Open any clinical research paper and you will find descriptive statistics in the first table: age, BMI, laboratory values, clinical scores. But look closely at how those numbers are presented, and you will often find researchers reporting the mean and standard deviation for variables that are clearly skewed — hospital length of stay, C-reactive protein, serum creatinine, waiting times — and sometimes the reverse: applying the median to roughly normal data without reason. Choosing between the mean and the median is not a stylistic preference. It determines whether your summary statistics honestly represent your patients or systematically mislead readers about what a "typical" result looks like. This guide explains when each measure is appropriate, why skewed distributions and outliers force the choice, and how this decision connects directly to your choice of statistical test.
Mean and Median: What They Actually Measure
The Mean
The arithmetic mean is the sum of all values divided by the number of values. It is what most people understand as the "average."
The Median
The median is the middle value when all observations are sorted in ascending order. It divides the dataset exactly in half: 50% of values fall below it and 50% above.
Mean — Use When
- Data is approximately normally distributed
- No influential outliers are present
- You need a measure compatible with parametric tests
- You will calculate SE or confidence intervals for the mean
- You need to pool results across groups (weighted mean)
Paired with: SD (standard deviation)
The Effect of Skewed Distributions
In a perfectly symmetric (normal) distribution, the mean and median are identical — both fall at the centre. Real medical data is rarely perfectly symmetric. The pattern of skewness determines which measure gives the more honest summary.
Left Skew (Negative)
Mean < Median
Tail: extends left
Mean: pulled DOWN
Median: more accurate
Example: age at death in
palliative care ICU
Symmetric (Normal)
Mean ≈ Median
Tail: equal on both sides
Mean: accurate
Median: accurate
Example: SBP in healthy
adult cohort
Right Skew (Positive)
Mean > Median
Tail: extends right
Mean: pulled UP (misleading)
Median: more accurate
Example: ICU LOS, CRP,
serum creatinine, costs
Right skew is the dominant pattern in medical data. Most clinical laboratory values, healthcare utilisation metrics, and time-based outcomes are right-skewed because they have a natural floor of zero but no upper ceiling. A patient cannot have negative creatinine or a negative length of stay, but a small number of patients can have extremely high values — severely ill patients stay for weeks or months, outlier biomarker readings can exceed the normal range by tenfold. These few extreme high values pull the mean upward and away from the typical patient's value.
The key diagnostic: if the mean is substantially higher than the median in your dataset, right skew is present and the median is the more appropriate central tendency measure. If the mean is substantially lower than the median, left skew is present and the median is again preferred. Only when the mean and median are close should you default to the mean — and even then, confirm with a formal normality test or histogram.
The Outlier Problem: A Concrete Illustration
Outliers are data points that lie far from the rest of the distribution. In medical research, they arise for legitimate clinical reasons — an exceptionally long hospital stay due to a complication, an extreme biomarker level in a severely ill patient, an unusually delayed treatment time during a system failure — and they are not errors to be deleted. They represent real patients. The question is: should those extreme patients drive the "typical" summary statistic for the whole group?
ICU Length of Stay — 10 Consecutive Admissions (days)
One patient with a 45-day stay — perhaps due to a serious post-operative complication — pulls the mean from approximately 5 days to 9 days. If a hospital administrator saw "mean ICU LOS = 9.0 days" in a quality report, they would conclude that typical ICU resource use is nearly double what it actually is for the overwhelming majority of patients. The median of 5 days honestly reflects the experience of 8 out of 10 patients in this dataset. The outlier patient's data is not ignored — it remains in the record — but it no longer distorts the summary of what is typical.
The Mean-Median Gap as a Skewness Signal
Before deciding which measure to report, always calculate both. If mean > median by more than 20–30% of the standard deviation, right skew is likely present and you should report the median. If mean < median by the same margin, left skew is present. If they are close (within 5–10% of each other), the distribution is approximately symmetric and the mean is acceptable — provided a normality test confirms this.
Clinical Dataset Examples
1
Hospital Length of Stay After Elective Colorectal Surgery
Retrospective cohort • n = 140 patients • LOS measured in whole days from admission to discharge
Mean LOS
8.7 days
Pulled by 12 patients with LOS >25 days
Median LOS
5.0 days
Reflects the typical uncomplicated course
IQR
4–8 days
Middle 50% of patients discharged in this window
Report Median + IQRLOS is structurally right-skewed. Reporting mean LOS = 8.7 days would mislead quality improvement teams into doubling capacity estimates for typical post-operative care.
Journal phrasing: "Median length of hospital stay was 5.0 days (IQR 4–8 days). Twelve patients (8.6%) required extended admission beyond 25 days due to post-operative complications and were reported separately."
2
Systolic Blood Pressure in a Community Hypertension Cohort
Cross-sectional study • n = 280 adults aged 45–70 with treated hypertension • SBP measured after 5 minutes seated rest
Mean SBP
142.3 mmHg
Shapiro-Wilk p = 0.38 (normal)
Median SBP
141.0 mmHg
Close to mean — symmetric distribution
SD
16.8 mmHg
Appropriate spread measure
Report Mean + SDSBP in a treated hypertension cohort within a defined age range is approximately normally distributed. Mean and median are nearly identical (142.3 vs 141.0 mmHg), confirming symmetry. Mean + SD is correct and enables direct comparison with published RCT data.
Journal phrasing: "Mean systolic blood pressure was 142.3 mmHg (SD 16.8 mmHg). Distribution normality was confirmed by Shapiro-Wilk test (W = 0.991, p = 0.38)."
3
Serum CRP in Patients Admitted with Community-Acquired Pneumonia
Prospective observational study • n = 95 adult inpatients • CRP measured on day of admission (mg/L)
Mean CRP
98.4 mg/L
Strongly inflated by 8 patients with CRP >300
Median CRP
54.2 mg/L
Typical patient's admission CRP
IQR
28–108 mg/L
Wide range reflects severity spectrum
Report Median + IQRInflammatory biomarkers such as CRP are universally right-skewed. Values span three orders of magnitude. Mean CRP of 98.4 mg/L significantly overstates the typical admission CRP and would misrepresent disease severity in most patients.
Journal phrasing: "Median serum CRP on admission was 54.2 mg/L (IQR 28.0–108.0 mg/L). Eight patients (8.4%) presented with CRP exceeding 300 mg/L, consistent with severe pneumonia."
4
Time to First Antibiotic in Sepsis (Emergency Department)
Quality improvement audit • n = 120 sepsis alerts • Time measured in minutes from triage to first antibiotic administration
Mean time
76 min
10 cases with delays >180 min inflate the mean
Median time
52 min
Representative of the protocol-compliant majority
IQR
38–74 min
Benchmark comparison target
Report Median + IQRTime-to-treatment data is almost invariably right-skewed in emergency research. Audit bodies and quality improvement benchmarks (e.g., Surviving Sepsis Campaign hour-1 bundle) specify target times against median performance, not mean performance.
Journal phrasing: "Median time from triage to first antibiotic administration was 52 minutes (IQR 38–74 minutes). Ten cases (8.3%) experienced delays exceeding 180 minutes, primarily due to diagnostic uncertainty at presentation."
5
BMI in an Obesity Intervention RCT
RCT baseline characteristics table • n = 110 per group • Participants recruited with BMI 30–40 kg/m²
Mean BMI
34.6 kg/m²
Shapiro-Wilk p = 0.14 (normal)
Median BMI
34.1 kg/m²
Close to mean — restricted range by design
SD
2.9 kg/m²
Tight SD reflecting eligibility criteria
Report Mean + SDBecause the trial recruited participants within a defined BMI range (30–40), the distribution is constrained and approximately normal. Mean + SD is appropriate, enables direct comparison across trials, and is consistent with the parametric tests used in the primary analysis.
Important lesson: BMI in a general population study (BMI 18–50+) might be right-skewed and require the median. But in a trial with inclusion criteria restricting the range to 30–40 kg/m², the restriction truncates the tails and produces approximate normality. Always check the distribution of your actual study sample — do not assume that a variable is normal or skewed based on its general population properties alone.
Parametric, Non-Parametric, and Consistency
The choice between mean and median does not stand alone. It is directly linked to your choice of statistical test, and these two choices must be internally consistent throughout your paper. Using the median in your Table 1 while applying a t-test to the same variable in your analysis is a methodological contradiction that peer reviewers will flag.
| Data Distribution | Central Tendency | Spread Measure | Comparison Test | Correlation |
|---|
| Normally distributed |
Mean |
SD |
Independent t-test
Paired t-test
One-way ANOVA |
Pearson r |
| Skewed or non-normal |
Median |
IQR (Q1–Q3) |
Mann-Whitney U
Wilcoxon signed-rank
Kruskal-Wallis |
Spearman rₛ |
| Ordinal scale |
Median |
IQR or Range |
Mann-Whitney U
Wilcoxon signed-rank |
Spearman rₛ |
| Count data (n < 20/group) |
Median |
IQR or Range |
Fisher's Exact
Wilcoxon signed-rank |
Spearman rₛ |
This table describes a fundamental principle: the distributional assumption that drives your choice of test is the same assumption that determines which central tendency measure correctly summarises your data. Parametric tests assume normal distribution and model behaviour using the mean and variance. Non-parametric tests make no distributional assumption and compare rank-based properties of the groups — which is why the median (itself a rank-based statistic) is the natural companion.
Decision Framework: Which to Report
• Step-by-Step Decision Guide for Choosing Mean vs Median
1
Know your variable type. Is it continuous, ordinal, or count data? Ordinal variables (Likert items, severity grades, ranked scores) should default to the median. Count data with many zeros or extreme values should use the median.
2
Is the variable structurally skewed? Some variables are always right-skewed by their clinical nature: length of stay, healthcare costs, serum CRP, creatinine, troponin, waiting times, time-to-event outcomes. Use the median for these by default regardless of what your normality test says — the skew is real, not an artefact of a small sample.
3
Plot a histogram. A visual inspection of the data distribution takes 30 seconds in SPSS (Analyze → Descriptive Statistics → Frequencies → Charts → Histograms) and immediately reveals whether the distribution has a long tail. If the tail extends clearly to the right or left, use the median.
4
Run a normality test. Shapiro-Wilk for n < 50 (Analyze → Descriptive Statistics → Explore → Normality plots with tests). Kolmogorov-Smirnov with Lilliefors correction for larger samples. If p < 0.05, the distribution significantly deviates from normality — use the median. If p ≥ 0.05, the distribution is consistent with normality — mean + SD is acceptable.
5
Compare mean and median. Calculate both. If mean > median by more than roughly 20–30% of the SD, right skew is influencing the mean and median is preferred. If they are very close, symmetry is likely.
6
Match your descriptive and analytic statistics. Whatever you decide — mean or median — use the same logic for your choice of statistical test. Mean + SD → t-test/ANOVA. Median + IQR → Mann-Whitney/Kruskal-Wallis.
Common Mistakes Researchers Make
Mistake 1: Reporting the mean for clearly skewed clinical variables
The most frequent error in descriptive statistics: reporting mean (SD) for length of stay, CRP, creatinine, troponin, costs, and time-to-event data. Every experienced reviewer recognises these variables as right-skewed and will question why the mean was reported. More importantly, readers who rely on the mean for clinical benchmarking will systematically overestimate what is "typical."
✓ Fix: Identify structurally skewed variables before analysis and commit to median + IQR for those. Build a protocol decision: if the variable has a natural floor of zero and is known from prior literature to be right-skewed, default to median without performing a formal test.
Mistake 2: Pairing the median with the standard deviation
"Median 54.2 mg/L (SD 38.6 mg/L)" — this pairing is statistically incoherent and appears regularly in published clinical papers. The SD is mathematically derived from the mean (it is the square root of the mean squared deviation from the mean). Reporting the SD alongside the median is like reporting the speed of a car alongside the distance driven by a different vehicle. The SD assumes symmetric distribution; when data is skewed enough to require the median, the SD is itself distorted by the same outliers and extreme values, making it uninformative as a spread measure.
✓ Fix: The correct pairing is median + IQR. If you have chosen the median, you must report the IQR (or range for small samples). Never mix measures from different distributional frameworks.
Mistake 3: Reporting the SEM (standard error of the mean) as the spread measure
The standard error of the mean (SEM = SD / √n) describes how precisely the sample mean estimates the population mean — it is a measure of estimation uncertainty, not data variability. Reporting "Mean SBP 142 mmHg (SEM 1.0 mmHg)" makes the data look far less variable than it is, because SEM shrinks with larger samples. For descriptive statistics tables (Table 1), always report SD. SEM belongs in inferential contexts — when you are constructing confidence intervals for the mean, not when describing what your sample looks like.
✓ Fix: Use SD for descriptive statistics in all Table 1 / patient characteristics tables. Use SEM only in inferential contexts: confidence intervals for the mean, standard error bars on graphs showing mean ± SE. Label every spread measure explicitly so readers know which is being reported.
Mistake 4: Not checking normality at all and defaulting to the mean
Many researchers report mean + SD for every continuous variable as an automatic default, without ever checking whether the distribution supports this. This approach is particularly common in theses and in specialties where statistical training is less emphasised. The result is baseline characteristics tables that systematically misrepresent skewed variables — a problem that flows through to mismatched statistical tests and biased conclusions.
✓ Fix: As a rule, run descriptive statistics and normality tests for every continuous variable before deciding how to report it. In SPSS: Analyze → Descriptive Statistics → Explore, which produces histograms, Q-Q plots, and Shapiro-Wilk results simultaneously for each variable. This takes ten minutes and prevents the most common errors in Table 1.
Mistake 5: Using the mean and median interchangeably in the text
"The average CRP was 54 mg/L" — does "average" mean mean or median? "Central tendency" is equally vague. In published research, every reference to a summary statistic must specify which measure is being cited. Editors and reviewers of major journals increasingly reject papers that use "average" without specifying the measure, because readers cannot interpret or compare the result without knowing whether it is mean or median.
✓ Fix: Never use "average" in a results section. Always write "mean" or "median" explicitly. Define in your statistical analysis section which variables will be described by mean + SD and which by median + IQR, and apply these definitions consistently throughout Results and any tables or figures.
Mistake 6: Applying a t-test to data described with the median
If your Table 1 shows "Median LOS 5 days (IQR 3–9)" and your Table 3 shows "t(138) = 2.14, p = 0.034" for the comparison of LOS between groups, you have an internal contradiction. The t-test tests whether population means differ — yet you have acknowledged (by reporting the median) that the data is skewed and the mean is not the appropriate summary. If the data requires the median, it requires a non-parametric comparison test.
✓ Fix: Match your statistical test to your distribution. If you report median + IQR for a variable, compare it between groups using Mann-Whitney U (two groups) or Kruskal-Wallis (three or more groups), and report accordingly: "Mann-Whitney U = 1842, p = 0.041."
Journal Reporting Examples
✓ Correct — Skewed Variable (LOS, CRP, Creatinine)
"Median ICU length of stay was 4.5 days (interquartile range 2.0–8.5 days) in the intervention group compared with 5.0 days (IQR 3.0–9.5 days) in the control group. Groups were compared using the Mann-Whitney U test (U = 3,412, p = 0.043)."
✓ Correct — Normal Variable (SBP, BMI, Age in restricted cohort)
"Mean systolic blood pressure was 142.3 mmHg (standard deviation 16.8 mmHg) in the treatment group and 143.1 mmHg (SD 17.2 mmHg) in the control group. There was no significant between-group difference at baseline (independent samples t-test: t(238) = 0.33, p = 0.74)."
✓ Correct — Methods Declaration
"Continuous variables were described as mean (SD) if normally distributed, as assessed by the Shapiro-Wilk test (α = 0.05) and visual inspection of histograms. Non-normally distributed variables are described as median (interquartile range, Q1–Q3). Normally distributed variables were compared between groups using the independent samples t-test; non-normally distributed variables were compared using the Mann-Whitney U test."
✕ Incorrect — Common Errors to Avoid
"The average CRP was 98.4 ± 76.2 mg/L." [Problems: (1) "average" is ambiguous; (2) SD of 76.2 on a mean of 98.4 implies negative values are possible, which is not biologically plausible for CRP — a telltale sign of right skew; (3) CRP requires median + IQR by the nature of the variable.]
The SD Must-Not-Be-Larger-Than-Mean Rule (For Bounded Variables)
For any variable that cannot be negative (CRP, creatinine, troponin, length of stay, platelet count, waiting times), if SD > mean/2, the distribution is almost certainly right-skewed and the mean is misleading. This is a fast bedside check: if the mean is 60 and the SD is 55, the distribution cannot be normally distributed because normal distributions have essentially no values below zero when mean/SD < 3. Always report the median for these variables.
Frequently Asked Questions
When should you use the median instead of the mean in medical research? +
Use the median when data is skewed, contains outliers, or is measured on an ordinal scale. In clinical research, structurally right-skewed variables include hospital length of stay, biomarker levels (CRP, troponin, creatinine, ferritin), waiting times, costs, and platelet counts in disease populations. For these, the median better represents the typical patient. Use the mean when data is approximately normally distributed — for example, blood pressure in a defined-age cohort, BMI within a restricted eligibility range, or standardised test scores.
What is the difference between mean and median? +
The mean is the arithmetic average (sum divided by n). It uses every data value and is sensitive to extreme values. The median is the middle value when data are sorted. It uses only positional ranking and is resistant to outliers. In a symmetric (normal) distribution, mean = median. In a right-skewed distribution, mean > median. In a left-skewed distribution, mean < median. The gap between mean and median is itself a quick diagnostic for skewness: a large gap signals that the mean is being pulled toward extreme values and the median is the more honest summary.
What should you report alongside the mean and the median? +
The mean must always be paired with the standard deviation (SD): "Mean 142.3 mmHg (SD 16.8 mmHg)." The median must always be paired with the interquartile range (IQR): "Median 4.5 days (IQR 2.0–8.5 days)." These pairings are statistically coherent because each pair comes from the same distributional framework. Never pair median with SD or mean with IQR — these combinations are meaningless and will flag your paper for statistical review. For small samples (n < 15), reporting the full range (minimum–maximum) alongside median and IQR adds useful clinical context.
How do outliers affect the mean vs the median? +
Outliers can dramatically inflate or deflate the mean while leaving the median nearly unchanged. Example: 10 ICU patients with stays of 3, 4, 4, 5, 5, 5, 6, 6, 7, and 45 days. Mean = 9.0 days (pulled 4 days above the typical patient by one outlier). Median = 5.0 days (unaffected by the 45-day case). Reporting the mean of 9 days as a summary of typical ICU utilisation would mislead clinicians and administrators into substantially overestimating standard resource use. The outlier patient's data remains in the record and should be reported separately (e.g., "one patient required 45 days due to severe post-operative complication") but should not distort the typical summary.
Does the mean or median relate to parametric vs non-parametric tests? +
Yes — they are directly linked and must be consistent. Normally distributed data → mean + SD → parametric tests (t-test, ANOVA, Pearson correlation). Skewed or non-normal data → median + IQR → non-parametric tests (Mann-Whitney U, Wilcoxon signed-rank, Kruskal-Wallis, Spearman). If you report the median in your Table 1, you must use Mann-Whitney for the same variable's group comparison in Table 2. Using median in descriptive statistics alongside a t-test in the analysis is an internal contradiction that indicates the distributional assumption of the t-test was not verified.
Is the mean or median better for skewed medical data? +
The median is consistently better for skewed medical data because it represents the middle patient rather than the mathematical average distorted by extreme values. Most real-world clinical variables are right-skewed: hospital LOS, biomarker concentrations, waiting times, costs. For these, a small minority of patients with very high values inflate the mean substantially. The median is unaffected by extreme values and gives a clinically accurate representation of what the typical patient experiences. This is why regulatory agencies, quality improvement organisations, and journals in intensive care, emergency medicine, and health economics require median reporting for these outcome types.
Can you use the mean for ordinal data such as Likert scales? +
This is debated. Strictly, the mean requires interval-level measurement (equal spacing between values), which Likert items do not formally provide. The median is technically more appropriate for ordinal data. However, many researchers report means for Likert scales with 5+ response categories, particularly when the distribution is approximately symmetric, and this is widely accepted in practice. The key question is whether the median or mean better represents the typical response in your sample. For composite Likert scores (summed across multiple items), the composite often approaches normality and the mean becomes more defensible. Always check the distribution before deciding.
How do I check whether my data is skewed enough to require the median? +
Use multiple complementary approaches: (1) Histogram: in SPSS, Analyze → Descriptive Statistics → Frequencies → Charts → Histograms. A right tail indicates right skew. (2) Shapiro-Wilk test: Analyze → Descriptive Statistics → Explore → Normality plots with tests. p < 0.05 = significant non-normality → use median. (3) Skewness statistic: from Descriptive Statistics output. Values > 1 or < −1 indicate meaningful skew. (4) Mean-median comparison: if mean substantially exceeds median, right skew is present. (5) Clinical knowledge: for variables known to be structurally skewed (LOS, biomarkers, costs), use the median regardless of formal test results in your sample.
What is the IQR and how do you calculate it? +
The IQR (interquartile range) is the range covering the middle 50% of data values: Q3 − Q1, where Q1 is the 25th percentile and Q3 is the 75th percentile. Report it as: "Median 4.5 days (IQR 2.0–8.5 days)" or "Median 4.5 days, Q1–Q3: 2.0–8.5 days." In SPSS, the IQR appears in Descriptive Statistics output or Explore. The IQR is resistant to outliers for the same reason the median is: values outside Q1–Q3 do not affect the calculation. Never substitute SD for IQR when reporting the median — SD assumes a symmetric distribution and becomes uninterpretable when data is skewed.
Should you report mean or median for patient age in medical research? +
In clinical trials with defined age eligibility criteria (e.g., adults aged 40–75), age is typically approximately normally distributed within the constrained range — mean + SD is appropriate. In general population cohorts or emergency department studies spanning from paediatric to elderly patients, age may be bimodal or skewed — check the histogram and consider median + IQR. In practice, most clinical trials report mean age because eligibility criteria produce a roughly normal age distribution within the enrolled cohort. When in doubt, report both and check normality formally. The median is rarely wrong; the mean can mislead when assumptions are not met.
Check normality and choose the right statistics
StatClinic's free tools help you run Shapiro-Wilk tests, choose between parametric and non-parametric methods, and format results correctly for your thesis or journal submission.
Open StatClinic →
Pulled 4 days above 8 of 10 patients