inferMM 0.1.1
- Fixed the profile estimator’s covariance inverse for singleton
clusters by explicitly retaining one-by-one diagonal matrix
dimensions.
- Clustered data cleaning now removes rows with a missing value in any
clustering column, with a warning reporting the number removed. Both
estimators, variance-model screening and reported cluster counts use the
same retained data. Added regression tests for singleton clusters,
missing and multi-column IDs, unused factor levels, NaN IDs and
refitting.
- Strengthened projected-cluster numerical diagnostics: non-finite
tail grid evaluations no longer discard an otherwise verified interior
root, and refinement warnings are consolidated and recorded. Both the
scalar and joint scaled scores are checked. Failures report their stage,
reason and numerical diagnostics without returning estimates or
intervals; there is no automatic switch to the profile method.
Non-finite variance shapes, unrepresentable scales and non-positive
joint covariance diagonals fail explicitly. Boundary variance moments
and incomplete searches have distinct structured warning reasons. Legacy
profile iteration-limit warnings now report the last parameter change
and explicitly mark the fit/intervals as unconverged.
- Added opt-in
cluster_mm(method = "projected"), an
MM-specific Ma-Genton-style conditional projected-score construction
with the identity instrument. It retains conditionally Gaussian
residuals but does not specify the random-effect density. A scalar Km
score is followed by explicit moment updates and a stacked
independent-cluster sandwich covariance, not the legacy mean-information
approximation or a general efficiency guarantee. Refits retain this
method; pointwise mean-confidence intervals are available, but
distribution-free prediction quantiles are not. The legacy profile
method remains the default. Root, boundary and covariance checks stop
invalid fits without returning estimates or intervals.
- Single-curve profile fits now verify standardized mean-score
residuals, compare detected acceptable roots by weighted residual sum of
squares, and reject boundary, non-positive, non-finite or numerically
degenerate solutions. Score minimization supplies candidates, not
automatic success. Single-curve covariance uses scaled unregularized
information and errors on singular or poorly conditioned matrices
instead of silently adding ridge. Diagnostics are stored in
convergence; failures carry class
inferMM_profile_error, also inheriting from
inferMM_convergence_error. No estimates or intervals are
returned after a validation error. Automatic variance selection records
failed candidates and selects only successful fits. Screening/grouping
and the higher-moment profile pilot share the new root checks; the
separate clustered solver is unchanged.
- Expanded higher-moment convergence errors to identify the stage,
stopping reason, iterations, scaled score norm, tolerance, and
continuation progress. Failures can be caught as
inferMM_convergence_error; no fit or confidence intervals
are returned. Non-Gaussian continuation failures suggest a manual retry
with normal working moments and the same variance function, without
automatic fallback or a guarantee of convergence or non-Gaussian
efficiency.
- Added the opt-in
fit_mm_moments() estimator for
independent curves. It jointly solves MM mean/variance estimating
equations weighted using third and fourth conditional moments, without
calling nls().
- Working standardized moments can be estimated from pilot residuals,
set using centered-gamma, Student-t or normal moments, or supplied as
numbers or concentration-dependent functions. Fixed user-defined
variance shapes are supported. This is not a parametric error-density
likelihood fit.
- Added full joint working-information (default) and optional sandwich
covariance, Wald intervals, pairs-bootstrap refits, and pointwise
mean-confidence bands. The nuisance scale is included before extracting
the mean covariance.
- Added continuation/root diagnostics and explicit failure for
unconverged, boundary, or ill-conditioned solutions. Higher moments do
not identify predictive quantiles, so this estimator does not report
prediction bands.
- Added standardized Student-t simulation errors with degrees of
freedom greater than four. The existing centered-gamma simulation
remains available.
- Existing
fit_mm(), variance screening/grouping, and
clustered estimators retain their defaults. The new estimator does not
claim automatic finite- sample coverage, robustness to an incorrect
variance shape, or dominance over all weighted or transformation-based
methods.
inferMM 0.1.0
fit_mm() accepts a fixed user-defined positive function
of concentration; variance_function() can attach a
descriptive label to it.screen_mm() and group_mm() accept named
variance_functions lists for comparison with built-in
candidates, including custom-only screening.- Custom functions are validated without silently clipping invalid
values and retained for prediction, plotting, and wild/pairs bootstrap
refits.
- Fixed custom shapes retain the scalar Km profile-score calculation
and closed-form Vmax and variance-scale updates. Shape parameters are
not fitted; inference and information criteria condition on the supplied
shape.
- The clustered fitting algorithm and built-in variance choices are
unchanged. User-defined shapes do not implement higher-moment efficient
scores.
- Corrected the help pages to call
set.seed() outside
package functions; random-number handling is unchanged.
inferMM 0.0.3
- Removed the
nls fitting path so all single-curve fits
now use the manuscript’s Michaelis-Menten profile-score estimator.
- Updated clustered initialization to use the pooled single-curve
profile-score fit without an
nls fallback.
- Refreshed package documentation to reflect the unified profile-score
implementation.
inferMM 0.0.1
- Initial development version.
- Added variance-aware Michaelis-Menten fitting and screening
functions.
- Added grouped workflows, simulation helpers, tests, and demo
data.