NEWS
tweedie 2026-11-15
Summary
Numerous changes, mostly internally. These changes substantially increase the
relative accuracy of the algorithms for dtweedie() and especially for ptweedie().
tweedie_profile() is much faster.
Crucially, the interpolation is no longer used for dtweedie() but will be
reinstated in a future update in a separate function; the purpose of this update
is improving accuracy. The result is much greater relative accuracy, but some
evaluations are slightly slower (though not painfully so). This will also mean
that the faster interpolation method will produce more accurate results when it
is reinstated.
Some of the important changes for improving accuracy include:
- new methods are introduced for computation (e.g., direct integration for the CDF);
- improved (corrected) saddlepoint approximations;
- more computations performed on the log-scale to prevent under- and over-flow;
- a re-scaling identity now is used for the CDF.
dtweedie() and ptweedie() have new inputs for log-outputs. More warnings and
errors are reported when accuracy targets are not met.
These changes have all been prompted by preparing the Tweedie CDF paper. For
example, the direct integration method inspired the more accurate computation
of the density.
New functions
- tweedie_convert_back(): convert from underlying to the GLM parameters.
- ptweedie_tail_integrate(): direct integration of density for tail probabilities.
- [dp]tweedie_regimes(): which method dtweedie() and ptweedie() use, and where.
Changed and fixed
- tweedie_profile(): prior weights are now used in the likelihood
(dispersion phi/weights); the MLE of p is the exact maximum of the spline (not
the best of 50 grid points), and CIs are found with uniroot(); several bugs
fixed (e.g., failed fits no longer stop the profile; returned p values match
L and phi); faster (the phi search is narrower, and L is reused from it).
- tweedie_convert(): updated and fixed error in gamma parameters
- ptweedie_inversion(): with details = TRUE, now also returns small_tail, flagging
tail probabilities below 1e-8. The inversion computes F = 1/2 -/+ I/pi, so its
error is absolute (a few times 1e-16), and such small tails can have poor relative
accuracy. The new Accuracy section of the help page explains this. No warning is
given (as for the other methods), and the values returned are unchanged; ptweedie()
already recomputes small tails by integrating the density.
- dtweedie_saddle(), dtweedie_series(), ptweedie_series(): new Accuracy sections in
the help pages, describing where each method is less accurate.
- ptweedie() and dtweedie(): Numerous improvements that improve accuracy and precision
- ptweedie() and dtweedie(): Numerous new warnings in the code when accuracy is not reached or fails
- ptweedie() and dtweedie(): Some refining of the inversion approach
- dtweedie() and ptweedie(): The thresholds that decide which method dtweedie() and ptweedie() use are now all in one place, tweedie_regime_settings() (R/tweedie_regimes.R); the internal functions dtweedie_regime() and ptweedie_regime() say which regime each value falls in. The results are unchanged
- ptweedie(): Use the re-scaling identity to simplify CDF computations
- ptweedie(): Increased a computation limit to allows greater precision
- ptweedie(): new log.p input; if TRUE, the log of the CDF is returned
- ptweedie(): Add lower.tail argument for ptweedie() functions and ptweedie_series()
- ptweedie(): New testthat tests for the log.p and lower.tail changes in ptweedie()
- ptweedie_series(): Changes in the ptweedie_series() approach as per paper
- ptweedie_series(): New input: max.terms, so that the truncation of the series can be controlled
- ptweedie(): using saddlepoint approx and direct integration, greatly improving accuracy
- ptweedie(): for p > 2, uses the Lugannani-Rice saddlepoint approximation when the dispersion is very small (xi = mu^(2-p)/{phi (p-2)} > 1e7, and in the body when xi > 1e6 and phi mu^(p-2) <= 1e-5), where the inversion and tail integration can fail
- ptweedie(): for p > 2 and xi > 1e4: tail probabilities below 1e-2 (not 1e-5) are computed by integrating the density, and the integrated value is used whenever the integration succeeds (before, between 1e-10 and 1e-5 the inversion value was kept if the two disagreed, which near p = 2 gave silent errors of up to 0.5%)
- ptweedie(): for p > 2 and 1e5 < xi <= 1e7: values in the body are also computed by integrating the density (the inversion's error there reached about 1e-8 near p = 2; now about 1e-9 or better), with the inversion where the integration returns no value; this is slower (about 0.1 s per value)
- ptweedie(): for p > 2: the saddlepoint approximation is also used at xi = 1e7 when phi mu^(p-2) <= 1e-5, where it is more accurate than the integration
- ptweedie(): for p > 2: where the inversion fails (near p = 2, for very small y and large dispersion it could return 0.5, or 0, whatever the true value), the density is integrated or, failing that, the saddlepoint approximation is used; these values are rough, and still flagged by a warning
- ptweedie(): y = NA gives NA (it gave an error)
- ptweedie_inversion(): a result is now flagged (exitstatus = 1, with a warning) when the pre-acceleration zone reaches its cap, since acceleration then starts where it is not valid
- dtweedie(): new argument log; if TRUE, the log of the density is returned, computed without underflow (each method works on the log scale), so it stays finite far into the tails where the density underflows to zero. tweedie_AIC(), logLiktweedie() and tweedie_profile() now use it, so an observation far in the tail no longer makes the log-likelihood -Inf. dtweedie_series() and dtweedie_inversion() return the log density too (details = TRUE: logf, logdensity)
- dtweedie(): no longer interpolates. For each value it uses the saddlepoint approximation with its asymptotic correction where z = phi y^(p-2) is small enough (its estimated relative error below 1e-13), otherwise the series (1 < p < 2; and p > 2 where its estimated error is below 1e-12), otherwise the Fourier inversion. The relative error is now below about 4e-13 everywhere tested (before, up to about 1e-5 for p near 1 at small dispersion, and about 2.5e-7 for 7 <= p <= 10)
- dtweedie_series() for 1 < p < 2: each term is computed with dpois() and dgamma(), and the terms are summed from each value's own peak, so it keeps full accuracy at large lambda (before, relative errors of about 1e-7 at lambda = 1e6) and no longer needs a huge matrix when the values of y are widely spread (it could exhaust the memory)
- dtweedie_series() for p > 2: only values that can be summed in double precision are tried, and the error estimate counts only the terms that matter; with details = TRUE it also returns the estimated relative error (relerr)
- dtweedie_inversion(): the method argument is now used (it was ignored, and the method with the smallest multiplier was always chosen; method = NULL does that); the default, method 3, keeps the relative accuracy in the tails. This also makes ptweedie() more accurate for p > 2 where it integrates the density
- dtweedie(): y = NA gives NA (it gave an error)
- tweedie_dev(): computed without cancellation, so it has full relative accuracy when y is close to mu (it lost most of its digits there); this also makes the saddlepoint and inversion densities more accurate at small dispersion
- Numerous new testthat tests
- Minor coding fixes and improvements
- Minor improvements and typo fixes in man pages throughout
- Fixed typos in README.Rmd.
- Change FORTRAN functions sin/cosPi to avoid clashes
tweedie 2026-05-20
Summary
Minor fixes and improvements
Changed
- Added details input to dtweedie_series(), that returns the lo and hi index for the summing of terms.
Other necessary changes made downstream.
- Added (tweedie_integrand(), tweedie_profile()) and improved (tweedie_plot()) the passing of plotting parameters.
Fixed
- Fixed tweedie_integrand: Plot 2 now plots cos( Im k(t) ) for the PDF (not sin).
tweedie 2026-04-26
Summary
Minor fixes and improvements
Changed
- Improvements to tweedie_plot() and passing plotting options
Fixed
- Fix error with accelerate when returning early: return properly(Thanks Jeonghwan Lee)
tweedie 2026-02-26
Changes:
- Fixed an error with xi = 1 and phi \ne 1 in qtweedie
(thanks Milan Bouchet-Valat).
- Added a test to prevent this again.
- Relocated hex.R and fixed the hex-producing script.
- Fix Makefile.
tweedie 2026-02-16
Changes:
- Improved the vignette.
- Some internal renaming.
- Fix some xi = 1 cases (thanks Milan Bouchet-Valat).
- Relocate some messages in tweedie_profile().
- Add poison example to vignette.
tweedie 2026-02-07
Changes:
- Trying to fix bugs that pop up (seemingly at random) with rhub etc. checks.
tweedie 2026-01-30
Changes:
- FORTRAN code restructured to make the similar flow in the three zones
(initial; pre-acceleration; acceleration) clearer
- Some fixes to documentation to pass tests.
- Some minor fixes to R code.
tweedie 2026-01-20
Changes:
- Some fixes to implementation of IGexact
tweedie 2025-11-29
Changes:
- Add IGexact for [dp]tweedie_inversion: whether to use exact values or inversion when p = 3.
- Fixed some comments
Changes:
- All code moved from FORTRAN77 to FORTRAN90. Almost no FORTRAN code remains from version < 3.
- PDF and CDF computations consolidated and code shared where possible, substantially reducing the amount of FORTRAN code.
- Separated FORTRAN code into different files for easier debugging.
- Improvements to the acceleration algorithm and root-finding algorithms, so should work better for more cases.
- Added verbose (shows what's happening behind the scenes) and details (reports on the fitting) as options for many user-facing R functions.
- Added ptweedie_inversion() to the man page for dtweedie.
- Tidied the man pages; added examples.
- Removed the almost-never used dtweedie.stable() function.
- Changed function names (e.g, tweedie.convert() to tweedie_convert()).
- Separated R functions into separate files depending on purpose (e.g., dtweedie.R and ptweedie.R).
- Moved the tweedie_Extra files into the main package.
- dtweedie.igrand() (now tweedie_igrand()) to plot the integrand for the DF also.
tweedie 2022-08-17
Changes:
- Added outputs gamma.mean and gamma.phi to tweedie.convert()
- Added more error checks to tweedie.convert() to prevent a model being provided
- Minor edits
tweedie 2017-11-15
Changes:
- Updated ptweedie.series() to fix a bug (reported by Lu Yang), where incorrect answers could sometimes be returned.
- Other minor fixes
tweedie 2017-11-06
Changes:
- Fixed an issue with AICtweedie(), where the incorrect AIC was given when prior weights used (reported by David Scollnik)
- Fixed a compilation error (in the subroutine smallp(), where variables were declared as initialised (thanks to Iñaki Úcar [email protected])
- Other minor edits.
tweedie 2017-08-23
Changes:
tweedie 2017-08-22
Changes:
- Fixed tweedie.f() to keep it quiet more often (sometimes, diagnostic reports meant for internal monitoring, were printed)
- Fixed a problem reported by Gustavo Lacerda, where ptweedie() returned NaN
- As a result, the series is now used in far more cases when 1<xi<2 for ptweedie()
- Changed rtweedie() algorithm for the case 1 < xi < 2 (thanks to Carlos J. Gil Bellosta)
- New function tweedie.convert() added
tweedie 2016-12-19
Changes:
- Added CITATION file
- Fixed an issue where the Tweedie cdf could return a value greater than one (reported by Jeremie Juste)
- Minor tidy of FORTRAN code
- Minor tidies in R code
tweedie 2014-06-06
Changes:
- Some administrative fixes for CRAN
tweedie 2013-09-10
Changes:
- Fixed an issue where ptweedie() would fail for very small y; set this to 0 when y<1.0e-300 (based on a report by Johann Cuenin)
tweedie 2013-01-15
Changes:
- Admin release (e.g. .First.lib() removed)
- Some minor edits to manual
- Added the control input to tweedie.profile() (thanks to Giri Khageswor, DPI Victoria)
- Minor fixes in the manual
tweedie 2012-11-01
Changes:
- Changed the example in tweedie-package to execute faster (CRAN requirement)
tweedie 2012-10-31
Changes:
- Fixed an error in dtweedie() that reported NA in the case when power=1 and phi != 1 (thanks to Dina Farkas)
- [dqpr]stable now in package stabledist rather then fBasics; fixed
- Fixed some typos in the help for tweedie-package (thanks to Peng Yu)
- rtweedie() reported an error if power = 1; fixed (thanks for Peng Yu)
- Edits to conform with new first argument of .Fortran (i.e. .NAME rather than name)
- Added NAMESPACE
tweedie 2011-06-09
Changes:
- Changed tweedie.profile() to ignore values of p/xi outside (1, 2) rather than report an error.
- In some unusual cases, when p/xi = 0 was used (with add0 = TRUE), the mle of p was between 0 and 1 (which is impossible). We report a warning message to check the data and the call to tweedie.profile(), but then set the mle to the value of p/xi giving the larger value of the likelihood
- Made the functions usually called by users able to accept xi or power
- A few minor edits to FORTRAN code; some variables not declared
tweedie 2011-06-08
Changes:
- Minor fixes to documentation
- Fixed the add0 input
- Some minor changes to the code to tidy up adding the zero
- If values of xi/p are given between 0 and 1, or less than 0, they are now omitted (with a message) rather than creating an error. If there are no values left after omitting the problem value, an error message is given.
tweedie 2010-09-30
Changes:
- Ensured tweedie.profile() does not use power = 1. This case (power=1 and phi not equal to 1) is too hard for me to deal with at present.
- Fixed an error introduced in version 2.0.5, where the value of xi.vec/p.vec was set to 1.2 (y >= 0) or 1.5 (y > 0) when not explicitly specified
- Fixed an error that reported the wrong mle of phi when the mle occurred at an endpoint of the given xi values.
tweedie 2010-08-27
Changes:
- Change to dtweedie.inversion() to ensure density = 0 is returned when y < 0 (1 < p < 2) or p <- 0 (when p > 2)
- Change to tweedie.profile() to fix a problem that p = 1 returned an error
- Changed so that tweedie.profile() works with p/xi = 0 when add0=TRUE (default is FALSE)
- Fixed some minor outputting messages (when verbose == 2)
- Location of CITATION file moved to correct location
tweedie 2010-07-12
Changes:
tweedie 2009-12-18
Changes:
- Changed default p.vec: There were too many values
- Added the facility to refer to p as xi in line with GLMs text
tweedie 2009-11-17
Changes:
- Changed the default p.vec when 1 < p < 2 to seq(1.2, 1.8, by = 0.05) (it was seq(1.2, 1.8, by = 0.1) )
- Slightly changed default output (added sep = "" to some paste commands)
- Added AICtweedie() to compute AIC for Tweedie glms
tweedie 2009-09-19
Changes:
- Corrected specification of GPL; no changes to functionality, code or documentation
- In dtweedie(), added power == 3 (inverse Gaussian) as a special case
- Minor change in documentation
- Added data, weights, and offset input arguments
- Fixed a bug where "=>" was used rather than ">="
- Fixed a bug in tweedie.profile(): exact zeros still allowed p = 2 (but not p > 2)
tweedie 2009-08-10
Changes:
- Made method = "inversion" the default (was "series")
- Slightly changed default output (added sep = "" to some paste commands)
- An error introduced earlier (unsure when exactly; prob v 1.6.1)
- In trying to identify and fix the error, tidied some of the FORTRAN code
- Made do.smooth = TRUE the default (was FALSE)
- If p.vec is not supplied, tweedie.profile() makes a sensible guess
- Made verbose = FALSE the default (was TRUE)
- Made minor changes to output when verbose = FALSE
tweedie 2009-07-18
Changes:
- Correct error in dtweedie.saddle() when y = 0 for 1 < p < 2 (Thanks to Glenn Meyers [email protected])
tweedie 2009-02-16
Changes:
- Fixed a bug in ptweedie.inversion() where variable cdf was undefined
- Fixed a bug in qtweedie() where any(power) < 1 should have been any(power < 1)
tweedie 2009-02-06
Changes:
- Fixed error in tweedie.f() (twice) where a float was used in a do-loop
- Fixed a bug in ptweedie(): It now returns 0 when y < 0 (rather than reporting an error)
- Fixed a bug in ptweedie.inversion(): y.len was undefined
- Fixed a bug in dtweedie.inversion(): if( any(phi < 0) ) written as if (any(phi) < 0 )
- Built for R version 2.6.0
- Changed the tweedie.plot() to produce no default x and y labels
tweedie 2007-09-01
Changes:
- Built for R version 2.5.1
- Removed the inversion computation from the example for dtweedie(); it was very slow and hence caused problems
tweedie 2007-05-03
- Built for R version 2.5.0
- Fixed citation details
- Added note in dtweedie() help page indicating where the methods are defined (Dunn and Smyth, 2007).
- Added tweedie-package.Rd
- Fixed a small bug in dtweedie.inversion() (replaced p with power)
- Built for R version 2.5
- Made change to use of pmatch() in function tweedie()
- Swapped the order of methods 1 and 2 in the call of tweedie.inversion to make consistent with a change in the associated paper Dunn and Smyth.
- Fixed an uninitialized variable error in the FORTRAN code which sometimes caused the inversion procedure to fail.
- Other minor changes (fixing spelling errors, updated Dunn and Smyth reference; etc.)
- Some very minor changes (some cosmetic) in the code
- Made the code more robust, exiting gracefully when Inf and NA appear in the computed log-likelihood
- Slightly changed the output plot in tweedie.profile(), to reflect that some value of p.vec may not have a corresponding likelihood computed accurately (or, indeed, at all)
- Added new function tweedie.plot() to make it easier to plot Tweedie densities
tweedie 2006-09-05
Changes:
- Built for R version 2.0.0
- Major fix: The mle estimate of phi was incorrect in tweedie.profile() (thanks Sarah Lennox, QDPI&F) * Some better warnings in tweedie.profile()
- tweedie.profile() now allows the actual computed points to be added to the plot using the do.points option (default is TRUE)
- The tweedie.profile() function now smooth and computes confidence intervals even when the computed likelihood values contain Infs or NAs.
tweedie 2004-09-30
Changes:
- Built for R version 1.9
- Fixed a bug in rtweedie(): it failed for p = 2 (thanks Sarah Lennox, QDPI&F)
- Fixed a bug in qtweedie(): it failed in some cases (thanks Gordon Smyth, WEHI)
- Added a proper .First.lib function (!) (thanks to James Wettenhall and Gordon Smyth, WEHI)
- Other minor bugs fixed
- The tweedie() family function removed, and package made dependent on the statmod package
- Updated for R version 1.9
tweedie 3.3
Summary
Changed and fixed
- Fixed some minor typos in the vignette.
- Corrections to README.
- tweedie_profile() has a new argument
glm.method, to choose the function that
fits the GLM at each value of p: "glm.fit" (the default, as before),
"glm.fit2" (from the glm2 package, now in Suggests), or a user-supplied
function. For some Tweedie models the usual IRLS algorithm cycles between two
values without converging, giving a jagged profile likelihood; glm.fit2
halves the step whenever the deviance increases, and converges in such cases.
- tweedie_profile() now gives a single warning listing the values of p where
the GLM fit did not converge (rather than one warning from glm.fit() for each).
tweedie 1.6.7 (2009-06-30)
Changes:
- Correct error in rtweedie() when power = 1 and phi \ne 1 (Thanks to Frederic Gosselin [email protected])