Gauge R&R: ANOVA Method, Formulas & Lab Guide

Q: Why does %Study Variation not sum to 100% like %Contribution?

%Contribution uses additive variance components (σ 2 ). %Study Variation uses standard deviations (σ = √σ 2 ), which do not sum linearly.

Q: How does ANOVA Gauge R&R differ from the Average and Range (X-bar/R) method?

The Average and Range method relies on empirical range multipliers and cannot isolate operator-by-part interaction. The ANOVA method mathematically decomposes all variance components, tests interactions, and provides exact variance estimates.

Gauge R&R (Gauge Repeatability and Reproducibility, also written as Gage R&R) is a core statistical method within Measurement Systems Analysis (MSA) that evaluates whether an inspection system produces reliable data. By isolating equipment variation from operator inconsistency, a Gauge R&R study determines whether observed process variance reflects actual part differences or measurement error.

Definition in One Paragraph

Gauge Repeatability and Reproducibility (Gauge R&R) is a statistical experimental methodology that evaluates the precision of a measurement system by decomposing total observed variation into part-to-part variance, equipment repeatability (within-operator error under identical conditions), and appraiser reproducibility (between-operator or between-setup variation). Using a balanced repeated-measures experimental design analyzed through analysis of variance (ANOVA), the study quantifies measurement error as a percentage of total process variation (%GRR) and product tolerance, verifying that the measurement tool possesses sufficient resolution and repeatability to support statistical quality control.

Formula and Symbols Explained

In measurement systems analysis, the crossed two-way random-effects ANOVA model with replication is the standard analytical technique. This approach builds upon the foundational difference between repeatability vs reproducibility in method precision and the classic distinction between accuracy vs precision in measurement.

Two-Way ANOVA Model and Variance Decomposition

For a crossed study with p parts, o operators, and r replicate trials, the linear model is:

Y_ijk = μ + P_i + O_j + (PO)_ij + ε_ijk

Where Y_ijk is the observation on part i by operator j during trial k; μ is the grand mean; P_i ∼ N(0, σ²_Part); O_j ∼ N(0, σ²_Operator); (PO)_ij ∼ N(0, σ²_Interaction); and ε_ijk ∼ N(0, σ²_{Repeatability}).

Degrees of freedom (DF) and Mean Squares (MS = SS / DF) partition as:

DF_Parts = p − 1
DF_Operators = o − 1
DF_Interaction = (p − 1)(o − 1)
DF_{Repeatability} = p × o × (r − 1)
DF_Total = (p × o × r) − 1

Variance components (σ²) are isolated from expected mean squares:

Repeatability: σ²_{Repeatability} = MS_{Repeatability}
Interaction: σ²_Interaction = max[0, (MS_Interaction − MS_{Repeatability}) / r]
Operator: σ²_Operator = max[0, (MS_Operators − MS_Interaction) / (p × r)]
Reproducibility: σ²_{Reproducibility} = σ²_Operator + σ²_Interaction
Gauge R&R: σ²_GRR = σ²_{Repeatability} + σ²_{Reproducibility}
Part Variation: σ²_Part = max[0, (MS_Parts − MS_Interaction) / (o × r)]
Total Variation: σ²_Total = σ²_GRR + σ²_Part

Pooling Condition: When F = MS_Interaction / MS_{Repeatability} yields p > 0.05, the interaction is pooled into repeatability: MS_{Repeatability(pooled)} = (SS_{Repeatability} + SS_Interaction) / (DF_{Repeatability} + DF_Interaction).

Key Evaluation Metrics and ndc

The AIAG MSA 4th Edition framework evaluates measurement capability through three standard ratios plus the category discrimination index:

%Contribution: 100 × (σ²_Component / σ²_Total) (additive, sums to 100%).
%Study Variation (%SV or %GRR): 100 × (σ_Component / σ_Total) (standard deviations, does not sum to 100%).
%Tolerance (%P/T): 100 × (6 × σ_GRR / Tolerance), where Tolerance = Upper Spec Limit − Lower Spec Limit.
Number of Distinct Categories (ndc): ndc = ⌊ 1.41 × (σ_Part / σ_GRR) ⌋ (truncated to nearest integer).

Worked Example with Real Numbers

An aerospace laboratory evaluated a digital micrometer measuring spacer thicknesses (nominal 10.25 mm, tolerance span = 0.50 mm). The study used p = 5 parts, o = 2 operators, and r = 2 replicate trials (N = 20 total observations).

Part	Op1, T1	Op1, T2	Op2, T1	Op2, T2	Part Mean
1	10.15	10.17	10.18	10.16	10.1650
2	10.35	10.34	10.37	10.38	10.3600
3	10.02	10.04	10.05	10.03	10.0350
4	10.48	10.50	10.52	10.51	10.5025
5	10.22	10.23	10.26	10.25	10.2400
Op Mean	10.2500 mm		10.2710 mm		Grand: 10.2605 mm

ANOVA Computations and Pooling

Sum of squares calculated around the grand mean (10.2605 mm):

SS_Total = 0.519295 (DF = 19)
SS_Parts = 4 × Σ(PartMean − 10.2605)² = 0.515420 (DF = 4, MS = 0.128855)
SS_Operators = 10 × Σ(OpMean − 10.2605)² = 0.002205 (DF = 1, MS = 0.002205)
SS_Interaction = 0.000420 (DF = 4, MS = 0.000105)
SS_{Repeatability} = 0.001250 (DF = 10, MS = 0.000125)

Because interaction testing yields F = 0.840 (p = 0.5304 > 0.05), interaction is pooled into repeatability:

SS_{Repeatability(pooled)} = 0.001250 + 0.000420 = 0.001670 mm²
DF_{Repeatability(pooled)} = 10 + 4 = 14
MS_{Repeatability(pooled)} = 0.001670 / 14 = 0.00011929 mm²

Variance Components and Derived Metrics

From the pooled ANOVA table, the variance components and standard deviation values are:

σ²_{Repeatability} = 0.00011929 mm² → σ_{Repeatability} = 0.010922 mm
σ²_{Reproducibility} = (0.002205 − 0.00011929) / 10 = 0.00020857 mm² → σ_{Reproducibility} = 0.014442 mm
σ²_GRR = 0.00011929 + 0.00020857 = 0.00032786 mm² → σ_GRR = 0.018107 mm
σ²_Part = (0.128855 − 0.00011929) / 4 = 0.03218393 mm² → σ_Part = 0.179399 mm
σ²_Total = 0.00032786 + 0.03218393 = 0.03251179 mm² → σ_Total = 0.180310 mm

Performance metrics:

%Contribution: GRR = 1.01%; Part = 98.99%.
%Study Variation (%GRR): 100 × (0.018107 / 0.180310) = 10.04%.
%Tolerance (%P/T): 100 × (6 × 0.018107 / 0.50) = 21.73%.
ndc: ⌊ 1.41 × (0.179399 / 0.018107) ⌋ = ⌊ 13.97 ⌋ = 13.

Interpretation and Acceptance Criteria

Decision criteria from AIAG MSA 4th Edition and ASTM E2782 guide measurement system acceptability:

Metric	Acceptable (Green)	Conditionally Acceptable (Yellow)	Unacceptable (Red)
%GRR (% Study Variation)	< 10%	10% to 30% (based on application importance/cost)	> 30% (system needs overhaul)
%Contribution (% Variance)	< 1%	1% to 9%	> 9%
%Tolerance (%P/T Ratio)	< 10%	10% to 30%	> 30%
Number of Distinct Categories (ndc)	≥ 5	n/a (strict threshold)	< 5 (insufficient discrimination)

In our worked example, the %GRR is 10.04% and %Contribution is 1.01%, placing it comfortably within acceptable production limits. The ndc of 13 easily satisfies the ≥ 5 requirement, confirming the micrometer can distinguish part variation for process control. In high-reliability facilities, these evaluations are frequently paired with installation qualification (IQ) protocols and ISO 22514-7 measurement uncertainty budgets.

Assumptions and Common Mistakes

Valid Gauge R&R conclusions depend on satisfying standard statistical and operational assumptions:

Calibrated Instrument: Gauge R&R isolates precision. It assumes instrument bias has been eliminated through calibration against traceable standards.
Process-Representative Parts: Parts must span the entire operating spread of the process. Narrow part sampling artificially suppresses σ_Part, inflating %GRR and reducing ndc.
Blind, Randomized Sequence: Operators must measure parts in randomized order without knowing the part identification to prevent memory bias.
Confusing %Contribution with %Study Variation: Do not report a 10% Study Variation as “1% error” or vice-versa. Always specify whether percentages denote variance or standard deviation.
Unaddressed Significant Interaction: When p ≤ 0.05 for Part × Operator interaction, pooling is prohibited. Investigate why specific appraisers obtain discrepant results on particular parts.
Destructive Testing Restrictions: Standard crossed ANOVA requires non-destructive testing of identical parts. Destructive testing requires a nested Gauge R&R model.

How to Report It (APA Example)

When documenting measurement validation in formal research or audits, follow APA 7th edition statistics reporting standards:

“A crossed two-way analysis of variance (ANOVA) was conducted to evaluate measurement system repeatability and reproducibility across five parts and two operators with two replications (N = 20). The part-by-operator interaction was not statistically significant, F(4, 10) = 0.84, p = .530. Following pooling into repeatability (MS_error = 0.000119 mm²), main effects showed significant part variation, F(4, 14) = 1080.22, p < .001, and significant operator variation, F(1, 14) = 18.49, p < .001. Gauge repeatability and reproducibility (GRR) accounted for 1.01% of total variance (%Contribution) and 10.04% of study variation (%SV, SD = 0.0181 mm). Across the 0.50 mm tolerance band, %Tolerance was 21.73%. The number of distinct categories was ndc = 13. Exceeding the benchmark of ndc ≥ 5 and %GRR ≤ 30%, the measurement system was accepted for quality assurance.”

How to Calculate It in Excel, R, and Python

Excel Calculation

Use Excel’s Analysis ToolPak:

Organize data in a two-factor table (rows = parts, columns = operators, sub-rows = trials).
Click Data > Data Analysis > Anova: Two-Factor With Replication.
Input your data range and enter Rows per sample = 2.
Extract Mean Squares: MS_Within represents repeatability; MS_Interaction represents interaction. If interaction p > 0.05, compute pooled repeatability: =(SS_Interaction + SS_Within)/(df_Interaction + df_Within).

R Script

# Crossed Gauge R&R in R
df <- data.frame(
  Part = factor(rep(1:5, each = 4)),
  Operator = factor(rep(rep(c("Op1", "Op2"), each = 2), 5)),
  Trial = factor(rep(1:2, 10)),
  Y = c(10.15, 10.17, 10.18, 10.16,
        10.35, 10.34, 10.37, 10.38,
        10.02, 10.04, 10.05, 10.03,
        10.48, 10.50, 10.52, 10.51,
        10.22, 10.23, 10.26, 10.25)
)

fit <- aov(Y ~ Part * Operator, data = df)
res <- summary(fit)[[1]]

p_inter <- res["Part:Operator", "Pr(>F)"]
if (p_inter > 0.05) {
  ms_rep <- (res["Part:Operator", "Sum Sq"] + res["Residuals", "Sum Sq"]) /
            (res["Part:Operator", "Df"] + res["Residuals", "Df"])
} else {
  ms_rep <- res["Residuals", "Mean Sq"]
}

var_rep  <- ms_rep
var_op   <- max(0, (res["Operator", "Mean Sq"] - ms_rep) / 10)
var_grr  <- var_rep + var_op
var_part <- max(0, (res["Part", "Mean Sq"] - ms_rep) / 4)
var_tot  <- var_grr + var_part

pct_grr  <- 100 * (sqrt(var_grr) / sqrt(var_tot))
ndc      <- floor(1.41 * (sqrt(var_part) / sqrt(var_grr)))
cat(sprintf("%%GRR: %.2f%% | ndc: %d
", pct_grr, ndc))

Python Script

import numpy as np
import pandas as pd
import statsmodels.api as sm
from statsmodels.formula.api import ols

data = [
    (1, "Op1", 1, 10.15), (1, "Op1", 2, 10.17),
    (1, "Op2", 1, 10.18), (1, "Op2", 2, 10.16),
    (2, "Op1", 1, 10.35), (2, "Op1", 2, 10.34),
    (2, "Op2", 1, 10.37), (2, "Op2", 2, 10.38),
    (3, "Op1", 1, 10.02), (3, "Op1", 2, 10.04),
    (3, "Op2", 1, 10.05), (3, "Op2", 2, 10.03),
    (4, "Op1", 1, 10.48), (4, "Op1", 2, 10.50),
    (4, "Op2", 1, 10.52), (4, "Op2", 2, 10.51),
    (5, "Op1", 1, 10.22), (5, "Op1", 2, 10.23),
    (5, "Op2", 1, 10.26), (5, "Op2", 2, 10.25),
]
df = pd.DataFrame(data, columns=["Part", "Operator", "Trial", "Y"])

model = ols('Y ~ C(Part) + C(Operator) + C(Part):C(Operator)', data=df).fit()
anova = sm.stats.anova_lm(model, typ=2)

p_inter = anova.loc['C(Part):C(Operator)', 'PR(>F)']
if p_inter > 0.05:
    ss_pool = anova.loc['C(Part):C(Operator)', 'sum_sq'] + anova.loc['Residual', 'sum_sq']
    df_pool = anova.loc['C(Part):C(Operator)', 'df'] + anova.loc['Residual', 'df']
    ms_rep = ss_pool / df_pool
else:
    ms_rep = anova.loc['Residual', 'sum_sq'] / anova.loc['Residual', 'df']

ms_op = anova.loc['C(Operator)', 'sum_sq'] / anova.loc['C(Operator)', 'df']
ms_part = anova.loc['C(Part)', 'sum_sq'] / anova.loc['C(Part)', 'df']

var_rep = ms_rep
var_op = max(0.0, (ms_op - ms_rep) / 10.0)
var_grr = var_rep + var_op
var_part = max(0.0, (ms_part - ms_rep) / 4.0)
var_tot = var_grr + var_part

pct_grr = 100.0 * (np.sqrt(var_grr) / np.sqrt(var_tot))
ndc = int(np.floor(1.41 * (np.sqrt(var_part) / np.sqrt(var_grr))))
print(f"%GRR: {pct_grr:.2f}% | ndc: {ndc}")

Frequently Asked Questions

What is the difference between repeatability and reproducibility?

Repeatability is equipment variation under constant conditions (same appraiser, same instrument, short timeframe). Reproducibility is appraiser variation across changing conditions (different operators, setups, or days).

When should the part-by-operator interaction term be pooled?

Under AIAG MSA 4th Edition rules, interaction is tested via an F-test at α = 0.05. If p > 0.05, the interaction is not statistically significant and is pooled into the repeatability residual error term.

Why does %Study Variation not sum to 100% like %Contribution?

%Contribution uses additive variance components (σ²). %Study Variation uses standard deviations (σ = √σ²), which do not sum linearly.

What is an acceptable ndc value in a Gauge R&R study?

An ndc ≥ 5 is required for process monitoring and Statistical Process Control (SPC). An ndc < 5 means the measurement system lacks discrimination to divide parts into meaningful process tiers.

How does ANOVA Gauge R&R differ from the Average and Range (X-bar/R) method?