Checking Group Sizes Before Comparing Summaries

Why Group Counts Belong Beside Summaries

A grouped mean or proportion is easier to interpret when the reader knows how many observations produced it. A mean of 80 based on four records and a mean of 80 based on four hundred records are numerically identical, but they do not carry the same amount of information. The smaller group’s summary is more sensitive to each individual observation.

When comparing outcome summaries across predictor groups, report the group size alongside each statistic. In R, a grouped summary can include a count such as n = n() as well as a mean or median. If outcome values may be missing, also count the records with an observed outcome: the count of rows in a group may not be the count used to calculate its mean.

Definition: The group size is the number of observations assigned to a particular predictor group. For a particular outcome summary, the usable sample size is the number of group observations with a recorded value for that outcome. A mean calculated after excluding missing outcomes uses the usable sample size, not necessarily the full group size.

Counts describe how much data contribute to a summary; they do not by themselves establish that a group is representative or that a difference is important. A large group can still be affected by biased data collection, and a small group can happen to have a summary close to the population value. Counts help calibrate interpretation, rather than settle it.

Read Each Summary Together with Its Count

A compact grouped table might report a predictor group, its count, and an outcome summary. For numeric outcomes, useful summaries may include the mean and median. For a binary outcome, a count of successes and the corresponding proportion make the denominator visible. A proportion without its denominator hides how much evidence supports it.

For example, if an outcome is recorded for every row, the group count and usable sample size are the same. If some outcomes are missing, report both when they matter. In R, a grouped summary using n() counts the rows in each group; a summary using mean(outcome, na.rm = TRUE) excludes missing outcome values from the calculation. Those two operations can therefore use different counts. Verify which count each reported statistic actually uses.

Worked Example: A Large Difference from Unequal Group Sizes

A hypothetical service team compares the recorded cost per case for two categories of requests. Category A has costs of $100, $110, $120, and $130. Category B has costs of $110, $120, $120, $125, $125, $125, $130, $130, $130, $130, $130, $135, $135, $135, $140, and $128. Calculate each count and mean, then explain what the unequal counts imply for interpretation.

Method. Count the observations in each category and divide the sum of its costs by that count. For Category A, there are four observations and their total is $100+110+120+130=460$. Thus its mean is $460/4=\$115$.

For Category B, there are 16 observations. Their sum is $110+120+120+125+125+125+130+130+130+130+130+135+135+135+140+128=2{,}048$. Its mean is $2{,}048/16=\$128$.

The summaries are therefore: Category A, $n=4$, mean $\$115$; Category B, $n=16$, mean $\$128$. The difference in observed means is $\$128-\$115=\$13$. Category A’s mean is based on only four cases, so each case has a comparatively large influence on that summary. The difference is a descriptive result; these counts and means alone do not show whether a difference persists in a broader population.

Common mistake avoided. Reporting only “Category B costs $13 more on average” leaves out the marked imbalance in group sizes. The counts do not erase the observed difference, but they give essential context for judging how confidently to interpret it.

Why Smaller Groups Tend to Give Less Stable Means

The link between sample size and precision can be justified for a simple model. Suppose $X_1,\ldots,X_n$ are independent observations from the same population, each with mean $\mu$ and finite variance $\sigma^2$. Their sample mean is $\bar{X}=(X_1+\cdots+X_n)/n$. Independence means the variance of the sum is the sum of the variances, so:

\operatorname{Var}(\bar{X}) =\operatorname{Var}\left(\frac{1}{n}\sum_{i=1}^n X_i\right) =\frac{1}{n^2}\sum_{i=1}^n\operatorname{Var}(X_i) =\frac{n\sigma^2}{n^2} =\frac{\sigma^2}{n}.

The standard deviation of the sample mean across repeated samples, called its standard error, is therefore $\sigma/\sqrt{n}$. If $\sigma$ is not known, it is commonly estimated by $s/\sqrt{n}$, where $s$ is the sample standard deviation. Under this model, increasing the sample size reduces the standard error, but it does not make it zero.

This result assumes independent observations with the same finite variance within the population being summarized. It does not guarantee that a group’s observed mean is close to its population mean, and it does not account for biased sampling, dependence, or differences in the populations being compared. It is a justification for treating small-group means as less stable on average, not a rule that can certify any particular summary.

Counts also matter when combining groups. The overall mean is calculated from all observations, so group means contribute in proportion to their group sizes. It is generally not the unweighted average of the group means.

Worked Example: Recovering the Overall Mean

A hypothetical equipment review records repair times in two types of workshops. In Workshop A, the three times are 180, 200, and 220 minutes. In Workshop B, the nine times are 80, 90, 100, 100, 100, 110, 120, 100, and 100 minutes. Find the two group means and the overall mean. Explain why simply averaging the group means gives the wrong overall result.

Method. Calculate each group’s total and count. Workshop A has $n_A=3$, total $180+200+220=600$, and mean $600/3=200$ minutes. Workshop B has $n_B=9$, total $80+90+100+100+100+110+120+100+100=900$, and mean $900/9=100$ minutes.

There are $3+9=12$ observations in all, with a total time of $600+900=1{,}500$ minutes. The overall mean is $1{,}500/12=125$ minutes. Equivalently, weight each group mean by its count:

\bar{x}_{\text{overall}} =\frac{3(200)+9(100)}{3+9} =\frac{600+900}{12} =125.

The unweighted average of the two group means is $(200+100)/2=150$ minutes. That calculation gives each workshop type equal weight, even though Workshop B contributes three times as many observations. It is not the mean of the 12 recorded times.

Common mistake avoided. Do not average group means equally unless the groups have equal counts or the question explicitly asks for an equal-weight average of group summaries. For the mean across all observations, use the counts as weights.

Counts and Group Proportions

For a binary outcome, a group proportion is the number of observed successes divided by the number of observations with a known outcome. A proportion of 0.80 could mean four successes out of five or 800 out of 1,000. The reported rate is the same, but the smaller denominator makes it more sensitive to individual outcomes.

Definition: For a group with $n$ observed binary outcomes and $x$ successes, the observed success proportion is $\hat{p}=x/n$. The denominator must be the number of outcomes actually observed, not simply the number of rows assigned to the group when outcomes are missing.

As a rough measure of how much a sample proportion can vary, suppose each outcome is an independent Bernoulli observation with the same success probability $p$. Then the estimated standard error of $\hat{p}$ is $\sqrt{\hat{p}(1-\hat{p})/n}$, using the observed proportion as an estimate of $p$. This approximation can be unreliable for small counts, particularly when the proportion is near zero or one. It is useful here as a reminder that both the rate and its denominator matter, not as a stand-alone significance test.

Worked Example: The Same Proportion with Different Counts

Two hypothetical outreach groups record whether each participant completes a follow-up. In Group A, 4 of 5 participants complete it. In Group B, 60 of 100 complete it. Calculate each observed proportion and its approximate standard error, then state what the comparison supports.

Method. Divide successes by observed outcomes. Group A has $n=5$, $x=4$, and $\hat{p}_A=4/5=0.80$, or 80%. Its estimated standard error is $\sqrt{0.80(1-0.80)/5}=\sqrt{0.16/5}=\sqrt{0.032}\approx0.1789$.

Group B has $n=100$, $x=60$, and $\hat{p}_B=60/100=0.60$, or 60%. Its estimated standard error is $\sqrt{0.60(1-0.60)/100}=\sqrt{0.24/100}=\sqrt{0.0024}\approx0.0490$. The observed rates differ by $0.80-0.60=0.20$, or 20 percentage points.

Group A’s approximate standard error is much larger, and the small sample makes the normal approximation especially questionable. The summaries show a 20-point difference in these recorded groups, but they do not establish a reliable population difference. Further analysis would require an appropriate method for the data and the question being asked.

Common mistake avoided. The 80% rate should not be treated as more convincing simply because it is larger. Its denominator is five, compared with 100 for the other rate, so the observed proportion in Group A can change sharply when one participant’s outcome changes.

Practical Checks Before Comparing Groups

Before interpreting grouped summaries, check that each statistic is paired with the count that actually supports it. For a numeric outcome, that is typically the number of nonmissing values used in the mean or median. For a binary outcome, report the successes and the observed denominator as well as the proportion. If group counts are unequal, say so plainly rather than letting a table of means imply equal amounts of evidence.

Count the rows in each group.
Confirm how many observations were assigned to each predictor category.

Check the outcome’s availability.
Find the number with a recorded outcome; it may be smaller than the group’s row count.

Pair each summary with its denominator.
Report counts beside means or medians, and successes beside proportions for binary outcomes.

Describe the difference at the right strength.
State what the observed summaries show, and avoid treating group size or a descriptive difference as proof of a population or causal effect.

A common reporting error is to show only the group means and describe the larger one as a dependable difference. The correction is to display the counts, check whether missing outcomes changed the usable sample sizes, and qualify the comparison when one or more groups are small. These checks make exploratory summaries more informative without turning them into a formal model or test.

Key takeaway: A grouped statistic is incomplete without the count behind it. Use the usable outcome count as the denominator, weight group means by group size when calculating an overall mean, and interpret small-group summaries cautiously.

Check Your Understanding

For each problem, calculate the requested summaries and use the counts to interpret them.

A small repair category has recorded times of 40 and 60 minutes. A larger category has times of 30, 40, 40, 50, 50, 60, 70, and 80 minutes. Find both group means and the overall mean. (Answer: Small group $n=2$, mean 50 minutes; larger group $n=8$, mean 52.5 minutes; overall mean 52 minutes.)
A group contains 18 records, but five have no recorded outcome. A mean is calculated after excluding missing outcomes. What is the group size, what is the usable sample size, and which count belongs with the mean? (Answer: Group size 18; usable sample size 13; report 13 as the count supporting the mean.)
Two groups have observed successes of 2 out of 4 and 18 out of 30. Calculate the proportions and approximate standard errors using $\sqrt{\hat{p}(1-\hat{p})/n}$. Which group’s proportion is less stable by this rough measure? (Answer: Proportions 0.50 and 0.60; standard errors 0.2500 and approximately 0.0894; the first group’s proportion is less stable.)
Group A has $n=2$ and mean 70; Group B has $n=8$ and mean 50. Calculate the overall mean and the unweighted average of the group means. (Answer: Overall mean $(2(70)+8(50))/10=54$; unweighted average $60$, which is not the overall mean.)
A binary outcome is observed for 3 of 5 records in one group and 48 of 80 records in another. Calculate both observed proportions. Explain why the equal proportions do not mean the summaries have equal precision. (Answer: Both proportions are 0.60; the first is based on 5 observed outcomes and is more sensitive to individual outcomes than the second, based on 80.)

What You'll Learn