Choosing Summaries for a Group Comparison

Why the Choice of Summary Matters

A group comparison often begins with a practical question: which group tends to have higher values, and how much do values vary within each group? A summary can help answer that question, but different summaries emphasize different features of a distribution. A few unusually large values can pull a mean upward, for example, while the median is determined by the middle position of the ordered data.

For a useful comparison, choose summaries with the data's shape and the question in mind. Means and standard deviations are informative for distributions that are reasonably balanced and do not have values far from the rest. Medians and interquartile ranges are often more representative when distributions are strongly skewed or include unusually distant observations. Then use the same type of summary for every group being compared, so the comparison is consistent.

Definition: For a nonempty group of \(n\) quantitative observations \(x_1,\ldots,x_n\), the mean is \(\bar{x}=\frac{1}{n}\sum_{i=1}^{n}x_i\). For \(n\ge 2\), the sample standard deviation is \(s=\sqrt{\frac{\sum_{i=1}^{n}(x_i-\bar{x})^2}{n-1}}\). The median is the middle value after the observations are ordered (or the average of the two middle values when \(n\) is even). The interquartile range is \(IQR=Q_3-Q_1\), where \(Q_1\) and \(Q_3\) are the first and third quartiles.

The mean and median describe center, but they do not mean the same thing. The mean is the arithmetic average; the median divides the ordered observations into lower and upper halves. The standard deviation and IQR both describe spread, but in different ways: standard deviation uses the distances of all observations from the mean, while the IQR measures the width of the middle half of the distribution.

These are descriptive summaries, not significance tests. For any nonempty finite quantitative data set, the mean can be calculated; for a data set with at least two observations, the sample standard deviation can be calculated. Neither calculation requires a normality assumption. That does not make every summary equally useful: the shape and purpose of the comparison determine which one communicates the data better. The observations should be measurements of the same variable in the same units, recorded in a comparable way across groups.

A Practical Choice for Comparing Groups

First look at the distributions, using the displays and shape descriptions introduced in earlier material on comparing groups. Ask whether the values are roughly balanced around a center, or whether one tail extends much farther than the other. Also notice whether a value lies far from most of the observations. An unusual value may be valid; its presence is a reason to consider how summaries behave, not by itself a reason to discard it.

If the distributions are reasonably symmetric and have no strongly influential distant values, a mean and standard deviation usually give a useful account of the center and spread. If the distributions are strongly skewed or have distant values, a median and IQR usually give a more resistant description of a typical value and the middle half. “Resistant” means that a change in the magnitude of a small number of extreme observations tends to affect the summary less than it affects the mean or standard deviation.

What you see in the group distributions	Often useful summaries	What to describe
Roughly symmetric, without strongly distant values	Mean and standard deviation	Average value and typical distance from the average
Strongly skewed or containing distant values	Median and IQR	Middle value and spread of the middle half
Different shapes across groups	Choose summaries that remain interpretable, and describe the shape difference	Do not let a single pair of summary numbers stand in for the distributions

The question matters too. If the question is about the average amount per observation, the mean directly answers it. If it is about the value at the center of the ordered group, the median answers it. If values are skewed, the median may better represent a typical observation, while a mean may still be relevant for a question specifically about an arithmetic average. A summary does not replace the question; it answers one particular version of it.

Worked Comparisons

Worked Example: A High Value Pulls the Mean

Suppose two hypothetical groups of seven people record the minutes needed to complete the same task. Group A's times are 10, 11, 12, 12, 13, 14, and 50 minutes. Group B's times are 8, 9, 10, 10, 11, 12, and 13 minutes. The groups use the same unit and measurement, so their summaries can be compared.

Group A has a much longer upper tail because one time is 50 minutes. The value may be genuine; the data alone do not establish otherwise. For Group A, the mean is

\bar{x}_A=\frac{10+11+12+12+13+14+50}{7} =\frac{122}{7}\approx17.43\text{ minutes}.

For its sample standard deviation, the sum of squared observations is \(3374\). Using \(\sum(x_i-\bar{x})^2=\sum x_i^2-n\bar{x}^2\), the sum of squared deviations is \(3374-122^2/7\approx1247.71\). Thus

s_A=\sqrt{\frac{1247.71}{7-1}} =\sqrt{207.95} \approx14.42\text{ minutes}.

The ordered Group A values have median 12 minutes. Using the median-of-halves convention, \(Q_1=11\) and \(Q_3=14\), so \(IQR_A=14-11=3\) minutes.

For Group B, \(\bar{x}_B=73/7\approx10.43\) minutes. Its sum of squared observations is \(779\), so its sum of squared deviations is \(779-7(73/7)^2\approx17.714\), and

s_B=\sqrt{\frac{17.714}{6}}\approx1.718\text{ minutes}.

Group B's median is 10 minutes, and its quartiles are 9 and 12 minutes, giving \(IQR_B=3\) minutes. The mean difference is about \(17.43-10.43=7.00\) minutes, whereas the median difference is \(12-10=2\) minutes. For a description of typical completion time, the median and IQR better convey that the middle portions of the groups have similar spreads; Group A's mean and standard deviation are strongly affected by its 50-minute observation. These summaries do not establish why the times differ.

Worked Example: Similar Shapes Favor Mean and Standard Deviation

Two other hypothetical groups record a measurement in centimeters. Group C's values are 18, 19, 20, 21, 22, and 24; Group D's are 21, 22, 23, 24, 25, and 27. Each group has six observations. The values in each group are fairly balanced, with no one observation far from the rest.

For Group C, the mean is \(124/6\approx20.67\) cm. Its squared observations sum to \(2586\), so the sum of squared deviations is \(2586-124^2/6\approx23.333\). Therefore

s_C=\sqrt{\frac{23.333}{6-1}} \approx2.160\text{ cm}.

For Group D, the mean is \(142/6\approx23.67\) cm. Its squared observations sum to \(3384\), giving a sum of squared deviations of \(3384-142^2/6\approx23.333\). Hence \(s_D\approx2.160\) cm as well. The mean for D is 3.00 cm higher than the mean for C, with equal sample standard deviations.

The ordered values also give medians of 20.5 cm for C and 23.5 cm for D. Those summaries tell a similar story here, but the mean and standard deviation are a clear choice because the distributions are reasonably balanced and lack strongly distant values. A good comparison states the group names, the difference in centers, the spread, and the measurement units.

Why the Summaries Behave Differently

The mean uses every value, including its distance from the rest of the data. If one observation is increased by \(\Delta\), while all other observations remain fixed, the mean increases by exactly \(\Delta/n\). For example, changing 50 to 60 in a group of seven raises the mean by \(10/7\) minutes. The middle rank does not change in this example, so the median remains 12 minutes. More generally, if the order of observations stays the same, changing the magnitude of an extreme value need not change the median or quartiles.

The standard deviation is also sensitive to distant values because it squares each deviation from the mean before averaging those squared deviations. Squaring makes a large distance contribute much more than a small one. The IQR, by contrast, is based on the central quartiles and is not directly enlarged just because the largest observation becomes more extreme, provided the quartile positions and ordering do not change. This is why median and IQR often describe the central portion of a skewed group more faithfully.

This reasoning explains the choice; it does not make either pair universally superior. A distant observation can be important to the actual question. If the goal is to report the arithmetic average amount per case, for instance, excluding or hiding a legitimate large value would give a misleading account. Examine the distribution and state which aspect of it the summaries describe.

Justification: Since \(\bar{x}=\sum x_i/n\), increasing one observation by \(\Delta\) changes the mean to \((\sum x_i+\Delta)/n=\bar{x}+\Delta/n\). The median and quartiles depend on the ordered positions used to find them. If the changed observation remains outside those positions and the ordering is unchanged, their values remain the same. This positional basis makes them less affected by the magnitude of an extreme value.

Worked Example: Let the Question Guide the Summary

Imagine a hypothetical clinic records appointment delays, in minutes, for eight visits: 5, 6, 7, 8, 9, 10, 11, and 40. The values are right-skewed. Use the median-of-halves convention for the quartiles.

The total delay is \(5+6+7+8+9+10+11+40=96\) minutes, so the mean delay per visit is \(96/8=12.00\) minutes. The sum of squared observations is \(2076\), and the sum of squared deviations is \(2076-8(12)^2=924\). The sample standard deviation is

s=\sqrt{\frac{924}{8-1}} =\sqrt{132} \approx11.49\text{ minutes}.

The median is halfway between the fourth and fifth observations: \((8+9)/2=8.5\) minutes. The lower half is 5, 6, 7, 8, so \(Q_1=(6+7)/2=6.5\) minutes. The upper half is 9, 10, 11, 40, so \(Q_3=(10+11)/2=10.5\) minutes. Thus \(IQR=10.5-6.5=4.0\) minutes.

If the question is “What is the delay at the center of the ordered appointments, and how wide is the middle half?”, report the median of 8.5 minutes and IQR of 4.0 minutes. If instead the question is “What was the arithmetic average delay per visit?”, report the mean of 12.00 minutes, while making clear that the 40-minute delay contributes to it. The mean and median answer different questions; choosing one does not make the other calculation incorrect.

Making a Consistent Group Comparison

When comparing two or more groups, use the same summary pair for each group unless there is a clear reason the displays or measurement process require different treatment. For example, reporting one group's mean and standard deviation beside another group's median and IQR makes the numbers difficult to compare: the centers and spreads have different meanings. A side-by-side display can help reveal whether the groups have similar shapes; summaries should support that picture, not replace it.

Quartile conventions can differ, particularly in small data sets. In the calculations above, the median-of-halves convention was stated so each IQR could be reproduced. When using software or a stated percentile rule, apply that same rule consistently across groups and report it if it matters to the result. Small groups also give limited information about distribution shape, so avoid treating a particular summary as a complete account of the data.

Keep units attached to the summaries: if the observations are minutes, the mean, median, standard deviation, and IQR are all in minutes. Standard deviation is not a range containing a specified proportion of observations, and the IQR is not the full spread from minimum to maximum. State the comparison in context—for example, “the median delay was 2 minutes higher”—rather than presenting numbers without identifying what they measure.

Common Mistakes

Automatically using the mean and standard deviation. A strongly skewed distribution or a distant value can make them poor descriptions of a typical group. Inspect the distributions and consider the median and IQR.
Automatically discarding a large value. A value that looks unusual may be valid. It can influence the choice and interpretation of summaries without being removed.
Mixing summary types across groups. Comparing one group's mean with another group's median is not a like-for-like comparison. Choose a consistent pair that fits the data.
Claiming the median answers every question. If the question is specifically about the arithmetic average, the mean is the relevant summary—even when it is influenced by skewness.
Calculating sample standard deviation for one observation. The formula divides by \(n-1\), so it is undefined for \(n=1\). A mean is defined for one observation, but a sample standard deviation is not.
Treating summaries as a test of group differences. Descriptive summaries characterize the observed groups; by themselves, they do not establish statistical significance or a cause for a difference.

Key takeaway: Use means and standard deviations to summarize reasonably balanced groups without strongly distant values; use medians and IQRs when skewness or extreme magnitudes make resistant summaries more representative. Let the question determine what “typical” and “spread” need to mean, and compare groups consistently.

Check Your Understanding

For each question, explain the choice or interpretation in context.

A group's measurements are strongly right-skewed. Which center and spread summaries would often describe its typical values more faithfully, and why?
For a nonempty finite quantitative data set, which of the mean and sample standard deviation can be calculated when \(n=1\)? Explain the difference.
A group of eight observations has a mean of 12 minutes and a median of 8.5 minutes. What does a substantial gap between them suggest, and what should you inspect before selecting summaries?
Why is it difficult to compare one group's mean with another group's median as if they were the same type of center?
If the research question asks for the arithmetic average delay per visit, should a valid long delay be hidden by reporting only the median? Explain.