Statistical analysis for GMP: control charts, capability, CPV, method validation and shelf life

What the module does

• Control charts of individual results with a moving range chart: alert limits at 2 sigma, action limits at 3 sigma, sigma from the average moving range or the standard deviation, and limits from all results, a baseline of the first results or an approved historical mean and sigma; fewer than 20 results for the limits are flagged as preliminary.
• Nelson's eight rules or the four Western Electric rules, each violation named with its points, and specification limits drawn apart from the statistical limits, so being in statistical control and being within specification are reported as separate questions.
• Capability (Cp, Cpk, CPU, CPL, Pp, Ppk, Cpm) with expected parts per million outside specification and the observed count, caveats when the process is out of control, has fewer than 30 results or varies more overall than within; normality by Shapiro-Wilk and Anderson-Darling with the probability plot.
• Hypothesis tests on attached data in Q Intelligence: Welch, paired and one-sample t-tests, Mann-Whitney, one-way ANOVA, Fisher's exact and chi-square, each with the null hypothesis, the statistic, p-value, effect size and confidence interval; the calculations are checked against SciPy.
• Statistical Analysis studies opened blank or from a workflow (continued process verification, capability, attribute monitoring, comparing groups or methods, method validation, stability shelf life, measurement systems, DOE, clinical, reliability), with data from CSV, Excel or PDF or pasted, each file fingerprinted and every dataset kept unedited.
• Q Intelligence as the study's statistician: it chooses the analyses and the assumption checks each needs, the engine runs them on every row, and the AI explains the conclusion, the size of the effect, what a failed assumption means and whether the result meets the guidance, cited by code and section.
• Menus as in a statistics package: descriptive statistics, tolerance intervals and outliers, t-tests, equivalence by two one-sided tests, power and sample size, ANOVA and regression, logistic regression, nonparametrics, reliability and survival, time series, principal components and clustering, each with its method guide.
• Process control for continued process verification: I-MR, Xbar-R, Xbar-S, P, NP, C and U charts, Laney P′ and U′, EWMA and CUSUM for small shifts, limits fixed from a historical baseline or set per stage, trend tests, capability for non-normal data by Box-Cox, lognormal or Weibull, and the process performance report.
• ICH Q2(R2) linearity, accuracy, precision and detection and quantitation limits; ICH Q1E shelf life with poolability at 0.25 and out-of-trend detection; Gage R&R and attribute agreement; acceptance sampling plans; designed experiments with the response optimizer; clinical sample size, randomization, Kaplan-Meier and bioequivalence.

How it fits the platform

A control chart can be made from a signal review in Signal detection and refreshed from it later, saved for routine trending and pinned to the Q Dashboard. In an investigation, the same hypothesis tests run on each fishbone cause and the result goes into the record. A continued process verification study links to the CPV plan in Validation, and the CPV report reads its capability studies.

Each analysis is read against the guidance a QA statistician cites with it, from FDA Process Validation stage 3 and EU GMP Annex 15 to ICH Q2(R2), ICH Q1E, AIAG MSA and ICH E9, with a link to read it in the regulatory library where it is indexed. A study is a record like any other: every upload, analysis, question, draft and edit is in its hash-chained audit trail, and its locked PDF carries the protocol, each dataset with its source and fingerprint, the analyses, the conversation, the report, the signatures and the audit trail.

Where people decide

The engine computes every estimate, interval, p-value and index; the AI chooses the analyses, explains them against the guidance and drafts the statistical protocol and the report, and never computes a number. People decide: the author edits the protocol the study follows, takes out an analysis that should not stand with a reason, and routes the study for review and approval, where the signer's electronic signature makes it the statistical record of a decision. A chart flags a point beyond an action limit or a run rule; whether it needs a deviation or an investigation is decided under your SOP.