Statistical outlier tests

Statistics can tell you that a value is unusual compared with the others. It cannot tell you why. The Grubbs test checks whether the most extreme value in a small, normally distributed data set is an outlier:

G = max |x - mean| / s

Compare G with a critical value for your n and confidence level. Commonly quoted two-sided 95 percent critical values:

Worked example. Six replicate results: 10.1, 10.3, 10.2, 10.0, 10.2, 11.0.

• Mean = 10.30, sample standard deviation s = 0.358.
• The most extreme value is 11.0: |11.0 - 10.30| = 0.70.
• G = 0.70 / 0.358 = 1.96, which is above 1.887.

So 11.0 is a statistical outlier at 95 percent. That is a reason to investigate, not by itself a reason to delete it. Limits on outlier tests:

• They assume normal data and usually test one outlier at a time.
• Small n gives them little power.
• FDA's OOS guidance says outlier tests are rarely appropriate for validated chemical assays and cannot by themselves invalidate a result in a GMP OOS investigation (they can have a role in highly variable biological assays). Follow your SOP.

Repeat testing done right

• Retest: re-analyse the same sample (or preparation). Resample: take a new sample from the batch. Resampling needs stronger justification in GMP, because it can hide a real batch problem.
• The number of retests, and how results are combined or reported, must be defined in the SOP before testing, not decided after you see the numbers.
• Keep all results, including the original, in the record.

Replicate acceptance

Many methods run samples in duplicate or triplicate and set an acceptance limit on agreement, for example "the range of replicates must not exceed 0.5 units" or "the CV must be 5 percent or less". If replicates fail, the SOP says what happens: usually the sample is repeated, and the failure is recorded. This is a predefined rule, so it is not testing into compliance.

Clinical checks

Clinical labs use extra checks that trigger review or repeats:

• Delta check: compares a patient's result with their previous result. A large change may mean a sample mix-up.
• Critical values: results that suggest immediate danger are verified and phoned to a clinician.
• Above the measuring range: the sample is diluted and rerun, and the dilution factor applied.
• Interferences: haemolysis, lipaemia or icterus indices may flag results as unreliable.

Common mistakes

• Repeating a test until the result passes and reporting only the passing one.
• Deleting a result because a Grubbs test flagged it, with no assignable cause.
• Deciding how many retests to do after seeing the first result.
• Averaging an OOS result with passing retests to "average into compliance".
• Not recording the original result.

Quick recap

• Reject a result only for a documented, assignable cause.
• Outlier tests (Grubbs: G = max |x - mean| / s) flag unusual values; they do not explain them.
• Retest rules, replicate limits and reporting rules are set in the SOP in advance; keep every result.
• Clinical labs add delta checks, critical values and dilution of over-range samples.

Try it

Open the Do tab. A SQL query groups replicate results by sample; you add the HAVING clause that finds samples whose replicates disagree by more than the acceptance limit and need repeat testing.