Degrees of Freedom Calculator

Degrees of freedom, in plain words

Degrees of freedom, usually written df, count how many values in a calculation are free to change once a constraint is fixed. Take five numbers with a mean of 10. You can pick four of them freely, say 8, 9, 11 and 12, but the fifth is forced: it has to be 10 for the mean to hold. Five values, one constraint, so 4 degrees of freedom.

That is the whole idea. Every statistic that estimates something from your data spends one degree of freedom. Estimate a mean from n observations and you have n - 1 left. Estimate a pooled variance from two samples and you spend two, leaving n1 + n2 - 2.

The count matters because test statistics are compared against distributions that change shape with df. A t statistic of 2.06 is borderline at 24 df and ordinary at 200 df, so the same number can lead two analysts to different conclusions if they read the wrong row.

The formula for each test

Each test spends degrees of freedom in its own way. The table below collects the formulas this calculator uses. The pattern to remember: subtract one for every mean or parameter you estimate from the data.

Two rows deserve a comment. The independence formula multiplies the shortfalls on each side, so a 3 by 4 table has (3 - 1) x (4 - 1) = 6 df, not 12. Regression with p predictors uses n - p - 1, which is why the simple version with one predictor is n - 2.

Worked example: 25 bags off the line

Suppose a machine is set to fill 50 gram bags. You weigh 25 bags and get a sample mean of 52.4 grams with a standard deviation of 6.8 grams. Is the machine off target?

• Standard error: 6.8 / 5 = 1.36
• t statistic: (52.4 - 50) / 1.36 = 1.76
• Degrees of freedom: 25 - 1 = 24
• Critical value at the 5% level, two sided: 2.064

The statistic of 1.76 sits below 2.064, so at 24 df you would not reject the target. Run the same test with 100 bags and the picture changes. The standard error falls to 0.68, the t statistic doubles to 3.53, and df rises to 99, where the critical value is close to 1.984. Same machine, same gap from target, and the conclusion flips because more data buys both a bigger statistic and a lower hurdle.

With two independent samples of 25 and 31 bags, the pooled test has 25 + 31 - 2 = 54 df, and the critical value at 5% is 2.005. The Penn State STAT 500 notes walk through the pooling step in detail.

Welch's correction when the spreads differ

The pooled two sample formula assumes both populations share one variance. When sample standard deviations differ a lot, say 4 and 9, that assumption is shaky, and the Welch Satterthwaite correction is the safer count:

df = (s1 squared / n1 + s2 squared / n2) squared / [ (s1 squared / n1) squared / (n1 - 1) + (s2 squared / n2) squared / (n2 - 1) ]

Run it with n1 = 12, n2 = 16, s1 = 4 and s2 = 9. The two variance terms are 16 / 12 = 1.333 and 81 / 16 = 5.063, so the numerator is 6.396 squared = 40.91. The denominator is 1.333 squared / 11 + 5.063 squared / 15 = 0.162 + 1.709 = 1.871. That gives df = 40.91 / 1.871 = 21.9, against 26 for the pooled version. The correction spends about 4 df to buy protection against the unequal spread, and the critical value moves from 2.056 to 2.080 as a result.

Software reports Welch df as a decimal, and that is fine. Round down when you read a printed table, so 21.9 is read at row 21.

Degrees of freedom in ANOVA and regression

Analysis of variance splits the total df across sources. With 4 groups and 60 observations in total:

The three numbers must add up: 3 + 56 = 59. If your ANOVA table does not add up, a group size was entered wrong or an observation was dropped.

Simple linear regression does the same split. Fit a line to 30 points and you spend 1 df on the slope and 1 on the intercept, leaving 30 - 2 = 28 residual df. Add a second predictor and residual df drops to 27. The NIST engineering statistics handbook lists the df breakdown for fitted models and for the t and chi square distributions used to judge them.

This is also where the F test gets its two df values. In a regression with 3 predictors and 40 observations, the model is tested on 3 and 36 df, because 40 - 3 - 1 = 36.

Why df changes your critical value

The t distribution has fatter tails when df is low, because the sample standard deviation is itself an estimate. Each extra observation sharpens that estimate, and the critical value slides toward the 1.960 of the normal curve.

Read the first and last rows together: a t statistic of 2.2 clears the bar at 60 df and fails at 5 df. Degrees of freedom are not bookkeeping. They are part of the verdict.

Chi square tables move the same way, with more room. A 5 category goodness of fit test has 4 df and a critical value of 9.488 at the 5% level, while 20 df needs 31.410. Report df next to every test statistic so a reader can check the comparison without guessing.

Four mistakes to avoid

• Using n instead of n - 1. Reporting df 25 for a one sample test on 25 observations reads the wrong row and understates the critical value.
• Reaching for the pooled formula with unequal spread. When one group has a standard deviation of 9 and the other 4, pooled df overstates what the data support. Welch costs a few df and is the honest count.
• Multiplying rows by columns. A 3 by 4 independence table has 6 df, because the formula is (r - 1)(c - 1).
• Forgetting predictors in regression. Ten observations with three predictors leave 10 - 3 - 1 = 6 residual df, not 8 or 9.

When in doubt, count constraints. Every estimated parameter subtracts one, and every margin in a contingency table subtracts one per row and one per column. The count falls out of the design, not out of the software output.

One last habit is worth the ten seconds it takes: write the df on the same line as the test statistic and the p value. A result pasted into a report without its df cannot be checked by anyone reading it later.