Use Gage R&R to Investigate Inconsistent Parts Measurements

When two inspectors report different values for the same feature, checking the gage's calibration certificate is necessary but incomplete. Calibration connects an instrument indication to reference values under stated conditions. A Gage R&R study examines variation in the measurement process as it is actually used, including the gage, fixture, method, operators, parts and environment.

The existing discussion of precision CNC machining tolerances for track rollers explains why controlled measurement matters. This article focuses on designing a study that represents the inspection task. It contains no fabricated dataset, universal study size or automatic acceptance threshold.

Define the measurement task before the study

Identify the part, feature and drawing revision and define the measurand and units. Record the tolerance and how the result is used: setup adjustment, process monitoring, product acceptance or another decision. A study of a convenient diameter does not establish performance for a different feature, range or geometry.

Freeze the measurement process. Name the gage, fixture, software and versions, the work instruction, resolution, part preparation, measurement location and orientation. Include routine zeroing, fixturing and calculation steps. The study should evaluate the real process rather than an ideal demonstration that operators will not use in production.

Define the operating population: operators, shifts, sites, equipment configurations and environmental conditions the conclusion is intended to represent. Record the known inconsistency that triggered the study. If reports disagree only on one geometry, one shift or one fixture, retain that pattern so the design can test it.

Decide whether measurement is non-destructive and meaningfully repeatable on the same physical item. Some tests alter or consume the sample; others change the surface, seating or condition enough that a repeat is not the same task. This decision controls whether a crossed design is possible.

Keep calibration and verification evidence with the study, but separate. A calibrated gage can still show operator, setup, fixture, geometry or short-term repeatability variation in use. The NIST/SEMATECH measurement-process handbook frames Gauge R&R among broader issues including operators, artifacts, configurations, repeatability, reproducibility, stability, bias, resolution and geometry.

Name the study owner and the people authorized to approve the design and interpret it. The owner should resolve missing task definitions before data collection begins.

Choose a crossed or nested design

Use a crossed structure when every appraiser can measure every selected part under meaningful repeat conditions. This layout helps separate part-to-part variation, within-appraiser repeatability and differences associated with appraisers. It assumes the same physical items remain suitable across measurements and that run order and handling do not create an uncontrolled change.

Use a nested structure when different but defensibly comparable parts must be assigned to different appraisers, as can happen with destructive or altered measurements. Part variation is then nested within appraiser and can be confounded with appraiser effects. The conclusion and model must acknowledge that limitation rather than describing the study as equivalent to a crossed design.

A destructive hardness section, tensile coupon or metallographic sample cannot be forced into a crossed design by calling adjacent nonuniform locations the same part. The study owner must justify the assumption that assigned pieces or regions are comparable. If that assumption is weak, redesign the study, obtain appropriate matched material or use qualified statistical guidance.

An expanded study can include additional factors such as site, gage, fixture or relevant interactions, including mixed or unbalanced structures when an appropriate analysis supports them. Adding factors is useful only when the data structure can estimate their effects. Do not let software availability dictate a design the physical process cannot support.

Minitab's factor guidance distinguishes crossed, nested and expanded designs according to how parts and operators are related. Before data collection, record whether operator is treated as fixed or random for the intended inference, identify all factors and interactions, state whether the plan is balanced, and select the analysis method.

Define repeatability as variation under the study's stated repeated conditions and reproducibility as variation across stated conditions such as appraisers or times. An Instron overview provides that conceptual distinction without establishing a universal design or threshold.

Prepare the randomized study inputs

Select parts that span the process variation the measurement system must handle. Avoid a set containing only consecutive pieces, visually easy parts, known good parts or known rejects. The goal is representation of routine geometry and range, not construction of an artificially quiet study.

Select appraisers who represent routine users, shifts or sites included in the intended conclusion. Record training status without coaching everyone into an unfamiliar special technique during the run. If a revised instruction is the subject of evaluation, identify that as a controlled factor or a separate study phase.

Choose repeats and total study structure based on the objective, expected variation, practical constraints and qualified statistical plan. There is no universal number of parts, operators or repeats in this article. Minitab's study guidance supports representative parts and randomized repeated measurements while leaving actual counts to the study design.

Randomize or blind run order where practical to reduce memory, drift and sequence effects. Reposition and re-fixture between repeats when routine inspection includes those actions. Otherwise the study can omit an important source of variation. Use the normal work instruction and a representative environment.

Record gage warm-up, zero and verification steps. Capture values at the system's native reporting precision and do not manually round before analysis. Define how failed or invalid readings are recorded and reviewed. If a part changes during the study, preserve the event and affected runs rather than deleting inconvenient data.

Study-brief state	Task and design	Inputs and execution	Action
Normal: design represents use	The measurand, population and crossed, nested or expanded rationale match the physical measurement process.	Representative parts and appraisers, justified repeats, randomized order, re-fixturing, raw data and deviations are controlled.	Run the preselected analysis and interpret it only for the stated population.
Missing: study brief incomplete	Part range, measurand, repeatability assumption, factors, model or intended decision is absent.	Randomization, repeats, appraiser coverage, re-fixturing, data precision or invalid-reading rules are not defined.	Close the design gaps before collecting or accepting a study result.
Conflict: structure contradicts process	A destructive or altered measurement is forced into a crossed design, or a calibration pass is offered instead of process-variation evidence.	Assigned parts are not demonstrably comparable, or actual execution departs from the frozen design.	Preserve the conflict, redesign under qualified guidance and do not invent missing results.

Store the randomized run plan, raw data, analysis file and software version under document control. Keep a deviation log. This is a design brief, not a completed study or a KTSU result.

Interpret the reported metrics in context

Read variance components before reducing the output to one percentage. Repeatability represents within-condition measurement variation under the study model. Reproducibility or appraiser-related components address defined between-condition effects. Where modeled, appraiser-by-part interaction indicates that appraiser differences change across parts.

Part-to-part variation describes the selected sample under the model; it is not measurement error. Total Gage R&R combines specified measurement-system components. Inspect graphs, residual patterns, run order, outliers and geometry groupings because a pooled summary can hide a problem confined to one feature or configuration.

Always identify the denominator and multiplier. Percent contribution, percent study variation and percent tolerance answer different questions. Percent tolerance is meaningful only when a valid tolerance for the studied measurand is entered. The same estimated measurement variation can appear different across these metrics because their denominators differ.

If the software reports the number of distinct categories, retain its model and input basis. Review confidence information or estimate stability where available; a small or unrepresentative study can produce unstable component estimates. Record software, analysis model and version so another reviewer can reproduce the output.

Do not apply familiar 10 or 30 percent bands as universal pass/fail thresholds. Acceptance criteria depend on the application, risk, customer requirements and quality system. The authorized owner must decide which metrics and thresholds govern before using them.

A standard Gage R&R study is also not a complete measurement-uncertainty budget. A NIST publication abstract on uncertainty and R&R makes that boundary explicit. Gage R&R remains valuable for studying repeatability and reproducibility; it simply does not capture every uncertainty contribution needed for a conformity statement.

Decide what to improve and retest

Use the observed pattern to develop candidates, not automatic corrections. Dominant repeatability variation may justify investigating fixture seating, method steps, gage behavior, resolution or environment. Appraiser-related variation may point to instruction clarity, setup interpretation or training. An interaction can show that the issue changes by geometry or part range.

Stability and drift studies remain separate over-time questions, while bias and linearity require their own references and designs. Do not label every unexplained component as operator error or assume a new gage is the remedy. Assign a corrective-action owner and investigate causes with the relevant technical teams.

Where feasible, make one controlled change, revise the work instruction and repeat a comparable randomized study using the same scope and analysis. If the scope must change, document why and limit direct before-and-after claims. Never delete the original study; link the follow-up data, analysis and approval.

If variation depends on geometry, configuration or site, stratify or redesign with qualified statistical guidance rather than hiding it in a pooled result. Keep the product-impact assessment separate. Study findings may trigger review of earlier measurements, but they do not automatically correct, accept, reject or release product.

The final status should state whether the study design represents routine measurement, lacks required inputs or conflicts with the physical process. Results and improvement decisions then remain open until their authorized owners act. This article does not establish a measurement-system-analysis capability for any supplier or release a lot; it provides the evidence structure needed to investigate inconsistency without inventing certainty.