Why ANOVA Calculations Go Wrong
A one-way analysis of variance (ANOVA) compares the means of \(k\) groups by relating variation among the group means to variation within the groups. The arithmetic is compact, but several quantities that look similar have different roles. The grand mean weights each group mean by its sample size; the degrees of freedom depend on the number of groups and total observations; and each mean square uses the degrees of freedom for its own source of variation.
A small error at any of these stages can change the F statistic. The checks in this tutorial are especially useful when group sizes are unequal or when an ANOVA table is reconstructed from summaries. Assume throughout that the observations come from independent samples, with normal populations having a common variance under the null hypothesis of equal group means. The usual one-way fixed-effects F test uses these assumptions; if they are not appropriate, its standard reference distribution need not apply.
A Short Derivation: The Grand Mean and the Sum-of-Squares Partition
The grand mean is the mean of all \(N\) observations, not generally the simple average of the \(k\) group means. Since group \(i\) contributes \(n_i\) observations with total \(n_i\bar{x}_i\), the total of all observations is \(\sum_i n_i\bar{x}_i\). Thus
For any observation, split its difference from the grand mean into a within-group difference and a group-mean difference:
Squaring and summing over all observations produces a cross-product term. For each group, that term is zero because deviations from that group’s mean sum to zero: \(\sum_j(x_{ij}-\bar{x}_i)=0\). Therefore the total squared deviation separates into within-group and between-group components:
The factor \(n_i\) in \(SS_B\) matters: a group mean based on more observations represents more observations in the total sum of squares. The partition also provides a useful audit: \(SS_T\) must equal \(SS_B+SS_W\), up to rounding.
Degrees of Freedom and Mean Squares
For \(k\) groups and \(N\) observations, the between-group, within-group, and total degrees of freedom are respectively \(k-1\), \(N-k\), and \(N-1\). They satisfy \((k-1)+(N-k)=N-1\). The mean square for a source is its sum of squares divided by that source’s degrees of freedom. The F statistic is the between-group mean square divided by the within-group mean square.
Under the null hypothesis that all group means are equal, and when the one-way ANOVA assumptions hold, this statistic is compared with an \(F\) distribution having numerator degrees of freedom \(k-1\) and denominator degrees of freedom \(N-k\). A mean square is not an additional sum of squares: it is a sum of squares scaled by its matching degrees of freedom.
Worked Example: Unequal Group Sizes and the Wrong Grand Mean
A hypothetical study compares a measured response across three treatment groups. The group sizes are \(2,4,3\), and the respective group means are \(4,7,10\). The within-group sums of squares are \(6,12,8\). Find the grand mean, \(SS_B\), \(SS_W\), and the F statistic. Identify the error made by using the unweighted average of the group means.
Method. Weight each group mean by its group size, then use the one-way ANOVA sums-of-squares and degrees-of-freedom formulas.
Step 1: Calculate the grand mean. The total sample size is \(N=2+4+3=9\), and the total of all observations is \(2(4)+4(7)+3(10)=8+28+30=66\). Therefore \(\bar{x}_{..}=66/9=22/3\), or about \(7.333\).
Step 2: Calculate \(SS_B\). The weighted contributions are \(2(4-22/3)^2=2(100/9)=200/9\), \(4(7-22/3)^2=4(1/9)=4/9\), and \(3(10-22/3)^2=3(64/9)=192/9\). Thus \(SS_B=(200+4+192)/9=396/9=44\).
Step 3: Calculate \(SS_W\) and the degrees of freedom. Add the given within-group sums: \(SS_W=6+12+8=26\). There are \(k=3\) groups, so \(df_B=3-1=2\), \(df_W=9-3=6\), and \(df_T=9-1=8\). The degrees of freedom check: \(2+6=8\).
Step 4: Calculate the mean squares and F. \(MS_B=44/2=22\). \(MS_W=26/6=13/3\). Hence \(F=22/(13/3)=66/13\approx5.0769\).
Common mistake and correction. The unweighted average of the three means is \((4+7+10)/3=7\), not the grand mean because the group sizes differ. Substituting \(7\) into the weighted between-group formula gives \(2(4-7)^2+4(7-7)^2+3(10-7)^2=18+0+27=45\), rather than \(44\). Use the weighted grand mean. As a check, \(SS_T=SS_B+SS_W=44+26=70\).
Worked Example: Mixing Degrees of Freedom
A hypothetical one-way ANOVA has \(k=4\) groups and \(N=15\) observations. A calculation gives \(SS_B=36\) and \(SS_W=44\). Construct the degrees-of-freedom and mean-square entries, and find \(F\). A draft calculation divides both sums of squares by the total degrees of freedom. Diagnose and correct it.
Method. First assign degrees of freedom by source; then divide each sum of squares by its own degrees of freedom.
Step 1: Find the degrees of freedom. \(df_B=k-1=4-1=3\). \(df_W=N-k=15-4=11\). \(df_T=N-1=15-1=14\). The component degrees of freedom add correctly: \(3+11=14\).
Step 2: Find the mean squares. \(MS_B=36/3=12\). \(MS_W=44/11=4\). The check on the sum of squares gives \(SS_T=36+44=80\).
Step 3: Find the F statistic. \(F=12/4=3\), with numerator degrees of freedom \(3\) and denominator degrees of freedom \(11\).
Common mistake and correction. Dividing both sums of squares by \(df_T=14\) gives \(36/14\) and \(44/14\); their ratio is \(36/44\approx0.8182\). This is not the ANOVA F statistic. Even though a common divisor cancels in that ratio, it does not produce the required ratio of source-specific mean squares. Divide \(SS_B\) by \(df_B\), and \(SS_W\) by \(df_W\), before forming \(F\). The corrected statistic is \(3\).
Worked Example: Omitting the Within-Group Size Weights
A hypothetical study has three groups with sizes \(3,5,4\), means \(5,7,8.5\), and sample variances \(2,3,4\), respectively. Calculate \(SS_B\), \(SS_W\), and \(F\). A draft calculation uses the sum of the three sample variances as \(SS_W\).
Method. Use the group sizes to obtain the grand mean and between-group sum of squares. Convert each sample variance into a within-group sum of squares using its own group’s degrees of freedom.
Step 1: Calculate the grand mean. Here \(N=3+5+4=12\), and the weighted total is \(3(5)+5(7)+4(8.5)=15+35+34=84\). Thus \(\bar{x}_{..}=84/12=7\).
Step 2: Calculate \(SS_B\). The group contributions are \(3(5-7)^2=12\), \(5(7-7)^2=0\), and \(4(8.5-7)^2=4(2.25)=9\). Therefore \(SS_B=12+0+9=21\).
Step 3: Calculate \(SS_W\). A sample variance in group \(i\) is the group’s sum of squared deviations divided by \(n_i-1\). Therefore its contribution to \(SS_W\) is \((n_i-1)s_i^2\): \( (3-1)(2)=4\), \((5-1)(3)=12\), and \((4-1)(4)=12\). Hence \(SS_W=4+12+12=28\).
Step 4: Complete the F calculation. \(df_B=3-1=2\) and \(df_W=12-3=9\). Thus \(MS_B=21/2=10.5\), \(MS_W=28/9\approx3.1111\), and \(F=10.5/(28/9)=94.5/28=3.375\).
Common mistake and correction. Adding the sample variances gives \(2+3+4=9\), but \(9\) is not \(SS_W\): sample variances have already been divided by their group degrees of freedom. Reconstruct the within-group sum of squares with \((n_i-1)s_i^2\) for each group. As a final partition check, \(SS_T=21+28=49\).
A Practical Error-Checking Sequence
When checking a one-way ANOVA calculation, work in a fixed order rather than trying to repair the F statistic first. A correct final ratio does not guarantee that the component calculations or degrees of freedom are right. Use these checks:
Calculate \(N=\sum_i n_i\) and identify the number of groups \(k\). Do not use \(k\) in place of \(N\), or vice versa.
For unequal group sizes, calculate \(\sum_i n_i\bar{x}_i/N\). An unweighted average is appropriate only when all group sizes are equal.
Use group-size weights in \(SS_B\), and ensure within-group variances are converted to sums of squares with their group degrees of freedom. Verify \(SS_T=SS_B+SS_W\).
Use \(k-1\) for between groups and \(N-k\) for within groups; verify they add to \(N-1\).
Calculate \(MS_B\) and \(MS_W\) separately, then divide \(MS_B\) by \(MS_W\). The numerator is the between-group mean square.
Common Traps to Remember
- Using the unweighted grand mean. This generally fails when group sizes differ. Weight every group mean by its sample size.
- Using the wrong degrees of freedom. The within-group degrees of freedom are \(N-k\), not \(N-1\); \(N-1\) belongs to the total.
- Dividing a sum of squares by the wrong degrees of freedom. \(SS_B\) uses \(k-1\), while \(SS_W\) uses \(N-k\).
- Adding sample variances as if they were sums of squares. Multiply each sample variance by \(n_i-1\) before adding contributions to \(SS_W\).
- Reversing the F ratio. The one-way ANOVA statistic is \(MS_B/MS_W\), not \(MS_W/MS_B\).
These checks support accurate calculations and interpretation of an ANOVA table, a skill aligned with the SOA VEE Mathematical Statistics syllabus. They do not replace checking whether the sampling design and model assumptions support the test.
Check Your Understanding
For each problem, calculate the requested ANOVA quantity and use the stated checks to catch a likely calculation error.
- A hypothetical study has two groups of sizes \(2\) and \(3\), with means \(6\) and \(9\). Find the grand mean, using the correct weights. (Answer: \(7.8\))
- An ANOVA has \(k=4\), \(N=18\), \(SS_B=27\), and \(SS_W=56\). Find the three degrees of freedom, both mean squares, and \(F\). (Answer: \(df_B=3,\ df_W=14,\ df_T=17;\ MS_B=9,\ MS_W=4,\ F=2.25\))
- Three groups have sizes \(2,4,2\) and means \(3,5,8\). Calculate the weighted grand mean and \(SS_B\). (Answer: \(\bar{x}_{..}=5.25,\ SS_B=25.5\))
- Three groups have sizes \(4,3,5\) and sample variances \(2,6,1\). Find \(SS_W\) and \(df_W\). (Answer: \(SS_W=3(2)+2(6)+4(1)=22,\ df_W=12-3=9\))
- An ANOVA has \(k=5\), \(N=20\), \(SS_B=40\), and \(SS_W=60\). Find \(MS_B\), \(MS_W\), and \(F\); state which degrees of freedom go with the F statistic. (Answer: \(MS_B=40/4=10,\ MS_W=60/15=4,\ F=2.5\), with numerator \(df=4\) and denominator \(df=15\))