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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721
|
\documentclass[11pt,a4paper]{article}
\usepackage{t1enc}
\usepackage[latin1]{inputenc}
\usepackage[english]{babel}
\usepackage{amsmath,amssymb}
\usepackage{graphics}
\usepackage[round]{natbib}
\bibliographystyle{jrss}
\pagestyle{plain}
\setlength{\parindent}{0in}
\setlength{\parskip}{1.5ex plus 0.5ex minus 0.5ex}
\setlength{\oddsidemargin}{0in}
\setlength{\evensidemargin}{0in}
\setlength{\topmargin}{-0.5in}
\setlength{\textwidth}{6.3in}
\setlength{\textheight}{9.8in}
\renewcommand{\thefootnote}{\fnsymbol{footnote}}
\begin{document}
\sloppy
\begin{center}
\LARGE
A User's Guide to the evd Package (Version 2.2) \\
\Large
\vspace{0.2cm}
Alec Stephenson \\
\normalsize
Copyright \copyright 2006 \\
\vspace{0.2cm}
Department of Statistics and Applied Probability,\\
National University of Singapore, \\
Singapore 117546. \\
\vspace{0.2cm}
E-mail: alec\_stephenson@hotmail.com \\
25th March 2006
\end{center}
%\nocite{frantiag84}
%\nocite{gumbgold64}
%\nocite{gumbmust67}
%\nocite{jenk55}
%\nocite{joe97}
%\nocite{kotzbala00}
%\nocite{mont70}
%\nocite{montotte78}
%\nocite{smit86}
%\nocite{sney77}
%\nocite{step:sim}
%\nocite{tawn93}
\section{Introduction}
\setcounter{footnote}{0}
\subsection{What is the evd package?}
\label{intro}
The evd (extreme value distributions) package is an add-on package for the R \citep{R} statistical computing system.
The package contains the following (user-level) functions.
It also contains the demo \verb+soe9+, giving examples from Chapter Nine of \citet{beirgoeg04}.
Univariate Distributions. Density, distribution, simulation and quantile (inverse distribution) functions for univariate parametric distributions.\\
\verb+ dgev dgpd dgumbel drweibull dfrechet dextreme dorder+\\
\verb+ pgev pgpd pgumbel prweibull pfrechet pextreme porder+\\
\verb+ rgev rgpd rgumbel rrweibull rfrechet rextreme rorder+\\
\verb+ qgev qgpd qgumbel qrweibull qfrechet qextreme+
Multivariate Distributions. Density, distribution, simulation and dependence functions for multivariate parametric extreme value models.\\
\verb+ dbvevd dmvevd pbvevd pmvevd rbvevd rmvevd abvevd amvevd+
Non-parametric Estimation. Calculate and plot non-parametric estimates of dependence functions and quantile curves.\\
\verb+ abvnonpar amvnonpar qcbvnonpar+
Stochastic Processes. Generate stochastic processes associated with extreme value theory, identify extreme clusters and estimate the extremal index.\\
\verb+ evmc marma mar mma clusters exi+
Fitting Models. Obtain maximum likelihood estimates and standard errors for univariate and bivariate models used in extreme value theory.\\
\verb+ fbvevd fgev fpot forder fextreme+
Pre-model Diagnostics. Threshold identification and dependence summaries.\\
\verb+ mrlplot tcplot chiplot+
Model Diagnostics. Model diagnostics for fitted models; diagnostic plots and analysis of deviance.\\
\verb+ plot.uvevd plot.bvevd anova.evd+
Profile likelihoods. Obtain profile traces, plot profile log-likelihoods and obtain profile confidence intervals from fitted models.\\
\verb+ profile.evd plot.profile.evd profile2d.evd plot.profile2d.evd+
The following datasets are also included in the package.\\
\verb+ failure fox lisbon ocmulgee oldage oxford lossalae+\\
\verb+ portpirie sask sealevel uccle venice sealevel2+
\subsection{Obtaining the package/guide}
The evd package can be downloaded from CRAN (The Comprehensive R
Archive Network) at \verb+http://cran.r-project.org/+.
This guide (in pdf) will be in the directory \verb+evd/doc/+
underneath wherever the package is installed.
\subsection{Contents}
This guide contains examples\footnote{All of the examples presented in
this guide are called with \texttt{options(digits = 4)}, and with the
option \texttt{show.signif.stars} set to \texttt{FALSE}.} on the use
of the evd package.
The examples do not include any theoretical justification.
See \citet{cole01} for an introduction to the statistics of extreme
values, and \citet{beirgoeg04} for a more detailed treatment.
Section \ref{uni} covers the standard (non-fitting) functions for
univariate distributions.
Sections \ref{biv} and \ref{mult} do the same for bivariate and
multivariate extreme value models.
Dependence functions of extreme value distributions are discussed in
Section \ref{depfun}.
Stochastic processes are discussed in Section \ref{stochproc}.
Maximum likelihood fitting of univariate models, peaks over threshold
models and bivariate extreme value models is discussed in Sections
\ref{unifit}, \ref{potfit} and \ref{bivfit} respectively.
Three practical examples using the data sets \verb+oxford+,
\verb+rain+ and \verb+sealevel+ are given in Sections \ref{egoxford},
\ref{egrain} and \ref{egsealevel} respectively.
This guide should not be viewed as an alternative to the documentation
files included within the package.
These remain the definitive source of information.
A reference manual containing all the documentation files can be
downloaded from CRAN.
\subsection{Citing the package}
Volume 2/2 of R-News (the newsletter of the R-project) contains an
article that describes an earlier version of the evd package.
To cite the package in publications please cite the R-News article.
The article and the corresponding citation can be downloaded from
\verb+http://www.cran.r-project.org/doc/Rnews/+.
\subsection{Caveat}
I have checked these functions as best I can but, as ever, they may
contain bugs.
If you find a bug or suspected bug in the code or the documentation
please report it to me at \verb+alec_stephenson@hotmail.com+.
Please include an appropriate subject line.
\subsection{Legalese}
This program is free software; you can redistribute it and/or
modify it under the terms of the GNU General Public License
as published by the Free Software Foundation; either version 2
of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful,
but without any warranty; without even the implied warranty of
merchantability or fitness for a particular purpose.
See the GNU General Public License for more details.
A copy of the GNU General Public License can be obtained from
\verb+http://www.gnu.org/copyleft/gpl.html+.
You can also obtain it by writing to the Free Software Foundation,
Inc., 59 Temple Place -- Suite 330, Boston, MA 02111-1307, USA.
\section{Univariate Distributions}
\setcounter{footnote}{0}
\label{uni}
The Gumbel, Fr\'{e}chet and (reversed) Weibull distribution functions are respectively given by
\begin{align}
&G(z) = \exp\left\{-\exp\left[-\left(\frac{z-a}{b}\right)\right]\right\},
\quad -\infty < z < \infty \label{gumbel} \\
&G(z) = \begin{cases}
0, & z \leq a, \\
\exp\left\{-\left(\frac{z-a}{b}\right)^{-\alpha}\right\}, & z > a,
\end{cases} \label{frechet} \\
&G(z) = \begin{cases}
\exp\left\{-\left[-\left(\frac{z-a}{b}\right)\right]^{\alpha}\right\},
& z < a, \\
1, & z \geq a,
\end{cases} \label{weibull}
\end{align}
where $a$ is a location parameter, $b > 0$ is a scale parameter and $\alpha > 0$ is a shape parameter.
The distribution \eqref{weibull} is often referred to as the Weibull distribution.
To avoid confusion I will call this the reversed Weibull, since it is related by a change of sign to the three parameter Weibull distribution used in survival analysis.
The Generalised Extreme Value (GEV) distribution function is given by
\begin{equation}
G(z) = \exp \left\{ - \left[ 1+ \xi \left( z-\mu \right) /\sigma \right]_{+}^{-1/\xi} \right\},
\label{gev}
\end{equation}
where ($\mu,\sigma,\xi$) are the location, scale and shape parameters respectively, $\sigma > 0$ and $h_{+}=\max(h,0)$.
When $\xi>0$ the GEV distribution has a finite lower end point, given by $\mu - \sigma/\xi$.
When $\xi<0$ the GEV distribution has a finite upper end point, also given by $\mu - \sigma/\xi$.
The parametric form of the GEV encompasses that of the Gumbel, Fr\`{e}chet and reversed Weibull distributions.
The Gumbel distribution is obtained in the limit as $\xi\rightarrow0$.
The Fr\'{e}chet and Weibull distributions are obtained when $\xi>0$ and $\xi<0$ respectively.
To recover the parameterisation of the Fr\'{e}chet distribution \eqref{frechet} set $\xi=1/\alpha>0$, $\sigma=b/\alpha>0$ and $\mu=a+b$.
To recover the parameterisation of the reversed Weibull distribution \eqref{weibull} set $\xi=-1/\alpha<0$, $\sigma=b/\alpha>0$ and $\mu=a-b$.
The generalised Pareto distribution (GPD) function is given by
\begin{equation*}
G(z) = 1 - \left[1 + \xi \left( z-\mu \right) /\sigma \right]_{+}^{-1/\xi},
\end{equation*}
for $z > \mu$, where ($\mu,\sigma,\xi$) are the location, scale and shape parameters respectively, $\sigma > 0$ and $h_{+}=\max(h,0)$.
The GPD has a finite lower end point, given by $\mu$.
When $\xi<0$ the GPD also has a finite upper end point, given by $\mu - \sigma/\xi$.
A shifted exponential distribution is obtained in the limit as $\xi\rightarrow0$.
It is standard practice within R to concatenate the letters r, p, q and d with an abbreviated distribution name to yield the names of the corresponding simulation, distribution, quantile (inverse distribution) and density functions respectively.
The evd package follows this convention.
Each of the five distributions defined above has an associated set of functions, as given in Section \ref{intro}. Some examples are given below.
They should be familiar to those who have had previous experience with R.
\begin{verbatim}
> rgev(6, loc = c(20,1), scale = .5, shape = 1)
[1] 23.7290 1.2492 19.6680 0.8662 19.7939 2.6512
> rgpd(3, loc = 2)
[1] 2.483681 3.666805 2.837809
> qrweibull(seq(0.1, 0.4, 0.1), 2, 0.5, 1, lower.tail = FALSE)
> qrweibull(seq(0.9, 0.6, -0.1), loc = 2, scale = 0.5, shape = 1)
# Both give
[1] 1.947 1.888 1.822 1.745
> pfrechet(2:6, 2, 0.5, 1)
[1] 0.0000 0.6065 0.7788 0.8465 0.8825
> pfrechet(2:6, 2, 0.5, 1, low = FALSE)
[1] 1.0000 0.3935 0.2212 0.1535 0.1175
> drweibull(-1:3, 2, 0.5, log = TRUE)
[1] -5.307 -3.307 -1.307 -Inf -Inf
> dgumbel(-1:3, 0, 1)
[1] 0.17937 0.36788 0.25465 0.11820 0.04737
\end{verbatim}
Let $F$ be an arbitrary distribution function, and let $X_1,\dots,X_m$ be a random sample from $F$.
Define $U_m=\max\{X_1,\dots,X_m\}$ and $L_m=\min\{X_1,\dots,X_m\}$.
The distributions of $U_m$ and $L_m$ are given by
\begin{align}
&\Pr(U_m \leq x) = [F(x)]^m
\label{maxdens} \\
&\Pr(L_m \leq x) = 1 - [1 - F(x)]^m.
\label{mindens}
\end{align}
Simulation, distribution, quantile and density functions for the distribution of $U_m$, given an integer $m$ and an arbitrary distribution function $F$, are provided by \verb+rextreme+, \verb+pextreme+, \verb+qextreme+ and \verb+dextreme+ respectively.
The integer $m$ should be given to the argument \verb+mlen+.
The distribution $F$ is most easily specified by passing an
abbreviated distribution name to the argument \verb+distn+.
If \verb+largest+ is \verb+FALSE+ the distribution of $L_m$ is used.
Some examples are given below.
\begin{verbatim}
> rextreme(1, distn = "norm", sd = 2, mlen = 20, largest = FALSE)
> min(rnorm(20, mean = 0, sd = 2))
# Both simulate from the same distribution
[1] -2.612
> rextreme(4, distn = "exp", rate = 1, mlen = 5)
> rextreme(4, distn = "exp", mlen = 5)
# Both simulate from the same distribution
[1] 2.2001 0.8584 4.5595 3.9397
> pextreme(c(.4, .5), distn = "norm", mean = 0.5, sd = c(1, 2), mlen = 4)
[1] 0.04484 0.06250
> dextreme(c(1, 4), distn = "gamma", shape = 1, scale = 0.3, mlen = 100)
[1] 0.3261328 0.0005398
\end{verbatim}
Let $X_{(1)} \geq X_{(2)} \geq \dots \geq X_{(m)}$ be the order statistics of the random sample $X_1,\dots,X_m$.
The distribution of the $j$th largest order statistic, for $j = 1,\dots,m$, is
\begin{equation}
\Pr(X_{(j)} \leq x) = \sum_{k=0}^{j-1} \binom{m}{k} [F(x)]^{m-k} [1 - F(x)]^k.
\label{orderdens}
\end{equation}
The distribution of the $j$th smallest order statistic is obtained by setting $j = m + 1 - j$.
Simulation, distribution and density functions for the distribution of $X_{(j)}$, for given integers $m$ and $j \in \{1,\dots,m\}$, and for an arbitrary distribution function $F$, are provided by \verb+rorder+, \verb+porder+ and \verb+dorder+ respectively.
The integer $m$ should again be given to the argument \verb+mlen+.
If \verb+largest+ is \verb+FALSE+ the distribution of the \verb+j+th smallest order statistic $X_{(m+j-1)}$ is used.
Some examples are given below.
\begin{verbatim}
> rorder(1, distn = "norm", mlen = 20, j = 2)
[1] 2.284
> porder(c(1, 2), distn = "gamma", shape = c(.5, .7), mlen = 10, j = 2)
[1] 0.5177 0.8259
> dorder(c(1, 2), distn = "gamma", shape = c(.5, .7), mlen = 10, j = 2)
[1] 0.7473 0.3081
\end{verbatim}
\section{Bivariate Extreme Value Distributions}
\setcounter{footnote}{0}
\label{biv}
The evd package contains functions associated with nine parametric bivariate extreme value distributions.
The univariate marginal distributions in each case are GEV, with marginal parameters ($\mu_1,\sigma_1,\xi_1$) and ($\mu_2,\sigma_2,\xi_2$).
There are three symmetric models, with distribution functions
\begin{align}
&G(z_1,z_2) = \exp\left\{- (y_1^{1/\alpha}+y_2^{1/\alpha})^\alpha \right\}, \quad 0<\alpha\leq1, \label{log} \\
&G(z_1,z_2) = \exp\left\{ - y_1 - y_2 + (y_1^{-r}+y_2^{-r})^{-1/r} \right\}, \quad r>0, \label{neglog} \\
&G(z_1,z_2) = \exp\left( - y_1\Phi\{\lambda^{-1}+{\textstyle\frac{1}{2}}\lambda[\log(y_1/y_2)]\} - y_2\Phi\{\lambda^{-1}+{\textstyle\frac{1}{2}}\lambda[\log(y_2/y_1)]\}\right), \quad \lambda>0, \notag
\end{align}
known as the logistic \citep{gumb60b}, negative logistic \citep{gala75} and H\"{u}sler-Reiss \citep{huslreis89} models respectively, where
\begin{equation}
y_j = y_j(z_j) = \{1+\xi_j(z_j-\mu_j)/\sigma_j\}_{+}^{-1/\xi_j}
\label{mtrans}
\end{equation}
for $j=1,2$.
Independence\footnote{
Independence occurs when $G(z_1,z_2) = \exp\{-(y_1+y_2)\}$.}
is obtained when $\alpha=1$, $r\downarrow0$ or $\lambda\downarrow0$.
Complete dependence\footnote{
Complete dependence occurs when $G(z_1,z_2) = \exp\{-\max(y_1,y_2)\}$.}
is obtained when $\alpha\downarrow0$, $r\rightarrow\infty$ or $\lambda\rightarrow\infty$.
The distributions functions \eqref{log} and \eqref{neglog} have asymmetric extensions, given by
\begin{align}
&G(z_1,z_2) = \exp\left\{ - (1-\theta_1)y_1 - (1-\theta_2)y_2 - [(\theta_1y_1)^{1/\alpha}+(\theta_2y_2)^{1/\alpha}]^\alpha\right\}, \quad 0<\alpha\leq1, \notag \\
&G(z_1,z_2) = \exp\left\{ - y_1 - y_2 + [(\theta_1y_1)^{-r}+(\theta_2y_2)^{-r}]^{-1/r}\right\}, \quad r>0, \notag
\end{align}
known as the asymmetric logistic \citep{tawn88} and asymmetric negative logistic \citep{joe90} models respectively, where the asymmetry parameters $0\leq\theta_1,\theta_2\leq1$.
For the asymmetric logistic model independence is obtained when either $\alpha = 1$, $\theta_1 = 0$ or $\theta_2 = 0$.
Different limits occur when $\theta_1$ and $\theta_2$ are fixed and $\alpha\downarrow0$.
For the asymmetric negative logistic model independence is obtained when either $r\downarrow0$, $\theta_1\downarrow0$ or $\theta_2\downarrow0$.
Different limits occur when $\theta_1$ and $\theta_2$ are fixed and $r\rightarrow\infty$.
The remaining four bivariate models are defined in Appendix A. Density, distribution and simulation functions for each of the nine models are provided by \verb+dbvevd+, \verb+pbvevd+ and \verb+rbvevd+ respectively.
The argument \verb+model+ denotes the specified model, which must be either \verb+"log"+ (the default), \verb+"alog"+, \verb+"hr"+, \verb+"neglog"+, \verb+"aneglog"+, \verb+"bilog"+, \verb+"negbilog"+, \verb+"ct"+ or \verb+"amix"+ (or any unique partial match).
The first argument in \verb+pbvevd+ and \verb+dbvevd+ should be a vector of length two or a matrix with two columns, so that each row specifies a value for $(z_1,z_2)$.
The parameters of the specified model can be passed using one or more of the arguments \verb+dep+, \verb+asy+, \verb+alpha+ and \verb+beta+.
The marginal parameters ($\mu_1,\sigma_1,\xi_1$) and ($\mu_2,\sigma_2,\xi_2$) can be passed using the arguments \verb+mar1+ and \verb+mar2+ respectively.
Gumbel marginal distributions are used by default.
The arguments \verb+mar1+ and \verb+mar2+ can also be matrices with three columns, in which case each column represents a vector of values to be passed to the corresponding marginal parameter.
Some examples are given below.
\begin{verbatim}
> rbvevd(3, dep = .8, asy = c(.4, 1), model = "alog")
[,1] [,2]
[1,] 0.07876 -0.7971
[2,] 0.01091 -0.8113
[3,] -0.10491 -0.8831
> rbvevd(3, alpha = .5, beta = 1.2, model = "negb", mar1 = rep(1, 3))
[,1] [,2]
[1,] 0.7417 1.085
[2,] 0.8391 1.825
[3,] 2.0142 2.280
> pbvevd(c(1, 1.2), dep = .4, asy = c(.4, .6), model = "an", mar1 = rep(1, 3))
[1] 0.173
> tmp.quant <- matrix(c(1,1.2,1,2), ncol = 2, byrow = TRUE)
> tmp.mar <- matrix(c(1,1,1,1.2,1.2,1.2), ncol = 3, byrow = TRUE)
> pbvevd(tmp.quant, dep = .4, asy = c(.4, .6), model = "an", mar1 = tmp.mar)
[1] 0.173 0.175
> dbvevd(c(1, 1.2), alpha = .2, beta = .6, model = "ct", mar1 = rep(1, 3))
[1] 0.1213
> dbvevd(tmp.quant, alpha = 0.2, beta = 0.6, model = "ct", mar1 = tmp.mar)
[1] 0.1213 0.0586
\end{verbatim}
%The logistic and asymmetric logistic models respectively are simulated using bivariate versions of Algorithms 1.1 and 1.2 in \citet{step02a}.
%All other models are simulated using a root finding algorithm to generate random vectors from the conditional distribution function.
%The simulation of the the bilogistic or negative bilogistic model is relatively slow (about 2.8 seconds per 1000 random vectors on a 450MHz PIII, 512Mb RAM) because each evaluation of either distribution function requires a root finding algorithm to evaluate $\gamma$.
\section{Multivariate Extreme Value Distributions}
\setcounter{footnote}{0}
\label{mult}
Let $z=(z_1,\dots,z_d)$.
The $d$-dimensional logistic model \citep{gumb60b} has distribution function
\begin{equation*}
G(z) = \exp\left\{-\left(\sum\nolimits_{j=1}^d y_j^{1/\alpha}\right)^\alpha\right\}
\end{equation*}
where $\alpha\in(0,1]$ and $(y_1,\dots,y_d)$ is defined by the transformations \eqref{mtrans}.
This distribution can be extended to an asymmetric form.
Let $B$ be the set of all non-empty subsets of $\{1,\dots,d\}$, let $B_1=\{b \in B:|b|=1\}$, where $|b|$ denotes the number of elements in the set $b$, and let $B_{(i)}=\{b \in B:i \in b\}$.
The multivariate asymmetric logistic model \citep{tawn90} is given by
\begin{equation*}
G(z)=\exp\left\{-\sum\nolimits_{b \in B} \left[\sum\nolimits_{i \in b}(\theta_{i,b}y_i)^{1/\alpha_b}\right]^{\alpha_b}\right\}
\end{equation*}
where the dependence parameters $\alpha_b\in(0,1]$ for all $b\in B \setminus B_1$, and the asymmetry parameters $\theta_{i,b}\in[0,1]$ for all $b\in B$ and $i\in b$.
The constraints $\sum_{b \in B_{(i)}}\theta_{i,b}=1$ for $i=1,\dots,d$ ensure that the marginal distributions are GEV.
There exists further constraints which arise from the possible redundancy of asymmetry parameters in the expansion of the distributional form.
Specifically, if $\alpha_b=1$ for some $b\in B \setminus B_1$ then $\theta_{i,b}=0$ for all $i \in b$.
Let $b_{-i_0}=\{i \in b:i \neq i_0\}$.
If, for some $b \in B \setminus B_1$, $\theta_{i,b}=0$ for all $i \in b_{-i_0}$, then $\theta_{i_0,b}=0$.
%The model contains $2^d-d-1$ dependence parameters and $d2^{d-1}$ asymmetry parameters (excluding the constraints).
%The logistic model \eqref{multlog} can be obtained by setting $\theta_{i,12 \dots d}=1$ for all $i = 1,\dots,d$ (which implies that $\theta_{i,b}=0$ whenever $|b|<d$) and $\alpha_{12 \dots d} = \alpha$.
%The density functions for the symmetric and asymmetric logistic models are given in \citet{step:phd}.
Density, distribution and simulation functions for these models are provided by \verb+dmvevd+, \verb+pmvevd+ and \verb+rmvevd+ respectively.
The argument \verb+model+ denotes the specified model, which must be either \verb+"log"+ (the default) or \verb+"alog"+ (or any unique partial match).
The argument \verb+d+ denotes the dimension of the model.
By default, \verb+d = 2+.
The first argument in \verb+pbvevd+ and \verb+dbvevd+ should be a vector of length \verb+d+ or a matrix with \verb+d+ columns, so that each row specifies a value for $(z_1,\dots,z_d)$.
The marginal parameters $(\mu_i,\sigma_i,\xi_i)$, for $i=1,\dots,d$, can be passed using the argument \verb+mar+.
Gumbel marginal distributions are used by default.
For the symmetric logistic model, the argument \verb+dep+ represents the parameter $\alpha$.
Some examples are given below.
%The simulation functions \verb+rmvlog+ and \verb+rmvalog+ use Algorithms 2.1 and 2.2 in \citet{step02a}.
\begin{verbatim}
> rmvevd(3, dep = .6, model = "log", d = 5)
[,1] [,2] [,3] [,4] [,5]
[1,] 0.1335 0.2878 1.07886 1.55515 1.310
[2,] 1.7100 0.9453 1.02070 -0.02553 1.527
[3,] -0.3376 -0.5814 0.07426 0.10906 2.827
> tmp.mar <- matrix(c(1,1,1,1,1,1.5,1,1,2), ncol = 3, byrow = TRUE)
> rmvevd(3, dep = .6, d = 5, mar = tmp.mar)
[,1] [,2] [,3] [,4] [,5]
[1,] 2.803 4.6415 1.8531 3.5569 8.854
[2,] 0.751 0.9704 2.3328 2.6537 1.233
[3,] 4.641 1.4321 0.5825 0.6041 2.021
> tmp.quant <- matrix(rep(c(1,1.5,2), 5), ncol = 5)
> pmvevd(tmp.quant, dep = .6, d = 5, mar = tmp.mar)
[1] 0.07233 0.16387 0.21949
> dmvevd(tmp.quant, dep = .6, d = 5, mar = tmp.mar, log = TRUE)
[1] -3.564 -6.610 -9.460
\end{verbatim}
For the asymmetric logistic model \verb+dep+ should be a vector of length $2^{\verb+d+}-\verb+d+-1$ containing the dependence parameters.
For example, when $\verb+d+ = 4$
\begin{equation*}
\verb+dep+ = \texttt{c}(\alpha_{12},\alpha_{13},\alpha_{14},\alpha_{23},\alpha_{24},\alpha_{34},\alpha_{123},\alpha_{124},\alpha_{134},\alpha_{234},\alpha_{1234}).
\end{equation*}
The asymmetry parameters should be passed to \verb+asy+ in a list with $2^{\verb+d+}-1$ elements, where each element is a vector (including vectors of length one) corresponding to a set $b \in B$, containing $\{\theta_{i,b}:i \in b\}$.
For example, when $\verb+d+ = 4$
\begin{align*}
\texttt{asy} = \texttt{list}&(\theta_{1,1}, \theta_{2,2}, \theta_{3,3}, \theta_{4,4}, \texttt{c}(\theta_{1,12},\theta_{2,12}), \texttt{c}(\theta_{1,13},\theta_{3,13}), \texttt{c}(\theta_{1,14},\theta_{4,14}), \texttt{c}(\theta_{2,23},\theta_{3,23}), \\
&\texttt{c}(\theta_{2,24},\theta_{4,24}), \texttt{c}(\theta_{3,34},\theta_{4,34}), \texttt{c}(\theta_{1,123},\theta_{2,123},\theta_{3,123}), \texttt{c}(\theta_{1,124},\theta_{2,124},\theta_{4,124}), \\
&\texttt{c}(\theta_{1,134},\theta_{3,134},\theta_{4,134}), \texttt{c}(\theta_{2,234},\theta_{3,234},\theta_{4,234}), \texttt{c}(\theta_{1,1234},\theta_{2,1234},\theta_{3,1234},\theta_{4,1234})).
\end{align*}
All the constraints, including $\sum_{b \in B_{(i)}}\theta_{i,b}=1$ for $i=1,\dots,d$, must be satisfied or an error will occur.
Some examples are given below.
The dependence parameters used in the following trivariate asymmetric logistic model are $(\alpha_{12},\alpha_{13},\alpha_{23},\alpha_{123})=(.6,.5,.8,.3)$.
The asymmetry parameters are $\theta_{1,1}=.4$, $\theta_{2,2}=0$, $\theta_{3,3}=.6$, $(\theta_{1,12},\theta_{2,12})=(.3,.2)$, $(\theta_{1,13},\theta_{3,13})=(.1,.1)$, $(\theta_{2,23},\theta_{3,23})=(.4,.1)$ and finally $(\theta_{1,123},\theta_{2,123},\theta_{3,123})=(.2,.4,.2)$.
\begin{verbatim}
> asy <- list(.4, 0, .6, c(.3,.2), c(.1,.1), c(.4,.1), c(.2,.4,.2))
> rmvevd(3, dep = c(.6,.5,.8,.3), asy = asy, model = "alog", d = 3)
[,1] [,2] [,3]
[1,] 0.52375 -0.8844 1.4898
[2,] 1.16174 -0.4368 -0.7404
[3,] -0.03737 1.5139 -0.5996
> dmvevd(c(2, 2, 2), dep = c(.6,.5,.8,.3), asy = asy, model = "a", d = 3)
[1] 0.006636
> tmp.quant <- matrix(rep(c(1,1.5,2), 3), ncol = 3)
> pmvevd(tmp.quant, dep = c(.6,.5,.8,.3), asy = asy, model = "a", d = 3)
[1] 0.4131 0.5849 0.7223
\end{verbatim}
The dependence parameters used in the following four dimensional asymmetric logistic model are $\alpha_b = 1$ for $|b| = 2$\footnote{
The values taken by $\alpha_b$ when $|b| = 2$ are irrelevant here because $\theta_{i,b}=0$ for all $i \in b$ when $|b|=2$.}
and $(\alpha_{123},\alpha_{124},\alpha_{134},\alpha_{234},\alpha_{1234})=(.7,.3,.8,.7,.5)$.
The asymmetry parameters are $\theta_{i,b}=0$ for all $i \in b$ when $|b|\leq2$, $(\theta_{1,123},\theta_{2,123},\theta_{3,123})=(.2,.1,.2)$, $(\theta_{1,124},\theta_{2,124},\theta_{4,124})=(.1,.1,.2)$, $(\theta_{1,134},\theta_{3,134},\theta_{4,134})=(.3,.4,.1)$, $(\theta_{2,234},\theta_{3,234},\theta_{4,234})=(.2,.2,.2)$ and finally $(\theta_{1,1234},\theta_{2,1234},\theta_{3,1234},\theta_{4,1234})=(.4,.6,.2,.5)$.
\begin{verbatim}
> asy <- list(0, 0, 0, 0, c(0,0), c(0,0), c(0,0), c(0,0), c(0,0), c(0,0),
c(.2,.1,.2), c(.1,.1,.2), c(.3,.4,.1), c(.2,.2,.2), c(.4,.6,.2,.5))
> rmvevd(3, dep = c(rep(1,6),.7,.3,.8,.7,.5), asy = asy, model = "alog", d = 4)
[,1] [,2] [,3] [,4]
[1,] -0.5930 -0.1916 1.0211 0.6113
[2,] 4.3522 -1.0050 2.3618 -0.1875
[3,] 0.5805 0.4443 -0.5958 0.9717
\end{verbatim}
%I will end this section with some examples that may be helpful in deciphering errors.
%\begin{verbatim}
%> asy <- list(.4, 0, .5, c(.3,.2), c(.1,.15), c(.4,.075), c(.2,.4,.25))
%> rmvevd(3, dep = c(.6,.5,.8,.3), asy = asy, model = "alog", d = 3)
%Error in mvalog.check(asy, dep, d = d) :
% `asy' does not satisfy the appropriate constraints
%
%# 0.5 + 0.15 + 0.075 + 0.25 does not equal one; the sum constraint on the third
%margin is not satisfied.
%
%> asy <- list(.4, 0, .6, c(.3,.2), c(.1,.1), c(.4,.1), c(.2,.4,.2))
%> rmvevd(3, dep = c(.6,1,.8,.3), asy = asy, model = "alog", d = 3)
%Error in mvalog.check(asy, dep, d = d) :
% `asy' does not satisfy the appropriate constraints
%
%# A dependence parameter is equal to one but the corresponding asymmetry
%parameters are not zero (the first `further constraint').
%# One possible alternative which preserves dep (and still satisfies the sum
%constraints) is
%
%> asy <- list(.4, 0, .6, c(.3,.2), c(0,0), c(.4,.1), c(.3,.4,.3))
%> rmvevd(3, dep = c(.6,1,.8,.3), asy = asy, model = "alog", d = 3)
% [,1] [,2] [,3]
%[1,] 4.627 2.9125 0.9414
%[2,] 1.200 0.1556 0.2048
%[3,] -1.159 -0.8749 -1.0340
%
%> asy <- list(.5, 0, .6, c(.3,.2), c(0,.1), c(.4,.1), c(.2,.4,.2))
%> rmvevd(3, dep = c(.6,.5,.8,.3), asy = asy, model = "alog", d = 3)
%Error in mvalog.check(asy, dep, d = d) :
% `asy' does not satisfy the appropriate constraints
%
%# The fifth element in asy contains exactly one non-zero asymmetry parameter
%(the second `further constraint').
%
%> asy <- list(.4, 0, .6, c(.3,.2), c(.1,.1), c(.4,.1), c(.2,.4,.2))
%> rmvevd(3, dep = c(.6,.5,.8,.3), asy = asy, model = "alog")
%Error in mvalog.check(asy, dep, d = d) :
% `asy' should be a list of length 3
%
%# the dimension has been mis-specified (the default dimension is 2).
%\end{verbatim}
\section{Dependence Functions}
\setcounter{footnote}{0}
\label{depfun}
Let $z=(z_1,\dots,z_d)$ and $\omega=(\omega_1,\dots,\omega_d)$. Any $d$-dimensional extreme value distribution function can be represented in the form
\begin{equation*}
G(z) = \exp\left\{ - \left\{\sum\nolimits_{j=1}^d y_j \right\} A\left(\frac{y_1}{\sum\nolimits_{j=1}^d y_j}, \dots, \frac{y_d}{\sum\nolimits_{j=1}^d y_j} \right)\right\},
\end{equation*}
where $(y_1,\dots,y_d)$ is defined by the transformations \eqref{mtrans}. It follows that $A(\omega)=-\log\{G(y_1^{-1}(\omega_1),\dots,y_d^{-1}(\omega_d))\}$, defined on the simplex $S_d =\{\omega \in \mathbb{R}^d_+: \sum_{j=1}^d \omega_j = 1\}$.
$A(\cdot)$ is known as the dependence function. The dependence function characterises the dependence structure of $G$.
It can be shown that $A(\omega)=1$ when $\omega$ is one of the $d$ vertices of $S_d$ (i.e.\ when one component of $\omega$ is equal to one, and all remaining components are equal to zero), and that $A$ is a convex function with $\max(\omega_1,\dots,\omega_d) \leq A(\omega) \leq 1$ for all $\omega \in S_d$.
The lower and upper bounds are obtained at complete dependence and mutual independence respectively.
In particular, $A(1/d,\dots,1/d)$ is equal to $1/d$ at complete dependence, and $1$ at mutual independence.
The dependence function of a \emph{bivariate} extreme value distribution is a special case (because the sets $S_2$ and [0,1] are equivalent), and is typically defined as follows.
Any bivariate extreme value distribution function can be represented in the form
\begin{equation}
G(z_1,z_2) = \exp\left\{ - (y_1 + y_2)A\left(\frac{y_1}{y_1+y_2}\right)\right\},
\label{bvdepfn}
\end{equation}
so that $A(\omega)=-\log\{G(y_1^{-1}(\omega),y_2^{-1}(1-\omega))\}$, defined on
$0\leq\omega\leq1$.\footnote{Some authors \citep[e.g.][]{pick81} use $A(\omega)=-\log\{G(y_1^{-1}(1-\omega),y_2^{-1}(\omega))\}$.}
It follows that $A(0)=A(1)=1$, and that $A(\cdot)$ is a convex function with $\max(\omega,1-\omega) \leq A(\omega) \leq 1$ for all $0\leq\omega\leq1$.
At independence $A(1/2) = 1$.
At complete dependence $A(1/2) = 0.5$.
Dependence functions for parametric bivariate and trivariate extreme value models can be calculated and plotted, at given parameter values, using the functions \verb+abvevd+ and \verb+amvevd+.
Non-parametric estimators of dependence functions can also be calculated and plotted, using the functions \verb+abvnonpar+ and \verb+amvnonpar+.
Some examples are given below.
The last lines of code produce Figure \ref{depfns}.
%Non-parametric estimators of dependence functions of bivariate extreme value models are constructed as follows.
%Suppose $(z_{i1},z_{i2})$ for $i=1,\dots,n$ are $n$ bivariate observations that are passed to \verb+abvnonpar+ using the argument \verb+data+.
%The marginal parameters are estimated (under the assumption of independence) and the data is transformed using
%\begin{align}
%y_{i1} &= \{1+\hat{\xi}_1(z_{i1}-\hat{\mu}_1)/\hat{\sigma}_1\}_{+}^{-1/\hat{\xi}_1} \notag \\
%y_{i2} &= \{1+\hat{\xi}_2(z_{i2}-\hat{\mu}_2)/\hat{\sigma}_2\}_{+}^{-1/\hat{\xi}_2}
%\label{transtoexp}
%\end{align}
%for $i=1,\dots,n$, where $(\hat{\mu}_1,\hat{\sigma}_1,\hat{\xi}_1)$ and $(\hat{\mu}_2,\hat{\sigma}_2,\hat{\xi}_2)$ are the maximum likelihood estimates for the location, scale and shape parameters on the first and second margins.
%If non-stationary fitting is implemented using the \verb+nsloc1+ or \verb+nsloc2+ arguments (see Sections \ref{unifit} and \ref{bivfit}) the marginal location parameters may depend on $i$.
%The estimator is specified using the argument \verb+method+. A number of different estimators are implemented. A short simulation study given in Appendix A compares the properties of these estimators. The default estimator is the estimator of \citet{capefoug97}, which is defined (on $0 \leq \omega \leq 1$) by
%which must be either \verb+"pickands"+, \verb+"deheuvels"+, \verb+"cfg"+ (the default), \verb+"tdo"+ or \verb+"hall"+ (or any unique partial match).
%These estimators are respectively defined (on $0 \leq \omega \leq 1$) as follows.
%\begin{equation*}
%\exp\left\{ \{1-p(\omega)\} \int_{0}^{\omega} \frac{H_n(x) - x}{x(1-x)} \, \text{d}x - p(\omega) \int_{\omega}^{1} \frac{H_n(x) - x}{x(1-x)} \, \text{d}x \right\}
%\end{equation*}
%\citet{pick81}
%\begin{equation*}
%A_p(\omega) = n\left\{\sum_{i=1}^n \min\left(\frac{y_{i1}}{\omega},\frac{y_{i2}}{1-\omega}\right)\right\}^{-1}
%\end{equation*}
%\citet{dehe91}
%\begin{equation*}
%A_d(\omega) = n\left\{\sum_{i=1}^n \min\left(\frac{y_{i1}}{\omega},\frac{y_{i2}}{1-\omega}\right) - \omega\sum_{i=1}^n y_{i1} - (1-\omega)\sum_{i=1}^n y_{i2} + n\right\}^{-1}
%\end{equation*}
%\citet{capefoug97}
%\begin{equation*}
%A_c(\omega) = \exp\left\{ \{1-p(\omega)\} \int_{0}^{\omega} \frac{H_n(x) - x}{x(1-x)} \, \text{d}x - p(\omega) \int_{\omega}^{1} \frac{H_n(x) - x}{x(1-x)} \, \text{d}x \right\}
%\end{equation*}
%\citet{tiag97}
%\begin{equation*}
%A_t(\omega) = 1 - \frac{1}{1 + \log n} \sum_{i=1}^n \min\left(\frac{\omega}{1+ny_{i1}},\frac{1-\omega}{1+ny_{i2}}\right)
%\end{equation*}
%\citet{halltajv00}
%\begin{equation*}
%A_h(\omega) = n\left\{\sum_{i=1}^n \min\left(\frac{y_{i1}}{\bar{y}_1 \omega},\frac{y_{i2}}{\bar{y}_2 (1-\omega)}\right)\right\}^{-1}
%\end{equation*}
%where $H_n(x)$ is the empirical distribution function of $x_1,\dots,x_n$, with $x_i = y_{i1} / (y_{i1} + y_{i2})$ for $i=1,\dots,n$, and $p(\cdot)$ is any bounded function on $[0,1]$, which can be specified using the argument \verb+wf+.
%By default $p(\cdot)$ is the identity function.
%In the estimator of \citet{halltajv00}, $\bar{y}_1 = n^{-1}\sum_{i=1}^n y_{i1}$ and $\bar{y}_2 = n^{-1}\sum_{i=1}^n y_{i2}$.
%Let $A_n(\cdot)$ be any estimator of $A(\cdot)$.
%$A_n(\cdot)$ will not necessarily satisfy $\max(\omega,1-\omega) \leq A_n(\omega) \leq 1$ for all $0\leq\omega\leq1$.
%An obvious modification is
%\begin{equation*}
%A_n^{'}(\omega) = \min(1, \max\{A_n(\omega), \omega, 1-\omega\}).
%\end{equation*}
%The function \verb+abvnonpar+ always implements this modification.
%Another estimator $A_n^{''}(\omega)$ can be derived by taking the convex minorant of $A_n^{'}(\omega)$.
%This can be achieved by setting the argument \verb+convex+ to \verb+TRUE+.
%Some examples of the functions described in this section are given below.
%The last eight lines of code produce Figure \ref{depfns}.
\begin{verbatim}
> bvlsm <- rmvevd(100, dep = 0.6, model = "log", d = 2)
> tvlsm <- rmvevd(100, dep = 0.6, model = "log", d = 3)
> abvevd(seq(0,1,0.25), dep = 0.3, asy = c(.7,.9), model = "alog")
[1] 1.0000 0.8272 0.7013 0.7842 1.0000
> abvnonpar(seq(0,1,0.25), data = bvlsm)
[1] 1.0000 0.8634 0.8158 0.8392 1.0000
> abvnonpar(data = bvlsm, plot = TRUE, blty = 1, lty = 2)
> abvevd(dep = .3, asy = c(.5, .9), model = "al", add = TRUE)
> abvevd(dep = 1.05, model = "hr", add = TRUE)
> amvnonpar(data = tvlsm, plot = TRUE, lower = 0.6)
\end{verbatim}
\begin{figure}
\hspace*{2.5cm}
\scalebox{0.25}{\includegraphics{depfns5.ps}}
\hspace{0cm}
\scalebox{0.3}{\includegraphics[0,100][260,580]{depfns6.ps}}
\vspace{-1cm}
\caption{Left: Parametric (solid lines) and non-parametric (dashed line) dependence functions for bivariate distributions. The triangular border represents the constraint $\max(\omega,1-\omega) \leq A(\omega) \leq 1$ for all $\omega \in [0,1]$. Right: non-parametric dependence function for a trivariate distribution. Darker colours depict smaller values, and hence stronger dependence.}
\label{depfns}
\end{figure}
\section{Stochastic Processes}
\setcounter{footnote}{0}
\label{stochproc}
The evd package contains four functions that simulate from stochastic processes associated with extreme value theory.
The functions \verb+marma+, \verb+mar+ and \verb+mma+ generate max autoregressive moving average processes, and the function \verb+evmc+ generates Markov chains with extreme value dependence structures.
The function \verb+clusters+ identifies extreme clusters of a stochastic process, and \verb+exi+ estimates a quantity known as the Extremal Index.
A max autoregressive moving average process $\{X_k\}$, denoted by MARMA($p$, $q$), satisfies
\begin{equation*}
X_k = \max\{\phi_1 X_{k-1}, \dots, \phi_p X_{k-p}, \epsilon_k, \theta_1 \epsilon_{k-1}, \dots, \theta_q \epsilon_{k-q}\}
\end{equation*}
where $(\phi_1, \dots, \phi_p)$ and $(\theta_1, \ldots, \theta_p)$ are vectors of non-negative parameters, and $\{\epsilon_k\}$ is a series of \emph{iid} random variables with a common distribution defined by the argument \verb+rand.gen+. The standard Fr\'{e}chet distribution is used by default.
A max autoregressive process $\{X_k\}$, denoted by MAR($p$), is equivalent to a MARMA($p$, 0) process, so that
\begin{equation*}
X_k = \max\{\phi_1 X_{k-1}, \dots, \phi_p X_{k-p}, \epsilon_k\}.
\end{equation*}
A max moving average process $\{X_k\}$, denoted by MMA($q$), is equivalent to a MARMA(0, $q$) process, so that
\begin{equation*}
X_k = \max\{\epsilon_k, \theta_1 \epsilon_{k-1}, \dots, \theta_q \epsilon_{k-q}\}.
\end{equation*}
The functions \verb+mar+, \verb+mma+ and \verb+marma+ generate MAR($p$), MMA($q$) and MARMA($p$, $q$) processes respectively.
Examples of calls to these functions are given below.
The \verb+n.start+ argument denotes the burn-in period, which can be specified so that the output series is not unduly influenced by the $p$ starting values, which are all zero by default.
\begin{verbatim}
> marma(100, p = 1, q = 1, psi = 0.75, theta = 0.65)
> mar(100, psi = 0.85, n.start = 20)
> mma(100, q = 2, theta = c(0.75, 0.8))
\end{verbatim}
The function \verb+evmc+ generates first order Markov chains.
Informally, a first order Markov chain $X_1, \ldots, X_n$ is a stochastic process such that at any given time $t$ the probability distribution of $X_{t+1}$ is independent the past $X_1, \ldots, X_{t-1}$, given the current state $X_t$.
The \verb+evmc+ function generates a first order Markov chain such that each pair of consecutive values has the dependence structure of a parametric bivariate extreme value model.
The main arguments of \verb+evmc+ are the same as those of \verb+rbvevd+.
The function \verb+evmc+ also has the argument \verb+margin+, which denotes the marginal distribution of each value.
This must be either \verb+"uniform"+ (the default), \verb+"rweibull"+, \verb+"frechet"+ or \verb+"gumbel"+ (or any unique partial match), for the uniform, standard reversed Weibull, standard Gumbel and standard Fr\'{e}chet distributions respectively.
Examples of calls to \verb+evmc+ are given below.
\begin{verbatim}
> evmc(100, alpha = 0.1, beta = 0.1, model = "bilog")
> evmc(100, dep = 10, model = "hr", margins = "gum")
\end{verbatim}
The function \verb+clusters+ identifies extreme clusters within (stationary) stochastic processes. A simple way of determining clusters is to specify a threshold $u$ and define consecutive exceedances of $u$ to belong to the same cluster.
It is more common though to consider a cluster to be active until $r$ consecutive values fall below (or are equal to) $u$, for some given clustering interval length $r$.
%If $r > 1$ the clusters may contain any arbitrarily low value.
%To avoid this problem a lower threshold $u_l < u$ can be specified so that a cluster is terminated whenever any values fall below (or are equal to) $u_l$.
The following code uses \verb+clusters+ to generate the plots depicted in Figure \ref{clust}.
These plots identify clusters graphically.
If the argument \verb+plot+ is \verb+FALSE+ (the default), then \verb+clusters+ returns a list of extreme clusters.
\begin{verbatim}
> set.seed(150)
> x <- evmc(50, dep = 0.55, model ="log")
> clusters(x, 0.8, plot = TRUE)
> clusters(x, 0.8, 4, plot = TRUE)
\end{verbatim}
\begin{figure}
\begin{center}
\scalebox{0.25}{\includegraphics{clust1.ps}}
\vspace{-1.5cm}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{clust3.ps}}
\end{center}
\vspace{0cm}
\caption{The identification of extreme clusters in a stochastic process. The clustering interval lengths are $r = 1$ (left) and $r=4$ (right). The threshold in each case is $u = 0.8$.}
\label{clust}
\end{figure}
The function \verb+exi+ returns estimates of the Extremal Index of a (stationary) stochastic process. The Extremal Index is defined in Chapter 3 of \citet{leadling83}. A more informal treatment is given in Chapter 5 of \citet{cole01}. The extremal index can be estimated using the inverse of the average size of extreme clusters, where the cluster size is defined as the number of exceedances that it contains.
\section{Fitting Univariate Distributions}
\setcounter{footnote}{0}
\label{unifit}
This section presents functions that produce maximum likelihood estimates for some of the distributions introduced in Section \ref{uni}.
Peaks over threshold models are discussed in Section \ref{potfit}.
Maximum likelihood estimates for bivariate extreme value distributions are discussed in Section \ref{bivfit}.
For illustrative purposes Sections \ref{unifit}, \ref{potfit} and \ref{bivfit} use only simulated data.
Three practical examples using the data sets \verb+oxford+,
\verb+rain+ and \verb+sealevel+ are given in Sections \ref{egoxford},
\ref{egrain} and \ref{egsealevel} respectively.
The function \verb+fgev+ produces maximum likelihood estimates for the GEV distribution \eqref{gev}.
The first argument should be a numeric vector containing data to be fitted.
Missing values are allowed.
If the argument \verb+start+ is given it should be a named list containing starting values, the names of which should be the parameters over which the likelihood is to be maximised.
If \verb+start+ is omitted the routine attempts to find good starting values for the optimisation using moment estimators.
If any of the parameters are to be set to fixed values, they can be given as separate arguments.
For example, the Gumbel distribution \eqref{gumbel} can be fitted using \verb+shape = 0+.
Arguments of the optimisation function \verb+optim+ can also be specified.
This includes the optimisation method, which can be passed using the argument \verb+method+.
Two examples of the \verb+fgev+ function are given below.
\begin{verbatim}
> data1 <- rgev(1000, loc = 0.13, scale = 1.1, shape = 0.2)
> m1 <- fgev(data1)
> m1
Call: fgev(x = data1)
Deviance: 3650
Estimates
loc scale shape
0.127 1.125 0.224
Standard Errors
loc scale shape
0.0400 0.0321 0.0248
Optimization Information
Convergence: successful
Function Evaluations: 51
Gradient Evaluations: 12
> m2 <- fgev(data1, loc = 0, scale = 1)
> fitted(m2)
shape
0.236
\end{verbatim}
In the first example the likelihood is maximised over (\verb+loc+, \verb+scale+, \verb+shape+).
In the second example the likelihood is maximised over \verb+shape+, with the location and scale parameters fixed at zero and one respectively.
The maximum likelihood estimators do not necessarily have the usual asymptotic properties, since the end points of the GEV distribution depend on the model parameters. \citet{smit85} shows that the usual asymptotic properties hold when $\xi > -0.5$.
When $-1 < \xi \leq -0.5$ the maximum likelihood estimators do not have the standard asymptotic properties, but generally exist.
When $\xi \leq -1$ maximum likelihood estimators do not often exist.
This occurs because of the large mass near the upper end point.
The likelihood increases without bound as the upper end point is estimated to be closer and closer to the largest observed value.
In terms of the reversed Weibull shape parameter $\alpha$, the usual asymptotic properties hold when $\alpha>2$, the asymptotic properties are not standard for $1<\alpha\leq2$, and maximum likelihood estimators do not often exist for $\alpha<1$.
When the usual asymptotic properties hold (as here) the standard errors of the maximum likelihood estimates, approximated using the inverse of the observed information matrix, can be extracted from the fitted object using
\begin{verbatim}
> std.errors(m1)
loc scale shape
0.03999 0.03214 0.02479
\end{verbatim}
%When the usual asymptotic properties do not hold the \verb+std.errors+ component will still be based on the inverse of the observed information matrix, but these values must be \emph{interpreted with caution} \citep{smit85}.
Likelihood ratio tests can be performed using the function \verb+anova+.
We can compare the two models \verb+m1+ and \verb+m2+ to test the null hypothesis that the location parameter is zero and the scale parameter is one.
\begin{verbatim}
> anova(m1, m2)
Analysis of Deviance Table
M.Df Deviance Df Chisq Pr(>chisq)
m1 3 3650
m2 1 3669 2 18.8 8.2e-05
\end{verbatim}
The deviance difference, \verb+deviance(m2)+ minus \verb+deviance(m1)+, is about $18.8$, which yields a p-value of $8.2 \times 10^{-5}$ when compared with a chi-squared distribution on two degrees of freedom. Diagnostic plots and profile traces for fitted models can be constructed using the functions \verb+plot+, \verb+profile+ and \verb+profile2d+ (see Section \ref{egoxford}).
By default the maximum likelihood estimates are calculated under the assumption that the data to be fitted are the observed values of independent random variables $Z_1,\dots,Z_n$, where $Z_i \sim \text{GEV}(\mu,\sigma,\xi)$ for each $i=1,\dots,n$. The \verb+nsloc+ argument allows non-stationary models of the form $Z_i \sim \text{GEV}(\mu_i,\sigma,\xi)$, where
\begin{equation*}
\mu_i = \beta_0 + \beta_1x_{i1} + \dots + \beta_kx_{ik}.
\end{equation*}
The parameters $(\beta_0,\dots,\beta_k)$ are to be estimated. In matrix notation $\boldsymbol{\mu} = \boldsymbol{\beta_0} + X \boldsymbol{\beta} $, where $ \boldsymbol{\mu}= (\mu_1,\dots,\mu_n)^T$, $\boldsymbol{\beta_0} = (\beta_0,\dots,\beta_0)^T$, $\boldsymbol{\beta} = (\beta_1,\dots,\beta_k)^T$ and $X$ is the $n \times k$ covariate matrix (excluding the intercept) with $ij$th element $x_{ij}$.
The \verb+nsloc+ argument must be a data frame containing the matrix $X$, or a numeric vector which is converted into a single column data frame with column name ``trend''.
The column names of the data frame are used to derive names for the estimated parameters.
This allows any of the $k+3$ parameters $(\beta_0,\dots,\beta_k,\sigma,\xi)$ to be set to fixed values within the optimisation.
The covariates must be (at least approximately) \emph{centred and scaled}, not only for numerical reasons, but also because the starting value (if \verb+start+ is not given) for each corresponding coefficient is taken to be zero.
When a linear trend is present in the data, the location parameter is often modelled as
\begin{equation*}
\mu_i = \beta_0 + \beta_1t_i,
\end{equation*}
where $t_i$ is some centred and scaled version of the time of the $i$th observation.
More complex changes in $\mu$ may also be appropriate.
For example, a change-point model
\begin{equation*}
\mu_i = \beta_0 + \beta_1x_i \qquad \text{where} \qquad
x_i =
\begin{cases}
0 & i \leq i_0 \\
1 & i > i_0
\end{cases},
\end{equation*}
or a quadratic trend
\begin{equation*}
\mu_i = \beta_0 + \beta_1t_i + \beta_2t_i^2.
\end{equation*}
See Sections \ref{egoxford} and \ref{egsealevel} for examples of non-stationary modelling.
The function \verb+fgev+ also has an argument called \verb+prob+.
If $\verb+prob+ = p$ is passed a value in the interval [0,1], \verb+fgev+ again produces maximum likelihood estimates for the GEV distribution, but the model is re-parameterised from $(\mu,\sigma,\xi)$ to $(z_p,\sigma,\xi)$, where $z_p$ is the quantile corresponding to the upper tail probability $p$. This argument can be used to calculate and plot profile log-likelihoods of extreme quantiles (see Section \ref{egoxford}).
If \verb+prob+ is zero/one, then $z_p$ is defined as the upper/lower end point $\mu - \sigma/\xi$, and $\xi$ is restricted to the negative/positive axis.
Under non-stationarity the model is re-parameterised from $(\beta_0,\beta_1,\dots,\beta_k,\sigma,\xi)$ to $(z_p,\beta_1,\dots,\beta_k,\sigma,\xi)$, so that $z_p$ is the quantile corresponding to the upper tail probability $p$ for the distribution obtained when all covariates are zero.
The \verb+fextreme+ function produces maximum likelihood estimates for the distributions \eqref{maxdens} and \eqref{mindens} given an integer $m$ and an arbitrary distribution function $F$.
The first argument should be a numeric vector containing the data to be fitted, which should represent maxima (if the argument \verb+largest+ is \verb+TRUE+, the default) or minima (if \verb+largest+ is \verb+FALSE+).
The argument \verb+start+ (which cannot be missing) should be a named list containing starting values, the names of which should be the parameters over which the likelihood is to be maximised.
If any of the parameters are to be set to fixed values, they can be given as separate arguments.
Arguments of the optimisation function \verb+optim+ can also be specified.
The example given below produces maximum likelihood estimates for the distribution \eqref{maxdens}, where $m = 365$ and $F$ is the normal distribution.
\begin{verbatim}
> d2 <- rextreme(100, distn = "norm", mean = 0.56, mlen = 365)
# Simulate yearly maxima using normal distribution
> sv <- list(mean = 0, sd = 1)
> nm <- fextreme(d2, start = sv, distn = "norm", mlen = 365)
> fitted(nm)
mean sd
0.685 0.959
\end{verbatim}
The \verb+forder+ function yields maximum likelihood estimates for the distribution \eqref{orderdens} given integers $m$ and $j \in \{1,\dots,m\}$, and an arbitrary distribution function $F$.
An example is given below, where $m = 365$, $j = 2$ and $F$ is the normal distribution.
\begin{verbatim}
> d3 <- rorder(100, distn = "norm", mean = 0.56, mlen = 365, j = 2)
> sv <- list(mean = 0, sd = 1)
> nm2 <- forder(d3, sv, distn = "norm", mlen = 365, j = 2)
> fitted(nm2)
mean sd
0.483 1.042
\end{verbatim}
\section{Fitting Peaks Over Threshold Models}
\setcounter{footnote}{0}
\label{potfit}
Suppose $X_1,\dots,X_n$ is a sequence of independent and identically distributed random variables, with $M_n = \{X_1,\dots,X_n\}$. Suppose that $n$ is large, so that (assuming certain regularity conditions) the distribution of $M_n$ is approximately GEV\@. Then for large enough $u$, the exceedances of the threshold $u$ are approximately distributed as generalised Pareto, with location parameter $u$. The function \verb+fpot+ fits this distribution to the exceedances, and hence produces maximum likelihood estimates for the shape and scale parameters. The value of the threshold $u$ must be specified by the user. It is typically chosen to be as small as possible, subject to the limit model providing a reasonable approximation.
The functions \verb+mrlplot+ and \verb+tcplot+\footnote{Both of these functions are heavily based on code by Stuart Coles.} produce diagnostic plots that facilitate the specification of $u$. The function \verb+mrlplot+ produces the empirical mean residual life plot, which is a plot of the empirical mean of the excesses of $u$ (i.e.\ the exceedances of $u$ minus $u$), plotted against $u$. If the exceedances of a threshold $u_0$ are generalised Pareto, the empirical mean residual life plot should be approximately linear for all $u > u_0$.
The function \verb+tcplot+ calculates maximum likelihood estimates for the shape and modified scale parameters using a number of different thresholds, and plots these estimates against $u$.
If the exceedances of a threshold $u_0$ are generalised Pareto, the shape and modified scale parameters should be approximately constant with respect to all thresholds $u > u_0$.
Threshold identification plots produced from the example given below are depicted in Figure \ref{threshid}.
In this case, the threshold $u = 1$ was chosen.
\begin{figure}
\begin{center}
\scalebox{0.25}{\includegraphics{threshid1.ps}}
\vspace{-1.5cm}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{threshid2.ps}}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{threshid3.ps}}
\end{center}
\caption{The identification of a threshold for the (generalised Pareto) peaks over threshold model. From left to right; the empirical mean residual life plot, modified scale parameter estimates and shape parameter estimates.}
\label{threshid}
\end{figure}
The following code generates $n = 500$ independent standard normal random variables and fits the (generalised Pareto) peaks over threshold model to the exceedances of the threshold $u = 1$.
The function \verb+fpot+ performs the fit.
Many of the arguments of \verb+fpot+ are similar to those of \verb+fgev+.
In particular, either of the \verb+scale+ or \verb+shape+ parameters can be set to fixed values by giving those parameters as arguments.
For example, an exponential distribution for the excesses (or equivalently, a shifted exponential distribution for the exceedances) can be fitted using \verb+shape = 0+.
\begin{verbatim}
> tmp <- rnorm(500)
> mrlplot(tmp, tlim = c(-1,1.5))
> tcplot(tmp, tlim = c(-1,1.5))
> pot1 <- fpot(tmp, 1)
> pot1
Call: fpot(x = tmp, threshold = 1)
Deviance: 40.5
Threshold: 1
Number Above: 76
Proportion Above: 0.152
Estimates
scale shape
0.593 -0.211
\end{verbatim}
The fitted model \verb+pot1+ gives the estimates for the scale and shape parameters of the generalised Pareto distribution fitted to the exceedances.
Also given is the proportion of values above the threshold, or equivalently, the maximum likelihood estimate for the probability of an exceedance.
Diagnostic plots and profile traces for fitted models can be constructed using the functions \verb+plot+ and \verb+profile+ (see Section \ref{egrain}).
The peaks over thresholds model is typically extended to stationary series via declustering, which corresponds to a filtering of dependent observations to obtain a set of threshold exceedances which are approximately independent.
An empirical rule is used to identify clusters of exceedances, and the generalised Pareto model is then fitted to the cluster maxima, assuming those maxima to be independent.
The empirical rule, as given in Section \ref{stochproc}, is defined by the function \verb+clusters+.
A model of this form can be implemented by setting the logical argument \verb+cmax+ to \verb+TRUE+.
The clusters are identified using the threshold of the peaks over threshold model.
An illustration of this technique is given below.
The argument \verb+r+ is the clustering interval length.
\begin{verbatim}
> tmp2 <- evmc(500, dep = 0.8, margins = "gum")
> pot2 <- fpot(tmp2, 1.5, cmax = TRUE, r = 3)
> pot2
Call: fpot(x = tmp, threshold = 1, cmax = TRUE, r = 3)
Deviance: 101.1
Threshold: 1.5
Number Above: 92
Proportion Above: 0.184
Clustering Interval: 3
Number of Clusters: 41
Extremal Index: 0.446
Estimates
scale shape
1.657 -0.272
\end{verbatim}
The Extremal Index is a quantity briefly discussed in Section \ref{stochproc}. The estimate of the Extremal Index is simply the number of clusters divided by the number of exceedances.
The function \verb+fpot+ also has an argument called \verb+mper+.
If $\verb+mper+ = m$ is passed a positive value, \verb+fpot+ again produces maximum likelihood estimates for the generalised Pareto model, but the model is re-parameterised from $(\sigma,\xi)$ to $(z_m,\xi)$, where $z_m$ is the $m$-period return level, defined as follows.
Let $G$ be the fitted generalised Pareto distribution function, with location parameter equal to the specified threshold $u$, so that $1 - G(z)$ is the fitted probability of an exceedance over $z > u$ given an exceedance over $u$.
The fitted probability of an exceedance over $z > u$ is therefore $p(1 - G(z))$, where $p$ is the estimated probability of exceeding $u$, which is given by the empirical proportion of exceedances.
The $m$-period return level $z_m$ satisfies $p(1 - G(z_m)) = 1/(mN\hat{\theta})$, where $N$ is the number of observations per period, and $\hat{\theta}$ is the estimate of the extremal index if cluster maxima are fitted, with $\hat{\theta} = 1$ otherwise. The value $N$ can be specified using the argument \verb+npp+. For example, if observations are recorded weekly and $\verb+npp+ = 52$, then $z_m$ is the $m$-year return level.
If \verb+mper+ is \verb+Inf+, then $z_m$ is defined as the upper end point $u - \sigma/\xi$, and $\xi$ is then restricted to be negative.
The argument \verb+mper+ can be used to calculate and plot profile log-likelihoods of return levels (see Section \ref{egrain}).
%The peaks over threshold model permits an alternative characterization in terms of point processes.
%Suppose again that $X_1,\dots,X_n$ is a sequence of independent and identically distributed random variables, with $M_n = \{X_1,\dots,X_n\}$, and that $n$ is large, so that (assuming certain regularity conditions) the distribution of $M_n$ is approximately \text{GEV}($\mu,\sigma,\xi$), with (possibly infinite) end points\footnote{If $\xi > 0$, $z_- = \mu - \sigma/\xi$ and $z_+ = \infty$. If $\xi < 0$, $z_- = -\infty$ and $z_+ = \mu - \sigma/\xi$. If $\xi = 0$, the expressions given are all defined by continuity, with $z_- = -\infty$ and $z_+ = \infty$.} $z_-$ and $z_+$. Then for large enough $u > z_-$, the sequence $\{X_1,\dots,X_n\}$ viewed on the interval $(u,z_+)$ can be approximated by a non-homogeneous Poisson process \citep{cole01}.
%The approximation leads to a likelihood for ($\mu,\sigma,\xi$), and hence maximum likelihood estimates can be obtained.
%The likelihood can be easily adjusted so that the maxima of a given (large) number $N \leq n$ of random variables is approximately distributed as \text{GEV}($\mu,\sigma,\xi$), so that e.g.\ if observations are recorded weekly and $N = 52$, then ($\mu,\sigma,\xi$) corresponds to the distribution of annual maxima.
%The point process characterization can be fitted using the \verb+fpot+ function with \verb+model = "pp"+.
%The value $N$ can by specified using the argument \verb+npp+.
%If \verb+npp+ is unspecified the default value $N = n$ is used.
%The following code uses the point process characterization to fit a peaks over threshold model to the simulated data \verb+tmp+.
%The models \verb+pot3+ and \verb+pot4+ are equivalent; the estimates in \verb+pot3+ correspond to the GEV distribution for the maxima of the data set, whereas those in \verb+pot4+ correspond to the GEV distribution for annual maxima, assuming the observations are recorded daily.
%\begin{verbatim}
%> pot3 <- fpot(tmp, 1, model = "pp", npp = 500)
%> pot4 <- fpot(tmp, 1, model = "pp", npp = 365.25)
%> fitted(pot3)
% loc scale shape
% 2.6839 0.2380 -0.2108
%> fitted(pot4)
% loc scale shape
% 2.6065 0.2542 -0.2108
%
%> fitted(pot1)
% scale shape
% 0.593 -0.211
%\end{verbatim}
%Also given above is the parameter estimates for the model \verb+pot1+, fitted using the generalised Pareto characterization. Let $(\tilde{\sigma}, \tilde{\xi})$ denote the scale and shape parameters of the \text{GPD}. The relationship between the two characterizations is then given by $\tilde{\xi} = \xi$ and $\tilde{\sigma} = \sigma + \xi(u - \mu)$, where $u$ is the threshold.
%This relationship can be seen in the above estimates.
%Under the generalized Pareto characterization, the parameter $\tilde{\sigma} - \tilde{\xi} u$ is referred to as the modified scale parameter, as plotted in the centre panel of Figure \ref{threshid}. Unlike $\tilde{\sigma} = \tilde{\sigma}(u)$, the modified scale parameter does not depend on the threshold $u$.
\section{Fitting Bivariate Extreme Value Distributions}
\setcounter{footnote}{0}
\label{bivfit}
The function \verb+fbvevd+ produces maximum likelihood estimates for nine bivariate extreme value models.
The first argument should be a numeric matrix (or a data frame) with two columns containing the data to be fitted.
Missing values are allowed.
If the argument \verb+start+ is given it should be a named list containing starting values, the names of which should be the parameters over which the likelihood is to be maximised.
If \verb+start+ is omitted the routine attempts to find good starting values for the optimisation using maximum likelihood estimators under the assumption of independence.
If any of the parameters are to be set to fixed values, they can be given as separate arguments.
Common marginal parameters can be fitted using the arguments \verb+cshape+, \verb+cscale+ and \verb+cloc+, and the dependence function can be constrained to symmetry using the argument \verb+sym+ (see the \verb+fbvevd+ help file for details).
The \verb+nsloc1+ and \verb+nsloc2+ arguments allow non-stationary modelling of the location parameters on the first and second margins respectively.
They should be used in the same manner as the \verb+nsloc+ argument of \verb+fgev+.
Examples of bivariate models with non-stationary margins are given in Section \ref{egsealevel}.
%For numerical reasons the parameters of each model are subject to the artificial constraints depicted in Table \ref{contab}. The scale parameters on each GEV margin are artificially constrained to be greater than or equal to $0.01$. These constraints only apply to the functions discussed in this section.
%\begin{table}
%\begin{center}
%\begin{tabular}{l|c}
%Bivariate Model & Constraints \\ \hline
%Logistic & $0.1\leq\alpha\leq1$ \\
%Asymmetric Logistic & $0.1\leq\alpha\leq1$, $0.001\leq\theta_1,\theta_2\leq1$ \\
%H\"{u}sler-Reiss & $0.2\leq\lambda\leq10$ \\
%Negative Logistic & $0.05\leq r \leq5$ \\
%Asymmetric Negative Logistic & $\quad0.05\leq r \leq5$, $0.001\leq\theta_1,\theta_2\leq1\quad$ \\
%Bilogistic & $0.1\leq\alpha,\beta\leq0.999$ \\
%Negative Bilogistic & $0.1\leq\alpha,\beta\leq20$ \\
%Coles-Tawn & $0.001\leq\alpha,\beta\leq30$ \\ \hline
%\end{tabular}
%\caption{For numerical reasons the parameters of each model are subject to the artificial constraints depicted here.}
%\label{contab}
%\end{center}
%\end{table}
The first example given below produces maximum likelihood estimates for the (symmetric) logistic model.
The second example constrains the model at independence (where $\texttt{dep} = 1$).
The estimates produced in the second example are the same as those that would be produced if \verb+fgev+ was separately applied to each margin.
\begin{verbatim}
> bvdata <- rbvevd(100, dep = 0.6, mar1 = c(1.2,1.4,0), mar2 = c(1,1.6,0.1))
> m1 <- fbvevd(bvdata, model = "log")
> m1
Call: fbvevd(x = bvdata, model = "log")
Deviance: 728.5
AIC: 742.5
Dependence: 0.3526
Estimates
loc1 scale1 shape1 loc2 scale2 shape2 dep
1.2121 1.3831 -0.1813 0.8404 1.4005 0.0834 0.7202
Standard Errors
loc1 scale1 shape1 loc2 scale2 shape2 dep
0.1540 0.1091 0.0673 0.1537 0.1144 0.0614 0.0624
Optimization Information
Convergence: successful
Function Evaluations: 47
Gradient Evaluations: 10
> m2 <- fbvevd(bvdata, model = "log", dep = 1)
> fitted(m2)
loc1 scale1 shape1 loc2 scale2 shape2
1.2231 1.3776 -0.1914 0.8367 1.4083 0.0868
> std.errors(m2)
loc1 scale1 shape1 loc2 scale2 shape2
0.1543 0.1089 0.0725 0.1565 0.1163 0.0670
> c(logLik(m2), deviance(m2), AIC(m2))
[1] -376 752 764
\end{verbatim}
The discussion in Section \ref{unifit} regarding the properties of maximum likelihood estimators for the GEV distribution also applies to all bivariate models.
The usual asymptotic properties will not hold if either of the marginal shape parameters are less than $-0.5$.
%When the usual asymptotic properties do not hold the \verb+std.errors+ component will still be based on the inverse of the observed information matrix, but these values must be \emph{interpreted with caution} \citep{smit85}.
The value in the output labelled \verb+Dependence+ is the fitted estimate of $\chi = 2\{1-A(1/2)\} \in [0,1]$ \citep{coleheff99}, where $A(\cdot)$ denotes the dependence function \eqref{bvdepfn}. At independence $\chi = 0$, and at complete dependence $\chi = 1$.
Diagnostic plots and profile traces for fitted models can be constructed using the functions \verb+plot+, \verb+profile+ and \verb+profile2d+ (see Section \ref{egsealevel}).
The function \verb+anova+ performs likelihood ratio tests.
The null hypothesis of the test performed below specifies that the margins are Gumbel distributions ($\texttt{shape1} = \texttt{shape2} = 0$).
The deviance of the constrained model is compared with the deviance of the unconstrained model, and the p-value is calculated to be $0.78$.
The hypothesis would not be rejected at any reasonable significance level.
\begin{verbatim}
> m3 <- fbvevd(bvdata, model = "log", shape1 = 0, shape2 = 0)
> anova(m1, m3)
Analysis of Deviance Table
M.Df Deviance Df Chisq Pr(>chisq)
m1 7 708
m3 5 708 2 0.5 0.78
\end{verbatim}
In the following example I attempt to fit the asymmetric logistic model to the simulated data set used above, which is known to be distributed as symmetric logistic.
\begin{verbatim}
> m4 <- fbvevd(bvdata, model = "alog")
> fitted(m4)
loc1 scale1 shape1 loc2 scale2 shape2 asy1 asy2 dep
1.2097 1.3928 -0.1853 0.8421 1.3831 0.0773 0.8331 0.9996 0.6925
\end{verbatim}
A boundary of the parameter space has been reached; the maximum likelihood estimate for the second asymmetry parameter is one.
This may cause difficulties for the optimiser.
There are two solutions to this problem: the second asymmetry parameter can be fixed at one, or the \verb+L-BFGS-B+ method can be used.
The \verb+L-BFGS-B+ method allows box-constraints using the arguments \verb+lower+ and \verb+upper+.
The following snippet illustrates both approaches.
\begin{verbatim}
> mb <- fbvevd(bvdata, model = "alog", asy2 = 1)
> round(fitted(mb), 3)
loc1 scale1 shape1 loc2 scale2 shape2 asy1 dep
1.212 1.385 -0.176 0.834 1.396 0.086 0.867 0.693
> up <- c(rep(Inf, 6), 1, 1, 1)
> mb <- fbvevd(bvdata, model = "alog", method = "L-BFGS-B", upper = up)
> round(fitted(mb), 3)
loc1 scale1 shape1 loc2 scale2 shape2 asy1 asy2 dep
1.212 1.385 -0.176 0.834 1.396 0.086 0.867 1.000 0.693
\end{verbatim}
\section{Example: Oxford Temperature Data}
\setcounter{footnote}{0}
\label{egoxford}
The numeric vector \verb+oxford+ contains annual maximum temperatures (in degrees Fahrenheit) at Oxford, England, from 1901 to 1980.
It is included in the evd package, and can be made available using \verb+data(oxford)+.
The data has previously been analysed by \citet{tabo83}.
I begin by plotting the data.
The assumptions of stationarity and independence seem sensible, given the plot (not shown) generated using the code below.
\begin{verbatim}
> data(oxford) ; ox <- oxford
> plot(1901:1980, ox, xlab = "year", ylab = "temperature")
\end{verbatim}
The following code fits two models based on the GEV distribution.
The first model assumes stationarity.
The second model allows for a trend term in the location parameter (even though the plot appears to show that this is unnecessary).
The \verb+nsloc+ argument is centred and scaled so that the intercept \verb+loc+ represents the location parameter in 1950 and the trend \verb+loctrend+ represents the increase in the location parameter (or decrease, if negative) over a period of 100 years.
\begin{verbatim}
> ox.fit <- fgev(ox)
> tt <- (1901:1980 - 1950)/100
> ox.fit.trend <- fgev(ox, nsloc = tt)
> fitted(ox.fit.trend)
loc loctrend scale shape
83.6617 -1.8812 4.2233 -0.2841
> std.errors(ox.fit.trend)
loc loctrend scale shape
0.5557 1.9675 0.3650 0.0707
\end{verbatim}
% Moved for graphics placement.
\begin{figure}
\begin{center}
\scalebox{0.25}{\includegraphics{graph3.ps}}
\vspace{-1.5cm}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{graph4.ps}}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{graph5.ps}}
\end{center}
\caption{Diagnostic plots for the model \texttt{ox.fit}.}
\label{oxdiag}
\end{figure}
The trend term not statistically significant (at any reasonable level).
The stationary model \verb+ox.fit+ is retained for further analysis.
\begin{verbatim}
> ox.fit
Call: fgev(x = oxford)
Deviance: 457.8
Estimates
loc scale shape
83.839 4.260 -0.287
Standard Errors
loc scale shape
0.5231 0.3658 0.0683
\end{verbatim}
The fitted shape is negative, so the fitted distribution is Weibull.
It is often of interest to test the hypothesis that the shape is zero (the Gumbel distribution).
The code \verb+confint(ox.fit)+ returns the 95\% Wald confidence intervals for the model parameters, roughly equal to the fitted estimates plus or minus twice their standard errors. The interval for the shape parameter is given by $(-0.42,-0.15)$.
The corresponding Wald test for $\xi = 0$ would be rejected at significance level $0.05$ since the 95\% confidence interval does not contain zero.
A likelihood ratio test for $\xi = 0$ is performed in the following snippet.
The hypothesis is rejected at any significance level above $0.00053$.
\begin{verbatim}
> ox.fit.gum <- fgev(ox, shape = 0)
> anova(ox.fit, ox.fit.gum)
Analysis of Deviance Table
M.Df Deviance Df Chisq Pr(>chisq)
ox.fit 3 458
ox.fit.gum 2 470 1 12 0.00053
\end{verbatim}
Diagnostic plots can be produced using \verb+plot(ox.fit)+.
The plots produced compare parametric distributions, densities and quantiles to their empirical counterparts (see the \verb+plot.uvevd+ help file for details). Selected diagnostics are depicted in Figure \ref{oxdiag}.
The small bars on the P-P, Q-Q and return level plots represent simulated (pointwise) 95\% confidence intervals.
The model \verb+ox.prof+ is seen to be a good fit.
The fitted density is close to the non-parametric estimator, and most points lie within the confidence intervals.
Profile log-likelihoods of the parameters can be plotted using
\begin{verbatim}
> ox.prof <- profile(ox.fit)
> plot(ox.prof)
\end{verbatim}
\begin{figure}
\begin{center}
\scalebox{0.25}{\includegraphics{graph7.ps}}
\vspace{-1.5cm}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{graph8.ps}}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{graph9.ps}}
\end{center}
\caption{Profile log-likelihoods for the model \texttt{ox.fit}.}
\label{oxprof}
\end{figure}
The profile log-likelihoods for the scale and shape parameters are the first two plots of Figure \ref{oxprof}. A horizontal line is (optionally) drawn on each plot so that the intersection of the line with the profile log-likelihood yields a profile confidence interval, with (default) confidence coefficient 0.95.
The end points of the intervals can be derived using \verb+confint(ox.prof)+.
The profile confidence intervals for the location and shape parameters are approximately the same as the Wald confidence intervals, since the profile log-likelihoods are approximately symmetric.
The profile log-likelihood for the scale parameter is asymmetric; both end points of the profile confidence interval $(3.64, 5.12)$ are larger than the corresponding end points of the Wald interval $(3.54, 4.98)$. The joint profile log-likelihood of the scale and shape parameters can be plotted using
\begin{verbatim}
> ox.prof2d <- profile2d(ox.fit, ox.prof, which = c("scale", "shape"))
> plot(ox.prof2d)
\end{verbatim}
This produces the image plot in the right panel of Figure \ref{oxprof}.
The colours of the image plot represent confidence sets with different confidence coefficients.
By default, the lightest colour (ignoring the background colour) represents a confidence set with coefficient 0.995; the darkest colour represents a confidence set with coefficient 0.5.
Let $G$ be the GEV distribution function, and let $G(z_p) = 1-p$, so that
\begin{equation*}
z_p =
\begin{cases}
\mu - \frac{\sigma}{\xi}[1 - \{-\log(1-p)\}^{-\xi}] & \xi \neq 0 \\
\mu - \sigma \log\{-\log(1-p)\} & \xi = 0,
\end{cases}
\end{equation*}
is the quantile corresponding to the upper tail probability $p$.
The profile log-likelihood for $z_{0.1}$ can be plotted using the following.
The argument $\verb+prob+ = p$ re-parameterises the GEV distribution so that \verb+fgev+ produces maximum likelihood estimates for $(z_p,\sigma,\xi)$.
\begin{verbatim}
> ox.qfit <- fgev(ox, prob = 0.1)
> ox.qprof <- profile(ox.qfit, which = "quantile")
> plot(ox.qprof)
\end{verbatim}
\begin{figure}
\begin{center}
\scalebox{0.25}{\includegraphics{graph10.ps}}
\vspace{-1.5cm}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{graph11.ps}}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{graph12.ps}}
\end{center}
\caption{Profile log-likelihoods for $z_{0.1}$, $z_{0.01}$ and $z_{0.001}$.}
\label{quantprof}
\end{figure}
Figure \ref{quantprof} shows profile log-likelihoods for $z_{0.1}$, $z_{0.01}$ and $z_{0.001}$.
The extent of the asymmetry in the profile log-likelihood increases for decreasing (small) $p$.
This is to be expected, since the data provide increasingly weaker information in the upper tail of the fitted distribution.
If $\verb+prob+ = p$ is zero, then $z_p$ is the upper end point of the GEV distribution, given by $\mu-\sigma/\xi$ when $\xi < 0$.
The profile log-likelihood for $z_0$ can be plotted using the following code.
\begin{verbatim}
> ox.qfit <- fgev(ox, prob = 0)
> ox.qprof <- profile(ox.qfit, which = "quantile", conf = 0.99)
> plot(ox.qprof)
> confint(ox.qprof)
lower upper
quantile 95.78 113.0
\end{verbatim}
The argument \verb+conf+ of the function \verb+profile+ controls the range of the profile trace.
The profile trace is constructed so that profile confidence intervals with confidence coefficients \verb+conf+ or less can be derived from it.
By default, $\verb+conf+ = 0.999$, though a smaller value is often appropriate when the profile log-likelihood exhibits strong asymmetry.
The 95\% profile confidence interval for the upper end point $z_0$ is derived as (95.8,113.0).
\section{Example: Rainfall Data}
\setcounter{footnote}{0}
\label{egrain}
The numeric vector \verb+rain+ contains 17531 daily rainfall accumulations at a location in south-west England, recorded over the period 1914 to 1962.
The data is not included in the evd package, but it is available in the ismev package, which can be downloaded from CRAN.
As usual, the package can be loaded using \verb+library(ismev)+, and the data can be made available using \verb+data(rain)+.
The plot of the data given in Figure 1.7 of \citet{cole01} shows that an assumption of stationarity is sensible.
The example given here follows \citet{cole01}, pages 84--86.
\begin{verbatim}
> mrlplot(rain, tlim = c(0,85), nt = 100)
> par(mfrow = c(2,1))
> tcplot(rain, tlim = c(0,50), nt = 20)
> potgp <- fpot(rain, 30, npp = 365.25)
> potgp2 <- fpot(rain, 30, npp = 365.25, cmax = TRUE, r = 7)
> clusters(rain, 30, r = 7, cmax = TRUE)
\end{verbatim}
The first three lines of code produce the threshold diagnostic plots given in pages 80 and 85 of \citet{cole01}, who subsequently decides to work with the threshold $u = 30$.
The model \verb+potgp+ reports that 152 observations lie above the threshold, giving an exceedance probability estimate of 0.00867.
The estimates and standard errors of the parameters of \verb+potgp+ agree with those given page 85 of \citet{cole01}.
In \verb+potgp2+ the peaks over threshold model is applied to cluster maxima, where clusters are defined using a clustering interval length of seven.
As there is little sign of clustering in the data, this leads to relatively small changes in the parameter estimates, and relatively small increases in the standard errors.
The final line of code calls the function \verb+clusters+ (see Section \ref{stochproc}) in order to produce the cluster maxima that were used for the fitting of model \verb+potgp2+.
Diagnostic plots can be produced using \verb+plot(potgp)+.
The plots compare parametric distributions, densities and quantiles to their empirical counterparts (see the \verb+plot.uvevd+ help file for details).
Selected diagnostics are given in Figure \ref{potdiag}.
The x-axis of the return level plot gives return periods in units of years, since we specified the number of observations per period as $\texttt{npp} = 365.25$.
The small bars on the P-P, Q-Q and return level plots represent simulated (pointwise) 95\% confidence intervals.
The model \verb+potgp+ is seen to be a good fit.
The fitted density tail is close to the non-parametric estimator, and most points lie within the confidence intervals.
\begin{figure}
\begin{center}
\scalebox{0.25}{\includegraphics{potdiag2.ps}}
\vspace{-1.5cm}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{potdiag3.ps}}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{potdiag4.ps}}
\end{center}
\caption{Diagnostic plots for the peaks over threshold model for daily rain data.}
\label{potdiag}
\end{figure}
Profile log-likelihoods of the shape parameter and the 100-year return level (not shown) can be plotted using the following code. The argument $\verb+mper+ = m$ re-parameterises the model so that \verb+fpot+ produces maximum likelihood estimates for $(z_m,\xi)$, where $z_m$ is the $m$ period return level, as defined in Section \ref{potfit}.
Horizontal lines denoting 95\% profile confidence intervals are depicted on each plot. The end points of profile confidence intervals can be derived using \verb+confint(prgp3)+.
\begin{verbatim}
potgp3 <- fpot(rain, 30, npp = 365.25, mper = 100)
prgp3 <- profile(potgp3)
plot(prgp3)
\end{verbatim}
%\begin{figure}
%\begin{center}
%\scalebox{0.18}{\includegraphics{potprof1.ps}}
%\vspace{-1.5cm}
%\hspace{0cm}
%\scalebox{0.18}{\includegraphics{potprof2.ps}}
%\end{center}
%\caption{Profile deviances for the shape parameter and 100-year return level in the peaks over threshold model for daily rain data.}
%\label{potprof}
%\end{figure}
\section{Example: Sea Level Data}
\setcounter{footnote}{0}
\label{egsealevel}
The \verb+sealevel+ data frame \citep{coletawn90} has two columns containing annual sea level maxima from 1912 to 1992 at Dover and Harwich, two sites on the coast of Britain.
It contains 39 missing maxima in total; nine at Dover and thirty at Harwich.
There are three years for which the annual maximum is not available at either site.
I begin by plotting the data, using the code below.
The plot of the Harwich maxima against the Dover maxima, given in the left panel of Figure \ref{seadata}, depicts a reasonable degree of dependence.
The outlier corresponds to the 1953 flood resulting from a storm passing over the South-East coast of Britain on 1st February.
The marginal plots (not shown) suggest that the Harwich and Dover maxima both increase with time. The last line of code\footnote{The function \texttt{chiplot} is heavily based on code by Jan Heffernan.} plots estimates of $\chi(u)$ and $\bar{\chi}(u)$ for $0 < u < 1$ \citep{coleheff99}, as depicted in Figure \ref{seadata}. For bivariate extreme value distributions, $\chi(u) = \chi$ is constant for all $0 < u < 1$, and $\lim_{u \rightarrow 1}\bar{\chi}(u) = 1$. The conditions do not seem unreasonable given the wide confidence intervals in each plot.
\begin{verbatim}
> data(sealevel) ; sl <- sealevel
> plot(sl, xlab = "Dover Annual Maxima", ylab = "Harwich Annual Maxima")
> plot(1912:1992, sl[,1], xlab = "Year", ylab = "Dover Annual Maxima")
> plot(1912:1992, sl[,2], xlab = "Year", ylab = "Harwich Annual Maxima")
> chiplot(sl)
\end{verbatim}
\begin{figure}
\begin{center}
\scalebox{0.25}{\includegraphics{bvgraph1.ps}}
\vspace{-1.5cm}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{chi.ps}}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{chibar.ps}}
\end{center}
\caption{From left to right; Harwich maxima vs Dover maxima, estimated values of $\chi(u)$ vs $u$, estimated values of $\bar{\chi}(u)$ vs $u$.}
\label{seadata}
\end{figure}
The following three expressions fit (symmetric) logistic models.
The first model incorporates linear trend terms on both marginal location parameters.
The second model incorporates a linear trend on the Dover margin only.
The third model assumes stationarity.
The \verb+nsloc1+ and \verb+nsloc2+ arguments are centred and scaled so that the intercepts \verb+loc1+ and \verb+loc2+ represent the marginal location parameters in 1950 and the linear trend parameters \verb+loc1trend+ and \verb+loc2trend+ represent the increase in the marginal location parameters (or decrease, if negative) over a period of 100 years.
\begin{verbatim}
> tt <- (1912:1992 - 1950)/100
> m1 <- fbvevd(sl, model = "log", nsloc1 = tt, nsloc2 = tt)
> m2 <- fbvevd(sl, model = "log", nsloc1 = tt)
> m3 <- fbvevd(sl, model = "log")
\end{verbatim}
%I'll leave you to analyse the models in detail.
%In particular, notice how the trend terms affect the parameter estimates.
%Marginal Weibull distributions (negative shapes) are estimated when the trends are not included, but marginal Fr\'{e}chet distributions (positive shapes) are estimated upon their inclusion.
The maximum likelihood estimates of the parameters can be compared with their standard errors to perform Wald tests. Wald confidence intervals can be derived using e.g.\ \verb+confint(m1)+.
Likelihood ratio tests are performed in the following snippet.
The p-values confirm the statistical significance of the linear trend terms.
\begin{verbatim}
> anova(m1, m2, m3)
Analysis of Deviance Table
M.Df Deviance Df Chisq Pr(>chisq)
m1 9 -36.5
m2 8 -29.2 1 7.26 0.007
m3 7 -9.7 1 19.56 9.7e-06
\end{verbatim}
Quadratic trends for the location parameter on either or both margins can be incorporated using the following code.
Further testing, using the models generated below, suggests that a quadratic trend may be implemented for the location parameter on the Harwich margin.
Despite this, I retain the model \verb+m1+ for further analysis.
\begin{verbatim}
> tdframe <- data.frame(trend = tt, quad = tt^2)
> m4 <- fbvevd(sl, model = "log", nsloc1 = tdframe, nsloc2 = tt)
> m5 <- fbvevd(sl, model = "log", nsloc1 = tt, nsloc2 = tdframe)
> m6 <- fbvevd(sl, model = "log", nsloc1 = tdframe, nsloc2 = tdframe)
\end{verbatim}
The code given below compares two logistic models that are nested within \verb+m1+. Model \verb+m7+ assumes independence.
The maximum likelihood estimates are the same as those that would be produced if \verb+fgev+ was separately applied to each margin.
The asymptotic distribution of the deviance difference between models \verb+m7+ and \verb+m1+ is non-regular because the dependence parameter in the restricted (independence) model is fixed at the edge of the parameter space.
\cite{tawn88} discusses non-regular cases, including this case, for which the asymptotic distribution is one-half of a chi-squared random variable on one degree of freedom.
In these cases the argument \verb+half+ should be set to \verb+TRUE+.
The resulting p-value is less than $10^{-6}$, and clearly the independence model is rejected.
Model \verb+m8+ assumes that both marginal shape parameters are zero (or equivalently, that both marginal distributions are Gumbel).
A likelihood ratio test of this hypothesis provides a p-value of $0.72$.
The hypothesis would not be rejected at any reasonable significance level.
\begin{verbatim}
> m7 <- fbvevd(sl, model = "log", nsloc1 = tt, nsloc2 = tt, dep = 1)
> anova(m1, m7, half = TRUE)
Analysis of Deviance Table
M.Df Deviance Df Chisq Pr(>chisq)
m1 9 -36.5
m7 8 -22.9 1 27.2 1.9e-07
> m8 <- fbvevd(sl, "log", nsloc1 = tt, nsloc2 = tt, shape1 = 0, shape2 = 0)
> anova(m1, m8)
Analysis of Deviance Table
M.Df Deviance Df Chisq Pr(>chisq)
m1 9 -36.5
m8 7 -35.8 2 0.67 0.72
\end{verbatim}
Diagnostic plots for the fitted (generalised extreme value) marginal distributions can be produced using \verb+plot+ with \verb+mar = 1+ or \verb+mar = 2+.
The plots produced are of the same structure as those given in Section \ref{egoxford}.
Diagnostic plots for the fitted dependence structure can be produced using \verb+plot+. There are six plots available (see the \verb+plot.bvevd+ help file for details). Two diagnostic plots are depicted within Figure \ref{seadiag}.
\begin{verbatim}
> plot(m1, mar = 1)
> plot(m1, mar = 2)
> plot(m1, which = 1:5)
\end{verbatim}
\begin{figure}
\begin{center}
\scalebox{0.25}{\includegraphics{bvgraph7.ps}}
\vspace{-1.5cm}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{bvqcurve.ps}}
\hspace{0cm}
\scalebox{0.25}{\includegraphics{bvgraph8.ps}}
\end{center}
\caption{From left to right; dependence function diagnostic plot, quantile curves diagnostic plot, profile log-likelihood of the dependence parameter.}
\label{seadiag}
\end{figure}
The model \verb+m1+ fits the data reasonably well.
There are some minor deviations within the conditional P-P plots (not shown), but they do not represent a serious departure of the empirical estimates from the fitted model.
The profile log-likelihood of the dependence parameter \verb+dep+, as given in the right panel of Figure \ref{seadiag}, can be plotted using the following.
The argument \verb+xmax+ denotes the upper bound of the parameter.
\begin{verbatim}
> m1.prof <- profile(m1, which = "dep", xmax = 1)
> plot(m1.prof)
> confint(m1.prof)
lower upper
dep 0.528 0.887
\end{verbatim}
A horizontal line is (optionally) drawn so that the intersection of the line with the profile log-likelihood yields a profile confidence interval, with (default) confidence coefficient 0.95. The interval is derived as $(0.53,0.89)$.
Further analysis with models other than the (symmetric) logistic yields the following conclusions.
The two models in Section \ref{biv} that include three parameters with which to describe the dependence structure (the asymmetric logistic and asymmetric negative logistic) are inappropriate.
In both cases, the maximum likelihood estimate for the parameter \verb+dep+ is at an artificial boundary, because the fitted model is close to a distribution (obtained in the limit) which contains a singular component.
This is clearly illustrated in the density plots of the fitted models, which both depict a ridge of mass extending towards the 1953 outlier.
The logistic and the bilogistic models have the lowest deviance of all one and two parameter models respectively.
The dependence structure of the fitted bilogistic model is almost symmetric.
At symmetry, the bilogistic model reduces to the logistic model, and so the latter would appear to be preferable.
A likelihood ratio test between the two (nested) models gives a p-value of $0.93$.
%Models that are not nested can be compared by adding penalty terms to the deviances.
%The penalty terms take into account the number of parameters fitted. (If both models have the same number of parameters the deviances can be compared directly.)
%Three commonly used penalty terms are $2p$ (Akaike's information criterion, or AIC), $p\log(n)$ (Schwarz's criterion, or SC) and $p\{1+\log(n)\}$ (Bayesian information criterion, or BIC), where $p$ is the number of parameters estimated and $n$ is the number of observations.\footnote{Since \texttt{fbvall} compares models for the dependence structure, $n$ is taken as the number of observations which are complete (i.e.\ not missing on either margin).}
%Any bivariate extreme value distribution function can be expressed as \citep{haan84}
%\begin{equation*}
%G(z_1,z_2) = \exp\left\{ - \int_0^1\max\{y_1f_1(x),y_2f_2(x)\} \, \text{d}x \right\}
%\end{equation*}
%where $(y_1,y_2)$ are again defined by the transformations \eqref{mtrans}, and where $f_1$ and $f_2$ are density functions with support [0,1].
%In particular, if we take the beta densities $f_1(x)=(1-\alpha)x^{-\alpha}$ and $f_2(x)=(1-\beta)(1-x)^{-\beta}$ we obtain
\section*{Appendix A: Additional Bivariate Parametric Models}
It can be shown, using a representation of \citet{haan84}, that
\begin{equation*}
G(z_1,z_2) = \exp\left\{ - \int_0^1\max\{y_1(1-\alpha)x^{-\alpha},y_2(1-\beta)(1-x)^{-\beta}\} \, \text{d}x \right\}, \quad \alpha,\beta < 1.
\end{equation*}
is a bivariate extreme value distribution function. If we further constrain the parameters to be non-negative we obtain the bivariate bilogistic model proposed by \citet{smit90}, which can also be expressed as
\begin{equation*}
G(z_1,z_2) = \exp\left\{ - y_1\gamma^{1-\alpha} - y_2(1-\gamma)^{1-\beta} \right\}, \quad 0 < \alpha,\beta <1,
\end{equation*}
where $\gamma=\gamma(y_1,y_2;\alpha,\beta)$ solves $(1-\alpha)y_1(1-\gamma)^\beta=(1-\beta)y_2\gamma^\alpha$.
The logistic model is obtained when $\alpha=\beta$.
Independence is obtained as $\alpha = \beta \rightarrow1$, and when one of $\alpha,\beta$ is fixed and the other approaches one.
Different limits occur when one of $\alpha,\beta$ is fixed and the other approaches zero.
Alternatively, if we constrain both parameters to be non-positive and set $\alpha_0=-\alpha > 0$ and $\beta_0=-\beta > 0$ we obtain the negative bilogistic model \citep{coletawn94}, which has the representation
\begin{equation*}
G(z_1,z_2) = \exp\left\{-y_1-y_2+y_1\gamma^{1+\alpha_0}+y_2(1-\gamma)^{1+\beta_0} \right\}, \quad \alpha_0,\beta_0 > 0,
\end{equation*}
where $\gamma=\gamma(y_1,y_2;-\alpha_0,-\beta_0)$.
The negative logistic model is obtained when $\alpha_0=\beta_0$ (with $r = 1/\alpha_0 = 1/\beta_0$).
Independence is obtained as $\alpha_0 = \beta_0 \rightarrow\infty$, and when one of $\alpha_0,\beta_0$ is fixed and the other tends to $\infty$.
Different limits occur when one of $\alpha_0,\beta_0$ is fixed and the other approaches zero.
The distribution function of the Coles-Tawn model\footnote{\citet{coletawn91} call this the Dirichelet model.}
\citep{coletawn91} is given by
\begin{equation*}
G(z_1,z_2) = \exp\left\{-y_1[1-\text{Be}(u;\alpha+1,\beta)] - y_2\,\text{Be}(u;\alpha,\beta+1) \right\}, \quad \alpha,\beta > 0,
\end{equation*}
where $u=\alpha y_2/(\alpha y_2+\beta y_1)$ and Be is the incomplete beta function, given by
\begin{equation*}
\text{Be}(u;\alpha,\beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\Gamma(\beta)} \int_0^u x^{\alpha-1}(1-x)^{\beta-1} \, \text{d}x.
\end{equation*}
Complete dependence is obtained in the limit as $\alpha = \beta \rightarrow\infty$.
Independence is obtained as $\alpha = \beta \rightarrow0$ and when one of $\alpha,\beta$ is fixed and the other approaches zero.
Different limits occur when one of $\alpha,\beta$ is fixed and the other tends to $\infty$.
The asymmetric mixed model \citep{tawn88} is typically defined using the corresponding dependence function \eqref{bvdepfn}, which is modelled as a cubic polynomial. Specifically, for $0 \leq t \leq 1$ the dependence function of the asymmetric mixed model is
\begin{equation*}
A(t) = 1 - (\alpha + \beta)t + \alpha t^2 + \beta t^3,
\end{equation*}
where both $\alpha$ and $\alpha + 3\beta$ are non-negative, and where both $\alpha + \beta$ and $\alpha + 2\beta$ are less than or equal to one. These constraints imply that $\beta \in [-0.5,0.5]$ and $\alpha \in [0,1.5]$, though $\alpha$ can only be greater than one if $\beta$ is negative. The (symmetric) mixed model is obtained when $\beta = 0$. Complete dependence cannot be obtained. Independence is obtained when $\alpha = \beta = 0$.
The asymmetric mixed model is often referred to in the literature because the dependence function has a simple form, and because the $\beta = 0$ case is historically important. However it cannot capture strong dependence, and hence it is of limited use as a statistical model. The extension to an $m$-degree polynomial can be made, but this is of no statistical interest because the additional parameters add little additional flexibility.
\bibliography{bibliog}
\end{document}
\section*{Appendix A: Simulation Study}
In this Appendix we use the tools in the package to perform a simulation study to examine the small sample properties of non-parametric estimators for the dependence function $A(\cdot)$ of the bivariate extreme value distribution.
The estimators referred to in this Appendix are defined in the documentation file for the function \verb+abvnonpar+.
Simulation studies of this form \citep[e.g.][]{halltajv00} typically use the known marginal parameters $(\mu_1,\sigma_1,\xi_1,\mu_2,\sigma_2,\xi_2)$ within the transformations \eqref{transtoexp}.
In practice, these parameters need to be estimated.
In this study we seek to replicate the behaviour of the estimators when applied to real data, and we have therefore estimated the marginal parameters by maximum likelihood.
Figure \ref{simfig} depicts the behaviour of the estimators of \citet{capefoug97}, \citet{pick81} and \citet{tiag97}, which we subsequently denote by $A_c$, $A_p$ and $A_t$ respectively. The estimators of \citet{dehe91} and \citet{halltajv00} are not considered, as they produce plots that are indistinguishable from those of $A_p$.
The first, second and third columns of the figure employ simulations from (symmetric) logistic distributions, with $\alpha$ equal to $0.5$, $0.75$ and $1$ respectively.
Standard Gumbel marginal distributions were used in each case.
The figure shows that the estimator $A_t$ is abysmal when estimating dependence functions with very strong ($\alpha = 0.5$) or very weak ($\alpha = 1$) levels of dependence.
The estimators $A_c$ and $A_p$ give more consistent performances across different levels of dependence.
The estimator $A_c$ appears to outperform $A_p$, as the estimates of the former appear to cluster more tightly around the true dependence function for each $\alpha = 0.5,0.75,1$.
The plots can easily be generated, using e.g.
\begin{verbatim}
> dep <- 0.5 ; method <- "cfg"
> abvevd(dep = dep, plot = TRUE, lty = 0)
> set.seed(44)
> for(i in 1:50) {
sdt <- rbvevd(100, dep = dep)
abvnonpar(data = sdt, add = TRUE, method = method, col = "grey")
}
> abvevd(dep = dep, add = TRUE, lwd = 3)
\end{verbatim}
\begin{figure}
\begin{center}
\scalebox{0.18}{\includegraphics{npsim11.ps}}
\vspace{-1.5cm}
\hspace{0cm}
\scalebox{0.18}{\includegraphics{npsim12.ps}}
\hspace{0cm}
\scalebox{0.18}{\includegraphics{npsim13.ps}}
\\
\scalebox{0.18}{\includegraphics{npsim21.ps}}
\vspace{-1.5cm}
\hspace{0cm}
\scalebox{0.18}{\includegraphics{npsim22.ps}}
\hspace{0cm}
\scalebox{0.18}{\includegraphics{npsim23.ps}}
\\
\scalebox{0.18}{\includegraphics{npsim31.ps}}
\vspace{-1.5cm}
\hspace{0cm}
\scalebox{0.18}{\includegraphics{npsim32.ps}}
\hspace{0cm}
\scalebox{0.18}{\includegraphics{npsim33.ps}}
\end{center}
\caption{Simulated non-parametric dependence function estimates. The grey lines represent estimates derived using the estimators $A_c$ (top row), $A_p$ (middle row) and $A_t$ (bottom row). The thick black lines represent the true dependence functions, which are (symmetric) logistic models with dependence parameters $0.5$ (first column), $0.75$ (second column) and $1$ (third column).}
\label{simfig}
\end{figure}
which generates the plot in the top left corner.
Only the first line of code needs to be changed in order to produce the remaining plots.
The second line of code establishes the plotting region.
The simulation is performed in the \verb+for+ loop, and the last line adds the true dependence function to the plot.
The \verb+set.seed+ function sets the seed of the random generator, which ensures that the simulated data sets used for each plot are comparable.
Let $A_n(\cdot)$ be any estimator of $A(\cdot)$.
Table \ref{simtab} gives median integrated absolute errors for various non-parametric dependence function estimators.
The table was constructed as follows.
For $\alpha = 0.5,0.75,1$ we simulated $1000$ datasets containing $n=25,100$ bivariate observations, using standard Gumbel margins.
Then for each of the $1000$ datasets we estimated the integrated absolute error $\int_0^1|A_n(x) - A(x)| \, \text{d}x$.
The table contains the median of the $1000$ values, for each value of $\alpha$ and $n$.
We have extended the number of estimators to include the convex minorants of $A_c$ and $A_p$, which we denote by $A_c^*$ and $A_p^*$.
The convex minorant of $A_t$ is identical to $A_t$, because $A_t$ is always convex.
The table again shows the poor performance of $A_t$ when $\alpha = 0.5$, and particularly when $\alpha = 1$.
$A_t$ is the best estimator when $\alpha = 0.75$, which is not surprising given that the estimator only yields adequate estimates at mid-range levels of dependence.
The estimator $A_c$ outperforms $A_p$, confirming the impression given by Figure \ref{simfig}.
Taking the convex minorant of $A_c$ or $A_p$ leads to an improvement for $\alpha = 0.5$ and $\alpha = 0.75$, but a considerable worsening for $\alpha = 1$.
This worsening is expected, since taking the convex minorant always leads to estimates of stronger dependence.
The values in the table can be generated using e.g.
\begin{verbatim}
> dep <- 0.5 ; n <- 25 ; method <- "cfg" ; cv <- FALSE
> nn <- 100 ; x <- (1:nn)/(nn + 1)
> a <- abvevd(x, dep = dep)
> iae <- numeric(1000)
> set.seed(44)
> for(i in 1:1000) {
sdt <- rbvevd(n, dep = dep)
anp <- abvnonpar(x, data = sdt, method = method, convex = cv)
iae[i] <- sum(abs(a - anp))/nn
}
> round(10^4 * median(iae))
\end{verbatim}
% FOR ENTIRE TABLE
%\begin{verbatim}
%method <- rep(c("cfg","cfg","pick","pick","tdo"), 6)
%cv <- rep(c(FALSE, TRUE, FALSE, TRUE, FALSE), 6)
%dep <- rep(rep(c(0.5, 0.75, 1), each = 5), 2)
%n <- rep(c(25, 100), each = 15)
%sim.all <- numeric(30)
%
%for(j in 1:30) {
% print(j)
% nn <- 100 ; x <- (1:nn)/(nn+1)
% a <- abvevd(x, dep = dep[j])
% iae <- numeric(1000)
% set.seed(44)
% for(i in 1:1000) {
% sdt <- rbvevd(n[j], dep = dep[j])
% anp <- abvnonpar(x, data = sdt, method = method[j], convex = cv[j])
% iae[i] <- sum(abs(a - anp))/nn
% }
% sim.all[j] <- median(iae)
%}
%round(10^4 * matrix(sim.all, nrow = 5, ncol = 6))
%\end{verbatim}
\begin{table}
\begin{center}
\begin{tabular}{|l|ccc|ccc|} \hline
& \multicolumn{3}{c|}{$n=25$} & \multicolumn{3}{c|}{$n=100$} \\
& $\alpha = 0.5$ & $\alpha = 0.75$ & $\alpha = 1$ & $\alpha = 0.5$ & $\alpha = 0.75$ & $\alpha = 1$ \\ \hline
$A_c$ & 210 & 415 & 110 & 104 & 198 & 62 \\
$A_c^*$ & 205 & 363 & 340 & 103 & 194 & 168 \\
$A_p$ & 243 & 469 & 211 & 134 & 242 & 113 \\
$A_p^*$ & 218 & 357 & 554 & 126 & 215 & 285 \\
$A_t$ & 393 & 189 & 983 & 334 & 155 & 830 \\ \hline
\end{tabular}
\caption{Median integrated absolute errors $\times$ $10^4$ for non-parametric estimates of the dependence function of the bivariate extreme value distribution, using datasets containing $n=25,100$ bivariate observations, simulated from the (symmetric) logistic model with dependence parameter $\alpha=0.5,0.75,1$. The estimators $A_c^*$ and $A_p^*$ are the convex minorants of $A_c$ and $A_p$ respectively.}
\label{simtab}
\end{center}
\end{table}
which generates the value in the top left corner.
Only the first line of code needs to be changed in order to produce the remaining values.
The integrated absolute error is estimated by evaluating the absolute difference between true dependence function and the non-parametric estimate at $\verb+nn+ = 100$ equally spaced points in the interval $[0,1]$.
The function \verb+numeric+ merely initializes the object \verb+iae+ to be a vector of $1000$ zeros.
|