Skip to main content
Home Improvement10 min read

How to Compare Two Data Sets: Center, Spread, and Relative Difference

A practical workflow for deciding whether two sets are comparable, describing their typical values and variation, and reporting differences without overstating them.

By Mohamed Zakrya

Updated · 10 min read

Share
Two sets, one average, and the order that keeps a comparison honest Comparing two data sets Three calculators return three numbers. The order, and what may be claimed, belong to nobody. BOTH AVERAGE 10 invented values · five observations each · both sum to 50 Set A range 4 · s ≈ 1.581 Set B range 16 · s ≈ 6.325 the same center, describing two unalike sets FOUR GATES 1 · Are they comparable? Same definition, units, conditions 2 · One center, both sets 3 · Spread beside it 4 · Relative difference, last and only if its denominator means something THE THREE THAT GO WRONG Percentage first A relative difference is built from summaries. Leading with it hides which center was chosen and how much variation it concealed. An unstated denominator The same gap of 10 is 18.18% read symmetrically and 20% read from a baseline of 50. Near zero, the denominator stops meaning anything. Describing, then claiming A difference in the observed sets is not evidence of one in a population, not a ruling on sampling variation, and not a cause. Write the conclusion so a reader can rebuild it Both sizes, the center and why it fits, a compatible spread, the difference in units, and the scope of what was tested.
A practical workflow for deciding whether two sets are comparable, describing their typical values and variation, and reporting differences without overstating them.

Comparing two data sets is not a matter of calculating two averages and choosing the larger one. A useful comparison has an order: establish that the sets are comparable, describe the center of each set, examine their spread, and only then express any difference in relative terms.

That order matters because every summary removes information. A mean hides the individual observations. A measure of spread restores some context, but it does not show every feature of the distribution. A percentage can make a difference easier to communicate, but it can also make a weak comparison look more authoritative than it is.

This guide is about describing observed data. It does not establish that a difference is statistically significant, rule out sampling variation, or show that one condition caused another.

Start by deciding whether the sets are comparable

Arithmetic cannot repair a comparison between measurements that mean different things. Before calculating a center, confirm that both sets measure the same variable under conditions close enough to support the intended conclusion.

Check the following:

  • Definition: The measured quantity must have the same operational meaning. “Response time” measured from submission to first reply is not directly comparable with response time measured to final resolution.
  • Units and scale: Both sets must use the same units, or one must be converted before comparison. Seconds cannot be compared directly with milliseconds, and a 1–5 rating scale is not interchangeable with a 0–100 score.
  • Collection conditions: Time period, equipment, inclusion rules, and other relevant conditions should be aligned. Weekday traffic and holiday traffic may describe different circumstances even if both are recorded as visits per hour.
  • Population and sampling method: The sets should represent comparable populations. Results from all customers in one set and only returning customers in another answer different questions.
  • Sample size: Unequal sizes do not automatically prohibit a descriptive comparison, but a very small set usually gives a less stable picture than a large one. Report both sizes and investigate why they differ.
  • Dependence: Measurements from two independent groups differ from measurements taken from the same subjects before and after an intervention. That distinction becomes important if a statistical test is later required.

A conversion or documented adjustment can sometimes make the sets comparable. A difference in definitions, populations, or selection rules usually requires narrowing the claim rather than adjusting the arithmetic.

Follow the comparison in a fixed order

A defensible comparison moves through a series of gates. If the sets fail the comparability check, stop and explain why. If they pass, choose a center that represents both sets honestly, place an appropriate measure of spread beside it, and then decide whether a relative difference adds useful information.

Comparability first, relative difference last Four gates, in this order Every summary removes information, so each gate decides what the next one may claim 1 · Comparable? Same definition, same units, aligned conditions, similar populations, sizes recorded No → stop here Narrow the claim. Arithmetic cannot fix a definition gap. 2 · Center Mean, median or mode — the same one for both sets 3 · Spread s beside a mean, IQR beside a median. State n or n − 1. 4 · Relative difference Only when the denominator has a clear interpretation A descriptive report Sizes, centers and spreads, plus the difference in units What it is not Not a significance test. Not a ruling on sampling variation. Not causation. Those need a test named and run elsewhere. Percentage difference is computed from summaries, so it inherits every choice made above it. Leading with it hides which center was picked and how much variation that center concealed.
A comparison workflow that begins with measurement validity and proceeds through center, spread, relative difference, and a qualified descriptive report.

This order prevents percentage difference from becoming the opening claim. A relative difference is calculated from summaries, so its meaning depends on how those summaries were selected and how much variation they conceal.

Choose a center that represents each set

Mean, median, and mode describe different aspects of a set. The mean distributes the total equally across all observations. The median identifies the middle position after sorting. The mode identifies the most frequently occurring value.

Choose according to the data and the question:

  • Use the mean when every value should contribute to the balance point and extreme observations are genuine, relevant parts of the quantity being summarized.
  • Use the median when the question concerns a typical ordered observation and skew or an unusually large or small value would pull the mean away from most observations.
  • Use the mode when the most common value or category is substantively useful. It is often less informative for measurements that rarely repeat exactly.

Apply the same kind of summary to both sets. Comparing the mean of one set with the median of another creates a difference between definitions as well as between data. If outliers make the mean unsuitable in one set, inspect both distributions and consider reporting both medians.

State the actual sets in a worked comparison. For example, suppose these are invented measurements used only to demonstrate the arithmetic:

Set A   8, 9, 10, 11, 12
Set B   2, 6, 10, 14, 18

Both sums are 50 and both sets contain five observations, so both means are 50 ÷ 5 = 10. Their medians are also 10. Center alone therefore reports no difference.

The Average Calculator can calculate the main measures of center, but selecting the appropriate one still depends on what the observations represent and how they are distributed.

Put spread beside center

The example sets have equal means, yet they are not alike. Set A stays close to 10, while Set B extends much farther in both directions. The range of Set A is 12 − 8 = 4; the range of Set B is 18 − 2 = 16.

Standard deviation incorporates every observation rather than using only the endpoints. For Set A, the deviations from 10 are −2, −1, 0, 1, and 2. For Set B they are −8, −4, 0, 4, and 8. Squaring each deviation and adding gives:

Set A   4 + 1 + 0 + 1 + 4      =  10
Set B   64 + 16 + 0 + 16 + 64  = 160

If these sets are treated as samples, that sum is divided by n − 1 before the square root is taken:

Set A   s = √(10 ÷ 4)   = √2.5  ≈ 1.581
Set B   s = √(160 ÷ 4)  = √40   ≈ 6.325

Set B’s observations are much more dispersed around the shared mean. Reporting only “both averages are 10” would omit the principal descriptive difference between the sets.

Equal center, unequal spread Both means are 10 Five observations each, both summing to 50 — invented values, checkable by hand mean = median = 10 Set A range 4 s ≈ 1.581 8 12 Set B range 16 s ≈ 6.325 2 18 0 5 10 15 20 “Both averages are 10” is true and omits the whole difference. Squared deviations sum to 10 for Set A and 160 for Set B — one center, two unalike sets.
Two example data sets centered at 10, with Set A clustered near the center and Set B extending across a wider interval.

Standard deviation works naturally beside a mean, but it can also be strongly affected by outliers. For skewed data summarized by the median, the interquartile range is often a more consistent companion because it describes the middle half of the ordered observations. A range is easy to read but depends entirely on the two endpoints.

Why the divisor reflects the role of the data

The standard deviation calculation depends on whether the observations constitute the entire population of interest or a sample used to describe a larger population.

For a complete population, divide the sum of squared deviations by n. In the example, the population standard deviations would be:

Set A   σ = √(10 ÷ 5)   = √2   ≈ 1.414
Set B   σ = √(160 ÷ 5)  = √32  ≈ 5.657

For a sample, variance uses n − 1. The reason is connected to how the sample mean is obtained. Once the sample mean and n − 1 deviations are known, the final deviation is forced, because all deviations from the sample mean must sum to zero. Only n − 1 of them can vary independently.

The sample mean is also fitted from the same observations whose spread is being measured. That fitted center tends to sit closer to its own observations than the unknown population mean would, so dividing the squared deviations by n would tend to underestimate population variance.

Under the usual independent-sampling assumptions, dividing by n − 1 corrects that bias for the variance estimate. Taking the square root introduces another distinction, so the resulting sample standard deviation is not itself exactly unbiased in general.

Use the version that matches the data’s role, and state which version you report. Do not select the smaller divisor merely because it produces a preferred result.

Add a relative difference after center and spread

An absolute difference retains the measurement’s units. If two comparable means are 50 units and 60 units, the absolute difference is 10 units. Whether 10 is large enough to discuss depends on the scale, variation, measurement context, and purpose of the comparison.

Percentage difference expresses the separation relative to the average magnitude of the two summaries:

percentage difference = |x₁ − x₂| ÷ ((|x₁| + |x₂|) ÷ 2) × 100%

For a second invented example, use these complete sets:

Set C   44, 48, 52, 56
Set D   54, 58, 62, 66

Their sums are 200 and 240. Because each set has four observations, the means are 50 and 60. Both sets have a range of 12, and their squared deviations from their respective means total 80 in each set, so both sample standard deviations are the same:

s = √(80 ÷ 3) ≈ 5.164

The symmetric percentage difference between the means is:

|60 − 50| ÷ ((60 + 50) ÷ 2) × 100%  =  10 ÷ 55 × 100%  ≈  18.18%
From two centers to a percentage, one step at a time Same spread, different centers Set C: 44, 48, 52, 56 · Set D: 54, 58, 62, 66 — invented; every figure follows from them CENTER 50 Set C mean 60 SPREAD √(80 ⁄ 3) ≈ 5.164 both sets, n − 1 ABSOLUTE 10 units, in the original measurement RELATIVE 10 ⁄ 55 18.18% symmetric THE DENOMINATOR IS THE CLAIM Neither is a baseline |60 − 50| ⁄ ((60 + 50) ⁄ 2) = 10 ⁄ 55 ≈ 18.18% Treats the two summaries alike. 50 is a defined baseline |60 − 50| ⁄ 50 = 10 ⁄ 50 = 20% Same gap, different question.
The two example sets move from means of 50 and 60, through equal sample standard deviations of about 5.164, to an absolute difference of 10 and a symmetric percentage difference of about 18.18%.

This 18.18% result treats the two means symmetrically. If 50 were a defined baseline and the question concerned movement from that baseline to 60, the calculation would instead be 10 ÷ 50 × 100% = 20%. State which relationship the denominator represents.

Relative measures can become misleading when values are near zero, cross zero, or can be negative. In those situations, the denominator may be tiny, zero, or difficult to interpret. Report the absolute difference or use a domain-specific scale rather than forcing a percentage.

Separate description from inference

A descriptive comparison reports what is present in the observed sets: their sizes, centers, spreads, and calculated differences. It does not show that the same difference exists in a larger population, that it would recur in another sample, or that a condition caused it.

If the data are samples and the question is whether an observed difference is compatible with sampling variation, an inferential procedure is needed. For two independent groups whose means are being compared, Welch’s two-sample t-test is often considered because it does not require equal variances. Measurements taken from the same subjects under two conditions may call for a paired t-test. Other designs and distributional questions require other methods and assumptions.

Those tests require more than the three calculators linked here, and this page does not perform them. A confidence interval or test result must also be interpreted in light of the study design, measurement quality, and practical importance. A statistically detectable difference is not automatically important, and a descriptive difference is not automatically evidence of an underlying population effect.

Write the conclusion so it can be checked

A compact comparison should let the reader reconstruct both the arithmetic and its limits. Include:

  1. What each set measures, including units and collection conditions.
  2. The number of observations in each set.
  3. The selected center and why it fits the distribution.
  4. A compatible measure of spread for each set.
  5. The absolute difference in original units.
  6. A relative difference only when its denominator has a clear interpretation.
  7. A statement that the results are descriptive unless an appropriate inferential analysis was performed.

For the second example, an accurate descriptive report would be: “In these invented sets of four observations each, Set C had a mean of 50 and Set D had a mean of 60. Both sample standard deviations were approximately 5.164. The mean difference was 10 units, corresponding to a symmetric percentage difference of approximately 18.18%. These calculations describe the displayed observations and do not establish statistical significance or causation.”

That wording gives center, spread, absolute difference, relative difference, sample size, and scope. It also avoids turning one calculated number into a claim the data cannot support.