Overview of the fitdistrplus package
fitdistrplus.RdThe idea of this package emerged in 2008 from a collaboration between JB Denis, R Pouillot and ML Delignette who at this time worked in the area of quantitative risk assessment. The implementation of this package was a part of a more general project named "Risk assessment with R" gathering different packages and hosted in R-forge.
The fitdistrplus package was first written by ML Delignette-Muller and made available in CRAN on 2009 and presented at the 2009 useR conference in Rennes. A few months after, C Dutang joined the project by starting to participate to the implementation of the fitdistrplus package. The package has also been presented at the 2011 useR conference and at the 2eme rencontres R in 2013 (https://r2013-lyon.sciencesconf.org/).
Four vignettes are available within the package:
a general overview of the package published in the Journal of Statistical Software (doi:10.18637/jss.v064.i04 ),
a document answering the most Frequently Asked Questions,
a document presenting a benchmark of optimization algorithms when finding parameters,
a document about starting values.
The fitdistrplus package is a general package that aims at helping the fit of univariate parametric
distributions to censored or
non-censored data. The two main functions are
fitdist for fit on non-censored data and
fitdistcens for fit on censored data.
The choice of candidate
distributions to fit may be helped using functions descdist and
plotdist for non-censored data and plotdistcens
for censored data).
Using functions fitdist and
fitdistcens, different methods can be used to estimate the
distribution parameters:
maximum likelihood estimation by default (
mledist),moment matching estimation (
mmedist),quantile matching estimation (
qmedist),maximum goodness-of-fit estimation (
mgedist).
For classical distributions initial values are automatically calculated
if not provided by the user.
Graphical functions plotdist and plotdistcens
can be used to help a manual calibration of initial values for parameters
of non-classical distributions. Function prefit is proposed
to help the definition of good starting values in the special case of
constrained parameters. In the case where maximum likelihood is chosen
as the estimation method, function llplot enables to
visualize loglikelihood contours.
The goodness-of-fit of fitted distributions (a single fit or multiple fits) can be explored
using different graphical functions (cdfcomp, denscomp,
qqcomp and ppcomp for non-censored data and
cdfcompcens for censored data). Goodness-of-fit statistics are also
provided for non-censored data using function gofstat.
Bootstrap is proposed to quantify the uncertainty on parameter estimates
(functions bootdist and bootdistcens)
and also to quantify the uncertainty on CDF or quantiles estimated
from the fitted distribution (quantile and CIcdfplot).