An outlier in a lab result is a value that sits far from what you expected, whether for a patient, a test run or a study population. Detecting one is the easy part. Deciding what it means is where practices, laboratories and trial teams go wrong. The value may be a real clinical signal, a specimen problem or an analytical error, and each one calls for a different response.
This guide is for physicians, practice managers, lab directors and research teams. It covers the statistical tools, the clinical thresholds regulators actually use, the checks that catch specimen errors, and the rule that matters most: decide how you will handle outliers before you see the data.
TL;DR
- Statistical rules such as the 1.5 x IQR rule or a z-score flag values that are far from the rest. Clinical thresholds, such as a creatinine rise of 1.5 times baseline, flag values that matter for safety. You need both.
- Define your outlier criteria and your handling procedure before testing or analysis begins. Post-hoc cleaning is how bias gets in.
- Never re-run a sample until it gives a pleasing number. Report every individual result and investigate the unexpected one.
- Delta checks are good at spotting physiological change, but weak at finding sample mix-ups.
- Your choice of outlier method can change reference intervals substantially, so standardize it and document it.
What an outlier means in a laboratory setting
In a general data science article, an outlier is a point that sits far from the distribution. In a medical laboratory, that definition is only a starting point. A potassium of 6.8 mmol/L is an outlier statistically, and it may also be a hemolyzed tube, a patient in danger, or both. The number alone does not tell you which.
It helps to sort outliers into three groups:
- Clinical outliers. The result is real and the patient is genuinely different from baseline. This is what you want to find quickly.
- Pre-analytical outliers. Something happened before the instrument ran: mislabeling, inadequate mixing, evaporation, a delayed transport, a contaminated tube. Research on iterative outlier removal lists exactly these non-laboratory processes as sources of extreme values.
- Analytical outliers. The instrument, reagent or calibration produced a bad number.
The reassuring part is that you do not need advanced mathematics to begin. You need a clear rule for each group and a habit of asking which group a flagged value belongs to before anyone changes or discards it.
Statistical rules for flagging unusual values
The pages that rank for this topic mostly teach the same handful of methods. They are useful, and each has a limit.
| Method | How it works | Best for | Watch out for |
|---|---|---|---|
| 1.5 x IQR rule | A value more than 1.5 times the interquartile range above the third quartile or below the first quartile is flagged | Skewed or unknown distributions, small to mid-sized sets | Will flag some legitimate values in naturally skewed analytes |
| Z-score or 3-sigma | Flags values more than 2 or 3 standard deviations from the mean | Roughly normal data | The mean and standard deviation are themselves distorted by the outliers you are hunting |
| Grubbs-type test | Tests whether the single most extreme value is an outlier in normally distributed data | One suspect value in a small sample | Assumes normality, finds one value at a time |
| Duplicate-difference limit | Flags a result when the difference between duplicate measurements is large | Method comparison and verification studies | Needs duplicate measurements built into the design |
| Delta check | Compares a patient's current result with the previous one | Routine patient testing | Better at physiological change than at sample swaps |
A worked example makes the IQR rule concrete. Suppose a clinic's ALT results in a screening group have a first quartile of 18 U/L and a third quartile of 34 U/L. The IQR is 16. The upper fence is 34 + (1.5 x 16) = 58 U/L. Anything above 58 is flagged for review. It is not deleted. It is reviewed.
The 2 x SD and 3-sigma rules work the same way with a different fence. Two standard deviations will flag more values and three will flag fewer. Pick one, write it down, and stay with it.
Clinical thresholds: how regulators define an outlier
Statistical distance is not the same as clinical importance. A result can be unusual for a group and still harmless, or sit near the middle of a distribution and still be dangerous for one patient.
This is why safety analyses in drug development use pre-defined clinical thresholds. FDA's integrated guide to standard safety tables and figures describes laboratory outlier analysis as identifying values that cross specified levels, for example creatinine at 1.5, 2.0 or 3.0 times baseline, rather than values that are merely far from the mean. The analyses are cumulative and draw on all scheduled and unscheduled visits. Thresholds can come from an established grading system such as the NIH Division of AIDS table, or from clinically relevant levels chosen on clinical opinion.
For a practice, the same idea applies without the regulatory paperwork. Ask which result would change what you do for this patient, and set your flags there. Our guide to four windows that fix safety lab monitoring walks through how to time those checks so a clinically important shift is not missed between visits.
Check the specimen before you question the patient
When an unexpected value arrives, the most common mistake is to treat it as either a diagnosis or a nuisance. A better first step is to ask whether the specimen is trustworthy.
Work through these questions in order:
- Does the value fit the patient? A result that contradicts the clinical picture deserves a second look before anyone acts on it.
- Is there a collection or handling explanation? Hemolysis, wrong tube type, delayed transport, temperature excursion or contamination can all produce extreme values.
- Is the identity secure? Check labels and accession records. Mislabeling is a classic cause of outliers.
- Did the run behave? Review quality control for that batch before blaming the sample.
- Is a recollection practical? When the answer is yes, a fresh specimen is often the cleanest way to resolve doubt.
The literature adds a caution here. Practitioners should look for iterative outliers that point to non-analytical problems, and relying only on statistical trimming can mask systemic errors that need a process fix instead of a data fix. If three tubes from one collection site in a week look odd, the problem is not three patients. It is the site.
This is also where specimen logistics matter. Clean outcomes begin with consistent collection, labeling and timely transport, which is why courier pickup and mobile phlebotomy are part of a laboratory relationship and not an afterthought.
Delta checks: useful, but not for everything
A delta check compares a patient's result with their previous one and flags a change larger than expected physiological variation. Many laboratories use them because they are cheap and automatic. There is published work on estimating change limits for delta checks that shows how a lab can derive limits from its own data.
The honest caveat is sensitivity. A stochastic-dynamic model of specimen mix-up found that delta checks often detect sample mix-ups poorly, with sensitivity below 20% except for certain analytes such as creatinine. They are better at catching real physiological shifts than swapped tubes. So a passed delta check is not proof that the specimen belongs to the patient, and a failed one is not proof of an error. Treat it as a prompt to look, not a verdict.
The rule that protects you: never re-run until it looks right
When a result is out of specification, the temptation is to repeat the test and keep the number that looks better. Resist it. FDA guidance on out-of-specification results says laboratories must not retest until an in-specification result appears, that averaging repeat tests hides the variability that signals a problem, and that every individual result must be reported. Even a single out-of-specification result is to be investigated by the quality control unit.
The same discipline serves a medical practice. Record the original value, the repeat, who ran it and why. If you are building a serial testing plan, define up front which repeats are scheduled and which are investigative. A three-month bone turnover markers plan for clinicians is one example of serial markers being read against a plan set in advance, instead of being judged one number at a time.
Outlier removal changes your reference intervals
If your lab or research group builds its own reference intervals, the method you use to remove outliers is not a minor detail. A study of outlier removal tools in healthy populations found that for 9 of 36 analytes, the upper reference limit differed by more than 20% depending on the method chosen, with the choice of method such as Reed-Dixon or Tukey's fences sometimes shifting limits dramatically.
Two practical conclusions follow:
- Standardize the method in your quality procedures, so results stay comparable over time and across studies.
- Do not assume the manufacturer's range fits your patients. A 2025 Clinical Chemistry abstract reports experts warning that manufacturer-provided intervals are often unsuitable for the local population and recommending ongoing evaluation using the laboratory's own historical results.
When several sites or labs share an interval, the same logic scales up. See our reference range harmonization checks and our guide to biomarker cutoff harmonization for the steps before values are compared across sources.
Outliers in clinical trials: decide first, then analyze
For sponsors and CROs, the rules are stricter, and they exist to protect the integrity of the study. The ICH E9 statistical principles guideline calls for exploring the influence of outliers, and where the protocol lacks a pre-defined handling procedure, the practical expectation is at least two analyses: one on the full data and one with outliers excluded, with the differences discussed and documented.
Three habits keep a trial defensible:
- Pre-define the rule. Write the outlier definition and handling in the Statistical Analysis Plan. A published example is the statistical analysis plan for one registered study, which states how abnormal laboratory values will be flagged.
- Document before unblinding. Record the justification for any exclusion before the database is finalized.
- Keep the raw record. Outlier analysis should surface clinically important signals. It is not a way to make data look tidier.
Data that arrives in consistent formats makes all of this easier. Our eight steps to analysis-ready lab data and first steps in lab data harmonization cover the groundwork.
Our perspective
We see outlier handling as a partnership question more than a statistics question. A flagged value is only useful if the person who ordered the test can quickly learn what happened to the specimen, what the lab verified and what the options are. That takes direct communication, which is what we aim to offer providers and research teams, rather than a result delivered with no one to ask.
Our advice is simple. Agree on criteria with your laboratory before testing starts. Ask how unexpected results are investigated, how repeats are reported, and how reference intervals and cutoffs are verified for your population. Ask how specimen pickup, labeling and electronic reporting are handled, because many outliers begin before the sample reaches the bench. A laboratory that answers these plainly is one you can rely on when a surprising number comes in.
— Kohealth Labs
A short checklist for practices and research teams
- Write down the statistical rule (IQR, 2 SD or 3 SD) and the clinical thresholds you will use.
- Decide who reviews flagged values and how quickly.
- Record every result, including repeats, with a reason for each repeat.
- Review quality control and specimen handling before attributing a value to the patient.
- Verify reference intervals against your own population on a schedule.
- For trials, put outlier criteria in the SAP and document exclusions before unblinding.
Working with a laboratory that supports this
If you want a lab partner for practice testing or study support, we provide clinical laboratory testing, courier specimen pickup, mobile phlebotomy, electronic results reporting and LIS and EHR integration. For sponsors and CROs, we also offer clinical trial laboratory services. Explore what we test or see how it works to start the conversation.
FAQ
What is the 1.5 x IQR rule for outliers?
The rule flags a value as an outlier when it falls more than 1.5 times the interquartile range above the third quartile or below the first quartile. It works without assuming a normal distribution, which makes it a common first screen for lab data. Treat flagged values as candidates for review, not for deletion.
Should I delete an outlier from my lab data?
Not by default. First check for specimen, labeling or analytical causes, and keep the original record. In studies, exclusions need a justification that is defined in advance in the analysis plan, and the standard approach is to analyze the data both with and without the outliers.
Can I re-run a sample if the result looks wrong?
You can investigate, but you should not repeat the test until a more acceptable number appears. Report every individual result and document the reason for each repeat. FDA guidance says averaging retests hides the variability that signals a problem.
Do delta checks catch sample mix-ups?
Only partly. Modeling work suggests their sensitivity for mix-ups is often below 20%, with exceptions such as creatinine. They are better at highlighting true physiological change, so pair them with careful labeling and identity checks.
Sources
- Iterative outlier removal: identifying outliers in laboratory recalibration studies
- FDA, Standard Safety Tables and Figures: Integrated Guide
- Estimation of change limits (delta check) in the clinical laboratory
- How useful are delta checks in the 21st century?
- FDA guidance on out-of-specification results
- Choice of statistical tools for outlier removal and reference intervals
- Clinical Chemistry abstract on reference intervals
- ICH E9 statistical principles guideline
- Example statistical analysis plan, NCT02657369
For more information about our diagnostic testing, clinical research laboratory services, or partnership opportunities, visit https://kohealthlabs.com or contact KoHealth Labs at (469) 792-8439 and ask for Amy Patterson.