MerQuR+Statistics Glossary

Statistics Glossary → Null hypothesis (H₀) and test value

Null hypothesis (H₀) and test valueconcept

The statement of 'no difference', 'no effect' or 'the parameter equals a given value' that a hypothesis test assumes, and against which the data are checked for incompatibility.

The null hypothesis states that the parameter of interest equals a specific value: H₀: μ = μ₀, H₀: p = p₀, H₀: ρ = 0 and so on. The number in that equality is the test value; in most tests it is 0 (no difference, no association), while for proportions or odds ratios it may be another meaningful value such as 0.5 or 1. The test statistic and the p value describe how often a result at least this extreme would occur if H₀ were true; the alternative hypothesis (H₁) is the claim the researcher is usually interested in.

When p < α, H₀ is rejected; otherwise it is 'not rejected'. Failing to reject H₀ does not show that it is true: with a small sample a real difference can easily be missed. To support a claim of 'no difference', an equivalence test (TOST) or a Bayes factor is more appropriate. The test value should be fixed in advance from theory or prior knowledge, not chosen after looking at the data.

An everyday example

A packing machine is supposed to put 500 g of flour in each bag. The mean of 40 bags is 496 g. The null hypothesis says 'the long-run mean of the machine is 500 g', so the test value is 500. A one-sample t-test checks whether a 4 g shortfall can be explained by chance variation alone.

In MerQur

The test value under H₀ is set by 'Test value μ (μ₀)' in the one-sample t-test, 'Expected proportion H₀ (p₀)' in the binomial test, 'Expected proportions' in the chi-square goodness-of-fit test, 'H₀ common OR value' in the Cochran-Mantel-Haenszel test and 'Test value (H0: ICC = rho0)' in the intraclass correlation. In analyses such as the independent-samples t-test, the ANOVAs, MANOVA, Kruskal-Wallis, Friedman and the crosstab, the confidence level sets the H₀ decision through α = 1 - confidence level. The multivariate tests in canonical correlation test H₀: all canonical correlations are 0, and the null hypothesis of the KPSS test in Time Series is stationarity.

Analyses where this concept appears

One-Sample t-TestIndependent-Samples t-TestOne-Way ANOVARepeated Measures ANOVAMANOVAKruskal-WallisFriedmanBinomial TestChi-Square Goodness of FitCohen's Kappa (including Fleiss κ and Krippendorff α)Cochran-Mantel-Haenszel (CMH) TestCrosstabCanonical CorrelationTrue Skill Statistic (TSS)Confusion Matrix MetricsIntraclass Correlation Coefficient (ICC)Bayesian t-Test (JZS Bayes Factor)Time Series

Related concepts

p-valueOne-sided and two-sided tests (hypothesis direction)