Mean, median, mode: when each one is the right average

“Average” in everyday speech usually means the arithmetic mean. In statistics it’s a fuzzier word covering at least three different metrics: mean, median, and mode. Each measures “the typical value” in a dataset, but they answer different questions and disagree dramatically on skewed data. This guide explains when each one is the right tool.

The three definitions, fast

• Mean — sum of all values divided by count. The default. Same units as the data.
• Median — the middle value of the sorted dataset. Half above, half below.
• Mode — the most frequently-occurring value. The only one that works on non-numeric data.

The classic divergence: income

US household income, 2023 (American Community Survey):

• Mean: $111,000
• Median: $80,610
• Mode: roughly $40,000-50,000 (the most populous bracket)

Three numbers that all answer “what does a typical US household earn?” — and disagree by up to a factor of three. The mean is pulled up by the top-1% earners; the median sits at the actual middle of the distribution; the mode reflects where the largest bracket is.

For policy discussions and news headlines, the median is almost always the right number — it’s robust against outliers and represents an actual achievable household. The mean misleads except for arithmetic budgeting (tax collected = households × mean income).

When to use each

Use the MEAN when

• The distribution is roughly symmetric.Heights, IQ scores, measurement error. The mean and median agree closely; mean is fine to report.
• You need to compute totals from averages.The mean (× count) gives you the sum exactly. The median doesn’t. For budgeting, accounting, and integrals, mean is mandatory.
• You’ll feed it into further statistics.Variance, standard deviation, confidence intervals all build on the mean. Pick the median and you lose the downstream toolkit.

Use the MEDIAN when

• The distribution is skewed. Income, house prices, response times, file sizes, hospital stays, project costs. The median is what real people experience; the mean is what one billionaire / one marathon outlier inflates.
• Outliers are likely and uncontrolled.Sensor data with occasional glitches, manually-entered numbers with typos, web analytics with bot traffic. The median filters them out automatically.
• You want the “typical experience”.Median wait time at the DMV, median commute, median response time at your customer service. These are questions about typical individuals, not aggregate throughput.

Use the MODE when

• The data is categorical. Favourite colour, browser used, country of origin. Mean and median are undefined for these; mode is the only summary.
• You want the most common case. Most popular product, most common error type, most-shipped UPS box size. Mode is the only metric that answers this directly.
• You suspect a bimodal distribution.Combined male+female heights, hot-cold weather data, on-peak/off-peak network traffic. Reporting two modes captures the structure that one mean would hide.

The skewed-distribution trap

News articles routinely report mean income, mean house price, mean response time. All three are right-skewed distributions; in all three the mean is consistently higher than the median; reporting the mean systematically overstates “typical”.

Quick test: if the dataset has a hard floor (zero) and no hard ceiling, it’s probably right-skewed. Use the median.

Examples by type:

How to spot the lie

Three sentences to watch for in claims using “average”:

1. “The average American earns $X.”If $X > $90,000, it’s the mean. Median is $80k-ish. The mean is technically correct but answers a different question.
2. “Average response time: 200ms.”For an API, this is almost certainly the mean, which a handful of slow requests inflated. The median is probably 50-100ms; the p99 might be 2000ms. The mean alone tells you little.
3. “Most popular X is...”That’s the mode. If used correctly, it’s fine; if used in place of mean or median, it’s misleading.

When to report all three

Honest data presentation usually shows mean, median, and a spread measure (standard deviation or interquartile range). The difference between mean and median tells the reader instantly how skewed the distribution is. A mean of $111k and median of $81k tells you more than either number alone — namely, the distribution has a long right tail.

Compute all three (plus percentiles, standard deviation, and a histogram) in one pass with our statistics calculator. For the working-statistician’s background on variance and standard deviation, see standard deviation explained.

Walkthrough: response-time monitoring

Six API requests this morning measured in milliseconds: [42, 51, 48, 55, 47, 2,800]. The last one is a cold-start outlier. Three summaries:

• Mean: (42+51+48+55+47+2800)/6 = 507 ms.
• Median: sort to [42, 47, 48, 51, 55, 2800], take the average of the two middle values = 49.5 ms.
• Mode: all values appear once, so the mode is undefined for this sample.

Reporting only the mean (“average response time: 507 ms”) would convince a stakeholder the API was broken. Reporting only the median (“typical response: 49.5 ms”) hides the cold-start problem. The honest summary is both numbers plus the p99: “median 50 ms, p99 ~2.8 s — fast in the common case, occasional cold-start spikes.” That’s actionable; either number alone isn’t.

Common mistakes

• Computing the “average of averages”. The mean of three groups’ means is not the mean of the combined data unless the groups are equally sized. Always recompute from the raw data, or use a weighted mean with group sizes as weights.
• Reporting median without a spread metric. Two datasets with identical medians can have completely different shapes. Add the interquartile range (Q3 − Q1) or a quartile summary; a five-number summary (min, Q1, median, Q3, max) is a cheap, honest single line.
• Calling a bimodal distribution “the average”. A combined male+female height distribution has two peaks at ~168 cm and ~178 cm. The mean (~173 cm) describes nobody. Disaggregate by group and report each subpopulation separately.
• Using mode for continuous data. For truly continuous measurements (heights, weights, response times) every value is technically unique; the mode is either undefined or an artefact of measurement granularity. Use a histogram and identify the densest bin instead.
• Comparing means across truncated samples. Means are extremely sensitive to whether outliers are included. If two studies report different means and one excluded outliers above the 99th percentile, the difference may be entirely artefactual.

For the related spread question (how dispersed the data is around the centre), continue with our standard deviation guide. For the percent-change framing that often follows summary statistics, see percentage vs percentage point.

Sources: US Census Bureau American Community Survey 2023; NIST/SEMATECH e-Handbook of Statistical Methods §1.3.5.3 (Measures of Location); Tukey, Exploratory Data Analysis (1977); OECD Income Distribution Database (2024 release).