MerQuR+Statistics Glossary

Statistics Glossary → Wald test and Wald confidence interval

Wald test and Wald confidence intervalconcept

A test that asks how many standard errors a parameter estimate lies from its null value, and the confidence interval of the form 'estimate ± z × standard error' built on the same idea.

For a single parameter the Wald statistic is z = (θ̂ − θ₀)/SE(θ̂); in large samples it follows a standard normal distribution, and its square follows a chi-square distribution with 1 degree of freedom. When several parameters are tested jointly (e.g. all dummies of a categorical variable), a multivariate Wald chi-square or Wald F is computed from the estimate vector and its covariance matrix. The Wald confidence interval is θ̂ ± z₁₋α/₂·SE. Most coefficient p values in regression output are Wald tests.

The Wald test is convenient because it needs only the fitted full model, but it becomes unreliable in small samples, when a parameter is near a boundary (e.g. a proportion close to 0 or 1) and, in logistic regression, for very large coefficients, where the standard error grows faster than the effect and the test can miss significance (the Hauck-Donner effect). In such cases the likelihood ratio test or a profile-likelihood interval, and for proportions the Wilson or Clopper-Pearson interval, are preferred. The Wald statistic also depends on the scale on which the parameter is written (e.g. log-odds versus odds).

An everyday example

A survey of only 20 people has 1 'yes' (p̂ = 0.05). The Wald interval is 0.05 ± 1.96 × 0.049, from −0.045 to 0.145, and its lower limit below zero makes no sense. For the same data the Wilson interval runs from about 0.009 to 0.236 and gives a far more realistic range.

In MerQur

In quantile regression, 'Test equality of slopes across quantiles (Wald)' tests whether an effect differs across quantiles with the Koenker-Bassett Wald test, and the McNemar test gives a Wald interval for the difference in proportions. In logistic, Poisson, multinomial and ordinal regression, GEE, Cox regression and structural equation modelling, the confidence intervals of the coefficients are Wald intervals (b ± z·SE); GEE gives Type III Wald tests, and the linear mixed model offers a Wald z option under 'Fixed-effect df method'. In the binomial test, Wald is one of the confidence-interval methods.

Analyses where this concept appears

Two-Way ANOVARepeated Measures ANOVAANCOVABinomial TestChi-Square IndependenceFisher's Exact TestMcNemar TestLog-Linear ModelCrosstabEffect Size CalculatorMeta-AnalysisMultiple Linear RegressionLogistic RegressionPoisson / Negative Binomial Regression (Count Regression)Multinomial Logistic RegressionOrdinal Logistic RegressionLinear Mixed Models (LMM)Nested Mixed Model (Nested LMM, Multi-Level Hierarchical Mixed Model)Crossed Mixed Model (Crossed LMM, Crossed Random Effects Mixed Model)Generalized Estimating Equations (GEE)Generalized Linear Mixed Models (GLMM)Panel Data RegressionInstrumental Variables Regression (2SLS)True Skill Statistic (TSS)Confirmatory Factor Analysis (CFA)Structural Equation Modeling (SEM)Weighted Frequency (Survey Freq)Weighted Linear Regression (Survey Regression)Weighted Logistic Regression (Survey Logistic Regression)Variance Components Analysis (VARCOMP)Generalized Additive Model (GAM, Nonlinear Regression)Nonlinear RegressionQuantile RegressionProbit RegressionTobit Regression (Censored Regression)Conditional Logit (Choice Model)Hierarchical Bayesian RegressionSAR (Spatial Autoregressive Lag)Spatial Error Model (SEM)Path AnalysisCox Regression (Proportional Hazards Model)Competing RisksSample Selection and Truncated Regression (Heckman)

Related concepts

Likelihood ratio test (LRT)Score (Lagrange multiplier) testStandard error