Age | Commit message (Collapse) | Author | Files | Lines |
|
|
|
|
|
|
|
The likelihood ratio test method is lrtest, in addition,
methods for update and residuals were added.
|
|
The cutoff now matches what is given by Venzon and Moolgavkar (1988).
Also, confidence intervals closely match intervals obtained with
stats4::confint in the test case where an stats4::mle object
is created from the likelihood function in one test case.
Static documentation rebuilt by pkgdown
|
|
Static documentation rebuilt by pkgdown
|
|
|
|
The confint method can do profile likelihood based confidence intervals!
|
|
|
|
|
|
|
|
|
|
in the hope that this makes plotting cross-platform also for this error
model
|
|
|
|
|
|
Especially on winbuilder (i386 and amd64)
|
|
Because on winbuilder obviously gcc was not found, so the Eigenvalue
based solution method was used, leading to a test failure when
comparing the summary, as the solution method is listed
|
|
|
|
Static documentation rebuilt by pkgdown
|
|
|
|
mkinfit objects now include an ll() function to calculate the
log-likelihood. Part of the code was refactored, hopefully making it
easier to read and maintain. IRLS is currently the default algorithm for
the error model "obs", for no particular reason. This may be subject
to change when I get around to investigate.
Slow tests are now in a separate subdirectory and will probably
only be run by my own Makefile target.
Formatting of test logs is improved.
Roundtripping error model parameters works with a precision of 10% when
we use lots of replicates in the synthetic data (see slow tests). This
is not new in this commit, but as I think it is reasonable this
closes #7.
|
|
generated with mkin < 0.9.49.5
|
|
|
|
Static documentation rebuilt by pkgdown
|
|
The test uses multiple cores in order to complete within a reasonable time
|
|
|
|
One of the tests exceeded the number of iterations when using the
d_3 error model algorithm, so only use "direct" in this case.
|
|
|
|
The new algorithm tries direct optimization of the likelihood, as well
as a three step procedure. In this way, we consistently get the
model with the highest likelihood for SFO, DFOP and HS for all 12
new test datasets.
|
|
All of them are working now and allow for comparison
Based on SFO, DFOP and HS fits to twelve test datasets, only
the combination of direct and threestep is needed to find
the lowest AIC
|
|
|
|
|
|
by plotting squared residuals against predicted values, and
showing the variance function used in the fitted error model.
Rebuild docs
|
|
Static documentation rebuilt by pkgdown
|
|
Now we have a three stage fitting process for
nonconstant error models:
- Unweighted least squares
- Only optimize the error model
- Optimize both
Static documentation rebuilt by pkgdown
|
|
- Skip long running tests on CRAN as well to avoid timeout on winbuilder
- Don't install benchmark results in the package, they are only needed
in the git repository
- Don't run example in man/add_err.Rd as it takes > 10 s on winbuilder
- Rebuild docs
|
|
- Improve control of the number of cores
- Reduce the precision of the correlation matrix in the test summary
output, as the exact results are platform dependent
|
|
- Also make it possible to specify initial values for error model
parameters.
- Run tests
- Rebuild docs
|
|
Remove skipped tests as I do not intend to reactivate them
|
|
Also reduce the digits in the representative half-live given by nafta()
|
|
- Write the NEWS
- Static documentation rebuilt by pkgdown
- Adapt mkinerrmin
- Fix (hopefully all) remaining problems in mkinfit
|
|
- No IRLS required
- Removed optimization algorithms other than Port
- Removed the dependency on FME
- Fitting the error model 'obs' is much faster for the FOCUS_2006_D
dataset and the FOMC_SFO model (1 second versus 3.4 seconds)
- Vignettes build slower. Compiled models needs 3 minutes instead of 1.5
- For other vignettes, the trend is less clear. Some fits are faster,
even for error_model = "const". FOCUS_Z is faster (34.9 s versus
44.1 s)
- Standard errors and confidence intervals are slightly smaller
- Removed code for plotting during the fit, as I hardly ever used it
- Merged the two cost functions (using transformed and untransformed
parameters) into one log-likelihood function
|
|
|
|
In order to be able to test cross-platform
|
|
Static documentation rebuilt by pkgdown
|
|
|
|
|
|
|
|
Also test the model specification via the link argument
|
|
Also:
- Change rounding in print.nafta
- Add dots argument to nafta()
- Use cores=1 in examples
- Restrict N in IORE model to values > 0
|