Outliers in Biological Data

An outlier is an observation that lies an unusually large distance from the rest of the data in a sample — quantitatively distant, not merely "surprising" or "unwanted." Distance is typically judged relative to the sample's own spread: a value several interquartile ranges (IQRs) from the nearer quartile, or several standard deviations from the mean, is flagged as a candidate outlier. The word "candidate" matters — statistical extremity is an objective, computable property of the number; whether that extremity reflects a real biological phenomenon or a procedural error is a separate, non-statistical question that requires biological judgment and investigation. It is essential to separate two things that students often conflate: outlier detection (a mathematical procedure applied to the numbers) and outlier interpretation (a scientific judgment about the numbers' origin). A value can be a strong statistical outlier and still be the most biologically important observation in a dataset — the patient with an extraordinarily rapid antibody response, the plant with unprecedented drought tolerance, the cancer cell line with an unusually high mutation burden. Equally, a value can be a modest statistical outlier and still be pure equipment error — a pipetting mistake that happens to land close to the rest of the distribution. Detection is objective and mechanical; interpretation is scientific and requires domain knowledge, metadata, and often repeating the measurement. A further subtlety: outlier status depends on the assumed shape of the underlying distribution. Tools built on the normal distribution (z-scores, Grubbs' test) will over-flag points in any biological dataset that is naturally right-skewed — viral titers, gene expression counts, enzyme kinetics, survival times, most things measured on a multiplicative rather than additive scale (see The Shape of Distributions and Skew). Log-transforming such data before applying outlier rules, or using rank/median-based methods that don't assume normality, prevents biologists from misclassifying an entirely ordinary heavy right tail as a swarm of "errors."

Step 1: Biological sources of genuine extreme values. Real biological variation generates extreme values through several distinct mechanisms:

• Genetic variance: a rare allele or de novo mutation produces a phenotype far outside the population range (e.g., an unusually tall individual due to a growth-hormone receptor variant).
• Physiological state: an animal in the middle of an immune response, molting, or fasting has metabolic values wildly different from its cohort's baseline.
• Developmental stage mismatch: a "juvenile" mistakenly measured alongside adults will appear as a low outlier in a size dataset — not an error, but a sampling/design issue.
• Rare exposure or event: in ecology, one plot may have experienced an unrecorded fire or flood, driving its community metrics far from neighboring plots.
• Multiplicative processes: gene expression, PCR amplification, and population growth compound multiplicatively, producing long right tails where extreme values are the expected mathematical consequence of the process, not anomalies.

Step 2: Technical/procedural sources of spurious extreme values.

• Instrument malfunction (a plate reader saturating, a balance not tared).
• Sample contamination or mix-up (mislabeled tube, cross-contaminated culture).
• Transcription or data-entry error (misplaced decimal, unit mismatch — mg vs. g).
• Reagent degradation or batch effects affecting a subset of samples.
• Statistical artifacts of small sample size, where a modest deviation looks extreme only because n is small.

Step 3: Detection procedures. Given a sample, several formal rules quantify "how far is too far": Boxplot / 1.5×IQR rule. Compute Q1 and Q3 (first and third quartiles), and IQR = Q3 − Q1. Any point below Q1 − 1.5×IQR or above Q3 + 1.5×IQR is flagged as a mild outlier; beyond 3×IQR, an "extreme" outlier. This method makes no normality assumption and is robust to the shape of the bulk of the data, which is why it is the default in boxplots (see Boxplots). Z-score rule. z = (x − mean)/SD. Points with |z| > 2 or 3 are flagged. This assumes approximate normality — badly misleading on skewed data, since the mean and SD themselves are outlier-sensitive (a single huge value inflates the SD, which can mask its own extremity — this is called masking). Modified z-score (median-based). Uses the median and the median absolute deviation (MAD) instead of mean/SD:

More robust because neither the median nor MAD is dragged around by the outlier itself. Grubbs' test. A formal hypothesis test for a single outlier in an approximately normal sample, testing whether the most extreme value is inconsistent with being drawn from the same normal population as the rest.

Step 4: The interpretive step — this is the part detection rules cannot do for you. Once flagged, an observation must be traced back through the lab notebook, the instrument log, and the sample metadata. Was this animal noted as sick? Was this plate read on a day the reader was recalibrated? Was this patient on a medication not recorded in the main dataset? Only this forensic step converts a statistical flag into a scientific decision: keep it as real biology, correct a documented error, or — as an absolute last resort with full disclosure — exclude it with a stated, pre-specified, non-circular rule.

There is no single "Meselson–Stahl" experiment for outliers, because outlier handling is a methodological principle rather than a single discovery — but its empirical foundations rest on well-documented statistical history and on cautionary biological case studies that shaped how the discipline now treats extreme values.

Frank Grubbs and the formalization of outlier testing (1950s). Working at the U.S. Army's Ballistic Research Laboratory, Frank E. Grubbs formalized a statistical test (published 1950, extended 1969) for deciding whether the single most extreme value in a normally distributed sample is statistically inconsistent with the rest. Grubbs's explicit motivation was measurement science: artillery and ballistics data occasionally contained values from equipment malfunction, and he wanted an objective, pre-specified rule rather than ad hoc judgment calls by individual analysts. Crucially, Grubbs also stressed — often lost in later textbook simplifications — that a statistically significant outlier test result is grounds for investigation, not automatic grounds for deletion. His original papers explicitly discuss cases where the "outlier" was retained after investigation confirmed it as genuine.

John Tukey and robust/resistant statistics (1970s). John Tukey, who also gave us the boxplot itself (see Boxplots), was central to developing "resistant" statistics — summary measures (median, IQR, trimmed means) that are not unduly influenced by a small number of extreme values. Tukey's broader research program (exploratory data analysis, published as Exploratory Data Analysis, 1977) demonstrated empirically, using real datasets, that mean-and-SD-based analysis routinely misrepresented skewed or contaminated real-world data, while resistant statistics gave stable, reproducible summaries even when a handful of extreme points were added or removed. This body of work is the direct ancestor of the modern advice: report robust statistics (median, IQR) alongside classical ones (mean, SD) whenever outliers are present, so readers can judge the sensitivity themselves.

A cautionary biological case: retraction risk from undisclosed outlier removal. Multiple documented cases in the clinical and preclinical literature show what happens when outlier removal is done non-transparently. A widely cited example is instances where preclinical rodent studies quietly excluded "non-responding" animals from treatment arms without pre-registered criteria, inflating apparent drug efficacy — a pattern flagged repeatedly in reproducibility audits of animal research (e.g., analyses accompanying the ARRIVE reporting guidelines, first published 2010, revised 2020). These audits found that studies excluding data post hoc, without a stated rule, systematically reported larger effect sizes than studies with pre-specified inclusion/exclusion criteria — direct empirical evidence that undisclosed outlier handling biases biological conclusions, usually toward the researcher's hypothesis.

Luria–Delbrück as an outlier-adjacent case study. The 1943 fluctuation test of Salvador Luria and Max Delbrück (establishing that bacterial mutations arise randomly, not in response to selection) is instructive here for a different reason: their data showed a distribution with occasional "jackpot" cultures — plates with vastly more phage-resistant bacteria than any other replicate. A naive analyst might have discarded these jackpots as technical outliers. Instead, Luria and Delbrück recognized that the extreme variance itself, including the jackpots, was the biological signal: it was exactly the Poisson-violating, heavy-tailed pattern predicted if mutations occur randomly before selection, rather than being induced by selection. Treating the extreme values as errors would have destroyed the very evidence that answered one of genetics' foundational questions.

Example 1 — Physiological: resting heart rate in a comparative physiology lab. A class measures resting heart rate (beats per minute) in 12 lab mice: ``` 520, 540, 535, 545, 528, 552, 610, 538, 522, 549, 531, 118 ``` Step 1 — order the data: ``` 118, 520, 522, 528, 531, 535, 538, 540, 545, 549, 552, 610 ``` Step 2 — compute quartiles (n = 12, using the median-of-halves method): Lower half (first 6): 118, 520, 522, 528, 531, 535 → Q1 = median = (522+528)/2 = 525 Upper half (last 6): 538, 540, 545, 549, 552, 610 → Q3 = median = (545+549)/2 = 547 IQR = 547 − 525 = 22 Step 3 — compute fences: Lower fence = 525 − 1.5(22) = 525 − 33 = 492 Upper fence = 547 + 1.5(22) = 547 + 33 = 580 Step 4 — flag: 118 bpm is far below the lower fence (492) → flagged outlier. 610 bpm exceeds the upper fence (580) → also flagged, though less extremely. Step 5 — investigate before deciding:

• 118 bpm is physiologically implausible for a healthy adult mouse (normal resting range roughly 310–840 bpm depending on strain and anesthesia depth, but values in the 500s cluster tightly here) — cross-checking the lab notebook shows this mouse was under deep anesthesia at time of recording, a known cause of severe bradycardia. This is a real biological/procedural confound, not a data-entry slip — the correct action is to exclude it with a documented reason (anesthesia depth not matched to other animals) and note the exclusion in methods, rather than silently drop it.
• 610 bpm has no such flag in the notebook; the mouse was awake, active, and otherwise normal. This is very plausibly a genuine excited/stressed physiological state — it should be retained, perhaps annotated, and its influence checked using a robust measure (median heart rate of the remaining 11 animals barely changes whether 610 is included or not).

Effect on statistics: including all 12 points gives a mean of 476.5 bpm and SD of 129.9 — badly distorted by the single 118 bpm value. Excluding the documented anesthesia-confound point (118) but retaining the 610 bpm point gives mean = 542.7, SD = 25.1 — a far more representative summary of the true underlying resting-heart-rate distribution.

Example 2 — Genetic cross with a skewed, over-dispersed count outlier. A geneticist crosses two lines of Drosophila differing at a marker gene and counts progeny with a recessive wing-shape phenotype across 10 vials (expected under simple Mendelian segregation: 1/4 of offspring per vial, families roughly independent): Vial phenotype counts (out of ~80 flies per vial): ``` 18, 21, 19, 22, 20, 17, 23, 19, 20, 61 ``` Expected proportion = 0.25, so expected count per vial of 80 ≈ 20. Step 1 — median and MAD (robust, since n is small and one value is extreme): Sorted: 17, 18, 19, 19, 20, 20, 21, 22, 23, 61 Median = (20+20)/2 = 20 Absolute deviations from median: 3, 2, 1, 1, 0, 0, 1, 2, 3, 41 MAD = median of these = (1+1)/2 = 1 Step 2 — modified z-score for the suspicious vial (61):

A modified z-score above roughly 3.5 is conventionally flagged as an outlier; 27.7 is enormous — an unambiguous statistical outlier. Step 3 — biological reasoning before removal. Two live hypotheses: 1. Technical error: this vial was scored by a different student who miscounted, confusing a similar-looking heterozygous phenotype for the homozygous recessive one — checking the raw scoring sheet (if kept) can directly confirm or refute this. 2. Genuine segregation distortion: this vial's parental pair happened to carry a linked meiotic drive element or a segregation-distorter allele, causing transmission ratio distortion so that the recessive allele is inherited in far more than 25% of gametes — a real, well-documented phenomenon in Drosophila (e.g., Segregation Distorter, SD, first characterized by Sandler and Novitski in the 1950s). If the scoring sheet reveals a transcription error (e.g., 61 should have been 21, with a digit doubled), correct it — this is not "outlier removal," it is error correction, and it should be documented as such. If the sheet confirms 61 flies were genuinely scored correctly, the correct action is to retain the vial and flag it as a candidate case of segregation distortion warranting follow-up crosses — deleting it would erase a potentially novel genetic finding. Step 4 — effect on the chi-square goodness-of-fit test. Pooling all 10 vials (n = 800 flies, observed recessive = 240) against expected 25% (200) gives χ² = (240−200)²/200 = 8.0, p ≈ 0.005 — apparently a significant deviation from 1:4. Removing vial 10 (n = 720, observed = 179, expected = 180) gives χ² ≈ 0.006, p ≈ 0.94 — no deviation at all. The entire "significant" departure from Mendelian ratios in the pooled dataset is being driven by a single vial. This is exactly the situation where reporting only the pooled p-value, without disclosing the sensitivity to one vial, would mislead readers about whether Mendelian segregation is holding.

Frank E. Grubbs (1913–2000) was an American statistician working primarily in industrial and military quality control and measurement science, most influentially at the U.S. Army's Ballistic Research Laboratory (Aberdeen Proving Ground, Maryland). His career sat at the intersection of applied engineering statistics and the emerging mid-20th-century movement to formalize decisions that experimentalists had previously made by eyeballing data — a movement that, transplanted into biology, gave rise to the modern discipline of biostatistics's treatment of anomalous values. Grubbs's 1950 paper, "Sample Criteria for Testing Outlying Observations," gave the field a reproducible, testable criterion where before there had only been informal judgment, and his 1969 follow-up broadened the framework to multiple suspected outliers and different underlying distributional assumptions. He explicitly warned against reflexive deletion, a nuance frequently lost when his test is taught purely as a mechanical filter. His work, together with John Tukey's parallel development of robust and exploratory statistics, forms the direct statistical ancestry of the outlier-handling conventions used across modern biology, medicine, and genomics.

Clinical trials and drug safety. Regulatory bodies (FDA, EMA) require pre-specified statistical analysis plans that state, before unblinding, exactly how outliers and protocol deviations will be handled — precisely to prevent researchers from excluding inconvenient patient responses after seeing which arm they favor. A "non-responder" in an oncology trial may reflect genuine tumor heterogeneity (a biologically real resistant subclone), not measurement error, and misclassifying it as noise can hide a safety signal or a resistance mechanism worth investigating.

Genomics and high-throughput screening. RNA-seq and single-cell expression datasets are dominated by long-tailed count distributions where a handful of genes or cells show expression orders of magnitude above the rest — often the most biologically interesting hits (a massively upregulated stress-response gene, a rare tumor-initiating cell). Pipelines routinely use robust normalization (median-of-ratios, trimmed mean of M-values) precisely because naive mean-based scaling would be wrecked by these legitimate extreme values — a direct link to Summary Statistics for Genomic Data.

Conservation biology and ecology. A single camera-trap record of an "extinct" species, or one plot with anomalously high biodiversity after a disturbance event, is often the scientifically critical observation in an entire dataset — ecologists are trained to flag but not discard such records without field verification.

Diagnostic medicine. A single patient's lab value flagged as a "critical" outlier by hospital reference-range software (e.g., extreme potassium or glucose readings) triggers immediate clinical action rather than data deletion — in medicine, the entire outlier-detection apparatus exists specifically to find the values that matter most.

Preclinical reproducibility movements. The replication crisis in biomedical science has driven adoption of pre-registration and mandatory disclosure of all exclusion criteria (data dredging and selective outlier removal are now recognized as major contributors to irreproducible preclinical findings), directly building on the Tukey/Grubbs-era insight that undisclosed data handling is itself a source of bias.

Ask the tutor: If I find a statistical outlier in my dataset but cannot trace its origin through any lab records, what is the most defensible way to report and analyze my data?