Friedman Test: Nonparametric Analysis in RCBD

• Friedman Test is a rank-based nonparametric method for detecting differences among treatments in randomized complete block designs, offering a robust alternative to traditional F-tests.
• It ranks observations within blocks to compute a statistic that approximates a chi-square distribution under large sample conditions, with explicit finite-sample corrections.
• Recent developments include improved F-transformations, graphical diagnostics (S-plot), and precise power/sample size calculations to guide practical applications.

The Friedman test is a rank-based nonparametric procedure employed for detecting differences among treatments in a randomized complete block design (RCBD), often functioning as a robust substitute for the traditional parametric F-test. The test is especially notable for its broad applicability to settings lacking normality or exhibiting non-homogeneous error structures. Its asymptotic justification, extensions for accurate finite-sample inference, and connections to graphical and post-hoc methodologies have been extensively analyzed, with recent research providing explicit guidelines for type I error control, power, and sample size determination (Jan et al., 21 Mar 2025, Gaunt et al., 2021, Elamir, 2022).

1. Statistical Model and Fundamental Hypotheses

In the RCBD context, the model is conventionally expressed as

$$X_{ij} = \mu + \theta_i + \gamma_j + \varepsilon_{ij}$$

where $X_{ij}$ denotes the observation for treatment $i$ in block $j$, $\mu$ is a grand mean, $\theta_i$ and $\gamma_j$ are fixed treatment and block effects, and $\varepsilon_{ij}$ are assumed to be independent and continuously distributed. The null hypothesis is

$H_0: \theta_1 = \theta_2 = \cdots = \theta_K = 0$

against the alternative that at least one $\theta_i$ differs. The error term distribution is unrestricted except for continuity, rendering the test nonparametric (Jan et al., 21 Mar 2025, Elamir, 2022).

2. Construction of the Friedman Statistic

Within each block $X_{ij}$0, observations are ranked among the $X_{ij}$1 treatments, resulting in ranks $X_{ij}$2 for treatment $X_{ij}$3. The sum of ranks for each treatment is

$X_{ij}$4

The classic Friedman statistic is

$X_{ij}$5

or, equivalently, for $X_{ij}$6 blocks and $X_{ij}$7 treatments,

$X_{ij}$8

where $X_{ij}$9 and $i$0 are functionally equivalent statistics depending on notation (Jan et al., 21 Mar 2025, Gaunt et al., 2021, Elamir, 2022).

3. Null Distribution, Asymptotics, and Chi-Square Approximation

Under $i$1, the $i$2 are exchangeable with mean $i$3 and variance $i$4 (Gaunt et al., 2021). By the central limit theorem for dependent variables, as $i$5 (large block count), the Friedman statistic converges in distribution to a chi-square with $i$6 degrees of freedom:

$i$7

However, this approximation is conservative: the true type I error rate is below nominal $i$8, especially as $i$9 increases. Gaunt & Reinert (2021) provide explicit finite-sample bounds:

• For smooth test functions, the error is $j$0, with the optimal rate being achieved only when $j$1.
• For the Kolmogorov distance, the error is $j$2, vanishing if and only if $j$3 (Gaunt et al., 2021).

Scenario	Bound on $j$4
General smooth functions	$j$5 (explicit $j$6 in (Gaunt et al., 2021))
Kolmogorov	$j$7

These finite-sample results enable practitioners to gauge the validity of the chi-square approximation—if $j$8, alternative approaches such as permutation tests or increasing $j$9 are warranted (Gaunt et al., 2021).

4. Improved Transformations and Power Analysis

To address conservativeness and enhance small-sample performance, transformations of the Friedman statistic to an $\mu$0 distribution have been proposed. The general transformation is

$\mu$1

with variants:

• "Kendall’s" $\mu$2: specific form using $\mu$3, $\mu$4.
• Proposed $\mu$5: aligns numerator degrees of freedom with classical ANOVA ($\mu$6), yielding

$\mu$7

The $\mu$8 has been shown to provide type I error rates tightly matched to nominal $\mu$9 (within $\theta_i$0) for a range of $\theta_i$1 and $\theta_i$2.

Noncentral $\theta_i$3 approximations under heterogeneous location shifts, incorporating explicit power functions, have also been derived. For the $\theta_i$4 variant:

$\theta_i$5

with closed-form expressions for the noncentrality parameter $\theta_i$6 and the means/variances of rank statistics under the alternatives for Uniform, Normal, Laplace, and Exponential distributions. Monte Carlo verification demonstrates power estimation error $\theta_i$7 for $\theta_i$8 with $\theta_i$9, in contrast to errors exceeding 20% for classic noncentral $\gamma_j$0 approaches (Jan et al., 21 Mar 2025).

5. Graphical and Single-Step Approaches

Recent work proposes consolidating global and post-hoc analyses through the "S-plot" approach, visualizing standardized group-rank contributions:

$\gamma_j$1

with the Friedman statistic decomposed as $\gamma_j$2. A gamma approximation for $\gamma_j$3 enables precise calculation of decision limits for family-wise error control (e.g., Bonferroni). This approach replaces classical cascades of Nemenyi or Conover pairwise tests:

• Plot $\gamma_j$4 across $\gamma_j$5; any $\gamma_j$6 exceeding its decision limit identifies the treatments most responsible for rejecting $\gamma_j$7.
• Simulation studies confirm type I error control close to nominal levels and efficiency in identifying significant effects, particularly for moderately large $\gamma_j$8 (Elamir, 2022).

Post-hoc Approach	Number of Tests	Error Control
Classical (Nemenyi)	$\gamma_j$9	Family-wise (FWER)
S-plot	$\varepsilon_{ij}$0 standardized ranks	FWER (Bonferroni)

6. Sample Size Determination and Practical Guidance

Explicit sample size calculations are enabled by the explicit power expressions for the $\varepsilon_{ij}$1 transformation. The minimum $\varepsilon_{ij}$2 achieving power at least $\varepsilon_{ij}$3 at level $\varepsilon_{ij}$4 is the smallest $\varepsilon_{ij}$5 such that

$\varepsilon_{ij}$6

where parameters incorporate the underlying effect size through closed-form probabilities (derivable for Uniform, Normal, Laplace, Exponential shifts). Usually, a grid search or Newton–Raphson is used, plugging in the relevant power function and distributional parameters (Jan et al., 21 Mar 2025).

7. Empirical Validation, Limitations, and Recommendations

• The chi-square approximation is conservative, with error worsening as $\varepsilon_{ij}$7 grows or $\varepsilon_{ij}$8 decreases; $\varepsilon_{ij}$9 furnishes type I error control much closer to the nominal level.
• The $H_0: \theta_1 = \theta_2 = \cdots = \theta_K = 0$0 transformation with noncentrality $H_0: \theta_1 = \theta_2 = \cdots = \theta_K = 0$1 yields the best overall accuracy for power and sample size estimation under a broad range of data-generating processes.
• S-plot graphical methods drastically reduce the number of post-hoc tests and provide an immediate, interpretable visualization for both global and specific treatment effects, contingent on moderately large block sizes for moment-based fit (Jan et al., 21 Mar 2025, Elamir, 2022).
• The classical method and S-plot both maintain type I error within Bradley’s interval $H_0: \theta_1 = \theta_2 = \cdots = \theta_K = 0$2 at 5% nominal, with accuracy improving as $H_0: \theta_1 = \theta_2 = \cdots = \theta_K = 0$3 increases.
• If block–treatment interaction or ties are present, additional adjustments or alternative methodologies may be necessary.

In summary, the Friedman test remains a foundational nonparametric tool for RCBD analysis. Recent developments on accurate $H_0: \theta_1 = \theta_2 = \cdots = \theta_K = 0$4-transformations, noncentral power approximations, explicit error bounds, and single-step graphical diagnostics extend its reliability and usability for modern experimental designs (Jan et al., 21 Mar 2025, Gaunt et al., 2021, Elamir, 2022).