MerQuR+Statistics Glossary

Statistics Glossary → Degrees of freedom

Degrees of freedomconcept

The number of independent pieces of information that are free to vary when a statistic is computed; it determines which theoretical distribution the test statistic is compared with.

Once the mean of n observations has been computed, the deviations from the mean must add up to zero, so only n - 1 of the n deviations are free to vary. That is why the sample variance is divided by n - 1 rather than n. As a general rule, the degrees of freedom equal the number of observations minus the number of parameters estimated from the data: n - 1 in a one-sample t-test, and k - 1 between groups and N - k within groups in a one-way ANOVA with k groups.

The shape of the t, F and chi-square distributions depends on the degrees of freedom. With few degrees of freedom the t distribution has heavy tails and a larger t value is needed for significance; as the degrees of freedom grow, the t distribution approaches the standard normal. In methods such as the Welch correction or the Satterthwaite approximation the degrees of freedom need not be a whole number (e.g. 23.4); this is not an error but a sign that an approximate distribution is being used. They are reported together with the test statistic, as in t(28) = 2.31 or F(2, 57) = 4.12.

An everyday example

Five friends are splitting a bill so that the average share is exactly 20 dollars. The first four can pick any amount they like (say 15, 25, 18 and 22), but the fifth person's share is no longer free: they must pay 20 to make the total 100. Because only four of the five amounts can be chosen freely, there are 4 degrees of freedom.

In MerQur

In most tests it is reported as 'df' next to the statistic. Fractional df from the Welch-Satterthwaite approximation are used in the independent-samples t-test and in the Welch variant of One-Way ANOVA, df multiplied by ε under a sphericity violation in Repeated Measures ANOVA, Satterthwaite or Kenward-Roger df in Linear Mixed Models, Barnard-Rubin df in Multiple Imputation and the design df (number of PSUs - number of strata) in the weighted (survey) analyses. In Partial Least Squares Regression the jackknife coefficient tests are given with segments - 1 degrees of freedom.

Analyses where this concept appears

Independent-Samples t-TestOne-Way ANOVATwo-Way ANOVARepeated Measures ANOVAMixed-Design ANOVA, Split-PlotANCOVAMultiple Comparison CorrectionKruskal-WallisChi-Square IndependenceChi-Square Goodness of FitCohen's Kappa (including Fleiss κ and Krippendorff α)Cochran-Mantel-Haenszel (CMH) TestLog-Linear ModelMR × Categorical CrosstabCochran's Q TestCorrelation AnalysisBland-Altman Agreement AnalysisEffect Size CalculatorMultiple Linear RegressionLogistic RegressionPoisson / Negative Binomial Regression (Count Regression)Multinomial Logistic RegressionOrdinal Logistic RegressionRidge RegressionLinear Mixed Models (LMM)Nested Mixed Model (Nested LMM, Multi-Level Hierarchical Mixed Model)Crossed Mixed Model (Crossed LMM, Crossed Random Effects Mixed Model)Panel Data RegressionCronbach's Alpha and Reliability AnalysisIntraclass Correlation Coefficient (ICC)Confirmatory Factor Analysis (CFA)Structural Equation Modeling (SEM)Weighted Mean (Survey Means)Weighted Frequency (Survey Freq)Weighted Total (Survey Total)Weighted Linear Regression (Survey Regression)Weighted Logistic Regression (Survey Logistic Regression)Variance Components Analysis (VARCOMP)Partial Least Squares Regression (PLS)Probit RegressionTobit Regression (Censored Regression)Bayesian t-Test (JZS Bayes Factor)Bayesian Linear RegressionPath AnalysisMultiple ImputationKaplan-Meier Survival AnalysisSurvey-PHREG (Weighted Cox)Cluster-Robust Cox (Marginal and Frailty Models)Time SeriesLOESS (Local Regression)PLS-SEM (Partial Least Squares Path Modeling)Unobserved Components Model (Structural Time Series)Point Pattern Analysis (Ripley K/L, G/F/J, CSR)

Related concepts

Satterthwaite and Kenward-Roger degrees-of-freedom approximationsWelch correction