MANOVA and MANCOVA (Multivariate Analysis of Variance and Covariance)

Description

Tests whether groups differ on several correlated outcome measures at once, optionally adjusting for covariates, before any single outcome is examined.

MANOVA extends ANOVA to two or more metric outcomes analysed together with one or more categorical factors. Instead of comparing single means it compares group mean vectors, taking the covariance among the outcomes into account, which controls the inflated Type I error of running a separate ANOVA per measure and can detect differences that only appear in a combination of measures. The overall test is reported with Wilks' lambda, Pillai's trace or Hotelling's T-squared (Pillai is the most robust when assumptions wobble). Designs may be one-way, factorial, between-subjects, within-subjects or mixed; the within-subjects multivariate approach is a separate use. MANCOVA adds interval covariates, so groups are compared on the outcomes after adjusting for, say, prior experience. A significant omnibus result is followed by comparisons of group centroids or univariate tests, ideally with discriminant function analysis to see which combination of outcomes separates the groups.

Attributes

Loss type
Siloed
Stage
Evaluation
Effort
Days

Steps

  1. Choose two or more conceptually related metric outcomes (for example time, effort rating and satisfaction) and the categorical factor(s); add interval covariates for MANCOVA.
  2. Check assumptions: approximate multivariate normality, linear relations among outcomes, homogeneity of covariance matrices (Box's M) and at least 15-20 observations per cell.
  3. Run the omnibus multivariate test and report Wilks' lambda, Pillai's trace or Hotelling's T-squared with its F approximation and an effect size.
  4. If significant, locate the differences with univariate follow-ups (with alpha adjustment), centroid comparisons or discriminant function analysis.
  5. For MANCOVA, check that covariate slopes are homogeneous across groups and interpret group differences on the adjusted outcomes.
  6. Report adjusted or unadjusted means per outcome and group together with the multivariate result.

Sources

  • Quantitative Social Research Methods (Singh, 2007)
  • A User's Guide to Principal Components (Jackson, 1991)
  • Methods of Multivariate Analysis (Rencher, 2002)
  • Multivariate Statistical Inference and Applications (Rencher, 1998)
  • Linear Statistical Inference and Its Applications (Rao, 1973)

Also known as

  • MANOVA
  • MANCOVA
  • Multivariate ANOVA
  • One-way MANOVA
  • Factorial MANOVA
  • Multivariate analysis of variance and covariance (MANOVA and MANCOVA)
  • Hotelling's T-squared test

Categories

  • Analytics & Behaviour
  • Statistics & Modelling