3. Data collection for RCA

Root cause analysis is only as credible as the data it consumes. A statistically sophisticated analysis built on a biased sample, a drifting sensor, or an undocumented unit change will produce a confident but wrong conclusion. This chapter treats data collection as an engineering discipline in its own right: selecting and validating measurement systems, choosing a sampling strategy that matches the failure mode, assessing data quality before analysis, and capturing metadata so that the dataset remains interpretable months later.

3.1 Measurement Systems as Part of the Process

Every measurement is a process with its own variation. Before collecting data for an RCA, characterize the measurement system using a Gage R&R (repeatability and reproducibility) study. Repeatability is variation when the same operator measures the same part with the same instrument; reproducibility is variation between operators or setups. The standard acceptance heuristic is that the combined Gage R&R standard deviation should be under 10% of the total process variation for critical decisions, 10–30% may be acceptable depending on cost, and above 30% is generally unacceptable.

Consider a shaft diameter problem where the specification is 25.00 ± 0.05 mm and the observed process standard deviation is 0.018 mm. A quick Gage R&R with two operators and ten parts yields a measurement standard deviation of 0.007 mm. The ratio is 0.007 / 0.018 ≈ 39%, which is unacceptable — the measurement system is consuming nearly 40% of the observed variation. Any RCA conclusion drawn from this data would be unreliable until the instrument is recalibrated or replaced.

For non-destructive, continuous signals (vibration, temperature, pressure), the equivalent checks are bias against a reference, linearity across the operating range, and stability over time. A thermocouple that reads correctly at 25 °C but drifts 3 °C at 180 °C will silently corrupt any correlation between temperature and a failure mode.

3.2 Sampling Strategies

The sampling strategy must match the question being asked. Four strategies dominate RCA work:

• Random sampling — every unit has an equal probability of selection. Appropriate for estimating overall defect rates, but inefficient when failures are rare.
• Stratified sampling — the population is divided into strata (shift, machine, lot, supplier) and samples are drawn from each. Essential when the suspected cause varies by stratum.
• Systematic sampling — every k-th unit is selected. Simple and cheap, but dangerous if the process has a periodic component that aligns with k.
• Rational subgrouping — samples are collected in small groups produced under similar conditions. This is the foundation of control charting and is the preferred approach when the goal is to detect a shift in the process mean.

Sample size should be justified, not guessed. For detecting a shift of d standard deviations with 80% power at α = 0.05, the two-sample t-test requires roughly n ≈ 16 / d² per group. To detect a 0.5σ shift, that is about 64 units per group. For attribute data, the rule of thumb is at least 50 defectives (not 50 units) in the sample to estimate a proportion with reasonable precision.

3.3 Data Quality Assessment

Before analysis, run a structured quality pass. The dimensions that matter most for RCA are:

A practical first pass is a missingness map and a range audit. If a temperature column has 12% nulls concentrated in one shift, that is not random missingness — it is a signal about the data collection process itself, and often about the failure mode.

3.4 Metadata Capture

Metadata is what makes a dataset reusable. For RCA, the minimum metadata set includes:

• Provenance — source system, instrument ID, calibration date, operator ID
• Context — production line, shift, lot number, material batch, ambient conditions
• Semantics — unit of measure, definition of each field, coding scheme for categorical values
• Transformations — any filtering, unit conversion, or aggregation applied after collection

Without provenance, a correlation between two variables may be an artifact of two different instruments being swapped mid-study. Without semantics, a column labeled "temp" is ambiguous — is it °C or °F, ambient or process, before or after the heat exchanger?

3.5 A Reproducible Collection Workflow

The following workflow is a concrete template for an RCA data collection effort. It is written as a sequence of decisions, not a rigid procedure.

1. Define the question. Write the RCA hypothesis in one sentence. Example: "Solder voids on QFN packages increase when the reflow peak temperature exceeds 245 °C."
2. Identify variables. Response: void percentage (from X-ray). Predictors: peak temperature, soak time, paste lot, stencil aperture.
3. Select measurement systems. X-ray void measurement requires a Gage R&R; thermocouple profiling requires calibration verification against a reference.
4. Choose sampling. Stratify by paste lot and stencil aperture; sample 30 boards per stratum across three shifts.
5. Collect with metadata. Each record carries board ID, shift, operator, paste lot, stencil ID, oven recipe ID, and profile timestamp.
6. Assess quality. Run completeness, range, and consistency checks before any modeling.
7. Freeze and document. Version the dataset, record the collection date range, and store the metadata alongside the data.

In a real engagement, step 6 often reveals that the oven recipe ID was not logged for 20% of boards, or that two different X-ray systems were used with different void-detection thresholds. Both findings change the analysis plan — and both are exactly the kind of insight that a pure statistical analysis would miss.

3.6 Practical Deliverable: Data Collection Plan Template

The following template can be filled in before collection begins. It forces the team to commit to measurement validation, sampling logic, and metadata fields up front, which prevents most post-hoc data quality problems.

RCA Data Collection Plan
========================
RCA ID:            RCA-2024-017
Question:          Does reflow peak temperature > 245 C increase QFN void %?
Response variable: Void percentage (X-ray, system XR-02, cal. 2024-03-01)
Predictors:        Peak temp (C), soak time (s), paste lot, stencil aperture (mm)

Measurement validation:
  - X-ray Gage R&R: 2 operators x 10 boards, target < 10% of process sigma
  - Thermocouple: verify against reference at 150, 200, 250 C before run

Sampling:
  - Strata: paste lot (3 levels) x stencil aperture (2 levels) = 6 strata
  - n per stratum: 30 boards (detects 0.5 sigma shift, 80% power)
  - Shifts: all three, balanced

Metadata fields per record:
  board_id, shift, operator_id, paste_lot, stencil_id, oven_recipe_id,
  profile_timestamp, xray_system_id, xray_cal_date, ambient_temp_C, ambient_RH_pct

Quality gates before analysis:
  - Completeness: < 5% nulls per field
  - Range: peak temp 200-270 C, void % 0-100
  - Consistency: units verified, categorical values from controlled vocabulary
  - Timeliness: profile timestamp within 5 min of board exit

Freeze:
  - Dataset version: v1.0
  - Collection window: 2024-04-01 to 2024-04-15
  - Stored with metadata in /rca/RCA-2024-017/data/

This plan is not bureaucracy — it is the difference between a dataset that supports a defensible root cause and one that produces a plausible but unverifiable story. The next chapter builds on this foundation by introducing the statistical tools used to analyze the collected data.