What Between-Group Variation Measures
In a one-way analysis of variance, observations are organized into groups defined by one categorical factor. The between-group sum of squares measures how far the group means lie from the overall mean, allowing for the number of observations in each group. A group mean based on many observations contributes more than the same-sized difference in a small group.
This quantity describes variation associated with differences among the observed group means. It does not, by itself, establish that group membership caused those differences. Nor does calculating it require a probability model. A distributional model becomes relevant when using sample results for inference, not for computing this descriptive sum of squares.
The grand mean must be calculated using the observations’ group sizes:
If group totals \(T_j\) are given instead of group means, use \(\bar{x}_j=T_j/n_j\), or equivalently calculate \(\bar{x}=(\sum_j T_j)/N\). The units of \(SS_B\) are the square of the response’s units. For example, if the response is measured in dollars, \(SS_B\) is measured in dollars squared.
Why the Group-Size Weight Appears
Each observation in group \(j\) has the same group-level displacement, \(\bar{x}_j-\bar{x}\). If that displacement is squared and counted once for each of the \(n_j\) observations in the group, its total contribution is \(n_j(\bar{x}_j-\bar{x})^2\). Adding these contributions over the groups gives \(SS_B\).
This interpretation also explains why a simple average of the group means is generally not the grand mean. The grand mean averages observations, not groups: a group containing 20 observations must count more than a group containing 3 observations. As a useful arithmetic check, the weighted deviations from the grand mean sum to zero:
A zero weighted sum of deviations does not mean \(SS_B=0\): the deviations are squared in \(SS_B\). It does confirm that the grand mean balances the group means when weighted by their sizes.
Worked Example: Three Groups With Different Sizes
A hypothetical benefits team compares the mean processing time, in hours, for three independent workflow categories. The category summaries are: Category A, \(n_1=4\) and \(\bar{x}_1=12\); Category B, \(n_2=3\) and \(\bar{x}_2=15\); Category C, \(n_3=5\) and \(\bar{x}_3=18\). Calculate the grand mean and \(SS_B\), and interpret the result. No interest rate, mortality assumption, or probability distribution is specified or needed for this descriptive calculation.
Method. First find the weighted grand mean, then multiply each squared group-mean deviation by its group size. This avoids treating the three categories as equally large.
The total number of observations is \(N=4+3+5=12\). The weighted sum of group means is \(4(12)+3(15)+5(18)=48+45+90=183\). Therefore,
Now calculate each contribution. Category A has deviation \(12-15.25=-3.25\), so its contribution is \(4(-3.25)^2=4(10.5625)=42.25\). Category B has deviation \(15-15.25=-0.25\), so its contribution is \(3(-0.25)^2=3(0.0625)=0.1875\). Category C has deviation \(18-15.25=2.75\), so its contribution is \(5(2.75)^2=5(7.5625)=37.8125\).
Check and interpretation. The weighted deviations are \(4(-3.25)=-13\), \(3(-0.25)=-0.75\), and \(5(2.75)=13.75\); their sum is \(0\). The between-group sum of squares is \(80.25\) hours squared. It summarizes the group means’ weighted separation from the overall mean; it is not itself an average difference in hours.
Common mistake avoided: Averaging \(12,15,18\) without weights gives \(15\), not the observation-level grand mean of \(15.25\). Using that unweighted value would calculate a different quantity.
Calculating From Unequal Group Sizes
Unequal sizes are ordinary, not exceptional. The formula works without modification: use the actual \(n_j\) for each group, both when finding the grand mean and when weighting its squared deviation. Do not give the group means equal influence simply because there are several groups.
Worked Example: Unequal Sizes and the Unweighted-Mean Trap
A hypothetical survey records the average annual number of service calls for three types of equipment. Type A has \(n_1=2\) units with mean \(20\) calls; Type B has \(n_2=5\) units with mean \(24\) calls; Type C has \(n_3=3\) units with mean \(30\) calls. Find the grand mean and \(SS_B\). The summaries are descriptive; no probability distribution or actuarial interest or mortality assumption is part of the calculation.
Method. Weight each group mean by its count to find the overall mean, then compute the weighted squared deviations.
The total number of units is \(N=2+5+3=10\). Their combined total number of calls is \(2(20)+5(24)+3(30)=40+120+90=250\). Thus,
The group contributions are as follows. Type A is \(2(20-25)^2=2(25)=50\). Type B is \(5(24-25)^2=5(1)=5\). Type C is \(3(30-25)^2=3(25)=75\). Hence,
Check and interpretation. The weighted deviations are \(2(-5)=-10\), \(5(-1)=-5\), and \(3(5)=15\), which sum to zero. The group means’ weighted squared separation from the overall mean is \(130\) calls squared.
Common mistake shown and corrected: The unweighted mean of the three group means is \((20+24+30)/3=74/3\). It is not the grand mean because the group sizes differ. If it is incorrectly used, the resulting sum is \(2(20-74/3)^2+5(24-74/3)^2+3(30-74/3)^2=1180/9\), or about \(131.1111\). The correct grand mean is \(25\), giving \(SS_B=130\).
Using Group Totals Instead of Group Means
Sometimes a table reports each group’s count and total response rather than its mean. Convert each total to a mean before finding its deviation, or find the grand mean directly from the combined total. Keeping the count attached to its corresponding total is essential.
Worked Example: Starting With Group Totals
A hypothetical administrator compares weekly completion counts in three teams. Team A has \(n_1=4\) workers and a total of \(28\) completed items; Team B has \(n_2=2\) workers and a total of \(18\); Team C has \(n_3=4\) workers and a total of \(40\). Calculate the grand mean and the between-group sum of squares. No distributional, interest-rate, or mortality assumptions are needed to calculate this summary.
Method. Convert each team total to a team mean, then apply the weighted formula.
The team means are \(28/4=7\), \(18/2=9\), and \(40/4=10\) items per worker. The total count is \(N=4+2+4=10\), and the combined response total is \(28+18+40=86\). Therefore, the grand mean is \(86/10=8.6\) items per worker.
Team A contributes \(4(7-8.6)^2=4(-1.6)^2=4(2.56)=10.24\). Team B contributes \(2(9-8.6)^2=2(0.4)^2=2(0.16)=0.32\). Team C contributes \(4(10-8.6)^2=4(1.4)^2=4(1.96)=7.84\). Adding them gives
Check and interpretation. The weighted deviations are \(4(-1.6)=-6.4\), \(2(0.4)=0.8\), and \(4(1.4)=5.6\), summing to zero. The between-group sum of squares is \(18.40\) items squared. It reflects the extent to which these three team means differ from the overall worker-level mean.
Common mistake avoided: The group totals \(28,18,40\) are not the group means. Substituting totals directly into the formula would compare quantities based on different group sizes and would not calculate the between-group sum of squares.
Interpretation, Checks, and Limits
The formula is always nonnegative because it adds group sizes multiplied by squared deviations. It equals zero exactly when every group with a positive size has a mean equal to the grand mean. Because the grand mean is a weighted average of the group means, that occurs when all the group means are equal.
A larger \(SS_B\) indicates greater weighted separation among the observed group means, but its numerical size depends on the measurement scale and sample size. It is not a scale-free measure of importance and should not be interpreted alone as evidence of statistical significance or causation. In a one-way ANOVA, it is one part of the total variation partition; this tutorial focuses on its calculation and meaning, not on the other sum of squares or the test statistic.
Before finalizing a calculation, verify the following:
- The grand mean is based on the group sizes, not the unweighted average of the means unless all group sizes are equal.
- Each deviation is group mean minus grand mean, and the deviation is squared before multiplication by group size.
- The group size used as a weight belongs to the same group as the mean in that term.
- The weighted deviations sum to zero, subject to rounding.
- The final units are squared response units, and the result is nonnegative.
This calculation is aligned with the SOA VEE Mathematical Statistics syllabus. For computation alone, no normality, equal-variance, independence, interest-rate, or mortality assumption is required. Those considerations matter when the sums of squares are used as part of an inferential analysis rather than simply calculated from the supplied group summaries.
Check Your Understanding
For each problem, calculate the weighted grand mean first, then use it to find the between-group sum of squares.
- A hypothetical study reports two groups: \(n_1=3\) with mean \(5\), and \(n_2=4\) with mean \(8\). Find the grand mean and \(SS_B\). (Answer: Grand mean \(=47/7\); \(SS_B=108/7\approx15.4286\).)
- Three hypothetical repair categories have sizes \(2,3,5\) and means \(10,12,16\), respectively. Find the grand mean and \(SS_B\). (Answer: Grand mean \(=13.6\); \(SS_B=62.4\).)
- Three hypothetical teams have sizes and total outputs of \(5\) and \(35\), \(3\) and \(30\), and \(2\) and \(25\), respectively. Find each group mean, the grand mean, and \(SS_B\). (Answer: Group means \(=7,10,12.5\); grand mean \(=9\); \(SS_B=47.5\).)
- Two hypothetical production lines have \(4\) observations with mean \(100\) and \(4\) observations with mean \(110\); a third has \(2\) observations with mean \(125\). Calculate the grand mean and \(SS_B\), in squared units of the response. (Answer: Grand mean \(=109\); \(SS_B=840\).)
- Three hypothetical service groups have sizes \(2,6,4\) and means \(3,5,8\). A candidate uses the unweighted mean of the group means. Calculate the correct grand mean and \(SS_B\), and explain why that candidate’s weighting is inappropriate. (Answer: Grand mean \(=17/3\); \(SS_B=116/3\approx38.6667\); the unweighted mean gives equal weight to groups with unequal sizes.)