Issues

[Factorial] A blank cell for a categorical factor is reported as "nan"

kgdunn/process-improve#1006 · opened

on Oct 27, 2026

analyze_experiment overflows on responses near 1e300 in the sums of squares

kgdunn/process-improve#1005 · opened

on Oct 27, 2026

variance_components reports an arbitrary split, with standard errors near 2e7, for a term with one reading per level

kgdunn/process-improve#1004 · opened

on Oct 27, 2026

reshape_to_long silently merges panelist labels that differ only by whitespace, or fails with a misleading message

kgdunn/process-improve#1003 · opened

on Oct 27, 2026

A response named "yield" is refused as an unsafe formula when the data also has a column "yield_"

kgdunn/process-improve#1002 · opened

on Oct 27, 2026

Binomial GLM: a trials column with a zero is not checked, and statsmodels fails with a NaN deviance

kgdunn/process-improve#1001 · opened

on Oct 27, 2026

analyze_experiment with no complete run fails inside numpy instead of saying so

kgdunn/process-improve#1000 · opened

on Oct 27, 2026

A one-component mixture region with any constraint crashes in extreme_vertices

kgdunn/process-improve#999 · opened

on Oct 27, 2026

recommend_next_step reads a shift between blocks as curvature when the centre runs are in a block of their own

kgdunn/process-improve#998 · opened

on Oct 27, 2026

analyze_experiment calls a model estimable when its interaction column is constant in the runs

kgdunn/process-improve#997 · opened

on Oct 27, 2026

Run the example notebooks in examples/ in CI, and fix the two that read data from URLs that return 404

kgdunn/process-improve#987 · opened

on Oct 26, 2026

Fit a factor named I or C under a stand-in name, so the name works instead of being refused

kgdunn/process-improve#986 · opened

on Oct 26, 2026

diagnose(): Hotelling's T² is a DataFrame from PCA and a Series from PLS; make the family consistent

kgdunn/process-improve#983 · opened

on Oct 25, 2026

TPLS: an all-NaN column is dropped silently, while a constant column gets a warning

kgdunn/process-improve#982 · opened

on Oct 25, 2026

TPLS: a fit with more than one Z group fails with a shape error

kgdunn/process-improve#981 · opened

on Oct 25, 2026

Hotelling's T² limits: rows used to fit the model and new rows need different laws; choose an approach

kgdunn/process-improve#978 · opened

on Oct 25, 2026

Mixture designs and models beyond the simplex lattice: the full simplex centroid, axial blends at a chosen distance, McLean and Anderson's design, U-pseudocomponents, ratio variables, component-amount designs, orthogonal blocks (Box and Draper chapters 16 to 18)

kgdunn/process-improve#973 · opened

on Oct 25, 2026

Design criteria over a region: integrated variance and squared bias, region and design moments, all-bias and minimum-bias designs, lack-of-fit detectability, continuous design measures (Box and Draper chapters 6, 7, 13, 14)

kgdunn/process-improve#972 · opened

on Oct 25, 2026

Canonical analysis: Box and Draper's A form, ridge forms with several small eigenvalues, a canonical summary, fitted canonical models, and reduction under linear restrictions (chapters 10, 11 and 20)

kgdunn/process-improve#971 · opened

on Oct 25, 2026

1

Ridge analysis: a focus, chosen radii or multipliers, all the stationary paths, and linear restrictions (Box and Draper chapters 12 and 19)

kgdunn/process-improve#970 · opened

on Oct 25, 2026

Lack of fit: zero pure error (or zero lack-of-fit df) returns "Insufficient degrees of freedom" instead of the analysis

kgdunn/process-improve#969 · opened

on Oct 25, 2026

Lack of fit: pure error pools replicates across blocks when the block is written `C(Block)`, so the lack-of-fit sum of squares goes negative with no warning

kgdunn/process-improve#968 · opened

on Oct 25, 2026

Plackett-Burman: 88 to 91 factors give 96 runs with no warning, though the 92-run design exists (and likewise at 116, 156, 172, 184 and 188)

kgdunn/process-improve#967 · opened

on Oct 25, 2026

`augment_design(..., "foldover", generators=...)` reports the defining words of the combined design without their signs

kgdunn/process-improve#966 · opened

on Oct 25, 2026

`canonical_analysis`: `b_canonical` is NaN, with divide-by-zero warnings, when two or more eigenvalues are zero

kgdunn/process-improve#965 · opened

on Oct 25, 2026

1

Navigation Menu

Repository navigation

Issues

[Factorial] A blank cell for a categorical factor is reported as "nan"

analyze_experiment overflows on responses near 1e300 in the sums of squares

variance_components reports an arbitrary split, with standard errors near 2e7, for a term with one reading per level

reshape_to_long silently merges panelist labels that differ only by whitespace, or fails with a misleading message

A response named "yield" is refused as an unsafe formula when the data also has a column "yield_"

Binomial GLM: a trials column with a zero is not checked, and statsmodels fails with a NaN deviance

analyze_experiment with no complete run fails inside numpy instead of saying so

A one-component mixture region with any constraint crashes in extreme_vertices

recommend_next_step reads a shift between blocks as curvature when the centre runs are in a block of their own

analyze_experiment calls a model estimable when its interaction column is constant in the runs

Run the example notebooks in examples/ in CI, and fix the two that read data from URLs that return 404

Fit a factor named I or C under a stand-in name, so the name works instead of being refused

diagnose(): Hotelling's T² is a DataFrame from PCA and a Series from PLS; make the family consistent

TPLS: an all-NaN column is dropped silently, while a constant column gets a warning

TPLS: a fit with more than one Z group fails with a shape error

Hotelling's T² limits: rows used to fit the model and new rows need different laws; choose an approach

Mixture designs and models beyond the simplex lattice: the full simplex centroid, axial blends at a chosen distance, McLean and Anderson's design, U-pseudocomponents, ratio variables, component-amount designs, orthogonal blocks (Box and Draper chapters 16 to 18)

Design criteria over a region: integrated variance and squared bias, region and design moments, all-bias and minimum-bias designs, lack-of-fit detectability, continuous design measures (Box and Draper chapters 6, 7, 13, 14)

Canonical analysis: Box and Draper's A form, ridge forms with several small eigenvalues, a canonical summary, fitted canonical models, and reduction under linear restrictions (chapters 10, 11 and 20)

Ridge analysis: a focus, chosen radii or multipliers, all the stationary paths, and linear restrictions (Box and Draper chapters 12 and 19)

Lack of fit: zero pure error (or zero lack-of-fit df) returns "Insufficient degrees of freedom" instead of the analysis

Lack of fit: pure error pools replicates across blocks when the block is written `C(Block)`, so the lack-of-fit sum of squares goes negative with no warning

Plackett-Burman: 88 to 91 factors give 96 runs with no warning, though the 92-run design exists (and likewise at 116, 156, 172, 184 and 188)

`augment_design(..., "foldover", generators=...)` reports the defining words of the combined design without their signs

`canonical_analysis`: `b_canonical` is NaN, with divide-by-zero warnings, when two or more eigenvalues are zero