Sum of Squares Calculator

The sum of squares is the total of the squared differences between observations and the point they are measured from. Change the point and the total changes, which is the whole difficulty in the subject: the name covers at least three different quantities and no default reference is assumed.

The form that matters in statistics is the one taken about the mean, SS = Σ(xᵢ - x̄)². It is the smallest total any reference can produce, it carries squared units, and dividing it by the degrees of freedom turns it into a variance. Everything else, including the raw Σx² and the ANOVA split, is either bigger or a partition of it.

Order of operations never matters. The mean is subtracted from each value, the result is squared, and the squares are added. Squaring is what makes the measure work: without it, deviations above and below the mean would cancel and spread would look like none.

Shifting the reference from the mean to c adds exactly n(x̄ - c)² to the total. That single term is the parallel-axis rule, and it explains why two analysts can report different sums of squares for identical data without either of them making an arithmetic mistake.

Rearranging the same algebra gives the computational form SS = Σxᵢ² - (Σx)²/n, which needs no mean. It is faster to compute and much worse conditioned: both terms on the right are typically far larger than their difference, so the subtraction cancels away significant digits.

In one-way ANOVA the total sum of squares about the grand mean decomposes exactly into the between-group part Σnᵢ(x̄ᵢ - x̄)² and the within-group part ΣᵢΣ(x - x̄ᵢ)². Dividing by k - 1 and n - k respectively gives mean squares whose ratio is the F statistic.

The per-value table is the real result. Being able to see that one observation contributes 9 of a total of 32 tells you something no total can, and it is how an outlier is caught before it reaches a report.

A total of zero is a fact, not a failure. It means every value is identical, and a variance of zero follows immediately rather than being an error to explain away.

Between and within sums of squares answer different questions. A large between figure means the groups differ; a large within figure means the data are noisy regardless of grouping. Neither one on its own is a test result.

The F ratio is a ratio, not a verdict. It needs a comparison against an F distribution with the same two degrees of freedom before it can be called significant.

• The reference point. Moving it off the mean adds n(x̄ - c)², so the same data can produce almost any total.
• The size of the numbers. Large values push Σx² and (Σx)²/n close together and hollow out the computational form.
• Outliers. A single far value is squared, so its contribution grows with the square of its distance rather than linearly.
• Group sizes in the ANOVA split. Between-group sums of squares weight each group mean by its size, so one large group dominates the total.

• Checking a variance or standard deviation someone else has reported.
• Finding how much of a data set one unusual value is responsible for.
• Building or verifying a one-way ANOVA table by hand.
• Comparing a sum of squares against a reference such as zero, a baseline, or a specification limit.

• The observations are assumed independent. Repeated measurements on the same subject inflate the total and make every downstream variance and F ratio optimistic.
• Nothing is assumed about the shape of the data, which is why the measure survives non-normal inputs, but a heavily skewed set still makes the sum of squares sensitive to its tails.
• The computational form assumes nothing and yet still loses precision; the warning is about arithmetic conditioning, not about the data being wrong.
• The groups mode implements the one-way layout only. Repeated measures, blocking, and unbalanced designs all need their own sums of squares beyond the three reported here.