Trending and analysis
Charts from data, normality and hypothesis tests
Upload a CSV or Excel file and ask Q Intelligence for the chart that shows the trend, test data for normality with p-values, and run statistical hypothesis tests (t-tests, Mann-Whitney, ANOVA, Fisher's exact, chi-square).
In the portal: /app/intelligence
Charts from a file
Attach the file to a Q Intelligence message and ask, for example "plot the trend of the assay results" or "make a Pareto of the deviation categories". Q Intelligence reads the column headers, picks the chart (or the one you name), draws it in the conversation and interprets it.
| Chart | Use it for | Analysis shown |
|---|---|---|
| Line | A value over time or sequence; several columns as series | n, mean, SD per series; optional linear trend (slope, R², p-value) and moving average |
| Bar | Values per category, or counts, totals or means by a category column | Highest and lowest |
| Pareto | The categories that matter most (deviation types, defect codes) | Cumulative % and the vital few (80%) |
| Histogram | The distribution of one parameter | Summary, normality tests, results outside specification |
| Box plot | One parameter across groups (line, shift, supplier) | Quartiles, outliers (1.5 × IQR), one-way ANOVA |
| Scatter | Two numeric columns | Regression line, slope, r, R², p-value |
- Several series share a chart only when they share a unit and scale; measures on different scales (assay and pH) get separate charts. No chart has two y-axes.
- Limits: 5,000 rows, 6 series, 60 categories or groups, 6 charts or analyses per answer.
- Each chart has Chart and Table views (and Normality for histograms), CSV and PNG downloads, and can be saved and pinned like a control chart.
Normality tests
| Test | Statistic | p-value |
|---|---|---|
| Shapiro-Wilk | W (Royston AS R94), 3 to 5,000 results | Royston's normalising transformation |
| Anderson-Darling | A² with the small-sample adjustment | D'Agostino and Stephens |
The data are reported normal when every p-value is at least 0.05 (no evidence against normality), and not normal otherwise, with a caution about normal-based capability and ppm. Skewness, excess kurtosis and a normal probability plot are shown. Notes warn when fewer than 8 results give little power, when 1,000 or more results reject trivial departures, and when coarse resolution causes rejection.
Hypothesis tests
Ask whether a difference or association is real, for example "is the defect rate higher on Line 2?", and Q Intelligence runs the test on the attached data: a value column grouped by a group column, two or more value columns as groups, or a count table cross-tabulated from two category columns. In an investigation, the same tests run on each fishbone cause (see Investigations).
| Test | Use it for | Reported |
|---|---|---|
| Two-sample t-test (Welch) | Two groups' means, unequal variances allowed | t, df, p, difference, confidence interval, Cohen's d |
| Paired t-test | The same units measured twice | t, df, p, mean difference, confidence interval |
| One-sample t-test | One group against a target | t, df, p, difference, confidence interval |
| Mann-Whitney U | Two groups without assuming normal data | U, p (exact for small samples), medians |
| One-way ANOVA | Three or more groups | F, df, p, highest and lowest means, η² |
| Fisher's exact test | 2 × 2 counts | p, odds ratio, the two rates |
| Chi-square test | Larger tables of counts | χ², df, p, Cramér's V, largest departure |
- Person
1. What are the data?
Measurements or counts
- Person
2. How many groups?
One (vs target), two, or more
- Person
3. Paired or independent?
Same units measured twice?
- System
4. Run the test
Computed on the server
- Person
5. Read p, effect and interval
Significance and size of the difference
- Measurements, two independent groups → Welch t-test (Mann-Whitney if skewed or very small).
- Measurements, same units twice → paired t-test. Three or more groups → one-way ANOVA.
- Counts in a 2 × 2 table → Fisher's exact test; larger tables → chi-square.
Tip: A p-value below alpha (0.05 by default) shows a difference or association, not its cause. A p-value above it does not prove the groups are the same, especially with small samples. The calculations match SciPy to about 1e-12.