Linearity Regression (Bias vs. Reference)
Core computation for MSA Type 4: regression of individual bias values against the reference value.
Description
For N individual measurements with references x_ref,i and readings y_ij, OLS fits b_ij = a · x_ref,i + b̂ + ε. Returns slope, intercept, standard errors SE(a) and SE(b̂), t-statistics, two-sided p-values, 95 % confidence intervals for both parameters and a point-wise 95 % confidence band around the fitted line. Feeds both the AIAG %Linearity KPI and the VDA 5 u_BI contribution.
Formulas
Assumptions
- Bias values are normally distributed conditional on x_ref
- Homoscedasticity: constant variance across the measurement range
- ≥ 5 distinct references and ≥ 3 repetitions per reference
- Repeatability conditions (one operator, one instrument, short time frame)
Limitations
- The regression only detects linear drift — quadratic or step-wise non-linearity is poorly captured
- Under heteroscedasticity the SE, p-values and confidence band become biased
References
- AIAG — Measurement Systems Analysis Reference Manual, 4th Edition (MSA-4), Kapitel III-B: Linearity
- VDA Band 5 — Prüfprozesseignung, 3. Auflage, Anhang zur Linearitäts-Bias-Studie
- Draper, N. R. & Smith, H. — Applied Regression Analysis, 3rd Edition, Wiley