StratumStatistical conventions
Where a statistic admits more than one reasonable definition, Stratum picks one and writes it down here. These rules apply across every analysis and chart, so you can always reconcile a Stratum result against another package.
Rather than restate them on every page, Stratum documents its global conventions once. Where a convention matches a well-known package, that's noted so you can line a Stratum result up against it.
Missing & non-finite values
A cell is missing when it's empty, or when a numeric cell holds a non-finite value — NaN or ±infinity. Stratum treats non-finite as missing everywhere: a value that can't be ordered or averaged isn't data.
| Situation | Behavior |
|---|---|
| Empty cell | Missing. Excluded from every statistic. |
| NaN or ±∞ in a numeric column | Treated as missing — never plotted, never entered into a sum, mean, or sort. |
| Single-variable statistic (mean, SD, quantile) | Missing values are dropped, then the statistic is computed on what remains. |
| Multi-variable analysis (correlation, regression, PCA) | Each method states whether it drops rows listwise or pairwise. Correlation is deliberately pairwise; the multivariate models use complete cases. |
Quantiles & percentiles
Quartiles, percentiles, the median, box-whisker fences, and the quantile() formula function all use one definition: linear interpolation between the two closest ranks. For n sorted values and a probability p in [0, 1], the position is
h = (n − 1) · p
quantile = v[⌊h⌋] + (h − ⌊h⌋) · ( v[⌊h⌋+1] − v[⌊h⌋] )where v is the ascending-sorted data and the median is simply p = 0.5. Requested probabilities are clamped to [0, 1].
numpy.quantile, method="linear"), pandas, and R (quantile(..., type = 7)). A Stratum quartile matches those tools exactly on the same data. Small differences against another tool usually trace to its quantile type, not an error.
Number formatting
Every user-visible number — report cells, tooltips, axis labels, exports — is rendered by one shared formatter, so the same value reads identically everywhere. How it's displayed is a presentation choice in Preferences ▸ Numbers; it never changes the value used in a computation.
| Mode | Meaning |
|---|---|
| Fixed | A fixed number of decimal places. Values too large or too small for the chosen precision spill to scientific notation so a real value never collapses to 0.0000. |
| Significant | A fixed number of significant digits, trailing zeros preserved so columns of differing magnitude share a consistent visible precision. |
| Scientific | Mantissa-and-exponent form, e.g. 9.7443e-09 — lowercase e, matching NumPy / pandas / matplotlib. |
Two rules override the global mode: a column given an explicit display format renders in that format, and integer-typed columns always render as integers (never 5.0000).
p-values
Any p below 0.0001 is shown as <.0001 rather than a long string of leading zeros. Larger values use the number mode above.
Grouping & high-cardinality limits
Splitting a chart or report by a categorical column with thousands of distinct values would fan out into thousands of series or panels. Stratum caps the number of groups, keeping the most frequent and folding the rest into a single Other (M) bucket.
| Grouping | Default cap | Range |
|---|---|---|
| Partition series (one color per group) | 12 | 2–17 (16-color palette + gray "Other") |
| Gallery panels (one panel per group) | 25 | 2–100 |
| Report groups (pivot / contingency) | 50 | 2–200 |
Each cap is the total number of buckets shown, including Other. The defaults are editable in Preferences ▸ Charts and Preferences ▸ Reports.
Randomness & reproducibility
Methods with a stochastic step — random-forest bagging, boosted-tree subsampling, cross-validation folds, K-Means initialization, point jitter — draw from a seeded generator. Re-running an analysis on the same data with the same settings reproduces the same result.
Related
- Glossary
- Quartiles & the IQR · P-Value · Standard Deviation
- Calculators
- Free statistics calculators
- In the app
- Open Stratum's built-in help (⌘/) → Statistical Reference for the per-method detail