MANOVA/MANCOVA: Zero to Hero Tutorial

This comprehensive tutorial takes you from the foundational concepts of Multivariate Analysis of Variance (MANOVA) and Multivariate Analysis of Covariance (MANCOVA) all the way through advanced test statistics, assumption checking, post-hoc analysis, effect size interpretation, and practical usage within the DataStatPro application. Whether you are a complete beginner or an experienced analyst, this guide is structured to build your understanding step by step.

Table of Contents

1. Prerequisites and Background Concepts
2. What are MANOVA and MANCOVA?
3. The Mathematical Framework
4. Assumptions of MANOVA and MANCOVA
5. Multivariate Test Statistics
6. MANCOVA: Adding Covariates
7. Effect Size Measures
8. Follow-Up Analyses
9. Power Analysis and Sample Size
10. Model Fit and Evaluation
11. Assumption Checking and Diagnostics
12. Contrast Analysis in MANOVA
13. Repeated Measures MANOVA (Profile Analysis)
14. Using the MANOVA/MANCOVA Component
15. Computational and Formula Details
16. Worked Examples
17. Common Mistakes and How to Avoid Them
18. Troubleshooting
19. Quick Reference Cheat Sheet

1. Prerequisites and Background Concepts

Before diving into MANOVA and MANCOVA, it is helpful to be familiar with the following foundational concepts. Do not worry if you are not — each concept is briefly explained here.

1.1 Univariate ANOVA Recap

Analysis of Variance (ANOVA) tests whether the means of a single continuous dependent variable differ across two or more groups defined by a categorical independent variable (factor). The core idea is to partition the total variability into:

• Between-group (treatment) variability: How much the group means differ from the grand mean.
• Within-group (error) variability: How much individual observations differ from their own group mean.

The F-ratio compares these two sources:

$F = \frac{MS_{between}}{MS_{within}} = \frac{SS_{between}/(k-1)}{SS_{within}/(n-k)}$

Where $k$ is the number of groups and $n$ is the total number of observations. A large $F$ provides evidence that at least one group mean differs from the others.

1.2 Vectors and Matrices

In multivariate analysis, each observation is described by a vector of measurements on $p$ dependent variables:

$\mathbf{y}_i = (y_{i1}, y_{i2}, \dots, y_{ip})^T$

The mean vector for group $j$ is:

$\bar{\mathbf{y}}_j = \frac{1}{n_j}\sum_{i \in \text{group } j} \mathbf{y}_i$

The grand mean vector across all observations is:

$\bar{\mathbf{y}} = \frac{1}{n}\sum_{i=1}^n \mathbf{y}_i$

1.3 The Covariance Matrix

The covariance matrix $\boldsymbol{\Sigma}$ ($p \times p$) contains:

• Variances of each dependent variable on the diagonal: $\sigma_{jj} = \text{Var}(Y_j)$.
• Covariances between pairs of dependent variables off-diagonal: $\sigma_{jk} = \text{Cov}(Y_j, Y_k)$.

$\boldsymbol{\Sigma} = \begin{pmatrix} \sigma_{11} & \sigma_{12} & \cdots & \sigma_{1p} \\ \sigma_{21} & \sigma_{22} & \cdots & \sigma_{2p} \\ \vdots & \vdots & \ddots & \vdots \\ \sigma_{p1} & \sigma_{p2} & \cdots & \sigma_{pp} \end{pmatrix}$

The covariance matrix is symmetric ($\sigma_{jk} = \sigma_{kj}$) and positive semi-definite ($\mathbf{v}^T\boldsymbol{\Sigma}\mathbf{v} \geq 0$ for any vector $\mathbf{v}$).

1.4 The Multivariate Normal Distribution

The multivariate normal distribution $\mathcal{N}_p(\boldsymbol{\mu}, \boldsymbol{\Sigma})$ generalises the univariate normal to $p$ dimensions. Its probability density function is:

$f(\mathbf{y}) = \frac{1}{(2\pi)^{p/2}|\boldsymbol{\Sigma}|^{1/2}} \exp\left\{-\frac{1}{2}(\mathbf{y} - \boldsymbol{\mu})^T \boldsymbol{\Sigma}^{-1} (\mathbf{y} - \boldsymbol{\mu})\right\}$

Where $|\boldsymbol{\Sigma}|$ is the determinant of $\boldsymbol{\Sigma}$ and $\boldsymbol{\Sigma}^{-1}$ is its inverse. MANOVA assumes observations follow this distribution within each group.

1.5 Matrix Determinants and Eigenvalues

The determinant $|\mathbf{A}|$ of a square matrix $\mathbf{A}$ is a scalar that summarises the matrix in terms of the "volume" it represents. A determinant of zero indicates a singular (non-invertible) matrix.

Eigenvalues $\lambda_1, \lambda_2, \dots, \lambda_p$ of a matrix $\mathbf{A}$ satisfy:

$\mathbf{A}\mathbf{v} = \lambda \mathbf{v}$

For covariance matrices, eigenvalues represent the variance explained in each orthogonal direction (principal component). They are always non-negative. The sum of all eigenvalues equals the trace of the matrix: $\sum \lambda_i = \text{tr}(\mathbf{A}) = \sum_i a_{ii}$.

1.6 The F-Distribution and Wilks' Lambda

The F-distribution with parameters $(df_1, df_2)$ arises from the ratio of two independent chi-squared variables divided by their degrees of freedom. It is the reference distribution for many univariate hypothesis tests.

Wilks' Lambda $\Lambda^*$ is the primary multivariate test statistic in MANOVA. It can be exactly or approximately converted to an $F$-statistic (see Section 5). Understanding that multivariate test statistics generalise the $F$-ratio to multiple dependent variables simultaneously is the key conceptual bridge from ANOVA to MANOVA.

2. What are MANOVA and MANCOVA?

2.1 MANOVA: Multivariate Analysis of Variance

MANOVA (Multivariate Analysis of Variance) is the multivariate extension of ANOVA. Instead of testing whether groups differ on a single dependent variable, MANOVA simultaneously tests whether groups differ on a set of dependent variables considered jointly.

Formally, MANOVA tests:

• $H_0$: The mean vectors of all groups are equal: $\boldsymbol{\mu}_1 = \boldsymbol{\mu}_2 = \dots = \boldsymbol{\mu}_k$.
• $H_1$: At least two group mean vectors differ on at least one dependent variable (or linear combination thereof).

2.2 MANCOVA: Multivariate Analysis of Covariance

MANCOVA (Multivariate Analysis of Covariance) extends MANOVA by adding one or more continuous covariates to the model. Covariates are variables that:

• Are related to the dependent variables but not of primary interest.
• Are included to statistically control for their effects, thereby:
  • Reducing error variance (increasing statistical power).
  • Removing confounding effects of baseline differences between groups.

MANCOVA asks: "After accounting for the covariates, do groups differ on the set of dependent variables?"

2.3 Why Use MANOVA Instead of Multiple ANOVAs?

A common question is: "Why not just run separate ANOVAs for each dependent variable?" There are several compelling reasons to prefer MANOVA:

⚠️ MANOVA is not always better than separate ANOVAs. If the dependent variables are conceptually unrelated and you have clear, directional hypotheses about each, separate ANOVAs with Bonferroni correction may be more appropriate and interpretable.

2.4 Real-World Applications

MANOVA and MANCOVA are used across many disciplines:

• Clinical Psychology: Comparing treatment groups on multiple symptom severity scales simultaneously (depression, anxiety, sleep quality).
• Education Research: Comparing teaching methods on multiple outcome measures (reading, mathematics, science scores).
• Pharmacology: Comparing drug dosages on multiple physiological endpoints (blood pressure, heart rate, cortisol level).
• Marketing: Comparing advertising strategies on consumer perceptions across multiple brand attributes.
• Sports Science: Comparing training protocols on multiple performance metrics (speed, endurance, strength).
• Neuroscience: Comparing patient groups on multiple cognitive measures (memory, attention, executive function).
• Environmental Science: Comparing sites on multiple ecological variables (species diversity, biomass, soil pH).

2.5 Design Terminology

3. The Mathematical Framework

3.1 The MANOVA Model

For a one-way MANOVA with $k$ groups, $p$ dependent variables, and $n_j$ observations in group $j$ ($n = \sum n_j$), each observation is modelled as:

$\mathbf{y}_{ij} = \boldsymbol{\mu} + \boldsymbol{\alpha}_j + \boldsymbol{\epsilon}_{ij}$

Where:

• $\mathbf{y}_{ij}$ ($p \times 1$) = observation vector for the $i$-th unit in group $j$.
• $\boldsymbol{\mu}$ ($p \times 1$) = grand mean vector.
• $\boldsymbol{\alpha}_j$ ($p \times 1$) = effect of group $j$, with constraint $\sum_{j=1}^k n_j \boldsymbol{\alpha}_j = \mathbf{0}$.
• $\boldsymbol{\epsilon}_{ij}$ ($p \times 1$) = random error vector, assumed $\boldsymbol{\epsilon}_{ij} \sim \mathcal{N}_p(\mathbf{0}, \boldsymbol{\Sigma})$.

The key assumption is that all groups share the same within-group covariance matrix $\boldsymbol{\Sigma}$ (homogeneity of covariance matrices).

3.2 The Matrix of Sum of Squares and Cross Products (SSCP)

MANOVA partitions the total variation in the multivariate data into between-group and within-group components using Sum of Squares and Cross Products (SSCP) matrices:

Total SSCP matrix $\mathbf{T}$ ($p \times p$):

$\mathbf{T} = \sum_{j=1}^k \sum_{i=1}^{n_j} (\mathbf{y}_{ij} - \bar{\mathbf{y}})(\mathbf{y}_{ij} - \bar{\mathbf{y}})^T$

Between-group (Hypothesis) SSCP matrix $\mathbf{H}$ ($p \times p$):

$\mathbf{H} = \sum_{j=1}^k n_j (\bar{\mathbf{y}}_j - \bar{\mathbf{y}})(\bar{\mathbf{y}}_j - \bar{\mathbf{y}})^T$

Within-group (Error) SSCP matrix $\mathbf{E}$ ($p \times p$):

$\mathbf{E} = \sum_{j=1}^k \sum_{i=1}^{n_j} (\mathbf{y}_{ij} - \bar{\mathbf{y}}_j)(\mathbf{y}_{ij} - \bar{\mathbf{y}}_j)^T$

Fundamental decomposition: $\mathbf{T} = \mathbf{H} + \mathbf{E}$

This is the multivariate generalisation of $SS_{total} = SS_{between} + SS_{within}$.

Degrees of freedom:

• $df_H = k - 1$ (between groups)
• $df_E = n - k$ (within groups)
• $df_T = n - 1$ (total)

3.3 The SSCP Matrices in Detail

The diagonal elements of $\mathbf{H}$ and $\mathbf{E}$ are the familiar $SS_{between}$ and $SS_{within}$ values from separate ANOVAs for each dependent variable. The off-diagonal elements capture the covariance structure — how variation in one DV covaries with variation in another, both between and within groups. This is what MANOVA uses beyond what separate ANOVAs provide.

For dependent variables $j$ and $k$:

$H_{jk} = \sum_{g=1}^G n_g (\bar{y}_{gj} - \bar{y}_j)(\bar{y}_{gk} - \bar{y}_k)$

$E_{jk} = \sum_{g=1}^G \sum_{i=1}^{n_g} (y_{igj} - \bar{y}_{gj})(y_{igk} - \bar{y}_{gk})$

3.4 The Hypothesis Test

MANOVA tests whether the between-group variation is large relative to the within-group variation, simultaneously for all $p$ dependent variables. This is assessed through the eigenvalues of the matrix $\mathbf{E}^{-1}\mathbf{H}$:

$\mathbf{E}^{-1}\mathbf{H} \mathbf{v}_s = \lambda_s \mathbf{v}_s, \quad s = 1, 2, \dots, \min(k-1, p)$

Where:

• $\lambda_s$ = the $s$-th eigenvalue (ratio of between-group to within-group variation in the $s$-th discriminant direction).
• $\mathbf{v}_s$ = the $s$-th eigenvector (the direction in multivariate space most discriminating between groups).
• $s^* = \min(k-1, p)$ = the number of non-zero eigenvalues.

The eigenvalues $\lambda_1 \geq \lambda_2 \geq \dots \geq \lambda_{s^*} \geq 0$ are the basis for all four multivariate test statistics (Section 5).

3.5 The Estimated Within-Group Covariance Matrix

The pooled within-group covariance matrix $\mathbf{S}_W$ (the multivariate analogue of $MS_{within}$) is estimated as:

$\mathbf{S}_W = \frac{\mathbf{E}}{n - k} = \frac{1}{n-k}\sum_{j=1}^k \sum_{i=1}^{n_j} (\mathbf{y}_{ij} - \bar{\mathbf{y}}_j)(\mathbf{y}_{ij} - \bar{\mathbf{y}}_j)^T$

This is an unbiased estimator of the common within-group covariance matrix $\boldsymbol{\Sigma}$ under the homogeneity assumption.

3.6 The Factorial MANOVA Model

For a two-way factorial MANOVA with factors A ($a$ levels) and B ($b$ levels):

$\mathbf{y}_{ijl} = \boldsymbol{\mu} + \boldsymbol{\alpha}_i + \boldsymbol{\beta}_j + (\boldsymbol{\alpha\beta})_{ij} + \boldsymbol{\epsilon}_{ijl}$

Where:

• $\boldsymbol{\alpha}_i$ = main effect of factor A at level $i$.
• $\boldsymbol{\beta}_j$ = main effect of factor B at level $j$.
• $(\boldsymbol{\alpha\beta})_{ij}$ = interaction effect of A × B.
• $\boldsymbol{\epsilon}_{ijl} \sim \mathcal{N}_p(\mathbf{0}, \boldsymbol{\Sigma})$.

Three separate SSCP matrices and hypothesis tests are conducted: for the main effect of A, the main effect of B, and the A × B interaction.

4. Assumptions of MANOVA and MANCOVA

MANOVA and MANCOVA rest on several critical assumptions. Violations of these assumptions can lead to inflated Type I error rates, reduced power, or misleading results.

4.1 Multivariate Normality

Assumption: Within each group, the $p$ dependent variables jointly follow a multivariate normal distribution $\mathcal{N}_p(\boldsymbol{\mu}_j, \boldsymbol{\Sigma})$.

Why it matters: The multivariate test statistics (Wilks' Lambda, Pillai's trace, etc.) are derived assuming multivariate normality. Severe departures — particularly heavy tails or outliers — inflate Type I error rates.

How to check:

• Mardia's tests: Test multivariate skewness ($b_{1,p}$) and kurtosis ($b_{2,p}$) — the most widely used multivariate normality tests.
• Royston's test: Extends the Shapiro-Wilk test to the multivariate setting.
• Henze-Zirkler test: Based on a consistent test of multivariate normality.
• Q-Q plots of Mahalanobis distances: If multivariate normality holds, the squared Mahalanobis distances $D^2_i = (\mathbf{y}_i - \bar{\mathbf{y}})^T \mathbf{S}^{-1} (\mathbf{y}_i - \bar{\mathbf{y}})$ should follow a $\chi^2_p$ distribution. Plot $D^2_{(i)}$ vs. quantiles of $\chi^2_p$ — departures from the diagonal indicate non-normality or outliers.

Robustness: MANOVA is fairly robust to moderate departures from multivariate normality when group sizes are large and equal, particularly for Pillai's Trace (the most robust statistic). With small or unequal group sizes, non-normality is more problematic.

4.2 Homogeneity of Covariance Matrices (Homoscedasticity)

Assumption: All $k$ groups share the same within-group covariance matrix:

$\boldsymbol{\Sigma}_1 = \boldsymbol{\Sigma}_2 = \dots = \boldsymbol{\Sigma}_k = \boldsymbol{\Sigma}$

Why it matters: This is the multivariate analogue of homoscedasticity in ANOVA. Violations lead to inflated or deflated Type I error rates depending on the pattern of inequality and the relative group sizes.

How to check:

• Box's M test: The standard test for equality of covariance matrices. Tests $H_0: \boldsymbol{\Sigma}_1 = \boldsymbol{\Sigma}_2 = \dots = \boldsymbol{\Sigma}_k$.

  $M = (n-k)\ln|\mathbf{S}_W| - \sum_{j=1}^k (n_j - 1)\ln|\mathbf{S}_j|$

  Box's M is converted to an approximate $\chi^2$ or $F$ statistic. A significant result ($p < 0.05$) indicates heterogeneous covariance matrices.

  ⚠️ Box's M is extremely sensitive to non-normality and detects trivial violations in large samples. A common guideline is to be concerned only when $p < 0.001$ rather than $p < 0.05$. Always pair Box's M with visual inspection of the covariance matrices.

What to do when violated:

• Use Pillai's Trace instead of Wilks' Lambda (Pillai's Trace is more robust to heterogeneous covariance matrices).
• If group sizes are equal, MANOVA results are robust to moderate violations.
• Use heteroscedasticity-robust MANOVA variants (e.g., James's test, Johansen's procedure).

4.3 Independence of Observations

Assumption: Each observation must be independent of all others. No observation should influence or be correlated with any other observation.

Why it matters: Correlation among observations (e.g., siblings in the same family, students in the same classroom, repeated measures from the same participant) violates the independence assumption and inflates Type I error.

How to check: Consider the study design. If observations are nested, clustered, or repeated, use appropriate models (multilevel MANOVA, repeated measures MANOVA).

4.4 Absence of Multivariate Outliers

Assumption: There are no extreme multivariate outliers — observations that are unusual on the combination of DVs even if they are not extreme on any single DV.

Why it matters: Multivariate outliers can distort the SSCP matrices and dramatically influence the test statistics.

How to check:

• Mahalanobis distance: $D^2_i = (\mathbf{y}_i - \bar{\mathbf{y}})^T \mathbf{S}^{-1} (\mathbf{y}_i - \bar{\mathbf{y}})$. Compare to $\chi^2_{p, \alpha}$ (e.g., $\chi^2_{p, 0.001}$). Observations exceeding this threshold are flagged as potential outliers.
• Robust Mahalanobis distance: Uses robust (MCD or MVE) location and scatter estimates, which are not themselves influenced by outliers.

4.5 Linearity Among Dependent Variables

Assumption: The relationships among the dependent variables are linear within each group.

Why it matters: MANOVA uses covariances (which measure linear association) to capture the multivariate structure. Non-linear relationships among DVs reduce the efficiency of MANOVA.

How to check: Scatter plot matrix (pairs plot) of the DVs within each group; look for non-linear patterns.

4.6 No Perfect Multicollinearity Among Dependent Variables

Assumption: No dependent variable is a perfect linear combination of other dependent variables.

Why it matters: Perfect multicollinearity makes the within-group covariance matrix $\mathbf{E}$ singular (non-invertible), preventing the computation of $\mathbf{E}^{-1}\mathbf{H}$.

How to check: Compute the condition number or determinant of $\mathbf{E}$. If $|\mathbf{E}| \approx 0$ or the condition number is very large, multicollinearity is a problem.

4.7 Additional Assumptions for MANCOVA

Homogeneity of regression (slopes): In MANCOVA, the regression of the DVs on the covariates must be the same across all groups — i.e., there is no interaction between the factor(s) and the covariate(s).

How to check: Fit a model including the factor × covariate interaction and test its significance. If significant, the homogeneity of regression assumption is violated and MANCOVA is inappropriate.

Covariate measured without error and not affected by treatment: Covariates should be measured reliably (low measurement error) and should not be caused by (a consequence of) the group membership — otherwise, the covariate adjustment is biased.

5. Multivariate Test Statistics

MANOVA provides four multivariate test statistics, all based on the eigenvalues $\lambda_1 \geq \lambda_2 \geq \dots \geq \lambda_{s^*}$ of $\mathbf{E}^{-1}\mathbf{H}$, where $s^* = \min(k-1, p)$. Each statistic summarises the eigenvalues differently and has different properties.

5.1 Wilks' Lambda ($\Lambda^*$)

Wilks' Lambda is the most widely used multivariate test statistic. It is the ratio of the determinant of the within-group SSCP matrix to the determinant of the total SSCP matrix:

$\Lambda^* = \frac{|\mathbf{E}|}{|\mathbf{H} + \mathbf{E}|} = \frac{|\mathbf{E}|}{|\mathbf{T}|} = \prod_{s=1}^{s^*} \frac{1}{1 + \lambda_s}$

Range: $0 \leq \Lambda^* \leq 1$.

Interpretation:

• $\Lambda^* = 1$: No group differences (all eigenvalues = 0; $\mathbf{H} = \mathbf{0}$).
• $\Lambda^* = 0$: Perfect group separation (at least one eigenvalue is infinite).
• Smaller $\Lambda^*$ → stronger group differences → more likely to reject $H_0$.

Conversion to F-statistic (Rao's approximation):

$F = \frac{1 - \Lambda^{*1/t}}{\Lambda^{*1/t}} \cdot \frac{df_2}{df_1}$

Where:

$t = \sqrt{\frac{p^2 df_H^2 - 4}{p^2 + df_H^2 - 5}}, \quad df_1 = p \cdot df_H, \quad df_2 = t\left(df_E - \frac{p - df_H + 1}{2}\right) - \frac{p \cdot df_H - 2}{2}$

Exact F when $p = 1$ or $p = 2$ or $df_H = 1$ or $df_H = 2$.

For $df_H = 1$ (two groups or testing one contrast):

$F = \frac{1 - \Lambda^*}{\Lambda^*} \cdot \frac{df_E - p + 1}{p} \sim F_{p,\, df_E - p + 1}$

Properties:

• Powerful when group differences are spread across multiple discriminant dimensions.
• More sensitive to violations of multivariate normality than Pillai's Trace.
• The most commonly reported MANOVA statistic in published research.

5.2 Pillai's Trace ($V$)

$V = \text{tr}\left[\mathbf{H}(\mathbf{H}+\mathbf{E})^{-1}\right] = \sum_{s=1}^{s^*} \frac{\lambda_s}{1 + \lambda_s}$

Range: $0 \leq V \leq s^* = \min(k-1, p)$.

Interpretation:

• $V = 0$: No group differences.
• $V = s^*$: Perfect group separation.
• Larger $V$ → stronger group differences.

Conversion to F-statistic:

$F = \frac{V/s^*}{(s^* - V)/s^*} \cdot \frac{df_2}{df_1} = \frac{V(2n_Y + s^* + 1)}{(s^* - V)(2n_X + s^* + 1)} \sim F_{s^* m, s^* n_Y}$

Where $m = df_H$, $n_X = (df_E - p - 1)/2$, $n_Y = (df_H - p - 1)/2$ (details of notation vary by software).

Properties:

• The most robust statistic to violations of homogeneity of covariance matrices and multivariate normality.
• Recommended when Box's M is significant or group sizes are unequal.
• Less powerful than Wilks' Lambda when group differences are concentrated on a single discriminant dimension.
• Preferred statistic when assumptions are questionable.

5.3 Hotelling-Lawley Trace ($T^2$ / $U$)

$U = \text{tr}(\mathbf{E}^{-1}\mathbf{H}) = \sum_{s=1}^{s^*} \lambda_s$

Range: $0 \leq U < \infty$.

Conversion to F-statistic:

$F = \frac{U}{s^*} \cdot \frac{df_E(df_E - p - 1)}{df_H(df_E + df_H - p - 1)} \cdot \frac{df_E - p - 1}{df_E}$

More precisely:

$F = \frac{df_E(df_E - p - 1) U}{s^* \cdot p \cdot df_H} \cdot \frac{1}{?}$

The exact approximation formula varies slightly by software; the statistic follows approximately an $F_{df_1, df_2}$ distribution.

Properties:

• The most powerful statistic when group differences are concentrated on a single dominant discriminant dimension.
• Most sensitive to outliers (least robust).
• Equivalent to Hotelling's $T^2$ test when $k = 2$.

5.4 Roy's Largest Root ($\theta$)

$\theta = \frac{\lambda_1}{1 + \lambda_1}$

Where $\lambda_1$ is the largest eigenvalue of $\mathbf{E}^{-1}\mathbf{H}$.

Range: $0 \leq \theta \leq 1$.

Conversion to F (approximate):

$F = \theta \cdot \frac{df_E}{df_H} \sim F_{p, df_E - p + 1} \text{ (upper bound)}$

Properties:

• The most powerful statistic when all group separation is on a single discriminant dimension — a condition that is rarely met in practice.
• Provides an upper bound to the distribution and is typically approximated.
• The least robust statistic; severely affected by heterogeneous covariance matrices and outliers.
• Most useful when there is a strong theoretical reason to expect a single dominant group difference.

5.5 Comparison of the Four Test Statistics

💡 In most applied research, report all four statistics. When they agree, the conclusion is robust. When they disagree, report Pillai's Trace as the most reliable result and investigate why they differ (often due to the structure of group differences).

5.6 Hotelling's $T^2$ (Two-Group Case)

When $k = 2$ (only two groups), MANOVA reduces to Hotelling's $T^2$ test — the multivariate generalisation of the independent-samples $t$-test:

$T^2 = \frac{n_1 n_2}{n_1 + n_2} (\bar{\mathbf{y}}_1 - \bar{\mathbf{y}}_2)^T \mathbf{S}_W^{-1} (\bar{\mathbf{y}}_1 - \bar{\mathbf{y}}_2)$

Converted to an $F$-statistic:

$F = \frac{n_1 + n_2 - p - 1}{(n_1 + n_2 - 2)p} T^2 \sim F_{p,\, n_1 + n_2 - p - 1}$

This is an exact $F$-test (not an approximation) and is equivalent to all four multivariate statistics when $k = 2$.

6. MANCOVA: Adding Covariates

6.1 The MANCOVA Model

MANCOVA extends MANOVA by including one or more continuous covariates $\mathbf{z}_i$ ($q \times 1$):

$\mathbf{y}_{ij} = \boldsymbol{\mu} + \boldsymbol{\alpha}_j + \mathbf{B}\mathbf{z}_{ij} + \boldsymbol{\epsilon}_{ij}$

Where:

• $\mathbf{B}$ ($p \times q$) = matrix of regression coefficients for the covariates (one row per DV, one column per covariate).
• $\mathbf{z}_{ij}$ = vector of covariate values for observation $i$ in group $j$.
• $\boldsymbol{\epsilon}_{ij} \sim \mathcal{N}_p(\mathbf{0}, \boldsymbol{\Sigma})$.

6.2 How MANCOVA Works: Partitioning SSCP with Covariates

MANCOVA computes adjusted SSCP matrices that remove the linear effect of the covariates from both $\mathbf{H}$ and $\mathbf{E}$:

Step 1: Compute the total SSCP for covariates: $\mathbf{E}_{ZZ}$ (SSCP of $\mathbf{Z}$ within groups) and $\mathbf{E}_{YZ}$ (cross-products between DVs and covariates within groups).

Step 2: Adjust the within-group SSCP for covariates:

$\mathbf{E}^* = \mathbf{E}_{YY} - \mathbf{E}_{YZ} \mathbf{E}_{ZZ}^{-1} \mathbf{E}_{YZ}^T$

Step 3: Similarly adjust the hypothesis SSCP to obtain $\mathbf{H}^*$.

Step 4: Compute test statistics using $\mathbf{H}^*$ and $\mathbf{E}^*$ as in MANOVA.

The adjusted matrices remove the variance explained by the covariates, leaving a more precise comparison of group means on the residualised DVs.

6.3 Purposes of Covariates

Purpose 1: Increasing Power (Noise Reduction) When the covariates are strongly related to the DVs, removing their variance from $\mathbf{E}^*$ reduces error variance, increasing statistical power for the group comparison.

Purpose 2: Controlling for Confounding (Bias Reduction) When groups differ on the covariate (e.g., groups have different pre-test scores), the covariate adjustment removes this pre-existing difference, providing a fairer comparison of group effects.

⚠️ Covariates should be selected based on theoretical grounds before data collection. Adding covariates post-hoc to achieve significance is p-hacking. Including covariates unrelated to the DVs can actually reduce power.

6.4 Adjusted Means (Estimated Marginal Means)

MANCOVA produces adjusted means (also called estimated marginal means or least-squares means) — the group means on the DVs after statistically controlling for the covariates, evaluated at the grand mean of the covariate(s):

$\bar{\mathbf{y}}_j^{adj} = \bar{\mathbf{y}}_j - \hat{\mathbf{B}}(\bar{\mathbf{z}}_j - \bar{\mathbf{z}})$

Where $\hat{\mathbf{B}} = \mathbf{E}_{YZ}\mathbf{E}_{ZZ}^{-1}$ are the within-group regression coefficients.

These adjusted means represent what the group means would have been if all groups had the same mean covariate value.

6.5 Testing the Covariate Effect

In addition to testing for group differences, MANCOVA also tests whether the covariates themselves have a significant multivariate relationship with the DVs. This is tested using the same four multivariate statistics applied to the SSCP for the covariate(s).

A significant covariate test ($p < 0.05$) confirms that the covariate meaningfully reduces error variance and that including it was justified.

6.6 Homogeneity of Regression Check

Before conducting MANCOVA, test the assumption that the covariate-DV regression slopes are the same across groups:

Test: Add the group × covariate interaction to the model and test its significance using the multivariate statistics.

$H_0: \mathbf{B}_1 = \mathbf{B}_2 = \dots = \mathbf{B}_k$

If the interaction is significant ($p < 0.05$), the homogeneity of regression assumption is violated and standard MANCOVA is inappropriate. Options include:

• Testing group differences separately at different covariate values (Johnson-Neyman regions).
• Using separate regression models per group.
• Modifying the research question to examine the interaction itself.

7. Effect Size Measures

Statistical significance tells us whether group differences are likely to be real; effect size tells us how large those differences are in practical terms. Multiple effect size measures are available for MANOVA/MANCOVA.

7.1 Multivariate Effect Sizes

7.1.1 Partial Eta-Squared ($\eta^2_p$)

The most widely reported multivariate effect size, computed separately for each test statistic:

From Wilks' Lambda:

$\eta^2_p = 1 - \Lambda^{*1/t}$

Where $t$ is defined as in Section 5.1. This represents the proportion of multivariate variance associated with the group effect.

From Pillai's Trace:

$\eta^2_p = \frac{V}{s^*}$

General formula:

$\eta^2_p = \frac{SS_{effect}}{SS_{effect} + SS_{error}} = \frac{df_H \cdot F}{df_H \cdot F + df_E}$

Interpretation:

⚠️ Partial eta-squared is a biased (upward) estimator of the population effect size, especially with small samples. Prefer omega-squared ($\omega^2$) or epsilon-squared ($\varepsilon^2$) for unbiased estimates.

7.1.2 Omega-Squared ($\omega^2$)

Omega-squared provides an unbiased (or less biased) estimate of the population effect size:

$\omega^2 = \frac{df_H(F - 1)}{df_H(F - 1) + n}$

Or equivalently using SSCP components:

$\omega^2 = \frac{SS_H - df_H \cdot MS_E}{SS_T + MS_E}$

Omega-squared can be negative (round to zero when this occurs) — this indicates the population effect is likely zero or very small.

7.1.3 Epsilon-Squared ($\varepsilon^2$)

Another less-biased alternative:

$\varepsilon^2 = \frac{SS_H - df_H \cdot MS_E}{SS_T}$

For multivariate settings, $\varepsilon^2$ is computed on the univariate follow-up ANOVAs.

7.1.4 Roy's $\theta$ as Effect Size

Roy's largest root $\theta$ directly represents the proportion of total multivariate variation explained by the strongest discriminant dimension:

$\theta = \frac{\lambda_1}{1 + \lambda_1}$

Interpreted on a 0–1 scale, $\theta$ is an effect size for the first (dominant) discriminant function.

7.2 Univariate Effect Sizes for Follow-Up Tests

After a significant MANOVA, follow-up univariate ANOVAs are conducted on each DV. Report univariate effect sizes for each:

Univariate Eta-Squared:

$\eta^2 = \frac{SS_{between}}{SS_{total}}$

Univariate Partial Eta-Squared:

$\eta^2_p = \frac{SS_{between}}{SS_{between} + SS_{within}}$

Cohen's d (for two-group comparisons within a DV):

$d = \frac{\bar{y}_1 - \bar{y}_2}{s_{pooled}}$

Benchmarks for Cohen's d:

| $|d|$ | Effect Size | | :---- | :---------- | | $0.20$ | Small | | $0.50$ | Medium | | $0.80$ | Large |

7.3 Discriminant Function Effect Sizes

After MANOVA, discriminant function analysis (DFA) can be used to characterise the nature of group differences. Each discriminant function has an associated canonical correlation $r_c$:

$r_{c,s} = \sqrt{\frac{\lambda_s}{1 + \lambda_s}}$

The canonical $R^2$ ($r_{c,s}^2$) represents the proportion of variance in the discriminant scores explained by group membership for the $s$-th function:

7.4 Multivariate Effect Sizes for MANCOVA

For MANCOVA, effect sizes are computed using the adjusted SSCP matrices $\mathbf{H}^*$ and $\mathbf{E}^*$ in place of $\mathbf{H}$ and $\mathbf{E}$. The adjusted $\eta^2_p$ reflects the group effect after controlling for covariates:

$\eta^2_{p,adj} = 1 - \Lambda^{*}_{adj}{}^{1/t}$

8. Follow-Up Analyses

A significant omnibus MANOVA result indicates that groups differ on the set of DVs, but does not specify which DVs or which groups are responsible. Follow-up analyses are needed to decompose and interpret the significant multivariate effect.

8.1 Strategy for Follow-Up Analysis

A principled approach to follow-up analyses after significant MANOVA:

Level 1 — Discriminant Function Analysis (DFA) Identifies which linear combination(s) of DVs best separate the groups. This is the most theoretically informative follow-up and should always be conducted first.

Level 2 — Univariate Follow-Up ANOVAs Tests each DV separately to identify which individual variables contribute to the group differences. Use a Bonferroni-corrected $\alpha = 0.05/p$ for each test.

Level 3 — Post-Hoc Group Comparisons For significant univariate ANOVAs, conduct pairwise comparisons to identify which specific groups differ.

8.2 Discriminant Function Analysis (DFA)

After MANOVA, DFA finds the linear combinations of DVs (discriminant functions) that maximally separate the groups:

$D_s = v_{s1}Y_1 + v_{s2}Y_2 + \dots + v_{sp}Y_p$

Where $\mathbf{v}_s$ is the $s$-th eigenvector of $\mathbf{E}^{-1}\mathbf{H}$ (the discriminant weights).

Standardised discriminant coefficients: Multiply raw weights by the within-group SD of each DV:

$v_{sj}^* = v_{sj} \times \sqrt{E_{jj}/(n-k)}$

Standardised coefficients indicate the relative importance of each DV in discriminating between groups (analogous to standardised regression coefficients).

Structure coefficients (discriminant loadings): Correlations between each DV and the discriminant function scores:

$r_{D_s, Y_j} = \frac{\text{Cov}(D_s, Y_j)}{\sqrt{\text{Var}(D_s)\text{Var}(Y_j)}}$

Structure coefficients $\geq 0.30$ in absolute value are generally considered meaningful.

Number of significant discriminant functions: Use a sequential likelihood ratio test:

For the $s$-th function (after removing the effects of all previous functions):

$\Lambda_s^* = \prod_{l=s}^{s^*} \frac{1}{1 + \lambda_l}, \quad \chi^2_s = -\left[n - 1 - \frac{p + k}{2}\right]\ln(\Lambda_s^*), \quad df_s = (p - s + 1)(k - s)$

8.3 Univariate Follow-Up ANOVAs

For each significant DV from the univariate follow-up ANOVAs, report:

• $F$-statistic and p-value.
• Partial $\eta^2_p$ (effect size).
• Group means and standard deviations.

Alpha adjustment for multiple comparisons:

💡 "Protected ANOVAs" — conducting univariate tests only if the omnibus MANOVA is significant — provides some protection against Type I inflation without requiring per-test corrections, but is controversial. The Bonferroni or Holm correction is safer.

8.4 Post-Hoc Pairwise Comparisons

After a significant univariate ANOVA on a specific DV, pairwise comparisons identify which group pairs differ. Common post-hoc tests:

8.5 Roy-Bargmann Stepdown Analysis

The Roy-Bargmann stepdown analysis is a theoretically motivated follow-up procedure for ordered DVs (when there is a theoretical ordering of importance):

1. Test the most important DV (DV1) as a univariate ANOVA.
2. Test the second DV (DV2) as an ANCOVA, with DV1 as a covariate.
3. Test the third DV (DV3) as an ANCOVA, with DV1 and DV2 as covariates.
4. Continue sequentially.

This tests whether each DV adds unique group discrimination beyond the information in all higher-priority DVs. Apply Bonferroni correction across the $p$ stepdown tests.

8.6 Planned Contrasts in MANOVA

Planned (a priori) contrasts test specific theoretical hypotheses about group comparisons formulated before data collection. They are more powerful than post-hoc tests because they do not require the omnibus MANOVA to be significant.

A contrast $\boldsymbol{\psi} = \sum_j c_j \bar{\mathbf{y}}_j$ (with $\sum c_j = 0$) is tested using:

$\mathbf{H}_\psi = \frac{n_h}{\sum_j c_j^2/n_j} \boldsymbol{\psi} \boldsymbol{\psi}^T$

Where $n_h$ is the harmonic mean of group sizes. The test statistic is Hotelling's $T^2$ for the contrast.

9. Power Analysis and Sample Size

9.1 Factors Affecting Statistical Power in MANOVA

9.2 Rules of Thumb for Sample Size

MANOVA requires adequate sample sizes to:

1. Estimate the within-group covariance matrix reliably.
2. Maintain the validity of asymptotic test statistics.

Minimum requirements:

• Total sample size $n > p + k$ (to ensure $\mathbf{E}$ is invertible).
• $n_j > p$ in each group (preferably $n_j \geq 2p$).
• $n_j \geq 20$ in each group for reasonable power with moderate effects.

Rule of thumb (Stevens, 2002): $n_j \geq 20$ per group for adequate power with medium effects and 5–8 DVs.

More precise guidance (Tabachnick & Fidell):

• Small effects: $n \approx 200+$ per group.
• Medium effects: $n \approx 50–100$ per group.
• Large effects: $n \approx 20–30$ per group.

9.3 Power Analysis Using Non-Central F

For a one-way MANOVA with $k$ groups and effect size $f^2$ (Cohen's multivariate $f$):

The non-centrality parameter for Wilks' Lambda approximately follows a non-central $F$ distribution. Exact power computations require specialised software (G*Power, SAS PROC POWER) that implements the non-central Wilks' distribution.

Cohen's multivariate $f$ from $\eta^2_p$:

$f^2 = \frac{\eta^2_p}{1 - \eta^2_p}$

Benchmarks:

9.4 Impact of Number of DVs on Power

The relationship between number of DVs and power is nuanced:

• Adding informative DVs (those that differ between groups) increases power.
• Adding uninformative DVs (those that do not differ between groups and are correlated with other DVs) decreases power by consuming degrees of freedom.
• Optimal strategy: Include only DVs that are theoretically relevant and expected to differ between groups.

10. Model Fit and Evaluation

10.1 Overall Multivariate Significance

The primary model fit assessment in MANOVA is whether the four multivariate statistics reach significance:

10.2 Univariate ANOVA Results

For each dependent variable, report:

10.3 Discriminant Analysis Summary

Report the canonical correlations, eigenvalues, and variance explained for each significant discriminant function:

10.4 MANCOVA-Specific Fit Information

For MANCOVA, additionally report:

• Multivariate test of the covariate effect (using all four statistics).
• Regression coefficients $\hat{\mathbf{B}}$ for each DV on each covariate.
• Adjusted group means (estimated marginal means) for each DV.
• Test of homogeneity of regression (covariate × group interaction).

11. Assumption Checking and Diagnostics

11.1 Multivariate Normality Tests

11.1.1 Mardia's Tests

Mardia's multivariate skewness:

$b_{1,p} = \frac{1}{n^2}\sum_{i=1}^n \sum_{j=1}^n \left[(\mathbf{y}_i - \bar{\mathbf{y}})^T \mathbf{S}^{-1}(\mathbf{y}_j - \bar{\mathbf{y}})\right]^3$

Under multivariate normality:

$\frac{n \cdot b_{1,p}}{6} \sim \chi^2_{p(p+1)(p+2)/6}$

Mardia's multivariate kurtosis:

$b_{2,p} = \frac{1}{n}\sum_{i=1}^n \left[(\mathbf{y}_i - \bar{\mathbf{y}})^T \mathbf{S}^{-1}(\mathbf{y}_i - \bar{\mathbf{y}})\right]^2$

Under multivariate normality: $E[b_{2,p}] = p(p+2)$.

Test statistic:

$z_{b_{2,p}} = \frac{b_{2,p} - p(p+2)}{\sqrt{8p(p+2)/n}} \sim \mathcal{N}(0,1)$

11.1.2 Mahalanobis Distance Q-Q Plot

Compute the squared Mahalanobis distance for each observation within each group:

$D^2_{ij} = (\mathbf{y}_{ij} - \bar{\mathbf{y}}_j)^T \mathbf{S}_j^{-1} (\mathbf{y}_{ij} - \bar{\mathbf{y}}_j)$

Plot the ordered $D^2_{(i)}$ values against the quantiles of $\chi^2_p$. Under multivariate normality, points should fall approximately on a straight line. Curved patterns or isolated points at the upper right indicate non-normality or outliers.

Critical value for outlier identification: $D^2_{ij} > \chi^2_{p, 0.001}$ flags a potential multivariate outlier.

11.2 Box's M Test for Homogeneity of Covariance Matrices

$M = (n - k)\ln|\mathbf{S}_W| - \sum_{j=1}^k (n_j - 1)\ln|\mathbf{S}_j|$

Approximate chi-squared statistic:

$\chi^2_{approx} = M\left(1 - c_1\right), \quad df = \frac{p(p+1)(k-1)}{2}$

Where:

$c_1 = \frac{2p^2 + 3p - 1}{6(p+1)(k-1)}\left(\sum_{j=1}^k \frac{1}{n_j - 1} - \frac{1}{n - k}\right)$

Interpretation:

• $p > 0.05$: No evidence against homogeneity → proceed with standard MANOVA.
• $0.001 < p < 0.05$: Marginal violation → use Pillai's Trace; check group sizes.
• $p < 0.001$: Significant violation → use Pillai's Trace; consider heteroscedastic MANOVA variants.

11.3 Outlier Detection

Univariate outliers: For each DV, compute standardised scores $z_{ij} = (y_{ij} - \bar{y}_j)/s_j$. Flag observations with $|z| > 3.29$ (Bonferroni-corrected at $\alpha = 0.05$).

Multivariate outliers: Mahalanobis distances exceeding $\chi^2_{p, 0.001}$ flag multivariate outliers. These are observations unusual on the combination of DVs even if not extreme on any single DV.

Leverage and influence in MANOVA:

The hat matrix for the multivariate design matrix:

$\mathbf{H} = \mathbf{X}(\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}^T$

Leverage $h_{ii}$ (diagonal of $\mathbf{H}$) measures how much observation $i$ influences its own fitted values. High leverage ($h_{ii} > 2(p+1)/n$) combined with large residuals indicates influential observations.

11.4 Checking Linearity Among DVs

Create a scatterplot matrix (pairs plot) of all DVs within each group. Non-linear relationships suggest transformations (log, square root) or that MANOVA's use of covariances is inefficient.

Look for:

• Approximately elliptical joint distributions (consistent with multivariate normality).
• No obvious non-linear (e.g., quadratic) trends.
• No severe restriction of range in any DV.

11.5 Checking Multicollinearity Among DVs

Compute the condition number and determinant of $\mathbf{E}$:

• If $|\mathbf{E}| \approx 0$: Near-singular — multicollinearity problem.
• Condition number $> 30$: Potential multicollinearity.

Compute pairwise correlations among DVs: correlations $> 0.90$ in absolute value suggest redundancy. Consider removing one of the highly correlated DVs or combining them into a composite.

💡 While some correlation among DVs is expected (and exploited by MANOVA), extreme multicollinearity creates numerical instability. DVs with $|r| > 0.90$ provide largely redundant information and one should be removed.

11.6 Checking Homogeneity of Regression (MANCOVA)

Test: Include the group × covariate interaction in the model:

$\mathbf{y}_{ij} = \boldsymbol{\mu} + \boldsymbol{\alpha}_j + \mathbf{B}\mathbf{z}_{ij} + (\boldsymbol{\alpha\beta})_j \mathbf{z}_{ij} + \boldsymbol{\epsilon}_{ij}$

Test the multivariate significance of $(\boldsymbol{\alpha\beta})_j$ using all four test statistics. A non-significant result ($p > 0.05$) supports the homogeneity of regression assumption.

12. Contrast Analysis in MANOVA

12.1 Planned Contrasts vs. Post-Hoc Comparisons

Planned (a priori) contrasts are comparisons specified before data collection based on theory. They:

• Have more statistical power than post-hoc tests.
• Do not require a significant omnibus MANOVA (though it is conventional to require it).
• Are limited in number (typically $k-1$ orthogonal contrasts for $k$ groups).

Post-hoc comparisons are exploratory comparisons conducted after the data are seen. They require stronger corrections for multiple testing.

12.2 Contrast Coding Schemes

For a factor with $k$ levels, $k-1$ contrasts can be specified. Common coding schemes:

Helmert Contrasts: Compare each level to the mean of all subsequent levels:

• Contrast 1: Level 1 vs. mean of Levels 2, 3, ..., $k$.
• Contrast 2: Level 2 vs. mean of Levels 3, ..., $k$.
• Etc.

Deviation Contrasts: Compare each level to the grand mean (all other groups):

• Contrast $j$: Level $j$ vs. grand mean of all other levels.

Repeated/Difference Contrasts: Compare adjacent levels:

• Contrast $j$: Level $j$ vs. Level $j+1$.

Orthogonal Polynomial Contrasts: Decompose a quantitative factor's effect into linear, quadratic, cubic trends.

Custom Contrasts: Any set of contrast coefficients $\mathbf{c} = (c_1, c_2, \dots, c_k)^T$ with $\sum_j c_j = 0$.

12.3 Testing a Single Multivariate Contrast

A single multivariate contrast $\boldsymbol{\psi} = \sum_j c_j \bar{\mathbf{y}}_j$ is tested using Hotelling's $T^2$:

$T^2 = \frac{\boldsymbol{\psi}^T \mathbf{S}_W^{-1} \boldsymbol{\psi}}{\sum_j c_j^2/n_j}$

$F = \frac{n - k - p + 1}{(n - k)p} T^2 \sim F_{p,\, n-k-p+1}$

12.4 Testing a Set of Multivariate Contrasts

For a set of $s$ planned contrasts defined by contrast matrix $\mathbf{C}$ ($s \times k$), the hypothesis SSCP is:

$\mathbf{H}_C = \bar{\mathbf{Y}}^T \mathbf{C}^T \left[\mathbf{C}(\mathbf{X}^T\mathbf{X})^{-1}\mathbf{C}^T\right]^{-1} \mathbf{C} \bar{\mathbf{Y}}$

The four multivariate test statistics are computed using $\mathbf{H}_C$ in place of $\mathbf{H}$.

13. Repeated Measures MANOVA (Profile Analysis)

13.1 Repeated Measures as MANOVA

When the same participants are measured on $p$ occasions (or $p$ related conditions), the data can be analysed as a one-sample or $k$-sample MANOVA — this is called profile analysis or doubly multivariate analysis.

Profile analysis treats the $p$ repeated measurements as $p$ correlated dependent variables in a MANOVA framework. It avoids the sphericity assumption required by repeated-measures ANOVA, making it more general and robust.

13.2 The Three Hypotheses in Profile Analysis

Profile analysis simultaneously tests three distinct hypotheses:

13.2.1 Test of Parallelism (Interaction)

$H_0$: The profiles of all groups are parallel — the pattern of change across occasions is the same for all groups.

This is the most important test. A significant result means the groups differ in their pattern of responses across occasions (interaction between group and occasion).

Tested using the $p-1$ difference scores $D_m = Y_{m+1} - Y_m$ ($m = 1, \dots, p-1$) as DVs in a MANOVA.

13.2.2 Test of Levels (Group Main Effect)

$H_0$: All groups have the same average level across occasions — the overall mean response is the same across groups.

Equivalent to a one-way ANOVA on the row means $\bar{Y}_i = \sum_{m=1}^p Y_{im}/p$. Only interpretable when the parallelism test is not rejected.

13.2.3 Test of Flatness (Occasion Main Effect)

$H_0$: The profile is flat — there is no change across occasions, averaged over all groups.

Tested using a one-sample MANOVA on the $p-1$ difference scores. Only interpretable when the parallelism test is not rejected.

13.3 Multivariate vs. Univariate Approach to Repeated Measures

💡 For datasets where the sphericity assumption is questionable and sample size is adequate ($n > p + k + 5$), profile analysis (MANOVA) is generally preferred over univariate repeated-measures ANOVA with Greenhouse-Geisser or Huynh-Feldt corrections.

14. Using the MANOVA/MANCOVA Component

The MANOVA/MANCOVA component in the DataStatPro application provides a full end-to-end workflow for conducting multivariate analysis of variance and covariance.

Step-by-Step Guide

Step 1 — Select Dataset Choose the dataset from the "Dataset" dropdown. The dataset should contain at least two continuous dependent variables and one categorical grouping variable.

Step 2 — Select Analysis Type Choose the analysis type:

• One-Way MANOVA (single factor, multiple DVs)
• Factorial MANOVA (two or more factors, multiple DVs)
• One-Way MANCOVA (single factor + covariates, multiple DVs)
• Factorial MANCOVA (two or more factors + covariates, multiple DVs)
• Profile Analysis (repeated measures MANOVA)
• Hotelling's $T^2$ (two-group special case)

Step 3 — Select Dependent Variables (DVs) Select two or more continuous dependent variables from the "Dependent Variables" panel.

💡 Select DVs that are theoretically related and expected to collectively represent the outcome construct. Avoid including DVs that are conceptually unrelated or nearly perfectly correlated ($|r| > 0.90$).

Step 4 — Select Independent Variable(s) / Factor(s) Select the categorical grouping variable(s) from the "Factor(s)" dropdown. For factorial MANOVA, select two or more factors. The application will automatically create all main effects and interactions.

Step 5 — Select Covariate(s) (MANCOVA Only) For MANCOVA, select one or more continuous covariates from the "Covariate(s)" dropdown. The application will:

• Automatically test the homogeneity of regression assumption.
• Compute adjusted means (estimated marginal means).
• Report covariate effects.

Step 6 — Select Contrast Type (Optional) For planned contrasts, select the coding scheme:

• None (omnibus test only)
• Helmert
• Deviation
• Repeated/Difference
• Polynomial (Trend)
• Custom (specify contrast coefficients manually)

Step 7 — Configure Post-Hoc Tests If applicable, select the post-hoc correction method for follow-up pairwise comparisons:

• Tukey HSD (recommended for equal group sizes)
• Bonferroni
• Holm
• Games-Howell (recommended for unequal group sizes or heterogeneous variances)
• Scheffé

Step 8 — Select Confidence Level Choose the confidence level for confidence intervals and effect sizes (default: 95%).

Step 9 — Select Display Options Choose which outputs to display:

• ✅ Box's M Test
• ✅ Multivariate Test Statistics (Wilks', Pillai's, Hotelling-Lawley, Roy's)
• ✅ Univariate ANOVA Results per DV
• ✅ Descriptive Statistics (means, SDs, $n$ per group per DV)
• ✅ Effect Sizes ($\eta^2_p$, $\omega^2$, Cohen's $d$)
• ✅ Discriminant Function Analysis
• ✅ Post-Hoc Pairwise Comparisons
• ✅ Multivariate Normality Tests (Mardia's, Royston's)
• ✅ Mahalanobis Distance Q-Q Plot
• ✅ Adjusted Means Plot (MANCOVA)
• ✅ Homogeneity of Regression Test (MANCOVA)
• ✅ Error Covariance Matrix $\mathbf{S}_W$
• ✅ Discriminant Function Loadings Plot
• ✅ Group Centroid Plot (Score Plot)
• ✅ Profile Plot (for Repeated Measures)

Step 10 — Run the Analysis Click "Run MANOVA/MANCOVA". The application will:

1. Compute group means, grand means, and the SSCP matrices $\mathbf{H}$, $\mathbf{E}$, $\mathbf{T}$.
2. Run Box's M test and multivariate normality tests.
3. Compute all four multivariate test statistics and their $F$ approximations.
4. Compute effect sizes.
5. Conduct univariate follow-up ANOVAs per DV.
6. Conduct discriminant function analysis.
7. Run post-hoc pairwise comparisons (if requested).
8. Compute adjusted means (MANCOVA).
9. Generate all selected visualisations and tables.

15. Computational and Formula Details

15.1 Computing the SSCP Matrices Step-by-Step

Step 1: Compute group means and grand mean

$\bar{\mathbf{y}}_j = \frac{1}{n_j}\sum_{i=1}^{n_j} \mathbf{y}_{ij}, \quad \bar{\mathbf{y}} = \frac{1}{n}\sum_{j=1}^k n_j \bar{\mathbf{y}}_j$