Graphical Diagnostics for Two-Group T Procedures

Why Look at Each Group Separately?

A two-sample t procedure uses sample means and standard deviations to draw conclusions about the difference between two population means. Those summaries do not show the full shape of the observed data. Two groups can have similar means but very different distributions: one may be fairly balanced, while the other has a long tail or an unusually extreme value. Looking at a graph of each group helps identify features that could make a t-based conclusion less reliable or less precise.

The key is to inspect the groups separately as well as compare them. A graph of the difference between the sample means cannot reveal whether one group is strongly skewed or contains an outlier. For a fair visual comparison, use the same measurement units and, where appropriate, the same horizontal scale and bin widths for both groups.

Conditions to keep in mind: Two-sample t inference concerns a difference in population means, \(\mu_1-\mu_2\), based on two independent groups of quantitative observations. The study design must support the intended inference, and observations within each sample should be independent. Graphs can help assess distribution shape and unusual values, but they cannot establish independence or show that a sample was randomly selected. For small samples, each group should be reasonably free of severe skewness and extreme outliers; larger samples can make mean-based inference more tolerant of moderate non-normality, but no sample-size rule makes every distribution harmless.

What the Main Graphs Reveal

A dotplot displays the individual observations along a number line, stacking dots when values repeat. It is particularly useful for small or moderate samples because the actual data values remain visible. A histogram groups values into bins and shows how many fall in each bin. It is useful for seeing broad shape in larger samples, although the appearance can change when bin width or bin boundaries change. A boxplot summarizes the center and spread using the median, quartiles, and whiskers. It makes group comparisons compact, but it hides much of the detailed shape.

In any of these displays, look for the overall balance of the data, tails that extend farther on one side, isolated points, and differences in spread. Skewness describes asymmetry: a right-skewed distribution has a longer or more extended right tail, while a left-skewed distribution has a longer or more extended left tail. Spread describes how dispersed the observations are; it is not the same thing as skewness.

A common boxplot convention draws a point separately when it lies below \(Q_1-1.5(IQR)\) or above \(Q_3+1.5(IQR)\), where \(Q_1\) and \(Q_3\) are the first and third quartiles and \(IQR=Q_3-Q_1\). Such a point is flagged for investigation, not automatically declared erroneous or removed. Boxplot whiskers and outlier markings can follow different conventions, so identify the convention being used.

Worked Example: A Flagged Value in a Small Group

Suppose two hypothetical groups each contain eight independent measurements in the same units. Group A has values \(8, 9, 9, 10, 10, 11, 12, 28\). Group B has values \(7, 8, 9, 10, 10, 11, 12, 13\). What might separate dotplots and boxplots tell us before a t procedure is interpreted?

Inspect the dotplots. Group B is spread fairly evenly around 10. Group A has seven values between 8 and 12, then a large gap before 28. The dotplot makes that isolated high observation immediately visible.

Check the boxplot rule for Group A. Its median is \((10+10)/2=10\). The lower half is \(8,9,9,10\), so \(Q_1=(9+9)/2=9\). The upper half is \(10,11,12,28\), so \(Q_3=(11+12)/2=11.5\). Thus \(IQR=11.5-9=2.5\), and the upper fence is \(11.5+1.5(2.5)=15.25\). Since \(28>15.25\), the value 28 is flagged by this convention.

Compare centers and consider sensitivity. The mean of Group A is \(97/8=12.125\), while the mean of Group B is \(80/8=10\). Group A’s median is still 10. Its mean is pulled upward by the high observation, so the choice of mean as the center of interest deserves attention. The flagged point is not automatically a mistake: its source and measurement should be checked, and a legitimate observation should not be discarded merely because it is extreme.

How Shape and Outliers Affect Mean Inference

The Welch t procedure uses the difference between sample means, with standard error \(SE=\sqrt{s_1^2/n_1+s_2^2/n_2}\). Its reference distribution is a t distribution with an estimated number of degrees of freedom. The procedure does not require the two population variances to be equal. However, the quality of its approximation can still be affected by the shape of the data, particularly when sample sizes are small.

A strongly skewed sample may be a poor guide to the sampling behavior of its mean when there are only a few observations. An extreme observation can also affect both the sample mean and the sample standard deviation. Since the mean difference is the numerator of the t statistic and the standard error depends on sample standard deviations, an extreme value can change both parts of the calculation. The direction and size of the net effect cannot be determined from the label “outlier” alone.

The mean’s response to changing one value is straightforward. If one observation \(x\) in a sample of size \(n\) is replaced by \(x+\delta\), the new mean is \((\sum x_i+\delta)/n\), so the change in the mean is \(\delta/n\). The original distance of \(x\) from the other observations does not by itself determine this change; what matters is the size of the replacement change, \(|\delta|\). An observation far from the rest may nevertheless be influential because it can contribute to the sample’s observed shape and variability.

\bar{x}_{\mathrm{new}}-\bar{x}_{\mathrm{old}} =\frac{\sum x_i+\delta}{n}-\frac{\sum x_i}{n} =\frac{\delta}{n}

This explains why visual diagnosis matters without turning it into a mechanical pass-or-fail test. A flagged observation or some asymmetry calls for scrutiny and context. It does not by itself prove that t inference is invalid. Conversely, a graph that looks reasonably balanced does not prove that the sample is independent or representative.

Worked Example: Histograms Show Different Shapes

Consider two hypothetical independent groups with 40 measurements each. The following table gives the number of observations at each recorded value; using these counts as histogram bars makes the shape comparison explicit.

Value	Group A count	Group B count
0	2	2
1	6	5
2	10	9
3	8	8
4	6	9
5	4	5
6	3	2
7	1	0

Check the sample means. The total for Group A is \(0(2)+1(6)+2(10)+3(8)+4(6)+5(4)+6(3)+7(1)=119\), giving \(\bar{x}_A=119/40=2.975\). The total for Group B is \(0(2)+1(5)+2(9)+3(8)+4(9)+5(5)+6(2)=120\), giving \(\bar{x}_B=120/40=3.000\).

Read the histograms. Group B’s counts are balanced around 3. Group A has many observations at 1 through 4 and a thinner extension to the right, including one value of 7; its histogram suggests some right-skewness. The means are close, but that does not make the shapes identical.

Relate the display to inference. With 40 observations in each group, moderate asymmetry may be less concerning than it would be with very small samples, but the graph alone cannot certify the approximation. Check for unusually extreme measurements and interpret the shape in context. There is no universal sample-size cutoff that guarantees safety for every skewed distribution.

Reading Spread and Comparing Groups

The vertical height of histogram bars concerns frequency; it should not be confused with variability on the measurement scale. To compare spread, examine the horizontal extent and concentration of the observations, or compare quantities such as the interquartile ranges and sample standard deviations. Boxplots are useful for a quick spread comparison, but dotplots or histograms may reveal clusters and gaps that a boxplot conceals.

In Welch inference, the standard error combines the estimated variability in each sample. Greater spread in one group generally contributes more to the standard error, other quantities held fixed, and therefore reduces the precision of the estimated mean difference. A graph can reveal this difference in spread, but does not by itself say whether it is important enough to change a conclusion. Unequal spread is not the same issue as skewness or an outlier.

Worked Example: Similar Centers but Different Spread

Suppose two hypothetical groups each have ten independent measurements, in score points. Group A is \(18,19,19,20,20,20,20,21,21,22\). Group B is \(12,14,16,18,19,21,22,24,26,28\). Compare their centers and spreads, then examine the contribution to the Welch standard error.

Compare the means. Group A sums to 200 and Group B also sums to 200, so \(\bar{x}_A=\bar{x}_B=200/10=20\) score points. The estimated difference is 0 score points.

Compare the boxplot summaries. Using the median-of-halves convention, Group A has \(Q_1=19\) and \(Q_3=21\), so its IQR is 2 points. Group B has \(Q_1=16\) and \(Q_3=24\), so its IQR is 8 points. Both medians are 20, but Group B is much more spread out. Neither group has a value beyond its 1.5-IQR fences: for A the fences are 16 and 24, and for B they are 4 and 36.

Calculate the Welch standard error. For Group A, the sum of squared deviations from 20 is 12, so \(s_A^2=12/(10-1)=1.333\) score-points squared. For Group B, that sum is 242, so \(s_B^2=242/9=26.89\) score-points squared. Therefore,

SE=\sqrt{\frac{s_A^2}{n_A}+\frac{s_B^2}{n_B}} =\sqrt{\frac{1.333}{10}+\frac{26.89}{10}} =\sqrt{2.822} \approx 1.680\text{ score points}

The broad spread in Group B contributes much more to the standard error than the narrow spread in Group A. The estimated difference is zero here, so the Welch t statistic is \(0/1.680=0\). The example shows why similar centers do not imply similar precision; in a different data set, a nonzero difference would be judged relative to the standard error.

A Practical Diagnostic Sequence

Before interpreting a two-sample t result, use a consistent sequence. Make separate plots for the two groups, and keep axes and scales comparable so that differences are not artifacts of presentation. Then describe each group’s shape, look for possible outliers or clusters, and compare spread. Connect any concern to the sample sizes and to the measurement context rather than making an automatic decision from one graph.

Choose a useful display.
Use dotplots when individual observations matter, histograms to see broad shape in larger samples, and boxplots for compact comparisons of center and spread.

Inspect each group on its own.
Note symmetry or skewness, gaps or clusters, isolated values, and the degree of spread.

Compare the groups fairly.
Use the same units and a suitable common scale, then assess whether the groups differ in shape or variability.

Judge implications in context.
Consider sample sizes and whether unusual observations are plausible. A graph informs the assessment of t inference; it does not replace checking the study design.

Common Pitfalls

Removing every flagged point. A boxplot flag is a prompt to check the observation and its context, not a deletion rule. Correct data that are extreme may need to remain in the analysis.
Assuming a graph proves the assumptions. A plausible-looking histogram cannot establish independence, random sampling, or representativeness.
Comparing plots made on different scales. Different axis ranges or histogram bin widths can exaggerate or hide differences. Use comparable settings before drawing conclusions.
Treating all non-normality as equally serious. A little asymmetry and a severe tail with an influential extreme value are not equivalent, and their implications depend partly on sample size.
Judging variability by the means alone. Groups with identical sample means can have very different spreads and therefore contribute differently to the standard error.

Key takeaway: Separate plots help assess the shape and spread of each group before two-sample t inference. Skewness and extreme observations can threaten the approximation, especially in small samples; unequal spread affects precision. Graphs guide judgment, but they do not establish design conditions or automatically determine whether an observation should be excluded.

Check Your Understanding

Answer each question using the diagnostic ideas in this tutorial.

Why should the two groups be plotted separately rather than relying only on their sample means?
A point lies above a boxplot’s upper 1.5-IQR fence. What does that tell you, and what does it not tell you?
How can severe right-skewness affect the reliability of t inference when a group is small?
What information about a group’s variability can a boxplot show, and what detail might it conceal compared with a dotplot?
Can a histogram establish that observations are independent? Explain.