-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathCHANGES
158 lines (126 loc) · 7.43 KB
/
CHANGES
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
---------------------------------- Version 1.1 ----------------------------------
This version fixes a few bugs and adds parametric bootstrap functionality for
gpd modelling.
Changes from version 1.0:
1. bootgpd and methods added.
2. gpd was not passing 'start' through to gpd.fit. Fixed.
3. Rationalized summary.gpd and its methods into a single script (no impact).
4. summary.gpd did the wrong simulation when covariates were in the model. Fixed.
5. Bug in qqgpd caused x and y values to be reversed. The y-axis is scaled to
accomodate the simulated envelope and this meant that outliers could be
excluded from the plot. Fixed.
6. print.summary.predict.mex was showing too many decimal places. Fixed
7. validate.texmex failed on Windows due to / in file path (all tests still
passed if executed individually). Fixed.
---------------------------------- Version 1.2 ------------------------------------
Changes from version 1.1:
1. rl.gpd had a bug in the way it computed standard errors. Fixed.
2. Original Heffernan-Tawn methodology used Gumbel marginal distributions.
Keef-Tawn-Papastathopoulos modified this to Laplace margins and identified
constraints that keep the model physically plausible. The modified and
constrained approach is implemented and is now the default.
3. Diagnostic plot of constrained Profile Log-likelihood surface added, to
help users check for convergence to mle on constrained parameter space.
4. New diagnostic plots for GPD models when there are covariates, as
part of the plot method for a gpd object.
5. A returned gpd object now has residuals attached.
---------------------------------- Version 1.3 ------------------------------------
Changes from version 1.2:
1. predict method added for objects of class gpd, bgpd, bootgpd. This ought to
make computation of return levels easier and more reliable. See the help
files.
2. Bug fix. The 'Metropolis algorithm' as it was implemented was not actually
a Metropolis algorithm due to a bug. The proposal distribution never changed
from iteration to iteration. (Checks on the effects of the bug suggest it
didn't produce distributions that were very different from those produced by
an actual Metropolis algorithm.)
3. Cauchy distribution added as a proposal distribution. Defaults to Gaussian.
4. Speed improvement due to simulation of random numbers (and thus lots of
matrix decomposition operations) being taken outside the simulation loop.
The speed gain is largely offset by the need to run the chains for longer
due to the bug fix described in item 2. Nevertheless, Dr. Metcalfe is a very
clever boy.
5. Vignettes added. One for modelling clinical trial lab safety data; one for
declustering.
6. NAMESPACE added.
7. Fixed bug in using linear predictor formulas when not passing a data.frame in.
---------------------------------- Version 2.1 ------------------------------------
Changes from version 1.2:
1. Main modelling function switched from gpd to evm.
2. Added Collate to DESCRIPTION.
3. Covariance of predictions for GPD model does now NOT take account of the
uncertainty in the rate, so estimated variances will be a little lower
than in previous versions.
4. GEV family added.
5. Default fitting procedure is now MLE rather than weakly penalized MLE.
6. Vignettes updated, switched from Sweave to knitr, multivariate vignette
added.
7. CITATION file added.
8. All test functions and validate.texmex removed. These will reappear in a
companion package.
9. Nasty bug in predict functions fixed - xlevels added to output object from
evm function.
10. Code refactoring.
11. NAMESPACE now keeps many internal functions internal.
12. An evm object now contains information in the true log-likelihood and the
penalized log-likelihood. AIC uses the true log-likelihood by default.
---------------------------------- Version 2.3 -----------------------------------
Changes from version 2.1
1. egp3 family added together with vignette. Additional threshold selection plot
based on egp3 added.
2. Test suite added back in, together with vignette.
3. Bug in u2gpd that was called from revTransform is now fixed. This affected
plot.mex bootmex and a couple of other things.
4. Bootstrap of evm models is not parallelized.
5. AstraZeneca's support for the development of texmex is now credited in various
places.
6. Available vignettes should now be listed from the package Index page.
7. Some instances of 'cat' and 'warning' have been changed to 'message' to allow
for easier suppression and less annoyance.
8. Added Huber sandwich estimator of parameter estimate covariance matrix for
dependent data, for gpd model only at this stage.
9. Added functionality to carry out estimation of Joint Exceedance Curves, from
fitted Heffernan and Tawn model, documented in multivariate vignette.
10.Plotting now available using ggplot2 methods.
---------------------------------- Version 2.4 -----------------------------------
Changes from version 2.3
1. weibull family added.
---------------------------------- Version 2.4.2 -----------------------------------
Changes from version 2.4
1. gcpd family added (constrained parameter space for xi) as suggested by Yee &
Stephenson (2007).
2. Profile likelihood based confidence interval estimation added for GPD models fit by
evmOpt, only for models containing no covariates.
3. Added option for user to supply reference marginal distributions for marginal
transformation required in conditional dependence model estimation. This is for use
in instances where the marginal distributions are not well represented in the data
prividing joint observation (for instance where data represent only observaitons with
a given variable above a threshold).
3. Bug in egp3RangeFit fixed (thanks to Emmanuel Flachire)
4. Bug fix: axis label in gpd range fit diagnostic plot.
5. Bug fix: plotting points in diagnostic plot for checkin independence of model residuals
from conditioning varibale in mex fitting.
6. Various improvements to documentation.
---------------------------------- Version 2.4.4 -----------------------------------
Changes from version 2.4.2
1. Bug fixes in gpdFitRange: calculation of log-scale parameter and also standard error
2. gpdIntCensored family added
3. glo (generalised logistic) family added
4. Generic cv function and cv.evmOpt added for finding shape parameter by cross-validation.
5. WAIC added
---------------------------------- Version 2.4.5 -----------------------------------
Changes from version 2.4.4
1. New 'chains' argument added to evmSim. Can now run multiple Markov chains in
parallel.
---------------------------------- Version 2.4.8 ------------------------------------
Changes from version 2.4.5 in 2.4.6 & 2.4.7 were around changes made to R.
Changes to Version 2.4.8 fix a longstanding bug in gpdRangeFit: the parameter \sigma*
is constrained to be > -u\xi, not > 0. Therefore, taking logs doesn't work. Switching
to working with \sigma*, updating plot functions. The function is now basically the
one written by Coles.
--------------------------------- Version 2.4.9 ------------------------------------
With covariates in the model ggplot.evmOpt jittered the data vertically by default,
but also used a fixed vertical range. Therefore, if the maximum residual was jittered
upwards, it was removed from the plot. Added jitter.height and removed fixed range.
Technical changes to avoid the package breaking under an upcoming change to R,
and to pass newer version of R CMD check.