Mathwords logoMathwords

Bessel's Correction — Definition, Formula & Example

Bessel's correction is the practice of dividing by n−1n - 1 instead of nn when you compute the variance of a sample. Without it, sample variances would come out too small on average.

For independent observations x1,…,xnx_1, \dots, x_n from a population with variance σ2\sigma^2, the sum of squared deviations from the sample mean has expected value E[∑(xi−xˉ)2]=(n−1)σ2E\left[\sum (x_i - \bar{x})^2\right] = (n-1)\sigma^2. Dividing that sum by n−1n - 1 makes s2s^2 an unbiased estimator of σ2\sigma^2. Dividing by nn instead gives an estimator with expected value n−1nσ2\frac{n-1}{n}\sigma^2, which is biased low. The factor n−1n - 1 is the number of degrees of freedom left after estimating the mean from the data.

Key Formula

s2=1n−1∑i=1n(xi−xˉ)2,E[s2]=σ2s^2 = \frac{1}{n-1}\sum_{i=1}^{n}(x_i - \bar{x})^2, \qquad E\left[s^2\right] = \sigma^2
Where:
  • s2s^2 = Sample variance with Bessel's correction
  • nn = Number of observations in the sample
  • xix_i = The i-th observation
  • xˉ\bar{x} = Sample mean
  • σ2\sigma^2 = Population variance being estimated
  • E[ ⋅ ]E[\,\cdot\,] = Expected value, the long-run average over all possible samples

Worked Example

Problem: A population consists of the three values 2, 4, 6. List all 9 ordered samples of size 2 drawn with replacement, so the two draws are independent. Average the sample variances computed with n−1n - 1, then average them again computed with nn. Compare both averages to the population variance σ2\sigma^2.
Step 1: Find the population mean and the population variance. Divide by N=3N = 3, because this is the whole population.
μ=2+4+63=4,σ2=(−2)2+02+223=83\mu = \frac{2+4+6}{3} = 4, \qquad \sigma^2 = \frac{(-2)^2 + 0^2 + 2^2}{3} = \frac{8}{3}
Step 2: For a sample (a,b)(a, b), the mean is a+b2\frac{a+b}{2}. The two deviations are ±a−b2\pm\frac{a-b}{2}, so the sum of squared deviations is SS=(a−b)22\text{SS} = \frac{(a-b)^2}{2}. Sort the 9 samples by their SS value.
(2,2),(4,4),(6,6)→0;(2,4),(4,2),(4,6),(6,4)→2;(2,6),(6,2)→8(2,2),(4,4),(6,6) \to 0;\quad (2,4),(4,2),(4,6),(6,4) \to 2;\quad (2,6),(6,2) \to 8
Step 3: With n=2n = 2, dividing by n−1=1n - 1 = 1 leaves each SS unchanged. Dividing by n=2n = 2 halves it. Average each version over all 9 samples.
3(0)+4(2)+2(8)9=249=83,3(0)+4(1)+2(4)9=129=43\frac{3(0) + 4(2) + 2(8)}{9} = \frac{24}{9} = \frac{8}{3}, \qquad \frac{3(0) + 4(1) + 2(4)}{9} = \frac{12}{9} = \frac{4}{3}
Answer: The n−1n - 1 version averages exactly 83\frac{8}{3}, which equals σ2\sigma^2, so it is unbiased. The divide-by-nn version averages 43\frac{4}{3}. That is only n−1n=12\frac{n-1}{n} = \frac{1}{2} of the true variance, so it systematically underestimates the spread.

Why It Matters

In AP Statistics, every one-sample and two-sample t-procedure uses ss computed with n−1n - 1. For one-sample t-procedures, the t-distribution has n−1n - 1 degrees of freedom for the same reason. On a TI-84, the 1-Var Stats output labeled Sx uses Bessel's correction, while σx does not. You need to pick the right one for sample data.

Common Mistakes

Mistake: Believing that Bessel's correction also makes the sample standard deviation ss an unbiased estimator of σ\sigma.
Correction: The correction only makes s2s^2 unbiased for σ2\sigma^2. Taking a square root does not preserve averages, so E[s]<σE[s] < \sigma is still true. The leftover bias is small for moderate nn, and AP Statistics procedures simply use ss as is.

Related Terms