ESTESTADISTICAADÍSTICA volumen 53 volumen 64
Junio y Diciembre 2012 número 182 y 183
REVISTREVISTAA DEL SEMESTRAL INSTITUTO DEL INSTITUTO INTERAMERICANOINTERAMERICANO DE DE EST ESTADÍSTICAADÍSTICA
BIANNUAL JOURNALJOURNAL OF THTHEE INTER-AMERICANINTER-AMERICAN ST ASTATISTICALTISTICAL INSTITUTE INSTITUTE
EDITORA EN JEFE / EDITOR IN CHIEF
CLYDE CHARRE DE TRABUCHI
Consultora/Consultant French 2740, 5º A 1425 Buenos Aires, Argentina
Tel (54-11) 4824-2315 e-mail [email protected] e-mail [email protected]
EDITORA EJECUTIVA / EXECUTIVE EDITOR
VERONICA BERITICH
Instituto Nacional de Estadística y Censos (INDEC) Ramallo 2846 1429 Buenos Aires, Argentina
Tel (54-11) 4703-0585 e-mail [email protected]
EDITORES ASOCIADOS / ASSOCIATE EDITORS
D. ANDRADE G. ESLAVA P. MORETTIN Univ. Fed. Sta. Catalina, BRASIL UNAM, MEXICO Univ. de Sao Pablo, BRASIL M. BLACONA A. GONZALEZ VILLALOBOS F. NIETO Univ. de Rosario, ARGENTINA Consultor/Consultant, ARGENTINA Univ. de Colombia, COLOMBIA J. CERVERA FERRI V. GUERRERO GUZMAN J. RYTEN Consultor/Consultant, ESPAÑA ITAM, MEXICO Consultor/Consultant, CANADA E. DAGUM R. MARONNA * S. SPECOGNA Consultor/Consultant, CANADA Univ. de La Plata, ARGENTINA ANSES, ARGENTINA E. de ALBA I. MENDEZ RAMIREZ P. VERDE INEGI, MEXICO IIMAS/UNAM, MEXICO University of Düsseldorf, ALEMANIA P. do NASCIMENTO SILVA M. MENDOZA RAMIREZ V. YOHAI ENCE/IBGE, BRASIL ITAM, MEXICO Univ. de Buenos Aires, ARGENTINA L. ESCOBAR R. MENTZ Louisiana State Univ., USA Univ. de Tucumán, ARGENTINA
* Asistente de las Editoras / Editors’ Assistant
ESTADÍSTICA (2012), 64, 182 y 183, pp. 5-22 © Instituto Interamericano de Estadística
JAMES DURBIN: IN MEMORIAM
JUAN CARLOS ABRIL Universidad Nacional de Tucumán, Facultad de Ciencias Económicas y Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET). Argentina [email protected]
RESUMEN
El Profesor James Durbin ha fallecido en la tarde del sábado 23 de Junio de 2012, en Londres, a la edad de 88 años. Fue una de las figuras más importantes de la Estadística. Sus contribuciones cubren áreas de muestreo, teoría de las distribuciones, estadística no paramétrica, procesos estocásticos y, principalmente, series de tiempo y econometría. Por su trascendencia científica, sus aportes a la Estadística y por la gran amistad de más de 38 años que me unía a él, presento este homenaje en donde se resaltan la trayectoria y la personalidad del Profesor Durbin, y sus importantes contribuciones a la ciencia.
Palabras Clave
Durbin; Econometría; Estadística; Series de Tiempo.
ABSTRACT
Professor James Durbin has died in the afternoon of Saturday June 23, 2012, in London, at the age of 88. It was one of the most important figures of Statistics. His contributions cover areas of sampling, theory of distributions, nonparametric statistics, stochastic processes, and mainly time series and econometrics. For his scientific significance, his contributions to Statistics and the great friendship of over 38 years that I had with him, I present this tribute which highlights the career and personality of Professor Durbin, as well as his important contributions to science.
6 ESTADÍSTICA (2012), 64, 182 y 183, pp. 5-22
Key words
Durbin; Econometrics; Statistics; Time series.
1. Introducción
Jim, como lo llamábamos los amigos, nació el 30 de Junio de 1923 en Wigan, Inglaterra. Fue educado en el Saint John’s College de la Universidad de Cambridge. Desde 1950, trabajó en The London School of Economics and Political Science (LSE) hasta su jubilación en 1988. Se inició allí como Ayudante de cátedra, se convirtió en Profesor Asociado en el año 1953 y en Profesor en 1961 sucediendo a Sir Maurice Kendall en la cátedra de Estadística. Después de su jubilación permaneció como Profesor Emérito de Estadística en la LSE.
Desde 2007 fue Profesor Honorario del University College de Londres (UCL), y donde también fue “Fellow” del Centre for Microdata Methods and Practice (CeMMAP). En 2001 fue nombrado “Fellow” de la British Academy. En 2008 por los logros de toda una vida en Estadística recibió la Medalla Guy de Oro de la Royal Statistical Society (RSS) después de haber recibido la de Bronce en 1966 y de Plata 1976. Fue Presidente del Instituto Internacional de Estadística (ISI) (1983- 1985), Miembro Honorario desde 1999 y Presidente de la Royal Statistical Society (1986-1987). Además, fue elegido “Fellow” del Instituto de Estadística Matemática (IMS) desde 1958, de la American Statistical Association (ASA) desde 1960 y de la Sociedad Econométrica desde 1967. También fue tesorero de la Sociedad Bernoulli para la Estadística Matemática y Probabilidad en los años 1975-1981. Sus deberes editoriales incluyen su papel como Editor Asociado de Biometrika (1964-1967), Annals of Statistics (1973-1975) y el Journal of the Royal Statistical Society (JRSS), Series B (1978-1981). En la comunidad internacional de estadísticos, J. Durbin fue un distinguido y laureado científico.
ABRIL: James Durbin: in memoriam 7
Figura 1: James Durbin. Foto tomada durante 1988 cuando era Presidente de la Royal Statistical Society
Durante sus años en la LSE, también fue activo como miembro del Consejo (1960- 1963) en el Institute of Statistics (equivalente a lo que conocemos como colegio de graduados) antes de que se fusionara con la Royal Statistical Society. Además de ser Presidente de la RSS en 1986-1987, fue miembro del Consejo durante 15 años en el período 1957-1989 y se le dio el rol de Vicepresidente durante una serie de años. También fue miembro durante muchos años de los comités de Exámenes y de Investigación de la RSS. Fue “Miembro” del Instituto Internacional de Estadística (ISI) desde 1955 y, durante los años comprendidos entre 1981 y 1987, fue sucesivamente Presidente Electo, Presidente y miembro del Consejo del ISI. También fue Presidente de comités del ISI durante varios años.
Sus trabajos de investigación han constituido importantes contribuciones en estadística y econometría, en particular, en los campos de la correlación serial (13 publicaciones), las series de tiempo generales (31), la econometría (4), la metodología de encuesta y muestreo (9), las pruebas de bondad de ajuste y las funciones de distribución muestral (13), la probabilidad (8), la teoría general de estadística (8) y la filosofía de las estadísticas (3). Sus publicaciones en revistas como Biometrika (14 publicaciones), Journal of Royal Statistical Society (8 en la Series A, 7 en la Series B), Journal of Applied Probability (4), Econometrica (3), Journal of the American Statistical Association (2), Annals of Mathematical Statistics (2) y Annals of Statistics (1). 8 ESTADÍSTICA (2012), 64, 182 y 183, pp. 5-22
James Durbin se casó con Anne en 1957 y tuvieron tres hijos, Joanna, Richard y Andrew. En 1950, Jim decidió convertirse en académico porque las largas vacaciones le permitirían practicar su deporte favorito, el alpinismo. Dado que Anne no estaba tan interesada en el montañismo, decidieron cambiar por el esquí como deporte para sus vacaciones en familia. Durante su vida realizó numerosas e importantes excursiones de montañismo a diversos lugares del mundo junto con el Club de Montañismo de la LSE. A la edad de 60 años subió el Kilimanjaro. Su preparación para escalar la montaña más alta de África, con más de 5.891 metros de altura, duró tres meses y consistió principalmente en trasladarse en bicicleta entre su casa en la zona londinense de Hampstead y su trabajo en el centro de Londres.
Conocí al Profesor James Durbin en Agosto de 1973 cuando visitó el Instituto de Investigaciones Estadísticas (INIE) de la Facultad de Ciencias Económicas de la Universidad Nacional de Tucumán, como parte de un programa de intercambio organizado entre el Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET) de Argentina y The Royal Society de Gran Bretaña. En esa oportunidad ofreció un curso sobre la teoría de distribución de tests basados en la función de distribución muestral que correspondía a los temas de su último libro publicado. Debo confesar que, por aquella época, el curso me resultó muy dificultoso pero altamente estimulante. Durante esa visita, me ofreció su apoyo para lograr una beca del British Council para estudiar una maestría en la LSE. Allí Durbin fue mi director de estudios y de tesis.
Luego regresé a Argentina, pero sin que se interrumpiera nuestra correspondencia. Así, en 1980 definimos un área de trabajo para mi doctorado que se inició ese mismo año. Para el primer trimestre de 1982, sucedió lo impensable: nuestros países entraron en guerra. Efectivamente, me tocó estar en Londres durante la Guerra de las Malvinas. Más allá de las posibles diferencias que pudiéramos haber tenido con respecto a la esencia del conflicto, el Profesor Durbin supo obrar conmigo como una gran persona y un perfecto caballero, ofreciendo todo el apoyo moral y económico que mi familia y yo pudiéramos necesitar.
Debido a ese conflicto regresé a la Argentina pero mi relación con Jim se mantuvo y gracias a su apoyo logré un lugar en la Universidad de Valencia, España. Desde allí, luego de reiterados viajes y, con un apoyo económico de la LSE obtenido nuevamente por sus recomendaciones, pude terminar mi doctorado en 1985.
El 16 de Mayo de 2001, la Universidad Nacional de Tucumán, Argentina, le otorgó el título de Doctor Honoris Causa.
ABRIL: James Durbin: in memoriam 9
Tuve la oportunidad de volver a la LSE en 1991, 1997, 2000 y por último en 2011, lo que me permitió mantener un fluido contacto académico, científico y de gran amistad con él.
Recibí la triste noticia de su fallecimiento el 24 de junio por medio de sendos correos electrónicos de parte de Neil Shephard y Siem Jan Koopman. Inmediatamente llamé por teléfono a la familia y pude hablar con su hija Joanna, expresándole mis condolencias y las de toda mi familia.
Escribir sobre la trayectoria académica del Profesor Durbin, es una tarea que considero nada fácil porque debo condensar en poco espacio lo que realizó en una larga trayectoria de sobresaliente actividad científica. Una versión extendida de este tributo puede ser encontrada en Abril (2012).
Sirva este trabajo como un justo homenaje al amigo, gran maestro y excelente científico.
2. James Durbin. Su vida y su obra
Su formación de grado se inició al ingresar a la Universidad de Cambridge, en el Saint John’s College, en los difíciles años de la Segunda Guerra Mundial. Obtuvo el título de grado de “Bachiller de las artes en tiempos de guerra” en matemáticas que incluía el servicio militar en el Grupo de Investigaciones de Operaciones del Ejercito. En esa universidad estaban Denis Sargan y Sir David Cox entre sus contemporáneos.
Al finalizar la guerra decidió continuar con sus estudios de postgrado en Cambridge. Por sugerencia de su tutor cursó la maestría en Estadística Matemática. Durante ese tiempo tuvo como supervisores a Richard Stone en Economía, Premio Nobel de Economía en 1984, y a Henry Daniels en Estadística. Durante esos años, Denis Lindley era docente auxiliar en el programa de postgrado, Wishart era el jefe del grupo de Estadística y Frank Anscombe era docente. Bartlett se había marchado a Manchester justo antes de que Jim ingresara.
Cuando James Durbin era estudiante de grado realizó algunos trabajos sobre estadística descriptiva. En aquella época nunca pensó que podría desarrollar un interés en la estadística matemática.
En 1948 se unió al nuevo Departamento de Economía Aplicada (DAE). Este fue fundado por Keynes quien insistía que en Cambridge debía haber un departamento de Economía Cuantitativa. En aquella época se realizó mucha investigación en 10 ESTADÍSTICA (2012), 64, 182 y 183, pp. 5-22
series de tiempo en el DAE con investigadores como Stone, Cochrane, Orcutt e investigadores visitantes que incluían a Hendrik Houthakker, Gerhard Tintner, Larry Klein (Premio Nobel de Economía de 1980), Michael Farrell, Theodore W. Anderson y Geoffrey Watson.
Cuando estaba finalizando la década de 1940, focalizó su atención en testar la correlación serial, influenciado por los trabajos que Stone, Cochrane y Orcutt estaban realizando. En este contexto conoció a Geoffrey Watson, joven australiano que llegó a Cambridge a realizar su doctorado, quien tenía similares inquietudes científicas. Tuvieron algunos intercambios de ideas preliminares y decidieron trabajar juntos en ese problema. De esta unión surgieron trabajos que aún siguen estando vigentes y que son ampliamente usados en las aplicaciones prácticas (Durbin and Watson, 1950, 1951). De hecho, ellos trabajaron juntos en el mismo departamento cerca de seis meses. Luego, Jim regresó a la LSE y se reunían esporádicamente en Cambridge o en Londres. Así fue como escribieron el primer trabajo sobre el test de correlación serial. En el segundo trabajo conjunto publicaron tablas que pueden ser usadas en el trabajo empírico. Veinte años después escribieron un tercer trabajo basándose en la teoría de la invarianza (Durbin and Watson, 1971). A pesar de haber previsto una cuarta, nunca pudo realizarse. Retrospectivamente, podemos ver que el test de Durbin-Watson tuvo un impacto extraordinario en la profesión, especialmente en el trabajo econométrico aplicado. Recientemente se lo revalorizó en su carácter de test de diagnóstico exacto y también parece tener una potencia difícil de mejorar, aún para otras hipótesis alternativas diferentes a la originalmente propuesta.
El Profesor James Durbin se unió en 1950 a una nueva unidad en la LSE, la de investigación estadística. El Profesor Maurice Kendall acababa de ser nombrado Profesor de Estadística de la LSE y quedó vacante su cargo anterior como docente. Kendall llamó a Daniels preguntándole si es que había alguien apto para ese trabajo y Daniels recomendó a Durbin. Jim lo aceptó porque además de incorporarse a esa prestigiosa institución, el trabajo sería temporario y eso le permitiría tomarse unas largas vacaciones para practicar montañismo, su deporte favorito. Aunque resulte difícil de creer, en aquel momento este deporte le interesaba más que la investigación académica.
En los años iniciales en la LSE, Durbin se interesó en diversas áreas. Sabía que los temas referidos a las encuestas por muestreo iban a aumentar su importancia en las aplicaciones de la estadística en las ciencias sociales y además se interesó en este tema por la influencia de Stone. Entonces una de las primeras actividades docentes en la LSE consistió en dictar un curso sobre la teoría de las encuestas por muestreo. Continuó con el curso de encuestas por muestreo por un cierto tiempo, ABRIL: James Durbin: in memoriam 11
posteriormente compartió el dictado de esta materia con Alan Stuart y luego se fueron turnando cada cierto número de años. Por ejemplo, su artículo en el Journal of the Royal Statistical Society (JRSS), Series B, de 1953, desarrolla una forma general para la estimación de las varianzas muestrales en muestreo múltiple con probabilidades desiguales (Durbin, 1953). Después trató de cubrir otras áreas como el análisis de la varianza y modelos lineales. Así como, en el DAE había sentido una dedicación especial hacia la economía y la econometría, en la LSE sintió una responsabilidad hacia la teoría estadística en general.
En esos primeros años en la LSE no enseñó series de tiempo porque Maurice Kendall era una autoridad internacional en ese campo. En esa época, Alan Stuart era su compañero más cercano y, dado que la carga docente en aquellos días era baja, pudieron realizar bastantes trabajos conjuntos en estudios experimentales y en correlaciones por rangos que fueron publicados en el Journal of the Royal Statistical Society (Durbin y Stuart, 1951a, 1951b, 1954).
Continuando con la reseña de la labor científica del Profesor Durbin es importante destacar que en 1953 fue invitado a dar una charla en el encuentro europeo de la Econometric Society en Innsbruck. La misma trató sobre errores en las variables y el uso de variables instrumentales en la estimación. Allí se encontraba Gil Goodswaard, que en ese tiempo era el editor de International Statistical Review, quien lo invitó a publicar su conferencia allí. Por lo tanto, parece que fue más bien accidental la publicación de ese trabajo. Jim refinó un poco el documento de su charla, le incorporó algunos agregados pero esencialmente era la conferencia (Durbin, 1954). Ese trabajo resultó muy importante porque contenía un test que, injustamente, ahora se conoce con el nombre del test de Hausman para la exogeneidad (Hausman, 1978). Algunos autores comenzaron recientemente a nombrarlo como el test de Durbin-Wu-Hausman, entendiendo que así se hace justicia para con su autor inicial.
Como ya se dijo anteriormente, en los años iniciales en la LSE, Durbin no enseñaba series de tiempo ya que lo hacía Kendall. Cuando este último quiso cambiar de área de trabajo, dejó el curso en manos de Jim y de Maurice Quenouille, quien a pesar de tener un cargo de tiempo completo como investigador, cooperó con él durante un tiempo en el dictado de este curso. Luego David Brillinger ingresó al departamento por cinco años y compartieron la enseñanza de series de tiempo.
El período de Quenouille en la LSE parece haber sido bastante productivo ya que escribió un libro sobre series de tiempo múltiples y desarrolló la teoría del jackknife. James Durbin se había interesado a finales de la década de 1950 en este 12 ESTADÍSTICA (2012), 64, 182 y 183, pp. 5-22
tema y escribió un trabajo acerca del jackknife que apareció en Biometrika en base a en un artículo anterior de Quenouille (Durbin, 1959b). Los resultados fueron muy interesantes ya que él junto con Quenouille, demostraron que mediante el método de jackknife no hace falta pagar un precio demasiado alto para reducir el sesgo. Intuitivamente esto no parecía un resultado tan obvio.
Después del trabajo de Biometrika sobre jackknife, Durbin desarrolló una metodología similar al bootstrapping pero, a fines de la década de 1950, el cómputo era un grave limitante. Según contó, había problemas interesantes en el diseño de las simulaciones sobre los que podría hacer progresos, pero nunca publicó el trabajo porque pensó que los cómputos no eran en realidad prácticos para el trabajo aplicado. Realmente es una pena que este último trabajo no haya sido publicado, aunque hoy en día la gente está usando las simulaciones para hacer cosas similares aprovechando que las computadoras son accesibles y el costo de grandes trabajos de computación es muy bajo. Durbin pensaba que muchas personas jóvenes no se dan cuenta de la enorme influencia que ha tenido la computadora en el trabajo estadístico tanto teórico como aplicado.
Más tarde, a finales de la década de 1950, regresó a los problemas de series de tiempo. En 1957, apareció su artículo de Biometrika sobre la prueba de Durbin- Watson para un sistema de ecuaciones simultáneas (Durbin, 1957). Se trata de un documento histórico ya que resuelve de manera inteligente un problema difícil y, por sobre todo, muestra la capacidad intelectual del Profesor Durbin. Debido a su fama y a que se puso de moda en la década de 1960, se hizo un uso incorrecto del test de Durbin-Watson, por ejemplo, en las regresiones basadas en modelos dinámicos con variables dependientes rezagadas. Para corregir esta situación, en Durbin (1970a), desarrolló el estadístico h como prueba de correlación serial con variables dependientes rezagadas usadas como regresores. El principio general de esta prueba fue reconocido más tarde como un procedimiento de multiplicadores de Lagrange. En ese mismo año desarrolló otro test alternativo de correlación serial (Durbin, 1970b).
En 1963 presentó un importante trabajo de estimación econométrica en el encuentro de Copenhague de la Econometric Society. Ese trabajo daba las ecuaciones de estimación para los estimadores “Full Information Maximum Likelihood” (FIML) en una nueva forma que facilitaba los lazos con otros procedimientos de estimación tales como variables instrumentales. Muchos econometristas se enteraron de este trabajo por comentarios de otros colegas. Desde entonces estos resultados fueron y son enseñados en todos los cursos de econometría a pesar que el artículo correspondiente fue publicado recién 25 años más tarde (Durbin, 1988). En la versión que presentó originalmente en Copenhague ABRIL: James Durbin: in memoriam 13
había algunos cálculos que resultaron ser incorrectos. Un asistente resolvió correctamente algunos de estos cálculos pero, antes de finalizar completamente la tarea, este asistente cambió de lugar de trabajo dejando el trabajo inconcluso. Jim tenía la idea de que no se debería publicar un nuevo resultado metodológico sin mostrar cómo se puede usar. Siempre mantuvo la costumbre de repetir los cálculos, pero las condiciones de uso de las computadoras en esa época eran tan complejas, que nunca pudo volver a hacerlos. Eventualmente, comenzó a trabajar en otros temas y dejó de lado para siempre esa área. Realmente ésta es una historia increíble: el trabajo anterior ya era usado como referencia en el texto de Edmond Malinvaud (1964). Así que, sin ninguna duda, poco después que Durbin lo presentó en Copenhague ya había mucha gente que estaba familiarizada con él. Se sospecha que David Hendry, hoy Sir David Forbes Hendry, lo haya usado ya que estudió en la LSE en esa época. Es posible que Hendry y Malinvaud hicieron que el trabajo fuera conocido.
En Bartlett (1955) se muestra cómo se puede usar la teoría de distribución de Kolmogorov-Smirnov para el periodograma acumulado como un test general de correlación serial en el caso de no estar ante un problema de regresión. Durbin se interesó inmediatamente en este procedimiento ya que permitía conocer las características de la correlación serial de la serie, especialmente mediante la representación gráfica del periodograma acumulado, lo cual a su criterio podría resultar atractivo para trabajadores aplicados. Posteriormente, durante la década del sesenta, retomó el tema de la correlación serial y escribió diversos trabajos desarrollando la teoría de los tests de periodogramas acumulados, el test h, el test t y otros.
Otro de sus intereses más importantes a lo largo de los años ha sido el relacionado con los procedimientos de ajuste estacional (Durbin, 1962, 1963) que se reavivó a partir de un llamado del gobierno nacional, a fines de los años 60, como consultor académico para estudiar el ajuste estacional de las series de desempleo. Trabajó en esos problemas por un año o dos con importantes estadísticos oficiales (Durbin, Brown y Cowley, 1970, 1971). El Primer Ministro Harold Wilson había trabajado como un estadístico económico en el servicio gubernamental durante la guerra y era bastante bueno en la interpretación de datos numéricos. El gobierno estaba crecientemente preocupado por el aumento del desempleo y Wilson estaba muy interesado en mirar los datos por sí mismo. Tuvo la idea, como Primer Ministro, que tal vez la razón por la que las series de desempleo parecían comportarse de una manera algo extraña se debía al procedimiento de ajuste estacional que estaba siendo usado. Llegaron a la conclusión que el Primer Ministro estaba en lo cierto y que había algo malo en el ajuste estacional. Es notable que un primer ministro se 14 ESTADÍSTICA (2012), 64, 182 y 183, pp. 5-22
haya fijado en un problema que es eminentemente técnico y que además haya estado en lo cierto.
En 1973, su trabajo sobre convergencia débil de una función de distribución empírica (Durbin, 1973a) apareció en los Annals of Statistics y se publicó su libro sobre la teoría de distribución de los tests basados en las funciones de distribución empíricas (Durbin, 1973b). Con estas contribuciones Durbin fue uno de los precursores en desarrollar la teoría de convergencia débil de procesos estocásticos para resolver los problemas asintóticos de convergencia.
Su trabajo sobre la distribución de estadísticos suficientes de Biometrika (Durbin, 1980a, 1980b), otorgó un nuevo enfoque a las aproximaciones de Edgeworth y de puntos de ensilladura, y a la teoría asintótica de alto orden (Ghosh, 1994; Taniguchi, 1991). Cuando comenzó a escribir el trabajo, necesitaba un teorema para expansiones de Edgeworth de variables dependientes y, para su sorpresa, encontró que no existía ese teorema. Consultando la literatura sobre probabilidad, le pareció que un argumento muy simple podía ser desarrollado usando el tratamiento Feller (1971) utilizaba para expansiones ordinarias de Edgeworth para variables aleatorias independientes. A Jim le pareció que se lo podía extender para variables dependientes. Así que desarrolló un teorema que lo demostraba (Durbin, 1980a).
En su trabajo, “Evolutionary Origins of Statisticians and Statistics” (Durbin, 1985) desarrolla una tesis fascinante acerca de la capacidad de la especie humana de hacer matemática y de la aplicabilidad de la teoría estadística en el mundo real. Por más de treinta años se interesó en analizar las razones por las cuales la especie humana puede hacer matemática así como por qué la matemática funciona tan bien cuando se la aplica en el mundo real. Jim consideraba sorprendente que los matemáticos no tuvieran un gran interés en esas cuestiones. Recibió una invitación para escribir ese trabajo en el volumen del centenario del International Statistical Institute (ISI) para el cual los editores le pidieron específicamente que no escribiera algo técnico en estadística porque ya tenían demasiados artículos técnicos. Antes de publicar el trabajo revisó en la biblioteca de la LSE todo lo referente a filosofía de la matemática. La LSE ha sido un centro importante para la filosofía de las ciencias y la matemática debido al trabajo de Lakatos y Popper aunque allí no se mencionaba la palabra evolución. Esto era extraordinario. Le parecía evidente a Jim que si se desea comprender los cimientos filosóficos de la matemática se debe empezar por los orígenes evolutivos del razonamiento humano.
Para Durbin la LSE siempre había sido un lugar de trabajo interesante y gratificante, ya que desde principios de 1950 pudo sostener un desarrollo continuo ABRIL: James Durbin: in memoriam 15
en el campo de las series de tiempo y la econometría. Actualmente podríamos decir que es más fuerte que antes puesto que hay un grupo muy importante de especialistas en esas áreas. Algunos profesores de econometría en la LSE, como Peter Robinson y Andrew Harvey, fueron antiguos estudiantes de esa universidad y, de hecho, puede ser tal vez sorprendente para algunos que ambos sean graduados en estadística y no en econometría. Lamentablemente para la LSE, Harvey se fue recientemente a Cambridge.
J. Durbin creía que el futuro de todas las ciencias sociales cuantitativas sería más interesante que el pasado por el mayor poder computacional que se tiende a tener. Pensaba que la mayoría de la gente joven que se dedica a la economía y otras ciencias sociales está interesada en el análisis cuantitativo y muchos de ellos son bastante sofisticados en matemática, estadística y computación. Opinaba que se puede esperar una mayor cooperación internacional en proyectos de investigación, donde cada persona trabajará en su propia terminal y colaborará de modo casi instantáneo con otros científicos de diferentes instituciones. De manera similar, se manejan modelos más poderosos basados en una percepción de la estructura real del área de interés en la que se trabaja, en vez de las estructuras aproximadas que se usaban antes. Jim afirmaba que Box y Jenkins (1976) hicieron una gran contribución al desarrollo del análisis de series de tiempo y le gustaba enseñar en sus cursos de la LSE el álgebra de estos modelos por su "limpieza", pero estaba convencido que la metodología basada en el enfoque estructural o de los componentes inobservables de series de tiempo era el camino a seguir en el trabajo aplicado. Andrew Harvey junto a otros colegas en la LSE han estado trabajando en este enfoque del análisis de las series de tiempo desde hace algunos años. Más aún, Andrew había desarrollado un marco metodológico completo para esta clase de modelos mientras estuvo en la LSE. Esta metodología de series de tiempo se basa en los modelos de espacio de estado y el filtro de Kalman asociado (Harvey, 1989). El Profesor Durbin apoyó firmemente estos desarrollos.
El Profesor Durbin creía que el debate acerca de la inferencia estadística a veces se plantea con miras estrechas ya que básicamente confronta el enfoque Bayesiano versus el enfoque clásico y deja de lado los aspectos más interesantes de la estadística aplicada. A su entender en la actualidad hay muchos factores que son relevantes, uno de ellos es el desarrollo de los paquetes estadísticos que contemplen los diferentes enfoques para que estén disponibles en caso de ser útiles para un trabajo posterior, de modo que sean accesibles. También consideraba desafortunado para el desarrollo de la probabilidad el planteo tan parcial de los Bayesianos, como de Finetti (1974) y Savage (1972), que insistían en que la inferencia estadística se basara únicamente en información subjetiva para el control de la incertidumbre. En su opinión la profesión como un todo debería reconocer 16 ESTADÍSTICA (2012), 64, 182 y 183, pp. 5-22
que la probabilidad tiene dos aspectos: variabilidad e incertidumbre. Una filosofía integrada, que contemple ambos aspectos, los abarcará y contendrá.
Consideraba que una buena parte de las antiguas ideas de la inferencia estadística estaban basadas en modelos paramétricos simples, que eran vistos como valederos en cierto sentido, también justificados por las limitaciones computacionales de esa época. En los días de Fisher, los modelos eran muy sencillos, había que suponer normalidad o alguna otra distribución conocida y las posibles formas de analizar un conjunto particular de datos eran extremadamente limitadas. En cambio actualmente, estaba convencido que, por el poder de cómputo disponible, existe una gran libertad para mirar los datos desde diferentes puntos de vista: la robustez y el manejo de los outliers y la posibilidad de modificar la presentación de la información para realizar un mejor análisis. Siempre pensó que muchas nuevas técnicas serán desarrolladas e incorporadas a los paquetes y que también que se debe educar a nuestros futuros estudiantes en un uso más pragmático de las teorías de inferencia estadística de manera tal que aprendan a usar las técnicas según la necesidad. Estaba más predispuesto a pensar en una filosofía de la estadística, un enfoque general de la materia, que en la aplicación de cualquier esquema específico de inferencia o sistema de inferencia y esperaba que la profesión como un todo se moviera en el futuro hacia lo que él llamó una filosofía única de la estadística.
Aunque casi todos los aportes del Profesor Durbin fueron impulsados por la teoría, sus trabajos surgieron sobre todo por el deseo de resolver problemas específicos y de obtener soluciones fáciles de poner en práctica. Por lo tanto, siempre consideró importante presentar un ejemplo numérico en sus investigaciones. Su importante artículo en el JRSS (Durbin, Brown and Evans, 1975) sobre residuos recursivos y la detección de los cambios estructurales en una serie de tiempo, es una buena ilustración de cómo identificó la importancia de un problema práctico, que no fue reconocido en su debido tiempo, pero que recibió la atención necesaria muchos años más adelante. Problemas empíricos y prácticos también fueron tomados muy en serio por el Profesor Durbin, lo que se evidencia en su trabajo sobre el ajuste estacional de series de tiempo en los proyectos conjuntos que tenía respectivamente con Murphy y Kenny (Durbin and Murphy, 1975; Durbin and Kenny, 1976, 1982). Apreciaba especialmente su influyente trabajo empírico con Andrew Harvey acerca de los efectos producidos por la legislación que obliga al uso del cinturón de seguridad sobre las víctimas de tráfico en Gran Bretaña (Durbin and Harvey, 1985, 1986). El quería demostrar que los métodos de series de tiempo se deben utilizar para estudiar problemas interesantes del mundo real y que son importantes en el análisis de políticas de estado.
ABRIL: James Durbin: in memoriam 17
En la entrevista realizada por Phillips (1988), Durbin comentaba algunos hechos interesantes de su carrera profesional que se detallan a continuación.
El proyecto sobre el uso del cinturón de seguridad desarrollado con Andrew Harvey incluyó un análisis de series de tiempo de “recuentos pequeños” del número mensual de accidentes graves con vehículos utilitarios en Gran Bretaña. Se despertó en él un interés por desarrollar métodos para el tratamiento de series temporales con características no Gaussianas. También le gustaban las investigaciones que, a principios de 1990, Siem Jan Koopman llevaba a cabo en la LSE bajo la supervisión de Harvey. Estaba dispuesto a participar y a colaborar. Las colaboraciones con Koopman y otros tuvieron como resultado sus publicaciones a partir de 1990 sobre los modelos de espacio de estado (Durbin, 1990, 1997, 2000, 2004; Durbin y Koopman, 1997, 2000a, 2000b, 2002, 2003; Durbin y Quenneville, 1997). Por otra parte, el Profesor Durbin estaba interesado en escribir un libro sobre métodos de espacio de estado con el objetivo de presentar una alternativa a la metodología de Box y Jenkins de análisis de series temporales (Durbin and Koopman, 2001, 2012).
Jim se retiró de la LSE después de casi 39 años de servicio activo en el Departamento de Estadística. Por ese motivo, el 15 de Diciembre de 1988, se organizó un seminario especial al que asistió Sir Maurice Kendall, entre muchos otros personajes destacados de la disciplina. Alan Stuart elogió la claridad de lenguaje en sus trabajos y sus habilidades para la enseñanza, y dijo en esa oportunidad: “En todas las conversaciones que he tenido con los estudiantes y otras personas, nadie se ha quejado que no podía entender lo que Jim quería hacer.” En el mismo seminario, Andrew Harvey elogió la originalidad de Jim en la investigación, y dijo también: “Las contribuciones de Jim han sido principalmente teóricas. Sin embargo, detrás de su trabajo, siempre ha habido un claro entendimiento de lo que es importante desde el punto de vista práctico. En otras palabras, nunca se siente en la lectura de su obra que él ha realizado el trabajo matemático por el mero hecho de hacer matemáticas. Está allí con un propósito, porque quiere resolver un problema que en la realidad tiene importancia práctica.”
En la Reunión Anual General de la Royal Statistical Society del 2 de Julio de 2008, Jim fue galardonado con la “2008 Royal Statistical Society’s Guy Medal in Gold”, por toda una vida de contribuciones muy influyentes que le han dado reconocimiento internacional destacándolo como líder en nuestro campo, teniendo especialmente en cuenta su trabajo pionero sobre los tests de correlación serial en regresión, en la estimación de ecuaciones, el movimiento browniano y otros procesos que cruzan fronteras curvas, en test de bondad de ajuste con parámetros estimados, y en muchos aspectos del análisis de series de tiempo, especialmente en 18 ESTADÍSTICA (2012), 64, 182 y 183, pp. 5-22
los relacionados con la econometría, así como su destacado y amplio servicio a la profesión en el escenario internacional.
Referencias
ABRIL, JUAN CARLOS (2012). "James Durbin: In Memoriam". Conferencia pronunciada en el X Congreso Latinoamericano de Sociedades de Estadística (X CLATSE). Córdoba, Argentina, 16 al 19 de Octubre de 2012. http://conferencias.unc.edu.ar/index.php/xclatse/clatse2012/paper/view/507/24
BARTLETT, M. S. (1955). An Introduction to Stochastic Processes. Cambridge University Press, Cambridge.
BOX, G. E. P. and JENKINS, G. M. (1976). Time Series Analysis: Forecasting and Control (Revised Edition). Holden-Day, San Francisco.
DE FINETTI, B. (1974). Theory of Probability. Wiley, London.
DURBIN, J. (1953). "Some results in sampling theory when the units are selected with unequal probabilities." Journal of the Royal Statistical Society, Series B. 15: 262-269.
DURBIN, J. (1954). "Errors in variables." Review of the International Statistical Institute. 22: 23-52.
DURBIN, J. (1957). "Testing for serial correlation in systems of simultaneous regression equations." Biometrika. 44: 370-377.
DURBIN, J. (1959a). "Efficient estimation of parameters in moving average models." Biometrika. 46: 306-316.
DURBIN, J. (1959b). "A note on the application of Quenouille’s method of bias reduction to the estimation of ratios." Biometrika. 46: 477-480.
DURBIN, J. (1962). "Trend elimination by moving-average and variate-difference filters." Bulletin of the International Statistical Institute. 39: 131-141.
DURBIN, J. (1963). "Trend elimination for the purpose of estimating seasonal and periodic components of time series." Time Series Analysis (M. Rosenblatt, Editor). Wiley, New York. ABRIL: James Durbin: in memoriam 19
DURBIN, J. (1970a). "Testing for serial correlation in least-squares regression when some of the regressors are lagged dependent variables." Econometrica. 38: 410-421.
DURBIN, J. (1970b). "An alternative to the bounds test for testing for serial correlation in least- squares regression." Econometrica. 38: 422-429.
DURBIN, J. (1973a). "Weak convergence of the sample distribution function when the parameters are estimated." Annals of Statistics. 1: 279-290.
DURBIN, J. (1973b). Distribution theory for tests based on the sample distribution function. Society for Industrial and Applied Mathematics, Philadelphia.
DURBIN, J. (1980a). "Approximations for densities of sufficient estimators." Biometrika. 67: 311-333.
DURBIN, J. (1980b). "The approximate distribution of partial serial coefficients calculated from residuals from regressions on Fourier's series." Biometrika. 67: 335-349.
DURBIN, J. (1985). "Evolutionary origins of statisticians and statistics." A celebration of Statistics: The ISI Centenary Volume. Springer-Verlag, New York.
DURBIN, J. (1988). "Maximum likelihood estimation of the parameters of a system of simultaneous regressions equations." Econometric Theory. 4:159-170.
DURBIN, J. (1990). "Extensions of Kalman modelling to non-Gaussian observations." Quadermi di Statistica e Mathematica Applicata. 12: 3-12.
DURBIN, J. (1997). "Optimal estimating equations for state vectors in non- Gaussian and nonlinear state space time series models." Selected Proceedings of Athens, Georgia, Symposium on Estimating Functions.
DURBIN, J. (2000). "The state space approach to time series analysis and its potential for official statistics (The Foreman Lecture)." Aust. and N. Zealand J. of Statistics. 42: 1-23.
DURBIN, J. (2004). "Introduction to state space time series analysis." State, Space and Unobserved Component Models. Cambridge University Press, Cambridge. 3-25. 20 ESTADÍSTICA (2012), 64, 182 y 183, pp. 5-22
DURBIN, J.; BROWN, R. L.; and COWLEY, E. H. (1970). "New method for seasonal adjustment of unemployment series." Economic Trends. 199: 16-20.
DURBIN, J.; BROWN, R. L.; and COWLEY, E. H. (1971). "Seasonal Adjustment of Unemployment Series." Studies in Official Statistics, Research Series 4. Central Statistical Office, London.
DURBIN, J.; BROWN, R. L.; and EVANS, J. M. (1975). "Techniques for testing the constancy of regression relationships over time (with discussion)." Journal of the Royal Statistical Society, Series B. 37: 149-192.
DURBIN, J. and HARVEY, A. C. (1985). "The effects of seat belt legislation on road casualties. Report on assessments of statistical evidence." Annex to Compulsory Seat Belt Wearing: Report by the Department of Transport. Her Majesty’s Stationery Office, London.
DURBIN, J. and HARVEY, A. C. (1986). "The effects of seat belt legislation on British road casualties: A case study in structural modelling (with discussion)." Journal of the Royal Statistical Society, Series A. 149: 187-227.
DURBIN, J. and KENNY, P. B. (1976). "Seasonal adjustment when the seasonal component behaves neither purely multiplicatively nor purely additive." Proceedings of Census Bureau/NBER Conference (A. Zellner, Editor). U.S. Government Printing Office, Washington.
DURBIN, J. and KENNY, P. B. (1982). "Local trend estimation and seasonal adjustment of economic and social time series." Journal of the Royal Statistical Society, Series A. 145: 1-41.
DURBIN, J. and KOOPMAN, S. J. (1997). "Monte Carlo maximum likelihood estimation for non-Gaussian state space models." Biometrika. 84: 669-684.
DURBIN, J. and KOOPMAN, S. J. (2000a). "Time series analysis of non-Gaussian observations based on state space models from both classical and Bayesian perspectives." Journal of the Royal Statistical Society, Series B. 62: 3-56.
DURBIN, J. and KOOPMAN, S. J. (2000b). "Fast filtering and smoothing for multivariate state space models." J. Time Series Analysis. 21: 281-296.
DURBIN, J. and KOOPMAN, S. J. (2001). Time Series Analysis by State Space Methods. Oxford University Press, Oxford. ABRIL: James Durbin: in memoriam 21
DURBIN, J. and KOOPMAN, S. J. (2002). "A simple and efficient simulation smoother for state space time series analysis." Biometrika. 89: 603-616.
DURBIN, J. and KOOPMAN, S. J. (2003). "Filtering and smoothing of state vector for diffuse state space models." J. Time Series Analysis. 24: 85-98.
DURBIN, J. and KOOPMAN, S. J. (2012). Time Series Analysis by State Space Methods. Second Edition. Oxford University Press, Oxford.
DURBIN, J. and MURPHY, M. J. (1975). "Seasonal adjustment based on a mixed additive-multiplicative model." Journal of the Royal Statistical Society, Series A. 138: 385-410.
DURBIN, J. and QUENNEVILLE, B. (1997). "Benchmarking by state space models." Int. Statist. Review. 65: 23-48.
DURBIN, J. and STUART, A. (1951a). "Differences in response rates of experienced and inexperienced interviewers." Journal of the Royal Statistical Society, Series A. 114: 163-206.
DURBIN, J. and STUART, A. (1951b). "Inversions and rank correlations coefficients." Journal of the Royal Statistical Society, Series B. 13: 303-309.
DURBIN, J. and STUART, A. (1954). "An experimental comparison between coders." Journal of Marketing. 19: 54-66.
DURBIN, J. and WATSON, G. S. (1950). "Testing for serial correlation in least square regression, I." Biometrika. 37: 409-428.
DURBIN, J. and WATSON, G. S. (1951). "Testing for serial correlation in least square regression, II." Biometrika. 38: 159-178.
DURBIN, J. and WATSON, G. S. (1971). "Testing for serial correlation in least square regression, III." Biometrika. 58: 1-19.
FELLER, W. (1971). An Introduction to Probability Theory and Its Applications. 2. Wiley, New York.
GHOSH, J. K. (1994). "Higher Order Asymptotics." NCF-CBMS Regional Conference Series in Probability and Statistics. 4. Institute of Mathematical Statistics, Hayward, California. 22 ESTADÍSTICA (2012), 64, 182 y 183, pp. 5-22
HARVEY, A. C. (1989). Forecasting, Structural Time Series models and the Kalman Filter. Cambridge University Press, Cambridge.
HAUSMAN, J. A. (1978). "Specification tests in econometrics." Econometrica. 46: 1251-1271.
MALINVAUD, E. (1964). Statistical Methods in Econometrics. North-Holland, Amsterdam.
PHILLIPS, PETER C. B. (1988). "The ET Interview: Professor James Durbin." Econometric Theory. 4: 125-157.http://www.jstor.org/stable/3532030
SAVAGE, L. J. (1972). The Foundations of Statistics. Dover, New York.
TANIGUCHI, M. (1991). "Higher Order Asymptotic Theory for Time Series Analysis." Lecture Notes in Statistics. 68. Springer-Verlag, Berlin.
Invited paper Received August 2012 Revised June 2013 ESTADISTICA (201x), 64, 182, pp. c Instituto Interamericano de Estad´ıstica
STATISTICAL INFERENCE WITH COMPUTER SIMULATION: AN INTRODUCTION TO BOOTSTRAP ANALYSIS WITH R
PABLO E. VERDE Coordination Center for Clinical Trials, University of Duesseldorf
Moorenstr. 5, D-40225, Duesseldorf, Germany [email protected]
ABSTRACT
Bootstrap methods are a general approach to make statistical inference using computer simulation techniques. They have made possible what was unthinkable some decades ago, like approaching complicated statistical problems where theo- retical analysis was hopeless. Although we live in the information and computer age, these techniques are still not part of the main statistical training. As a conse- quence they are usually neither well understood nor widely used in routinely sta- tistical applications. The aim of this article is to review bootstrap computations with R in a tutorial style. The presentation is written informally omitting most of the technical details and concentrating on the use of bootstrap techniques. Key words
Bootstrap; Standard Errors; Confidence Intervals; Empirical Likelihood; R. RESUMEN
Los m´etodos “bootstrap” son un enfoque general para hacer inferencia estad´ıstica utilizando t´ecnicas de simulaci´onpor computadora. Han permitido hacer lo que hasta hace algunas d´ecadasera impensable, tal como resolver problemas estad´ısticoscomplicados para los cuales una resoluci´onan´alitica te´oricaser´ıaim- practicable. Si bien vivimos en la era de la informaci´ony la inform´atica,estas t´ecnicasno forman parte a´unde la formaci´onestad´ısticacl´asica.En consecuencia, estos m´etodos no son ni bien entendidos ni utilizados rutinariamente en las aplica- ciones estad´ısticas.El objetivo de este art´ıculoes revisar los m´etodos “bootstrap” con el software R en un estilo tutorial. La presentaci´ones informal omitiendo gran parte de los detalles t´ecnicosy concentr´andoseen el uso de estas t´ecnicas.
Palabras clave
Bootstrap; Errores Est´andar;Intervalos de Confianza; Funci´onEmp´ıricade Verosimil- itud; R.
2
1. Introduction
Bootstrap methods were invented by Efron (1979) as a general approach to make statistical inference using computer simulation techniques. The advantage of us- ing computer simulation in statistics is that it can be applied to complicated situations where theoretical analysis is hopeless or where a sample size is too small for a procedure to work properly, or when an on-the-fly statistical solution is required. The editors of Estad´ıstica have kindly invited me to write a tutorial paper on bootstrap methods using R (R Development Core Team, 2012). During the last years R has become the platform of knowledge exchange between people perform- ing statistical analysis, doing methodological research and developing statistical software. R offers an unmatched computer environment to explore and present bootstrap ideas. So, I hope that by using R a data analyst may better understand bootstrap techniques and also be encouraged to use them in practice.
A large amount of statistical research has been produced to explore theoretical properties of the bootstrap methods and to show the wide scope of their appli- cations. The most relevant publications will be cited in the following sections. For the readers interested in learning more about bootstrap we recommend two comprehensive references: the introductory book by Efron and Tibshirani (1993) and the most advanced volume by Davison and Hinkley (1997). Two additional tutorial papers of bootstrap analysis with R are Canty (2002) and Davison and Kuonen (2002). Another introduction to bootstrap calculations with R is pre- sented by Venables and Ripley (2002, page 133). This paper is organized as follows: in Section 2 we will review basic ideas on bootstrap methods, in Section 3 we will present some working examples of boot- strap analysis with R. Bootstrap confidence interval calculations will be covered in Section 4 and bootstrap empirical likelihood in Section 5. We will briefly cover bootstrapping complex data structures, like regression analysis and hier- archical modeling, in Section 6. A summary is presented in Section 7. An R package called bootcamp implements the full R script of the paper. The package is available from CRAN (The Comprehensive R Archive Network) at http://cran.r-project.org/.
2. Basic Ideas of Bootstrap Methods
2.1 Non-parametric Estimation of Standard Errors and Bias
Non-parametric estimation of standard errors and bias are probably the most pop- ular application of bootstrap methods. In this case, our observed data y1, . . . , yn have been modeled as a realization of a random sample drawn from an unknown distribution function F . In the simplest case observations are scalar values, but
3 in principle the sample space Y can be more general, e.g., multivariate data or a mixture of discrete and continues variables, etc. The data are used to estimate a parameter θ = t(F ) of particular interest. For example the median of F is estimated by the median of the data θb = t(Fb). In the non-parametric setup, the distribution function F is estimated by Fb, the empirical distribution function, which puts probability mass 1/n at each y1, . . . , yn and the function θb = t(Fb) is assumed to be a symmetric function of the data, this means it does not depend on the sample order.
Now, once we compute θb we wonder how accurate it is as an estimate of θ. Let σ(F ) denote the standard error of θb, as a function of the unknown sampling distribution F . The bootstrap estimate of σ(F ) is based on the plug-in principle, which replaces F by Fb in σ(F ):
σb = σ(Fb). (1) In many statistical applications there is not a simple expression of σb, the original Efron’s idea was to use Monte Carlo simulation to approximate (1). The bootstrap approximation corresponds to the following algorithm:
1. Assign probability 1/n to each y1, . . . , yn. That is, estimate F by Fb.
∗ ∗ 2. Simulate a bootstrap sample, say y1, . . . , yn from Fb by sampling with re- placement from the data y1, . . . , yn. The asterisk is used to denote a real- ization of a bootstrap sample.
3. Calculate θb from the bootstrap sample, say θb∗. 4. Repeat steps 2 to 3 a large number of times, say B. This generates bootstrap ∗ ∗ ∗ values θb1, θb2,..., θbB.
∗ ∗ ∗ The bootstrap values θb1, θb2,..., θbB are used to make statistical inference for θb. ∗ The standard error σb, is estimated by the variability of θb as:
v u B u 1 X ∗ ∗ 2 σB = t (θb − θb ) , (2) b (B − 1) B (·) r=1 ∗ PB ∗ where θb(·) = 1/B r=1 θbB. In the same way, assessment of bias of θb can be approximated by ∗ bias = θb(·) − θ.b (3)
4
The number of bootstrap samples B controls the Monte Carlo error in the numer- ical approximation of σbB. It is easy to show that σbB converge in probability to σb as B → ∞. In practice B is typically taken between 50 to 200 for estimation of standard errors.
2.2 General Bootstrap Setup
In general we can summarize the bootstrap methods in the following schematic form:
Statistical Model Bootstrap Computations
F → y =⇒ {F,Sb } → y∗ ↓ ↓ ↓ θ = t(F ) θb Analytics → θb∗ & . ↓ . {σ(F ),CI} {σbB, CIc } We wish to estimate the accuracy of statistics θb for estimating a parameter of in- terest θ. Accuracy may include the estimation of standard errors, σbB, confidence intervals, CIc , and so on. The point estimates Fb for F and a simulation proce- ∗ ∗ dure S deliver the bootstrap data {y1, . . . , yn} which are used to calculate the ∗ ∗ ∗ bootstrap replications θb1, θb2,..., θbB. A bunch of statistical procedures, that we called Analytics above, are applied to the bootstrap replications to get accuracy measures. In practice, the success of the bootstrap analysis may depend on many factors: 1. The choice of the model Fb which mimics the hypothetical model F . In sim- ple applications like those presented in Section 3 this may not be a problem, but for complex data structures like regression or hierarchical modeling the choice of Fb is an issue. In Section 6 we will present two examples. 2. By construction the distribution of θb∗ is a discrete distribution, however the distribution of θb is usually continuous. Therefore, we should decide which simulation procedure S we use to make the distribution of θb∗ similar to the sampling distribution of θb. In Section 3 we will show that the smooth bootstrap can mitigate this problem. 3. The presence of outliers influences bootstrap results. Outliers in the boot- strap behave differently than in estimation. Even for robust estimates the bootstrap distribution can be affected. A graphical sensitivity analy- sis called Jackknife-after-bootstrap (Efron, 1992) is presented in Section 4 to assess the influence of outliers in the bootstrap distribution.
5
4. The smoothness of θb = t(Fb) is particularly important in confidence intervals construction. As well as the use of θb∗ − θb as pivotal quantity. In Section 4 we give a worked example using double bootstrap method to analyze these issues in the construction of confidence intervals. There is no general solution to all of these issues, but there are a series of diagnostic techniques that help us to understand the use of bootstrap methods in a particular problem. We are going to illustrate these techniques in the following sections.
3. Bootstrapping in R R has powerful and easy to use functions that simplify the use of bootstrap methods in practice. One is the function sample(x), which takes a sample, either with or without replacement, from the elements of a vector x. For example, > set.seed(123) > sample(1:10, size = 5, replace = TRUE) [1] 3 8 5 9 10
takes a random samples of size 5 with replication from the set of integers of 1 to 10. The other useful function is replicate(n, expression), which replicates and evaluates n times the argument expression. For example, > set.seed(123) > replicate(10, median(rexp(10, rate=1))) [1] 0.315 0.536 1.258 0.960 1.206 1.000 0.290 0.555 1.451 0.795 replicates 10 times a random sample of size 10 from an exponential distribution and calculates the median for each replication. These two functions can be used for our own implementation of a bootstrap analysis. There are two main packages in R to make bootstrap analysis. One is the boot package, which is linked to the volume of Davison and Hinkley (1997) and the other package is bootstrap which is associated to the introductory book by Efron and Tibshirani (1993). In general the package boot is a more flexible implementation, but we are going to use both packages for confidence intervals computations.
Example: Inference for the Median
Table 1 presents the results of a fitness test performed on 24 professional athletes.
The outcome measurement is the maximal oxygen uptake in ml/kg/min (VO˙ 2 max). This parameter is considered the golden standard in cardiovascular fitness. As
6 a simple example of bootstrap computations, suppose that we are interested in median value of VO˙ 2 max between professional athletes. The estimated median
˙ Table 1. Maximal OXYGEN UPTAKE in ml/kg/min (VO2 max) of 24 Professional Athletes (Source: Olympic Training Center Buenos Aires, Argentina) 62.90 56.50 43.30 61.50 45.90 58.60 56.60 57.00 63.80 63.20 63.70 40.00 57.00 51.00 61.00 52.90 60.00 63.00 50.50 50.50 53.80 62.80 58.80 58.10
˙ VO2 max from Table 1 is 58 ml/kg/min. How accurate is this estimator? We use the function sample() to make a bootstrap analysis of this problem:
> #Boostrapping with sample() function > mvo2 <- c(62.9, 57, 56.5, 51, 43.3, 61, 61.5, 52.9, 45.9, 60, 58.6, 63, 56.6, 50.5, 57, 50.5, 63.8, 53.8, 63.2, 62.8, 63.7, 58.8, 40, 58.1) > > set.seed(123); m <- 5000; b.res.1 <- numeric(m) > for(i in 1:m) b.res.1[i] <- median(sample(mvo2, replace=T) ) > > #Bias measure > mean(b.res.1 - median(mvo2)) [1] 0.13 > > # Standard deviation of the median > sd(b.res.1) [1] 1.5
The estimated bias of 0.13 is very small for this medical problem and the stan- dard deviation of 1.5 indicates a very accurate estimation of the median. We can compare these results with those based on large-sample techniques. The asymptotic distribution of the sample median θb is normal with standard devia- tion σ = 1/p4nf 2(θ).
We estimate σ by assuming that the data in Table 1 follows a normal distribution, in R: > dnorm(58, mean(mvo2), sd(mvo2)) [1] 0.058 > 1/(2*sqrt(24)*0.058) [1] 1.8
7 which confirms that bootstrap and large-sample theory give similar results in this problem. The left panel of Figure 1 presents the bootstrap distribution of θb∗ which
Figure 1. Bootstrap Distribution of the Median Maximal Oxygen Consumption in ml/kg/min of 24 Professional Athletes. Left Panel: Bootstrap Distribution by Resampling Each Observation With Probability 1/n. Right Panel: Bootstrap Distribution by Smoothing the Observations Probability With a Normal Kernel With Bandwidth of 2. Data Source: Olympic Training Center Buenos Aires, Argentina.
Nonparametric bootstrap Smoothed bootstrap
0.8
0.3
0.6
0.2 0.4 Density Density
0.1 0.2
0.0 0.0 50 52 54 56 58 60 62 52 54 56 58 60 62 b.res.1 b.res.2 is clearly non-normal and affected by discreteness of the empirical distribution of the data. The last issue can be problematic for confidence interval computations. One remedy is to smooth the sample probability of each data point, which is known as the smoothed bootstrap. We can do this by using a normal kernel with bandwidth of 2 as follows:
> # Smooth bootstrap > b.res.2 <- numeric(m) > for(i in 1:m) b.res.2[i] <- median(sample(mvo2, replace=T) + rnorm(n=24)*.5)
8
> #Bias measure > mean(b.res.2 - median(mvo2)) [1] 0.13 > > # Standard deviation of the median > sd(b.res.2) [1] 1.5 the bias and standard errors estimates are identical to the non-parametric boot- strap. The bootstrap distribution presented on the right panel of Figure 1 is now corrected by discreteness.
Example: Are the Data Normally Distributed? Assessing normality of a data set is one of the most common tasks in data analysis. Usually, a qq-plot is built to display the possible deviations of a data set with respect to the standard normal distribution. Alternatively, we may ask which are the consequences of lack of normality if we are going to use our data analysis in a predictive way, i.e., what can we expect about new data if we assume normality?
Parametric bootstrap can be used to assess this type of question. First, we sim- ∗ ∗ ulate a bootstrap sample y1, . . . , yn from a normal distribution with meany ¯ and standard deviation σ. Second, we estimatey ¯∗ and variance σ∗ from the bootstrap b b ∗∗ ∗∗ sample. Then, a predictive bootstrap sample is generated by simulating y1 , . . . , yn ∗ ∗ from a normal distribution with meany ¯ and standard deviation σb . The pre- dictive bootstrap sample can be used to investigate particular features of the original data. We define the following features between the observed data y and the predictive data y∗∗:
T ∗ = min(y∗∗),T ∗ = max(y∗∗),T ∗ = q (y∗∗) − q (y∗∗) 1 2 3 75 25 and ∗ ∗∗ ∗ ∗∗ ∗ T4 = |q90(y ) − y¯ | − |q10(y ) − y¯ |. These measures are compared with the corresponding values based on the observed data: T1 = min(y),T2 = max(y),T3 = q75(y) − q25(y) and T4 = |q90(y) − y¯| − |q10(y) − y¯|. Here T1 and T2 measures how extreme a new observation could be, T2 measures the variability of a new sample and T4 measures asymmetry. The parametric double bootstrap for this example is implemented as follows:
9 min.y <- max.y <- asy1 <- asy2 <- inter.q <- rep(0,1000) for(b in 1:1000) { # Parametric bootstrap data y.boot <- rnorm(length(mvo2), mean = mean(mvo2), sd = sd(mvo2)) y.pred <- rnorm(length(mvo2), mean = mean(y.boot), sd = sd(y. boot))
# Data features min.y[b] <- min(y.pred) # min max.y[b] <- max(y.pred) # max
# Asymmetry predicted data asy1[b] <- abs(quantile(y.pred, prob=0.90) - mean(y.pred)) - abs( quantile(y.pred, prob=0.10) - mean(y.pred)) # Asymmetry original data asy2[b] <- abs(quantile(mvo2, prob=0.90) - mean(y.pred)) - abs( quantile(mvo2, prob=0.10) - mean(y.pred)) # Variability inter.q[b] <- quantile(y.pred, prob=0.75) - quantile(y.pred, prob=0.25) }
Figure 2 presents the predictive analysis. For example, the upper left panel shows the analysis of the predictive minimum, the vertical line indicates that the ob- served value is located almost at the center of the predictions. All in all, normality looks reasonable for the VO˙ max data of Table 1, but one warning is displayed 2 by the upper right plot which predicts future VO˙ 2 max values that may be not realistic for this population of athletes. If this were an important aspect then a normal distribution should be replaced by a model which gives less chance to ˙ predict large values of VO2 max.
4. Bootstrap Confidence Intervals In the construction of a confidence interval we want to asses the uncertainty about a scalar parameter value θ, by a random interval, say C1−2α with nominal coverage 1 − 2α such that if θ is a true parameter value, then
Prob [θ ∈ C1−2α] = 1 − 2α. (4) There is a small number of cases in applied statistics where exact confidence inter- vals can be calculated, e.g., the use of the t-distribution to calculate the confidence
10
Figure 2. Predictive Bootstrap Analysis. Top Panels: Predictive Bootstrap Distribution for the Minimum on the Left and for the Maximum on the Right. The Vertical Line Shows the Observed Value. Lower Panels: Predictive Bootstrap Distribution for the Range on the Left and Asymmetry on the Right
Minimum y* Maximum y* Frequency Frequency 0 10 20 30 40 50 0 10 20 30 40 50 60
25 30 35 40 45 50 60 65 70 75 80 85
min.y max.y
Variability y* Asymmetry
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ●●●● ●● ●● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ●● ● ● ● ● ● ●●● ● ●● ● ●● ●● ●●●●●●● ● ● ● ● ●●● ● ●● ●● ●●●● ● ● ● ●●●●●● ● ● ● ● ● ● ●●●●●●●●● ● ● ●●● ● ●● ●●● ● ● ● ● ● ● ●●● ● ●● ● ● ●● ●● ● ● ● ● ●●●● ● ● ● ●●●●● ● ●● ●● ● ●● ● ● ● ● ● ●●● ●● ●● ●●●●●●● ●●● ● ● ● ● ● ● ● ●●●● ●● ●●● ●●●●● ●● ●●●● ● ●● ●● ● ● ● ● ● ● ●●●●● ●● ●●●● ● ●●●●●●● ●●●●● ●● ●● ● ● ●● ● ● ● ● ●● ●● ●●●● ●● ● ●● ● ● ●●● ●●●● ● ●● ●●● ● ● ● ● ● ● ● ● ●● ● ●●●● ●●●●●●●● ●●●●●● ● ●●●● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ●●●● ● ●●●●● ● ●●●● ● ●●● ●● ● ● ● ●● ●●●●● ●●●● ● ●●●● ●●● ●●● ● ● ● ● ● ● ●● ● ● ● ●●●●●●●●●●●●●●●● ●●● ●●● ● ●● ●● ● ● ● ● ●●● ● ● ●● ● ●● ●●●●● ●● ●● ●● ●● ●● ● ●● ●●● ● ● ● ● ● ●●● ●●●●●●●● ● ● ● ● ●● ●●●●●● ● ●● ●● ●●●● ● ● ● ● ● ●●● ●● ●● ● ●● ●●●●●●●●● ●● ● ● ● ● ● ● ● ●●● ●●● ● ●● ● ●● ● ● ● ● ● ●● ● ● ● ●●●● ●●● ● ●●● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ●● ● ●●●●● ●● ● ● ● ● ● ● ● ●● ●● ● ●●●●● ● ●●●● ● ● ● ● ● ● ●●● ●● ● ● ● ● ●● ● ● ● ●●● ● ● ● ● Frequency ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ●● ●● ●● ●● ● ● ●●● ● ● ●●● ●● ●●●●● ● ●● ●●● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●●● ● ● ● ●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● Asymmetry original data ● ●● ● ● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 20 40 60 80 100 0 2 4 6 8 10 12
5 10 15 4 6 8 10 12 14
inter.q Asymmetry predicted data
interval of the mean. In most applied problems these exact results are not possi- ble and confidence intervals are calculated approximately. The use of bootstrap methods is particular convenient for this task. They try to automatically encap- sulate sophisticated statistical thoughts that sometimes provide good solutions to complicated statistical problems.
Example: Confidence Intervals for the Correlation Coefficient
Table 2 presents 20 measurements of newborn infants with low weight (< 1.500g). These data were kindly provided by Professor Hoehn at the University Clinic in Duesseldorf. They were used to determine non-invasive parameters in the treatment of a common functional cardiovascular disease in pre-term infants. We use these data to illustrate different types of bootstrap confidence intervals using 11 the package boot. Now, suppose that we are interested in calculating bootstrap
Table 2. Measures of 20 Neonates: The Variable Weight is Measured in Grams and Concentration of Urinary Protein Level of NGAL is Presented in Logarithmic Scale. Data Source: University Clinic Dusseldorf,¨ Germany
weight 1210 815 1120 700 660 680 520 l.NGAL 9.43 10.36 10.04 12.21 11.94 12.21 11.09 weight 573 710 1415 1090 1230 1340 495 l.NGAL 13.09 12.56 9.98 12.41 11.38 6.91 12.96 weight 1140 495 870 1100 980 850 l.NGAL 9.74 11.40 11.37 8.87 11.90 6.91 confidence intervals for the correlation coefficient between weight and l.NGAL. The first step to make bootstrap calculations with the package boot is to write the statistics of interest in the sampling form: # Bootstrap function: boot.cor <- function(data, ind) cor(data[ind, ])[1,2]
The first argument of the function boot.cor is the R data frame containing the observations, the second argument is the rows index ind. The function boot() is used to generate the bootstrap replicates by combining the data frame dat.b, containing the data of Table 2, the function boot.cor and by indicating the number of simulations R=10000: > library(boot) > boot1 <- boot(dat.b, boot.cor, R = 10000) > boot1 ... Bootstrap Statistics : original bias std. error t1* -0.57 -0.0071 0.15
The object boot1 belongs to the class boot and can be used to make further bootstrap analysis. For example the function boot.ci calculates different types of bootstrap confidence intervals: > boot.ci(boot1, conf = c(0.9, 0.95), type = c("norm", "perc", "bca")) ... Intervals : Level Normal Percentile BCa 90% (-0.81, -0.32 ) (-0.80, -0.32 ) (-0.77, -0.25 ) 95% (-0.86, -0.28 ) (-0.83, -0.26 ) (-0.80, -0.19 ) 12
In this example we have calculated intervals which are only based on the bootstrap values of θb. We can see that the Normal and Percentile methods gave similar results, but the BCa interval differs from the others. We can graphically appreciate this correction in Figure 3. In the next subsections we review some technical details on these confidence intervals.
Figure 3. Bootstrap Distribution of the Correlation Coefficient Between Weight and Urinary Protein Level of NGAL in 20 Newborns. The Dashed Vertical Lines Correspond to the 95% CI Based on the Percentiles of the Bootstrap Distribution. The Solid Vertical Lines Display the Correction Introduced by the BCa Method
Bootstrap distribution Frequency 100 150 200 250 50 0
−0.8 −0.6 −0.4 −0.2 0.0
theta.s
The Bootstrap Normal Interval
The Normal distribution confidence interval is the most simple approach. In this ∗ ∗ procedure we assumed that the distribution of θb1,..., θbR is perfectly normal with,
∗ 2 θbr ∼ NOR(θ,b σb ), (5) then a confidence bound with level α is given by
(α) θbNOR[α] = θb+ z σ.b (6)
(α) (0.95) Here, z is the 100αth percentile of a normal deviate, e.g., z = 1.645 and σb is (2). 13
Clearly this method is not suitable for the correlation coefficient example: We can see in Figure 3 that for a sample size of n = 20 the bootstrap distribution is not normal. Furthermore, the method assumes that θb∗ − θb is a pivotal quantity, i.e. its distribution does not depend on θ. Figure 5 in bootstrap-t confidence intervals shows that this is not the case.
The Bootstrap Percentile Interval
The Percentile confidence interval is a more natural way to construct a confidence ∗ ∗ interval for θ based on the empirical distribution of the values θb1,..., θbB:
n ∗ o #θbb ≤ c Gb(c) = . (7) B The α confidence bound is defined as −1 θbPERC[α] = Gb (α), (8) which corresponds to the (B × α)th value in the ordered list of B replications of ∗ θb . For example if α = 0.05 and B = 2000, θbPERC[0.05] corresponds to the 100th ordered value of bootstrap replications. If B × α is not an integer, we take the kth largest value such that k ≤ (B + 1)α. The percentile method generalizes the normal confidence interval to allow asym- metry in the distribution of θb. It is based on the assumption that there is a monotonic increasing function φ = m(θ) that perfectly normalizes the distribu- tion of θb: 2 φb − φ ∼ NOR(0, σφ). (9) Then under this scale a confidence bound of level α is (α) φb[α] = φb + z σφ, (10) which back transforming to the original scale of θ with m−1(·) gives
−1 −1 (α) θbPERC[α] = m (φb[α]) = m (φb + z σφ). The breakthrough of the percentile method is that in practice we do not need to know m(·). This transformation is implicitly constructed by computational brute force from the bootstrap values θ∗. The percentile method has two important properties. First, it is transformation- invariant, that is the confidence interval for a parameter ψ resulted from a mono- tonic transformation g(θ) = ψ is the percentile confidence interval for θ mapped by g(θ):
(ψbPERC[α], ψbPERC[1 − α]) = (g(θbPERC[α]), g(θbPERC[1 − α])). 14
Second, the percentile interval is range-preserving, that is the confidence bounds fall within the range of values where θ is defined. This method, however, does not correct for bias and it may be sensitive to in- fluence observations or outliers. Figure 4 displays a sensitivity analysis called jackknife-after-bootstrap (Efron, 1992) for the bootstrap distribution of the cor- relation coefficient of Table 2. This plot shows the sensitivity of the percentiles of the bootstrap distribution to deletion of individual observations. Figure 4 is produced as follows: jack.after.boot(boot1, main = "Jackknife-after-bootstrap")
The horizontal axis represents a scale of influence with mean zero and standard deviation one. The vertical axis represents the bootstrap distribution centered at the estimated value θb. The connected sawed lines show the quantiles estimates by removing each observation in turn. In this analysis we can spot out that observation number 20 influences the bootstrap calculations.
Figure 4. Jackknife-after-bootstrap Plot. This Plot Shows the Sensitivity of the Percentiles of the Bootstrap Distribution to Deletion of Individual Observations
* * * *** * * * * ** * * * * * * * * * * ***** * * * 0.2 * * * * * * * * * * * ***** * * * ** * * * * * *
0.0 * * * * ***** * * * ** * * * * *
* * * * * * * * ***** * * ** * * * * * * * * * * ***** * ** * * * * * −0.2 * * *** * * * * * * ** * * *
5, 10, 16, 50, 84, 90, 95 %−iles of (T*−t) * *
13 4 3 19
14 6 17 7
8 15 10 12
1 9 2 11 −0.4 18 5 16 20
−2 −1 0 1 2 3
standardized jackknife value
The BCa Interval
BCa stands for bias corrected and accelerated. This bootstrap confidence interval has been proposed by Efron (1987) to improve the performance of the percentile confidence interval. The BCa interval corrects the percentile method when the 15 estimate θb is biased and when its standard error σb depends on the value of θb. Biased estimates with non-constant standard errors are commonly encountered in applied problems (odds ratios, correlation coefficients, etc.), making the BCa method particularly attractive for practical purposes.
The BCa method is transformation-invariant and range-preserving like the per- centile method, but their limits are second-order accurate and they are also second- order correct. By second-order accurate we mean that for a nominal confidence level of α we expect that, in average, the confidence level in repeated samples has an error of O(1/n). Secondary exactness means that the difference between the 3/2 theoretical exact confidence limit and the estimated one differs by Op(1/n )( Hall (1988)).
The BCa interval has a peculiar model construction which is far from being intu- itive, but it is well motivated by the transformation theory that we describe in this section. Just as the percentile method, the BCa postulates the existence of a monotonic increasing function φ = m(θ) that perfectly normalizes the sampling distribution of θb, with φb = m(θb) having distribution:
2 φb ∼ NOR(φ − z0 σ(φ), σ(φ) ), σ(φ) = 1 + aφ. (11)
Here the constant z0 plays the roll of bias correction factor and is estimated from the bootstrap distribution as: ! #θ∗ ≤ θb z = Φ−1 b , (12) b0 B + 1
−1 ∗ where Φ (·) is the standard-normal quantile function and #θb ≤ θb is the num- ber of bootstrap estimates that are lower than the original estimate θb. The coefficient a is a skewness correction factor called acceleration constant, which is estimated as 3 Pn ¯ i=1 θb(−i) − θ a = , (13) b 23/2 Pn ¯ 6 i=1 θb(−i) − θ
¯ where θb(−i) is the value of θb when the ith observation is omitted and θ is their average value. We can directly calculate a and z0 for our correlation coefficient running example: > # estimation of z0 > z0 <- qnorm(sum( boot1$t < boot1$t0 )/ 10001) > z0 16
[1] 0.126 > > # centered influence values > uu <- empinf(data = dat.b[1:20, ], statistic = boot.cor, + type = "jack", stype="i") > # estimation of a > acc <- sum(uu * uu * uu)/(6 * (sum(uu * uu))ˆ1.5) > acc [1] 0.035 >
The constant a is typically |a| < 0.2 and the same applies to z0 (Efron, 1987, Section 3). So the estimated values for our example show that both parameters introduce an important correction in the confidence interval computation.
Under the previous conditions the α level confidence limit for the BCa method is given by: z + z(α) θ [α] = G−1 Φ z + b0 . (14) bBCa b b0 (α) 1 − ba (zb0 + z ) Formula (14) looks intimidating at first sight, but we can see that for the case z0 = a = 0 the confidence limit defined by (14) is
−1 θbBCa [α] = Gb (α) the 100αth Percentile of the bootstrap distribution. If in addition Gb is perfectly normal, then θ [α] = θ + z(α)σ, bBCa b b the Normal interval.
Bootstrap-t Intervals
Bootstrap-t confidence are conceptually simple, its name comes from the analogy with the Student’s t-statistic. This interval needs an estimate of the standard error ∗ ∗ σb of the statistic θb for each bootstrap sample. It is based on the studentized statistic ∗ ∗ θb − θb T = ∗ . (15) σb The bootstrap distribution of T ∗ is used to estimate the distribution of
θb− θ T = , (16) σb 17 which is unknown in most situations. Usually these pivotal quantities are pre- sented with n1/2 in front of the right-hand equation, here this constant is absorbed ∗ by σb and σb respectively. By analogy of the Student-t confidence interval, the end points of a 1 − 2α bootstrap-t confidence interval are defined as
(1−α) (α) θbT [α] = θb− Tb σ,b θbT [1 − α] = θb− Tb σ.b (17)
(α) ∗ Here Tb is obtained by αth ordered value of the simulated Tr for r = 1,...,R. (0.025) ∗ For example if R = 1000 and α = 0.025 then Tb is the 25th ordered Tr . This method was originally proposed by Efron (1979), but poor numerical behav- ior reduced its interest. Babu and Singh (1983) gave the first proof of second-order accuracy for the bootstrap-t. Hall (1988) showed that the bootstrap-t limits are second-order correct and revived its interest. Davison and Hinkley (1997) present extensive use of this technique in several applied problems. Venables and Ripley (2002, p.137) recommend its use in general applications.
The bootstrap-t is computationally very intensive. It requires that we estimate ∗ σb for each bootstrap sample. If we use a second level bootstrap to calculate ∗ ∗ σb with R2 bootstrap replications, then the number of evaluations of θb will be R2 × R. This computational burden is one of the drawbacks of this method. One remedy is to use the jackknife estimate of σ∗ in each bootstrap sample (see b below). Another drawback is that, unlike the percentile method and the BCa, this method is not transformation invariant. More dangerous in practice, the bootstrap-t algorithm could be very unstable. Its numerical problem is produced ∗ ∗ by the fact that σb could be very small compared to θb − θb, this artifact produces an artificially heavy tailed distribution of T ∗ resulting in a very long confidence interval. This is particular dangerous in situations where the confidence limits must be bounded to the range where θ is defined.
Example: Bootstrap-t Interval for the Correlation Coefficient
In order to apply a bootstrap-t confidence interval in R we need to modify the ∗ ∗ resampling function to deliver in each bootstrap sample θb and σb . In this case, the resampling function is: # Bootstrap function for t-intervals cor.boot.t <- function(data, ind) { # Calculate theta in each bootstrap sample theta <- cor.boot(data, ind) # Calculate the Jackknife SE for theta in each bootstrap sample
18
var.theta <- var.linear( empinf(data = data[ind, ], statistic = cor.boot, type = "jack", stype="i") ) return(c(theta, var.theta)) }
This function calls the var.linear function in the boot package, which esti- ∗ mates σb in each sample by using the Jackknife method. The empinf function calculates the empirical influence values for a statistic applied to a data set.
Now we can directly apply the boot function to calculate the bootstrap-t confi- dence interval:
> boot2 <- boot(dat.b[1:20,], boot.cor.t, R = 10000) > boot.ci(boot2, conf = c(0.9, 0.95), type = c( "bca", "stud")) BOOTSTRAP CONFIDENCE INTERVAL CALCULATIONS ... Intervals : Level Studentized BCa 90% (-0.816, -0.044 ) (-0.770, -0.250 ) 95% (-0.881, 0.091 ) (-0.800, -0.185 )
We can see that the length of the bootstrap-t is substantially greater than the BCa. The length of the bootstrap-t interval could be a dangerous feature in practical applications, where confidence bounds may fall outside the range of values where θ is defined.
Example: Bootstrap-t Intervals and Variance Stabilization
Now, the 95% CI based on bootstrap-t interval covers the zero correlation, the question is how much can we trust this result? One diagnostic tool to investigate the validity of this interval in a particular problem is the variance plot. This plot ∗ ∗ shows the relationship between θb and σb for each bootstrap sample.
The left panel of Figure 5 shows the resulting variance plot for this example. ∗ ∗ Clearly, there is a positive association between θb and σb , with a correlation of 0.71. The bootstrap computations have uncovered a problem with the bootstrap- t method, which is that T ∗ is not an approximate pivotal quantity of θ. The previous analysis suggests that if we have a function h(·) where, h(θb∗) − h(θb) T ∗ = , (18) pVar(h(θ∗))
19
Figure 5. Variance Plot of the Bootstrap Distribution. Left Panel: Estimated Variance Against Estimated Value in Each Bootstrap Sample. The Scatter Shows a Negative Correlation Between Variance and Estimate. Right Panel: Variance Plot of the Bootstrap Values After Applying a Logistic Transformation of the Estimate and its Variance
● 10 10
●
● ● ●
● ● ● ● ● ● ● ● ● ● ● 8 8 ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●● ● ● ● ● ● ● ● ● ●●● ●●● ● ● ● ●●● ●● ●●● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ●● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●●●●● ●●●●● ●● ● ● ● ● ● ● ● ● ●● ● ●●●●● ●●●●● ●● ●●● ●●●● ● ●● ● ● ● ●●●●●●● ●● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ●●● ●●●●●● ● ● ● ● ● ● ● ● ●●● ● ●●● ●● ● ●● ● ● ● ●● ● ●●●● ●●● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ●● ● ● ● ● ●●●●● ● ●● ●●● ● ●●●●●●● ●●● ●●●● ● ● ● ● ●● ● ●●● ● ●●●●●●●●● ●●● ● ●● 6 ● ● 6 ● ● ●●● ●● ● ●● ● ●●●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ●●●●●● ● ● ● ● ● ● ● ● ● ●● ●●●●●● ●● ●●● ●●●●●●●● ●● ●●● ●● ● ● ●● ●●●●●●●●●● ●●●● ●●●●●● ● ● ● ●● ● ● ● ●● ●●●● ● ● ● ●● ● ●● ●●● ● ●●●●● ●●● ●● ● ● ● ● ● ●● ●●●● ●●● ●●●●●● ●● ●● ● ●●● ●● ● ●● ● ●● ●● ●●● ●●●●●●●● ●●●● ●●●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●●●● ● ● ● ● ● ● ●● ● ● ●●● ●●●●●●● ●●●● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ●●●●●●●●●●● ●●●● ● ●● ●●● ● ● ● ● ● ● ● ● ●●●●●●● ●●●● ●● ●●●●● ●●● ● ● ● ● ● ● ●● ● ●●●●●●● ●●●●● ● ●● ●●●●●●●●● ●● ●● ●● ● ●● ● ● ●● ●●●● ● ● ● ●● ●●● ●●●●●●● ●●● ● ● ● ● ● ●●●●●●●●●●●●● ●● ●● ●●●●●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●●●●● ● ● ●●● ●●●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●●●●●●●●●● ●●●● ●●●●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●●● ●●●●●●● ●● ● ● ●● ●● ● ● ● ● ● ●● ●●●●●●● ● ●● ●● ●●●● ●●●●●●● ● ● ● ●●● ●●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●●● ●● ●●●●●● ● ●●● ● ●●● ●●●●●● ● ●● ● ● ● ●● ● ●● ● ● ● ● ●●●● ● ● ●●● ● ●●●●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●●●●●● ●●● ●●● ●●●● ● ●●●●●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●●●●● ● ●●●● ● ●●●●●●●●●●●●●●● ●● ● ●● ● ● ●● ●● ● ● ● ●● ● ● ● ● ●● ●● ● ●●●● ●● ●●●●●●●●●●●●●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●●●● ●● ●●● ●● ● ● ● ● ●● ●● ●● ● ●● ● ● ● ● ● ● ●● ●● ● ●●●●●● ●● ●●●●● ●●●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●●● ● ●● ● ●● ●●● ●● ●● ●● ●● ● ● ● ●● ● ● ●● ● ●●● ●● ●●●● ● ●● ● ●●● ●● ● ● ● ● ●●●● ● ●● ● ●● ● ● ● ● ● ● ●● ●● ●● ● ● ● ●● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ●●● ●● ●● ●● ●● ●● ●●●●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ● ●● ● ● ● ●●● ● ● ● ●● ● ●● ● ● ● ● ● ●● ●● ●● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ●●● ●● ●●●● ●● ●●● ● ●●● ● ●● ●● ● ●● ● ●● ● ●●●●● ●●● ●● ● ● ● ● ● ● ● ●●● ●● ● ●●●●● ●●● ●● ● ●● ●● ● ● ● ●● ● ●●●●●● ●● ●● ● ● ●● ●● ● ●●● ●●● ● ●● ● ●● ● ● ● ●● ●●●●● ● ● ● ●●● ●●●●●● ●●●● ● ● ●●● ● ● ● ●●● ● ●●● ● ●● ● ●●●● ● ● ● ● ● ●●● ●●● ●●●●●●●● ●●● ●● ● ● ● ● ● ●● ● ● ●● ● ●● ●● ● ●●●● ●● ● ● ●●● ●● ● ● ●●●●● ●●● ● ● ● ●● ●● ●● ● ● ● ● ●● ●●●● ●● ● ● ● ● ●● ● ● ●● ● ●● ●● ● ● ●● ●●●●●●●●●●●●● ● ● ● ●●●●● ● ● ● ● ● ●● ● ●● ● ●● ●● ●● ● ●● ● ● ● ● ●● ● ● ● ●● ●● ●●●●●●●●● ●● ●●●●●● ●●●●●● ● ● ● ● ● ● ●● ● ●● ●●● ● ● ● ● ●● ● ● ● ●● ● ●● ●●●●●●●●● ●●●●●●● ●●●● ●● ●●●●●●● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●●●●●● ●●●●● ●●●●● ● ●●●● ●●●●●●● ● ●● ●● ● ● ● ●● ●●● ● ● ● ●●●● ● ●●●● ●●● ● ●●●● ● ● ● ● ● ● ● ●● ●●● ● ● ●●●● ● ●●● ● ●● ● ●● ● ● 4 ● 4 ●●●● ●● ●●● ●● ● ● ●●● ●●●●●●●●●●● ● ●● ● ●● ●● ● ●● ● ● ● ● ●● ● ● ●● ●●●●● ●●●●●●●●●●●●● ●● ●●●●● ● ●● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ●● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ●●●●●●● ●●● ●● ●●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Var(theta.star)/Var(theta) ● ●●● ●● ●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●●●●●●●●●●● ●●●●● ●●● ●● ● ● ●●● ●●● ● ● ● ● ●●● ●● ●●● ● ● ● ● ● ● ● ●● ● ●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●● ● ● ● ● ● ● ● ● ●● ●● ● ●●● ● ● ● ● ●● ● ●●● ●●●●●●●●●●●●●● ● ●●●●● ●●● ●● ●● ● ● ● ● ● ● ●● ● ●●●● ● ● ● ● ● ● ●● ● ●●●●● ●● ●●●●●●●●●●● ●● ●● ●● ●● ● ● ● ●● ●● ●● ● ● ● ● ●● ● ● ●● ●●●● ●●●●●●●●●●●●●●●●● ● ●●● ● ●●●● ● ● ● ● ● ●● ●●●●●● ● ●● ● ● ● ● ●● ●●●●●●●● ● ●●●●● ●● ● ● ●● ● ● ● Var(h(theta.star))/Var(h(theta)) ● ●● ● ● ● ● ● ●● ●● ●●●●●●●●●●●●●●●● ● ●●● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ●●● ●●●●●●● ●●● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●●●● ●● ● ●●●●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●●●● ●●●●● ●●● ●●● ●●●● ● ●●● ● ● ● ● ● ● ● ● ●● ● ●● ●●●● ●●●●●● ●●●● ●●●●●●●●●●●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●●●●●●●●●●● ●●●●●●●●● ●●● ●● ● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ● ● ●●● ●●●●●●●●●● ●● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ●●●●●●●●●● ●●●●●●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●●●●●● ●●●● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ●● ●● ● ●● ● ●● ●●●●● ●● ●●●●● ● ● ●● ● ● ● ● ● ● ●●●● ● ●●●●● ●●●●●●● ● ● ● ● ● ● ● ●● ●●●● ●● ●●●●● ● ● ● ● ● ● ●●●●●● ● ●●●● ●● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ●●●●●● ●●●●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ●● ● ● ●● ●●● ● ● ●● ●●● ●● ● ● ● ● ● ● ● ●● ●●●●●●●● ●●●● ● ●●●●●●● ● ● ● ● ● ● ● ●●●● ●● ●● ●●● ●●●● ● ● ●● ● ●●● ●●●●●● ● ● ● ● ●● ● ● ●● ●● ● ●● ● ●●●●●● ● ● ● ● ● ● ● ●●●● ●●●●●●● ● ●●●● ●●●● ● ●● ● ● ● ● ● ● ●●●● ● ● ●●●● ● ● ● ●●●●●●● ●● ●● ●●●●● ●●● ● ● ● ●● ● ● ●●●●●●●●●● ● ●● ● ● ● ●● ● ●●●● ●●●● ● ●● ●● ●●●● ● ● ● ●●●●●● ● ●●● ●●● ●● ● ● ●●●●●●●● ● ●●●●● ● ●●●●●● ●●● ● ● ● ● ● ● ●●●●●●●● ●● ●● ●●● ● ● 2 ● ● ● ●●●●●● ● ● ● 2 ● ●●● ●●●●● ●●●●● ● ● ●●●●●● ●● ●● ●● ● ● ●●● ●● ● ● ●●● ●● ● ●● ● ●●● ● ● ● ● ● ● ● ●●●● ●●●●●●●●●●●●● ● ● ●●●●●●● ●●●●●●● ●● ● ●●●●●●●●●●●●● ● ● ● ● ● ● ●●●●●●●●● ● ●●●● ● ●●●● ● ●●● ●● ●●● ●●● ●●● ● ● ●● ● ●● ●●●● ● ● ●● ● ● ● ● ● ● ●●●●●● ● ● ● ●●● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● 0 0 −0.8 −0.6 −0.4 −0.2 0.0 −1.5 −1.0 −0.5 0.0
theta.star h(theta.star)
does not depend on θ then we can improve the construction of this confidence interval. In this case the α confidence bound based on T ∗ is calculated as q −1 (1−α) θ[α] = h h(θb) − Tb Var(hd(θ)) (19)
(1−α) where Tb is the (1 − α)th ordered value of the simulated T , but this time calculated in the scale of h(·). These results are backward transformed by h(·)−1 to the scale of θ.
The right panel of Figure 5 presents the variance plot after applying Fisher’s z-transformation to θ∗, 1 1 + θ∗ h(θ∗) = log 2 1 − θ∗ we can see that this transformation has a variance stabilization effect in this case, the correlation of the scatter is now -0.088. To calculate the bootstrap-t confidence intervals under the transformation h(·) we define three functions: the transformation h(·), the inverse transformation h(·)−1 and the first derivative of h(·):
20
# Transformation function f.tr <- function(x)0.5*log((1+x)/(1-x))
# First derivative f.tr. <- function(x) 2/((1+x) (1-x)) * # Inverse transformation function inv.f <- function(x) (exp(2*x) - 1 )/( exp(2*x) + 1) then, we can directly use the function boot.ci with the following arguments: > # Results with transformation > boot.ci(boot2, h = f.tr, hdot = f.tr., hinv = inv.f, conf = 0.95, type = c( "bca", "stud")) ... Level Studentized BCa 95% (-0.827, -0.001 ) (-0.799, -0.185 ) Calculations on Transformed Scale; Intervals on Original Scale
The bootstrap-t intervals are automatically calculated in the scale of h(·) and presented in the scale of θ. The effect of the transformation in this case was to shorten the length of the intervals and to ensure the bounds values are in the scale of θ.
Example: Automatic Computation of the Variance Stabilization Function
One major problem is that the variance stabilization function h(·) is unknown, the Fisher-z transformation is a good candidate when the data is multivariate normal, but any deviation from normality may affect h(·).
One way to reveal the shape of h(·) is by plotting each bootstrap replicate θb∗ ∗ against its linear approximation θbLin in each bootstrap sample, any deviation of linearity is a hint of the shape of h(·). The function linear.approx() in ∗ boot implements the non-parametric delta method to calculate θbLin (Davison and Hinkley, 1997, Sections 3.10.2), in our example we have: # Linearity of the correlation L.reg <- empinf(boot.out = boot2, type = "reg") plot(boot2$t[,1] , linear.approx(boot2, L.reg), col="magenta", xlab ="theta.star", ylab="Linear-approx-theta.star")
21
Figure 6 shows the resulting scatter plot. The scatter is clearly non-linear, the dashed line corresponds to the Fisher-z transformation, which does not seem to be a good choice. The smooth curve in the middle of the scatter is the automatic computation of h(·) proposed by Tibshirani (1988) and implemented in the function boott in the package bootstrap. The confidence intervals and the estimated transformation function are calculated as follows:
library(bootstrap) cor.bootstrap <- function(x, dat.b){ cor(dat.b[x,1],dat.b[x,2]) } cor.tt <- boott(1:20, cor.bootstrap, dat.b[1:20,], VS=TRUE) > cor.tt$confpoints[c(3, 9)] # gives confidence points [1] -0.789 -0.216 > points(cor.tt$theta, cor.tt$g, type="l", lwd=3) # add h() to the plot
Figure 6. Scatter Plot of Bootstrap Replicates of the Correlation Coefficient and Their Linear Approximation. The Solid Line is the Empirical Estimation of the Variance Stabilization Function h(·) and the Dashed Line is the Fisher-z Transformation
● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●●●●● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ●●●●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ●●●●●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ●●● ●● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ●● ●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ●●● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ●●●●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ●● ● ● ●● ●●● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ●●● ● ● ●●●●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ●● ●● ●● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ●●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ●●● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ●● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ●●● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●●●● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ●●●● ●●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ●● ● ●● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ●●●● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●●●● ● ● ●● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ●●●●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●●● ●●● ●● ● ●●● ● ●●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ● ● ●● ●● ● ● ●● ●● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ● ●● ●● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●●● ●● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●●●●●● ●● ● ● ●●●●● ●●● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●●● ●● ●● ●●●● ●● ● ●●●●●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ●●● ● ● ● ● ● ● ●●●●●● ●●●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ●●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ●●●●●● ●● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ●● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ●● ● ● ●● ●● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●● ● ●● ●●● ● ●●● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●●●● ●●●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ●● ● ● ● ● ●● ●● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●●●●●●●● ●●●● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ●●● ●● ● ● ● ●●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●●● ● ● ● ● ● ●●●●● ●● ● ● ●● ● ● ● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ●●● ●●●● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●●● ●● ●● ● ● ●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●●●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●●●● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ●● ●● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●●● ● ● ● ● ●●●●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ●● ●● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ●● ●● ● ● ● ● ●● ●●● ●●● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●●● ● ●●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ●● ● ●● ●● ● ● ●● ●●● ●● ●●● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●●●●● ● ●● ●●●●●●● ●● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●●●●●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ●● ●● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●●●● ● ●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ●● ● ●● ●● ●● ●● ● ●● ●● ●●● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ●●● ●● ● ● ●● ●●● ● ● ●●●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ●● ●●● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ●● ●●● ●●● ●●●●●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ●● ●● ● ● ●●●●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●●●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ●● ● ●●● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ●●● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ●● ●● ● ●●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●●●● ● ● ●● ● ●● ●● ●●●● ● ●● ● ●●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●●● ● ●●●● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●●●●● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ●●●● ●● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●●●●●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ●● ● ● ●● ●● ● ● ● ● ● ● ●●● ● ● ● ● ●●● ●● ●●● ● ● ● ●●●● ●●●● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ●● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ●● ●● ● ● ● ●● ●●●●● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●●●●●● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ●● ●● ●●● ●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ●●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ●● ● ● ● ● ●● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ●● ●● ●● ● ● ●● ● ● ● ● ●● ● ●● ● ●● ● ● ●● ● ● ● ●● ● ●●● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●● ●●●● ● ● ● ● ●●●●● ● ●●●●● ● ● ● ● ● ● ● ●● ●●● ● ●● ● ●● ●●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ●● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●●● ●●● ●●●● ●● ●● ●● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●●●●● ● ●●● ●●●● ● ●●● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ●● ● ● ● ●●● ●●● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ●● ●● ● ● ● ● ● ●●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●●● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●●●● ● ● ●● ● ● ●● ● ●● ● ●● ●● ●● ● ● ● ● ●● ●● ● ● ● ●●● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●●● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ●● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ●●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ●●● ●● ●●● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ●●●●●● ●● ● ●● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●●●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●●●●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●●●● ● ● ● ● ● ● ● ● ● ● ●● ● ●●●● ● ●● ● ●●● ● ● ●● ● ●● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ●● ●● ●●●● ● ●● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ●●● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●●● ●● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●● ● ●● ● ● ● ●● ●● ●● ●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ●●●● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ●●● ● ●● ● ●● ● ● ● ● ● ● ● ●●●● ● ● ● ● ●●● ● ● ● ●●● ● ● ● ●●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ●●● ●●●● ●● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●●● ●● ●●● ● ● ●● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ●● ●● ● ●●● ●● ● ● ●●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●●● ● ●●●● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ●● ● ● ●●● ● ●● ●● ● ● ● ● ● ●● ● ●●●● ●●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ●●●● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ● ● ● ●●●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●●● ●● ●● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ●● ●● ● ● ●●● ● ● ● ● ● ● ●● ●●● ●●● ● ●● ● ●● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●●● ● ●●●● ● ● ● ● ● ● ● ● ● ●●●● ●●● ● ● ●● ● ● ● ●●● ● ● ● ●● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●●● ● ● ● ●● ● ● ● ●● ● ● ●● ● ●●●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ●● ● ● ●● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● Linear−approx−theta.star ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● Tibshirani transformation ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Fisher transformation ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ●
● ●
●
−1.0 −0.9 −0.8 −0.7 −0.6 −0.5 −0.4 −0.3 −0.9 −0.8 −0.7 −0.6 −0.5 −0.4 −0.3 theta.star
For further discussion about this approach, see DiCiccio and Romano (1995),
22
Efron and Tibshirani (1993, Section 12.6) and DiCiccio et al. (2006) and further applications given by Davison and Hinkley (1997, Sections 3.9 and 5.2).
5. Bootstrap Computations of Empirical Likelihoods
The likelihood function plays a central role in statistical inference. It is the common contact point between classical and Bayesian statistics and is the natural device to combine information across multiple experiments, such as the case of meta-analysis.
Briefly, the statistical problem at hand is the following: Suppose that we have a random sample y1, . . . , yn ∼ F and we are interested in a parameter θ(F ). In this section we are interested in computing Lpro(θ) the profile likelihood of θ. In a full non-parametric setting where we don’t have any idea about a good candidate for F , the likelihood for F is define as:
n Y L(F ) = F (yi), (20) i=1 where F (yi) is the probability of the set {yi} under F . The profile likelihood for θ is Lpro(θ) = sup L(F ), (21) F :t(F )=θ which requires to maximize L(F ) for each θ over all distributions satisfying t(F ) = θ. This is clearly un-practicable, so we simplify the problem by restrict- ing attention to the set of distributions with support on {yi}. Let (w1, . . . , wn) and define Fw to be the discrete distribution putting probability mass wi on yi, Qn i = 1, . . . , n. The probability of obtaining our sample {yi} under Fw is 1 wi, we define the empirical likelihood for θ by
n Y Lemp = sup wi. (22) F :t(F )=θ i=1
Replacing F into L (θ) by F is to cast the problem in terms of the multinomial pro w distribution having support on the data points, in this way the Lemp is a profile likelihood.
The bootstrap likelihood LB is a numerical approximation of Lemp based on com- puter simulation. There are different strategies to use bootstrap results in the construction of likelihoods, nested bootstrap with addition kernel smoothing or saddle point approximation is used by Davison and Kuonen (2002), pivotal quan- tities are applied by Boss and Monahan (1986) and Hall (1987), confidence sets by Efron (1993). The use of bootstrap computations to approximate nonparametric
23 likelihoods and their connection to Bayesian inference are reviewed in (Efron and Tibshirani, 1993, Chapter 24) and in (Davison and Hinkley, 1997, Chapter 10).
In this section we present another bootstrap likelihood approach, that we called
LBCa . This method is based on the theory of the BCa intervals of Section 4. The LBCa likelihood was introduced by Pawitan (2000), who showed that the resulting −1/2 likelihood agrees with the Lemp up to order O(n ), which is essentially the same result of Davison and Kuonen (2002).
The idea of LBCa is straightforward, we construct a likelihood function of φ assum- ing that the underline model is (11), then we use the fact that BCa is invariant under transformation to define the likelihood for θ. If the graph {φ, L(φ)} is the likelihood of φ where,
(φb − φ + z0σφ) log L(φ) = − log σφ − 2 , 2σφ
−1 then {m (φ),L(φ)} is the graph of the likelihood of θ. The BCa likelihood is defined as h −1 i LBCa = L Φ {Gb(θ)} , where φb = Φ−1{Gb(θb)} is used in L(φ).
Example: Bootstrap Empirical Likelihood for the Correlation
Let’s illustrate the computations of LBCa for the correlation coefficient example of Section 4. The empirical likelihood is calculated as follows:
# BCa likelihood (Pawitan, 2000) z0 <- 0.126 acc <- 0.035 nr <- 200 rhoboot<- seq((min(boot1$t)+.05),(max(boot1$t)-.02), len=nr) phi<- rep(0, nr) for (i in 1:nr) phi[i]<- qnorm(sum(boot1$t < rhoboot[i])/nb) s <- 1 + acc*phi pmean <- -z0*s + phi lik <- dnorm(phi0, mean = pmean, sd=s) plot(rhoboot, lik/max(lik), type="l", col="blue", lwd=3, xlab="Correlation coefficient", ylab="Likelihood", main="Bootstrap Distribution and Empirical Likelihood", xlim=c (-1, 0)) hboot <- hist(boot1$t, breaks=100, plot=F)
24
Figure 7. Bootstrap Likelihood of the Correlation Coefficient. The Smooth Line Corresponds to the Empirical Likelihood Calculated by the BCa Method. The Histogram Corresponds to the Bootstrap Distribution
Bootstrap Distribution and Empirical Likelihood
Liklihood
0.0 0.2 0.4 0.6 0.8 1.0 −0.8 −0.6 −0.4 −0.2 0.0 Correlation coefficient
points(hboot$mids, hboot$counts/max(hboot$counts), type="s", lwd=3)
Figure 7 displays the resulting likelihood of the correlation coefficient together with the bootstrap distribution. Most of the correction of the likelihood is made by shifting the bootstrap distribution to the right.
6. More Complex Data Structures
In the previous sections we have applied bootstrap methods with R for a com- plex statistical problem, but with a simple data structure. More complex data structures arise in regression modeling and in hierarchal data analysis. For these types of problems, we need to define a resampling method that mimics the data
25 generating mechanisms as much as possible before we apply the confidence inter- vals that we reviewed in the previous sections. In this section we present two examples, one in non-linear regression and one in hierarchical data analysis.
6.1 Bootstrapping a regression problem
Carlin and Gelfand (1991) consider data on length yi and age xi measurements for 27 dugongs (sea cows) and use the following nonlinear growth curve with no inflection point and an asymptote as xi tends to infinity: 2 yi ∼ NOR(µi, σ ) xi µi = α − βγ , where α, β > 0 and γ ∈ (0, 1). The data is given in the following vectors: x <- c( 1.0, 1.5, 1.5, 1.5, 2.5, 4.0, 5.0, 5.0, 7.0, 8.0, 8.5, 9.0, 9.5, 9.5, 10.0, 12.0, 12.0, 13.0, 13.0, 14.5, 15.5, 15.5, 16.5, 17.0, 22.5, 29.0, 31.5) y <- c(1.80, 1.85, 1.87, 1.77, 2.02, 2.27, 2.15, 2.26, 2.47, 2.19, 2.26, 2.40, 2.39, 2.41, 2.50, 2.32, 2.32, 2.43, 2.47, 2.56, 2.65, 2.47, 2.64, 2.56, 2.70, 2.72, 2.57) plot(x, y, ylim = c(1.7, 3), xlab = "Age in years", ylab ="Length in mts")
The left panel of Figure 8 presents the data and the fitted model curve by maxi- mum likelihood. The model is fitted with the nlm() function in R as follows: # Initial values ... > ini.par <- c(alpha=max(y), beta =max(y) - min(y), gamma = 0.5) > ini.par alpha beta gamma 2.72 0.95 0.50 > # Fit the model ... > summary(fit <- nls(y˜alpha-beta*gammaˆx, start = ini.par)) ... Parameters: Estimate Std. Error t value Pr(>|t|) alpha 2.6581 0.0615 43.2 < 2e-16 *** beta 0.9635 0.0697 13.8 6.3e-13 ***
26 gamma 0.8715 0.0246 35.4 < 2e-16 *** .. Residual standard error: 0.0952 on 24 degrees of freedom
Now, the question is how to implement a resampling method for this statistical model if we are interested in making inference on the vector parameter θ = (α, β, γ, σ). There are basically four resampling methods to this problem: One is to sample the pairs of data (xi, yi) with replacement, like we did in the correlation example. The other one is to resample the residuals of the model and generate the bootstrap samples by: y∗ = f(x; θb) + r∗. The wild bootstrap is a variate of sampling residuals, but instead of directly sampling residuals, each residual is multiplied by a random variable with mean 0 and variance 1. This method may be useful in small sample problems. Another alternative is to implement a full parametric resampling plan and simulate the bootstrap data by y∗ = f(x; θb) + e∗, where each e∗ is simulated from a normal distribution with mean zero and variance 2 σb . In this section we illustrate how to sample from the residuals. In the next section we show how to implement a full parametric resampling schema in a more complex situation.
To sample from the residuals of the model, we define a data frame with the observations and the model fit, here the column fit contains the values of f(xi; θb) for each pair (xi, yi). The bootstrap function for θ should have as arguments a vector of residuals and a sampling index as we did in the previous sections. library(boot) d <- data.frame(y, x, fit=fitted(fit)) boot.fun <- function(rs, i ) { d$y <- d$fit + rs[i] # Generate the bootstrap data m1 <- nls(y˜alpha-beta*gammaˆx, data=d, start = coef(fit)) tmp <- summary(m1) theta <- tmp$coef[,1] # Extract coefficients estimates sigma <- tmp$sigma # Sigma hat cbind(theta, sigma) } 27
The boot function is applied as usual but we pass as data the scaled residuals. In the R workspace, the function boot.fun will find the data frame d: rs <- scale(resid(fit), scale=F) #remove the mean boot.res <- boot(rs, boot.fun, R = 5000)
The right panel of Figure 8 shows the results of the first 50 bootstrap samples. These results are generated by boot.d <- data.frame(boot.res$t) names(boot.d) <- c("alpha","beta", "gamma","sigma") # Right panel plot(x, y, ylim = c(1.7, 3), xlab = "Age in years", ylab ="Length in mts") for(i in 1:50) { curve(mu(x, alpha=boot.d[i,1], beta = boot.d[i,2], gamma = boot .d[i,3]), from = 1, to =max(x), lwd=1, col="blue", add = TRUE, lty =2) }
The bootstrap distribution of this regression model has dimension 4, Figure 9 presents the marginal distribution for each parameter on the diagonal and the pairwise bivariate distributions on the upper and lower diagonal panels. Clearly, the most striking feature is the association between the asymptotic growth value α and the rate of growth γ. This plot is generated with: library(car) scatterplotMatrix(˜alpha + beta+ gamma + sigma, reg.line=FALSE, smooth=TRUE, spread=FALSE, span=0.5, diagonal = ’histogram ’, lwd= 2, data=boot.d,cex=0.5)
The confidence intervals for θ = (α, β, γ, σ) can be calculated with the function boot.ci by using the argument index, where the number indicates the com- ponent of θ. For example the confidence interval for γ is calculated as: > boot.ci(boot.res, index =3, type = c("norm", "bca")) ... Intervals : Level Normal BCa 95% ( 0.8267, 0.9199 ) ( 0.8186, 0.9113 ) Calculations and Intervals on Original Scale 28
Figure 8. Left panel: Data from Carlin and Gelfand (1991). The Data Correspond to the Length yi and Age xi Measurements for 27 Dugongs (Sea Cows). The Smoothed Line is the Fitted Nonlinear Growth Curve With no Inflection Point and an Asymptote as xi Tends to Infinity. Right Panel: The Scatter Corresponds to the Original Data and the Lines are the Resulting Fitted Model From 50 Bootstrap Samples 3.0 3.0 2.8 2.8
● ● ● ●
● ● ● ●
2.6 ● 2.6 ● ● ● ● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● 2.4 2.4
●● ●●
● ● Length in mts ● ● Length in mts ● ●
● ● 2.2 2.2 ● ●
● ● 2.0 2.0
● ● ● ●
● ● 1.8 ● 1.8 ●
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Age in years Age in years
In this example the normal and BCa intervals give similar results. It is worth mentioning that this example is calculated with a full Bayesian method with non-conjugate priors in the WinBUGS (Spiegelhalter et al., 2004). The Bayesian analysis gave almost identical results as the BCa, however their interpretation is of course different.
6.2 Bootstrapping Hierarchical Data
In recent years, there has been an increasing interest in the analysis of hierarchi- cal data in a wide range of applied problems, e.g., multilevel data in sociology, longitudinal analysis, frailty modeling, meta-analysis, etc. In these problems each experimental unit is measured several times, e.g., patients participating in a clin- ical study are measured in different periods of the trial. The common feature of this type of data is that measurements within units can not be considered statisti- cally independent. Therefore, a special modeling technique should be considered. For a gentle introduction to this statistical area see Gelman and Hill (2007). In this section we illustrate the parametric bootstrap for hierarchical data, where 29
Figure 9. Scatter Matrix Plot of the Bootstrap Distribution of the Parameters of a Nonlinear Growth Curve With no Inflection Point. The Marginal Distribution For Each Parameter on the Diagonal and the Pairwise Bivariate Distributions on the Upper and Lower Diagonal Panels
0.8 0.9 1.0 1.1 0.06 0.08 0.10 0.12
● ●●● ●● ● ●●●● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ●●●●● ● ● ●●● ● ● ● ● ● ●●● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● alpha ● ●● ●● ●● ● ● ●●●●●●● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ●● ●● ●●●●● ● ● ●●● ● ● ● ● ● ● ●●● ●● ● ●●●●●●● ● ● ●● ●●● ● ● ● ● ●● ● ●●● ●●● ●●●●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●●●● ●●●● ● ● ● ● ●●● ● ● ● ●● ● ●●●● ● ● ●●●●●● ●● ● ● ● ●● ● ●● ● ● ● ● ●●● ●●●●●●●●● ● ●●●●●●●●●●●●● ● ● ●●● ●●●●●● ●●● ●●● ● ● ● ● ●●●● ●●●●●●● ●● ● ●● ●●●●●●●●● ●●● ●●●● ●●●● ● ●●● ●● ● ● ● ●● ●● ●●●●●●●● ● ●● ● ● ● ●●●●●●●●●●●●● ●● ● ● ● ●●●● ● ● ●● ● ● ● ●● ● ●● ●●● ● ●●● ● ●●●●●●●● ● ● ●●●●●●●●●●● ● ●● ●● ● ● ● ●●● ●● ● ● ● ● ● ●●●● ● ●●●●● ●●● ● ● ● ●●●●●●●● ●●● ● ● ● ● ● ● ● ●●●● ●●●●●● ●●● ●●●●●● ●●●●●●●●●●● ● ● ● ●● ●●●●●●●●●●●●●● ●●●● ●● ●●●●● ●●● ● ●● ● ● ● ● ● ●●●●●●●●●●● ●●● ●●● ●●● ● ● ●●● ● ●●●●●●●●●●●●●●●●●● ●● ●●● ●●●●●● ●●●● ●●● ●●●● ● ● ● ●● ●● ●●●●●●●●●●●●●●● ● ●●● ● ● ● ●●● ●●●●●●●●●●●●●●● ●● ● ● ● ● ●● ●●●●● ●●●●●● ●●● ● ●●● ● ● ●● ● ●●● ●● ●●● ●●●●●●●●●●●●●● ●●●●●●● ● ● ●●●●●●●●●●●●●●●●●● ●●●●●●● ● ●● ●●●●●●●●●●●●●● ●●●●●●●●● ● ●● ●● ● ●● ●● ●●●●●●●●●●●●●●●●● ●● ● ●●●● ●● ● ●●●●●●●●●●●●●●●●●●●● ● ● ●● ●● ●●●●●●●●●●● ●●●●●●●●●●●● ●●● ●● ● ● ● ● ●● ●●● ●● ●●● ● ● ● ● ● ● ●●●●●●●●●●●●●●● ● ●●● ●●●●●● ●● ●●●● ●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●● ● ● ●●●●●●●●●●●●●●●●●●●●● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ● ● ●● ● ● ●●● ●●● ●●●●●●●●●●●●●●●● ●● ●●●●● ● ● ●● ● ● ●●● ●●●●●●●●●●●●●●●●● ●●●●● ● ●● ● ●● ●●●●●● ●●●●●●●●● ●●●● ● ● ●●●● ● ● ●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●● ●●●●●●● ● ● ● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ●●●●●●●●●●●● ●●●●●●●●●●●●●●● ●●●●●● ● ●● ●● ●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ● ●●●●●●●● ●●●●●● ●● ●●●●●●● ●●●●● ● ● ● ● ● ● ●●●●●●●●●●●●●●● ●●●●●●●●●● ●●●●● ●●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●● ● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ● ●● ● ●●●●●●●●● ●●●●● ●●●●●●●●●●●●● ●● ●● ● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●●●●●● ●●●●●●● ●●●●●●●●●●●●●● ● ● ● ● ● ●●● ●● ●●●● ●●●●●●●● ●●●●●●●●●●●●●● ●●●●● ● ● ●● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●● ● ● ●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ● ●●● ● ● ●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ●● ●●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●●● ●●●●●●●●●●●●●●●●●●●●●●● ●●●●●● ●●●●●●●●● ● ●●● ● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ●●●● ● ●● ●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●● ●●●● ● ● ● ●●●● ●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●● ● ● ● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ● ● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ● ● ● ● ● ● ● ● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ● ● ● ● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●● ●●●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ●● ● ●●● ●●●●● ●● ●●●●●●●●● ●●●●●●●●● ●●●●●●●●●●●●●●●●●●●● ● ● ●●●● ●●●●●●●●● ●●●●●●●●●●●●●●●●●●●● ●● ●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●● ●●●● ●● ●●●●●●●● ●●●●●●●●●●●●●● ●●● ●●●●● ●●● ● ● ● ●● ● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●●● ● ●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●●●●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●● ●● ●● ● ●● ●● ●●●●●●●● ●●● ●●●●●●●●●● ●●●●●●●●●●● ● ●●● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●● ● ●●● ●●●●●●●●●●●●●●● ●● ●●●●●●●● ●●●●●● ● ● ● ● ●● ● ●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●● ● ● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●●●● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ●●● ●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●● ● ●●● ●●● ●● ● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●● ● ● ● ● ●●●● ●●●●●●● ●● ●●●●●●●●●●●●●●●● ● ● ● ●●● ●● ●● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●●●●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●● ●● ●● ● ●●●●●●●●●●● ●● ● ●●●●●●●●●●●●●●●●●●●● ●●● ●●● ●● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●● ●● ●●● ● ●●●●●●●●●● ● ●●●●●●●●●●●●●●●●●●● ●●● Frequency ● ●●●● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●●● ●●● ●●●●●●●●●●● ● ●●●●●●●●●●●●●●●● ●● ● ● ●●●●●●●●●●●●●●●● ●●●●●●●●● ●● ● ● ● ●●●●● ●●●●●●●●●●●●●●●●●●● ●●● ●●●●●●● ● ● ●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●● ●● ● ● ● ●●● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ●● ●●● ●●●● ●●●●● ●●● ●●● ●● ●●●●●●●●●●● ●● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●●●● ●●●●●●●●●●●●●● ●●●●●●●●● ●● ●●●●● ● ● ● ● ●●●●●●●●●●●●●●●●● ●●●●●● ● ●●● ●●●● ● ● ●● ●●●●●●●●●●●●●●●●●●●●● ●● ●● ●● ● ● ●● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●● ● ●●● ● ● ●● ●●●●●● ●●●●●●●● ● ● ● ●● ● ● ● ●●●●●● ●●●●●●●●● ●●● ● ● ● ●●● ● ●● ● ●●● ● ●●●● ●●●●●●●●●● ● ● ● ● ● ●●●●●● ●●●●●●● ● ●●●●● ●● ● ● ● ●● ●●●●●●●●●●●●●●●●●●● ●● ● ● ●●● ● ● ●●●●●● ●●●●●●●● ●● ● ● ● ●● ● ●●● ●●● ●●●● ●● ● ● ● ●●●● ● ●●●● ● ●●●●● ●● ●●●●●● ● ● ● ●●●●●●● ●● ●● ●● ●● ● ● ● ●● ● ●●● ● ●●● ●● ●●● ●● ● ●● ● ●●●●●●●●●●● ●●●●● ● ● ●● ●●●●●● ●●●●●● ●● ●●● ●● ● ● ●● ● ● ●●●●● ● ●●●●● ● ● ●● ● ● ● ●●●●●●●● ●●● ●●●● ● ● ●●● ● ●●●●● ●●●●●●● ● ●● ● ● ●●●● ●●● ●●● ● ●● ● ● ●●●● ●● ●● ●●●● ● ● ● ●● ● ● ●● ●●●● ●● ●● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ●● ● ● ● ● ●● ●● ●●● ● ● ● ● ● ●● ● ●●● ● ●● ●● ● ● ●● ● ● ●● ●● ● ● ● ● ● ●● ●●● ●●●● ● ●● ● ●● ● ●● ●●●●●●● ●● ●● ● ●● ● ● ● ● ●● ● ●●● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● 2.5 2.6 2.7 2.8
● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ●● ●●●● ● ● ● ● ● ● ● ● ● ●●● ●●● ●●● ● ● ●● ● ● ●● ●● ●● ●● ● ● ● ●● ●●● ●●● ● ● ● ● ●● ● ● ● ●● ● ● ●●●●● ●● ● ●●●●● ● ● ● ●● ● ● ●● beta ● ● ●●● ● ●●●● ● ● ●● ● ● ● ● ●●● ●● ● ●● ●● ● ● ● ●● ● ●● ● ● ●●● ●●● ● ●● ● ● ● ●● ●● ● ●● ●●● ●● ●● ● ● ●●● ●●● ● ● ●●● ●● ●●● ● ● ●●●●●● ●● x●●●● ● ● ● ●●● ●●●●● ●●●●● ● ● ● ●● ●● ●●●●● ●● ● ● ● ●● ● ● ●● ● ● ●● ●●● ● ●● ● ●● ● ● ● ●● ● ●●●●●●●● ●● ●●● ● ● ●● ●●●●● ● ●●● ● ● ● ● ● ● ● ● ● ●●●●●● ●●●●● ● ● ● ● ● ●● ● ●● ●● ● ●●●●●● ● ●● ● ●●● ●●● ● ●●●●●● ● ● ● ● ●●●● ●●● ●●●●●● ●● ●●●●● ●● ● ● ● ● ●●● ●●●● ●● ●● ●●●●● ●●● ●● ● ● ● ●●●●●●● ●●●●● ● ●●●●● ●● ● ● ● ● ●● ● ●●● ●● ●●● ●● ●● ●●●● ● ● ● ●●●● ● ● ●● ●● ●●●● ● ● ●● ●●●●●●●●●●●● ●●●● ●●●● ● ●●●●● ●● ●●●●●●●●●●●● ●●● ●● ●● ● ● ●●●●●●● ●●●●●● ●●● ● ●●●● ● ●●●●● ● ●● ● ● ● ● ●●● ●●●●●●●● ●●●●●●● ●●●●●●●●● ● ● ● ● ●● ●●●●●●● ● ●●●●●● ●● ● ●●●●● ● ●●●●●● ● ●● ●● ●●●●● ● ● ●●●●●●●●●●●●●●● ● ● ● ●●● ●●●●●● ●●●●● ●●●●●●●●●●● ●● ●● ● ●●●● ● ●●●●●● ●●●●●● ●● ● ●●● ● ● ● ●● ●●●● ●●● ●●●●●●●●●●●●●● ● ●●●●● ●●● ● ● ●● ● ●●●● ●●●●●●●●●●●● ●●●●●●●● ●●●●●●● ●● ● ●● ●●● ● ● ●●●●●●●●●●●●●●●●●●● ●●● ●● ● ● ●●●● ●● ●●●●●●● ●●●●● ●● ● ●● ● ● ● ● ●● ● ●●●● ●●●●●●●●●●●● ● ● ● ● ●●●●●●●●● ●●●●●●●● ● ●● ●● ● ● ●●●●●●●●●●●●●●●● ● ● ●●●●●●● ●●●●● ● ● ● ●●●● ●●●●●● ●●●●●● ●● ●●●●●●●●●●● ● ● ● ● ●●●●●●●●●● ●●●●●● ●●●●●●●●●●● ●● ● ●● ●●● ●● ● ●●●●●●●●●●●●●●●● ●●●●●●●●● ●●●●●● ● ● ● ● ●● ● ● ●●●●●● ●●● ●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ● ● ●●● ●●●●●●●●●●●●●●● ●●●●●●● ●●●● ●●●●●●● ● ● ●● ●●●● ●●●●●●●●●●●●●● ● ●●●● ●●●●● ●●● ● ●● ● ● ●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●● ●●● ● ●● ● ● ● ● ●● ●●● ●●●●●●●●●●●●● ●●●●●●●●●●●●● ● ● ● ● ● ● ● ●●●●●●●●●●●●●●●●● ●●●●●●●●● ●● ●●●●●●●●● ●●● ● ● ● ●●● ● ●● ●● ● ●●● ●●●●●●●●●●●●●●●●● ●● ●● ● ●●●●● ●●●●●● ●●●●●●●●●●●●●●●● ●●●●●●● ●● ●●● ●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ●●●●●●●● ●● ● ● ● ● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●●●● ●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●● ●●●●●●● ● ● ●● ●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ● ● ●● ●● ●●●● ●●●●●●●●●●●● ●●●●●●● ●● ● ● ● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●● ● ● ●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●● ●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ●● ●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●● ●● ● ● ●●● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ●●● ● ●● ● ● ● ● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ●● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ● ●● ●● ●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●● ●●● ● ● ● ●●● ●●●●●●●●●●●●●●●●●●●●●● ●●●●●● ●●●●●●●● ●● ● ●● ● ● ● ●●● ●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●● ●●● ● ●● ● ●●● ●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ●●●●●●● ● ●●●● ●●●● ●●●●●●● ●●●●●●●●●●●●●● ● ●●● ●● ●●●● ● ● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ● ●●● ● ●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●● ●● ●● ● ● ●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ●●●●● ● ● ● ● ● ●●●●● ●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ● ● ●●●●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ●● ●●● ●● ●●●●●●●●●●●●●●●●●●●●●● ●●●●●●● ●● ●●●● ● ● ● ●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●●● ●● ●●● ●●●●●●●● ●●●●●●●●●● ●●●●●●●●●●●●●● ●●●● ●●●●●●●●● ● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●● ●● ● ● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ●● ● ●●● ● ● ●● ● ●●●●●●●● ●●●●●●●● ●●●●●●●●●●●●●● ●● ●●● ● ● ● ●● ●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ●● ●●● ● ●●●●●●●●●●●●●●●●● ●●●●●●●●● ●●●●●●●● ● ● ● ● ●● ●●● ●● ●●●●●●●●●●●●●● ●●●●●●●●●●●● ●●●●● ● ● ● ● ●●●●●●●●●●●●●● ●●●●● ●●●●● ●●●●●●●● ●● ● ● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ● ● ● ● ●●●●● ●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●● ●● ● ●●●●●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ●●● ●● ●●●●●●●●●● ●●●●●●●●●● ●● ● ● ●● ●●●● ●●●●●●●●● ●●●●●●●●●●●●●● ●●●●●● ● ●●● ● ● ● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ● ● ● ● ●●● ●●●●●●●●●●● ●●●●●●●●●●●●●●●● ●●●●●●● ● ● ●●● ●●● ●● ● ●●●●●●●●●●●●●●●●● ●●● ● ●●● ● ● ●●●● ●●● ●●●● ●●●●●●●●● ●●●●● ●●● ●●● ● ● ● ●●● ●●●● ●●● ●●●●●●●●●●●●●● ●●● ● ● ● ● ●●●●●●●●●●●●●●● ●●●● ●●● ● ●●● ● ● ● ●● ● ●●●●●●●●●●●●●●●●●● ●●●●●● ●● ● ● ● ●●● ●●●●●●●●● ●●●●●●●●●●●●●●● ● ●●● ● ● ●● ●●●●●● ●● ● ●●●●●●●●●●●●●●●●●●● ●● ● ● ●●●●●●●●●●●●●● ●●●●●●●●●●●●●● ●●●● ● ●● ●●●●●●●●●●●●●●●●●●● ●●●● ●●●●●●● ● ● ● ● ● ●● ●●● ●●●●●● ●●●●● ●●●●●●●●●●●● ●● ●● ●● ●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ● ● ● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●● ●● ● ● ●● ● ●●●●●● ●●●●●●●●●●●●●● ●●●●●●●● ● ● ●● ●●● ● ●●●● ●● ●●● ●●● ●● ● ●● ● ● ●● ● ● ●● ●●●●● ●●●● ●●●●● ● ● ● ● ●●● ●●● ● ●●● ●●●●● ● ●●●●● ● ● ● ● ●●●●●●● ●●●●●●●●●●●●● ●● ● ● ● ●● ● ● ●● ●●●●●●● ●●●●●●●●●●●●●●●● ● ● ● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●● ●● ●●●● ● ● ●● ●●● ● ●●●●●●● ●●●●●●●●●●●● ● ● ●●● ●● ● ●●● ●●●●●●●● ●●●●●●●● ●●●●●●●●● ● ●●● ●●●●●● ●●●●●●●●●●●●●●●● ●●●●●● ● ● ● ●● ● ● ● ●● ●●●●●● ● ● Frequency ● ● ● ● ●●●●● ● ● ●●●●● ● ● ●● ●● ● ● ● ●● ●●●●●●● ●● ● ● ●●● ●●●●●● ●●●●●●●●● ●●●● ● ● ● ● ●●● ●●●●●●●● ● ●●● ●●●●● ● ● ● ●● ● ● ●●●●● ●● ●●● ●●●●● ●● ●● ● ● ●●● ●● ●●●●●●●●●●●●●●● ● ● ● ● ● ● ●● ● ● ●●●●●● ●●●●● ● ● ● ● ● ● ● ●● ● ●●● ●● ●●●●● ●●●●●● ● ● ●● ●●●●●●●● ●●●● ● ● ●● ●●●● ● ●● ● ● ●● ● ●● ●● ● ●●●● ●● ● ●● ● ●● ● ●●●● ●● ●● ●●● ● ●● ● ● ●●●● ●● ●●● ● ● ● ●● ●●● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ●●● ●● ● ●● ●●● ●● ● ● ● ● ●●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●●● ● ● ● ●● ● ●● ● ● ● ● ● ●●● ●● ● ● ● ● ●●●● ●●● ●● ● ● ● ● ● ●●● ●● ● ● ● ●● ●●●● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ●●● ● ● ● ● ● ●●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
0.8 0.9 1.0 1.1 ● ● ● ● ● ● ● ● ●
●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●●● ● ● ● ● ● ● ●●● ● ●● ● ●● ● ●●● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ●● ●●● ● ●● ●● ● ●●● ● ● ● ●● ● ● ●●● ● ● ●●● ● ● ● ● ●●● ● ● ●● ●● ●● ● ● ● ●●●● ●●●●● ●●●●●●● ● ● ● ● ● ● ●● ●● ●● ● ● ●● ● ● ●● ● ● ●●●●●●● ● ● ●●● ● ● ● ●●● ●● ●● ●●●●●●●●●● ● ● ● ● ● ●●●● ●●●●●●●●● ●●●● ● ●●●●●● ●●●●● ●●●●●●● ●● ● ● ● ● ● ● ● ●●●●●●●●●●●●●●● ●●● ● ● ●●●●● ●●● ●●●●●●● ●●●●●● ●● ● ● ● ● ● ● ●●●● ●●●●●●● ●●● ●● ●● ●●● ● ● ● ● ●●●● ● ●●●●●●●●● ●● ● ●●● ● ● ●● ●●●●● ●●● ●● ● ● ● ● ● ● ●●● ●●●●●●●●●●●● ●● ● ● ●●●●● ●●●●●●●●●● ● ● ●● ● ●●● ● ● ●● ● ●●●● ● ●● ● ● gamma ● ● ●● ●●●●●●●●● ● ●● ●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●● ●● ● ●●● ●●●● ●●● ●●● ●● ● ●●●●● ●● ● ● ● ●●●●●●●●● ●●●●●●●●● ● ●●●●● ●● ● ● ● ●●●●● ●●●● ●●●●●●●●●●●●●● ● ● ● ●● ● ●● ●●●●●●●●●●● ●●●●● ●●● ●●●●●● ● ● ● ● ●●●●●●●●●●● ●●● ●●●●●●●●●●●●● ●●●● ● ● ●●● ●●●●●●●●●●●●●●●●●●●●●● ●● ● ● ● ● ●●●●●●●●●● ●●●●●●● ●●● ●●●●●●● ● ● ●● ● ●●● ●●●●●●● ●●●●●●●●●●●●●●●●● ●● ● ●●●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ● ● ● ●● ● ●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ●●● ●●●●●● ●●●●●●●x●●●●● ●● ●● ●●●●●●●●●●● ●●● ●● ● ●●● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ● ● ●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●● ● ● ● ● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●● ● ●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●● ● ●●●●●●●●●●●●●●● ●●●●●●●● ●●●●●●● ● ● ● ● ● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●●●● ●● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ● ● ● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ●● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ● ● ●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●● ●●●● ●●●●●●●●● ● ● ● ●●●● ●●● ● ●●●●●●●●●● ●●●●●●●●●●●●● ●●●●●●●●●●●●●●● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ●●●● ●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●● ●● ●● ● ●● ●● ●●●●●● ●●●●●●●●●●●● ●●●●●●●●●●●● ●●●● ●● ●●●●● ● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●●● ●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ●●● ●●●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ●●●● ●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ● ●● ●● ● ●● ●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●● ●●● ●● ● ●●● ●● ●●● ● ● ●● ●●●●● ● ●●● ●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●● ●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●● ●● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●●●● ● ● ● ●● ● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●● ●●●●●●●●● ●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●●●●●● ●●● ●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ● ●●●●●● ●● ●●●●●●●●●●●●● ●●●●●●●● ●●●●●● ● ●●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●●●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●●●●●●●● ●●●● ●●●●●● ● ●●●●●●●●● ● ● ●● ● ● ●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●●● ●●●●●●●●●●●●●●●●●●●●●● ● ●●● ●●●●● ●●●●●●●●●●● ●●●● ●●●●●●● ●● ● ●● ● ●●●●●●●●●●●● ● ●●● ●●●●●● ●●● ●●● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ● ● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ●● ●● ● ● ● ●● ●●●● ●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●● ● ●●●●●●●●●●● ● ●●● ● ●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●● ●●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ● ●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●● ●●●●● ●● ● ● ●● ●●● ●●●●●● ● ●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ●● ●●● ● ●●●●● ●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●● ●● ●● ● ● ●●●●●●●●●●●●●●●●●●●● ● ●●●●●●●●●●●●●●●●● ● ● ● ●● ●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ● ●●●● ●●●●●● ● ●● ●●●●●●●●●●●●●● ●●● ● ● ● ● ● ●●●●● ●●●●●●●●●●●●●●●●●● ●● ●●● ● ● ●●●●●●●●●●●●●●●●●● ● ● ● ●● ● ●●●●●●●● ●●●●●● ●● ●● ●●● ● ● ● ●● ● ●● ● ●●●●●●●● ●●● ●●●●●● ● ● ● ● ● ●●●●●●●●●●●●●●●●● ● ●●● ● ●●●●●● ●●●●● ● ●● ●● ● ● ●● ● ● ●● ●●●● ●●●●●●●●●●● ● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●●● ●●●●●●●●●● ●●●●●● ●●● ●●●● ●● ● ● ● ●●●● ●● ●●●●●●●●●● ● ●●●●●●●● ● ● ●●● ●●●●●●●●●●● ●●●●● ● ● ● ● ● ●●●● ● ● ●●●●● ●●● ● ● ● ● ● ●●●●●●●●●●●●●●●●●● ● ● ● ● ●● ●●●● ●●●●●● ●● ● ● ● ●●●●●● ●●● ● ●● ● ● ● ● ●● ●●●●●● ● ● ● ●● ● ●●●●●●●●●●●●● ●● ●● ●● ● ● ● ●● ●●●● ● ● ●●●● ● ● ● ●●●●● ● ●● ● ●●● ●●●● ●●● ●● ●●● ●●●●●●●●●●●● ●●● ●●● ● ● ●●●●● ● ● ●● ●●● ● ● ● ●●● ●● ●●● ●●●●● ● ● ● ●●●● ●● ● ● ● ● ●●● ● ●●● ●● ● ● ●●●●●●● ● ● ● ● ●●●●●●● ●●●● ●● ● ●● ●● ● ● ●● ●●● ●●● ● ● ● ● ● ●● ● ● ● ●●●● ●● ●●● ●● ● ● ●● ●●● ● ● ● ●● ●●● ● ●● ●● ● ● ● ● ●● ●●● ● ● ●●● ●●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●●●●●●● ● ● ●● ●● ●●● ● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● Frequency ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ●●● ● ● ●● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ●●● ●● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ● 0.80 0.85 0.90
●● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●●● ● ● ● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ● ●●●● ●● ● ●● ●●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ●●● ● ●●● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●●●● ● ● ● ● ● ●● ● ● ● ● ● ●● ●●●● ●● ● ● ●●● ●● ●● ● ● ● ● ● ● ●●●● ● ●● ● ● ● ●● ● ●● ●●● ● ● ●● ● ●●● ●●●●●● ●●● ●●● ● ● ● ●●●●●●● ● ●● ● ●●●●●●● ● ● ●●●● ●●●●●●● ●●● ● ● ● ● ● ● ● ● ● ●●●●●●●● ●● ●● ● ● ●●●● ●●● ●●● ● ●● ● ● ● ● ●● ●● ● ●●●●●●●● ● ● ● ● ●● ● ●●●●● ● ● ● ●●● ● ●● ● ● ●● ●●● ●●●● ●●●●●● ●● ● ● ● ●● ●●● ●●●●●●● ●●● ●●●● ● ●●●● ●● ●●●●●●●●● ● ● ●● ● ● ●● ● ● ●●●●● ●● ●●●●● ● ●●● ●● ● ● ●● ● ● ● ●●●●●●●●●●●●●●●● ●● ●● ●●● sigma ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●●● ●●●●●●●●● ●● ● ●●●● ● ● ● ● ●●● ●● ●●●●●●●●● ●●●● ●● ●● ● ● ● ● ●● ● ●●●●●●●●● ●● ●●●● ●●●● ● ● ● ●●●●●●●●●●● ●●●●● ●● ● ● ● ● ● ●●●● ● ●● ●● ● ●●●●●●●● ● ● ● ● ● ● ●● ●●● ●●●●●●● ● ● ●●● ● ● ● ● ●●● ●●●●●●● ●●●●●●●●●●●●●●● ●●●● ●● ●● ● ● ● ● ● ●●●● ●●●●●●● ●●●●●●● ●●● ● ●● ●●●● ●● ● ●● ●●●●x●●●●●●● ●●●●●●● ●●● ● ●●●●●● ● ●● ●●●●●●●● ●●●●● ● ●●●● ● ● ● ●●●●●● ● ● ●●●●● ● ●●● ●● ● ● ● ●● ● ●● ●● ●●●●●●●●●● ● ●●● ● ● ● ●● ● ●●●●●●●●●●●● ● ●●●●●●●●●● ● ●●● ●● ● ● ●● ● ● ● ●● ●●● ●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●● ● ● ●● ● ● ● ●●● ●●●●● ●●●●●●●●●●●●●●●● ●●●●●●●●●● ●●●●● ●●●●● ●●●●●●●●●●●●●●●●●●● ● ●● ●● ● ● ● ●●● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●● ● ●● ● ● ● ●●●● ●●●● ●●●●●●●●●●● ●●●●●●●●●● ● ●● ● ● ●● ●● ●●●●●●●●●●●●●●●●●●●●●● ●●●●● ●● ● ●● ● ●●● ●● ●●● ●●●●●●●●●●● ●●● ●●●●●●● ● ●●●● ● ● ● ●● ●●●●●●●●●●●● ●●● ●●●●●●● ●● ● ●● ●●●●●●●● ●●●●●●●●●●● ●● ●●●●●●●●●●●●● ●● ● ● ● ● ● ● ●● ●● ● ●●●●●●●●●●● ●●●●● ●●●●●● ●●● ● ●● ● ● ● ●● ●● ● ●●●●● ● ●●●●● ●●●● ● ●●●●●●●● ●●●●●●●● ●● ●● ●●●●●●●●●● ●●●● ●●●●●●●● ●●●●●●●●● ● ● ●● ●● ● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●●● ● ● ●●●●● ●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●● ●● ● ● ●●●●●●●●●● ●●●●● ●● ●●●●●●●●●●●●●● ● ●●●● ●● ● ● ● ●●●● ●●●●●●● ●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●● ● ● ● ●●●●●●●●●●●●●●●●●●●●●● ●●●●●●● ●●●●●●●●●● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●● ● ●●● ● ●●●● ●●● ●●●●●●●●●●●●●●●●●●●●●●● ● ●●●● ● ●●●● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●● ● ●● ● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ●●● ● ●●●● ● ● ● ● ● ●● ●●● ●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●● ● ● ●●●● ●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●● ● ● ● ● ● ● ● ●●●●●●●●●● ●●●● ●●●●●● ●●●●●●●●●●●●● ● ● ●● ● ●● ●●●● ●●●●● ●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ● ●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ● ●● ● ● ● ●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●● ●● ●●●●●●●● ● ● ● ● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ● ● ●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●● ●●● ● ● ● ● ● ● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●● ●●● ● ●● ● ● ●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●● ● ● ● ● ● ● ●●● ●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●● ● ● ● ● ●● ●● ● ●●● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ● ● ●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●● ●●●●●● ●●●● ●●●●●●●●●●● ●●●●●●●●●●● ●●●●●●●● ● ● ●● ● ● ● ● ●●●●● ● ● ●●●●●●●●●●●●● ●●●●●●●●●●●● ●●●●● ● ● ● ● ●●●●●●●●●●●●●● ●●●●●●●●●●● ●●●●●●●●●●●●●●● ●●●● ● ● ●● ●●●● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ● ● ● ●● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●● ●●●●●●●● ● ● ● ● ●●●●●●● ●●● ●●●●● ●●●●●●●●●●●●●●●● ●●●●● ●●● ●● ●●●●●●●●● ●●●●●●●●●● ●●●●●●●●●●● ●●●●●●●●● ●●● ●●● ●● ● ● ●●●●●●● ●●●●●● ● ●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ● ● ● ● ● ●● ●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●● ●●●● ●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●● ●●● ● ● ●● ●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ● ● ● ● ●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●● ● ●●● ● ● ● ● ● ●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●● ● ●●●● ● ● ● ● ●●● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●● ● ● ● ● ●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●● ● ● ●● ● ●● ●●● ●● ●●● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●● ●●● ●● ● ● ● ●● ● ●●●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●● ●● ●●● ● ●●●●●● ●●●●● ●●●●●●●●●●●●●●●●●●●● ●● ●●●● ● ● ●● ●● ●● ● ● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●● ●● ● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●● ● ● ●●●●●●●●●●●●● ●●●● ● ●●●●●●●●● ●● ●● ● ● ● ● ●●● ●● ● ●● ●●●●●●●●●●●● ●●● ●●● ●●●● ● ● ● ● ● ● ●●●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ●●●●●● ●●● ● ● ●●●●●●●●●●●●●●●● ●●● ● ●● ● ● ● ● ● ●●●● ●●●●●●●● ●●●●●●●●●●●●●●● ● ● ● ● ● ● ● ● ● ●● ●●●● ● ●●● ●●●●●●●●● ●●●● ● ●●● ●● ●●●●●●● ●●● ●●●●●●●●●● ●●●●●● ●● ●●●●● ●● ● ●● ●● ● ●● ●●●●●●●●●●●● ●●●●●●●●●●● ●●●●●●●●●●● ● ●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ●●●●●● ●●●● ●●●●●●●●●●●● ● ●● ●●●●●● ● ● ●●● ●● ●●● ●●●●●●●●●●●●●●●●●●●●● ●● ● ●● ● ● ● ●●●●●●●●● ●●●●●●● ● ●●●●●●●●● ● ● ● ● ● ● ● ●●●●●●● ●● ●●●●●●●●●●● ●● ● ● ● ●● ● ●●●●●●●●●●●●●●●● ●●● ●● ● ●● ● ● ●● ●●●● ●●●●●●●●●●●●●● ●●●● ● ● ● ●●● ●●●● ● ●● ● ● ● ●● ●● ●●●● ● ● ● ● ● ● ● ●●● ●●●●● ● ●●● ●●●●● ●●●● ●● ●● ● ● ● ●●●● ● ●●●● ●●●●●● ●●●● ●●●● ● ● ● ● ● ● ● ● ● ●●● ●●●●● ●● ●●●●● ●● ● ●● ● ●●●●●● ● ● ●●●●● ● ● ● ● ● ● ●●●●● ● ●●●●●●●● ●● ●● ● ● ●●●●● ●●●● ●●●●●●● ● ● ● ●●●● ●●● ● ● ●●● ● ● ●●●●●●● ● ●●● ● ●●● ● ● ●●● ● ● ● ● ● ● ● ●●● ●●●●●●●●●●●● ●● ●● ● ● ● ●●● ●●● ●●●● ●●● ● ● ● ● ● ●● ● ● ● ●●●●●● ●●● ● ●● ●● ● ● ● ● ● ●● ● ●●●●● ● ● ● ● ● ● ● ●● ● ●●●●● ● ● ●●● ●●● ● ● ● ● ● ●● ●● ● ● ● ●●● ● ●●● ●●● ●● ● ●● ● ●● ●●● ● ●● ●● ●●●●● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ●● ●●● ●●●● ●●● ● ●● ● ● ●●●● ●●● ● ● ●●● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ●● ● ●●● ● ● ● ● ● ● ●●● ●●● ● ● ●● ● ● ● ● ●●● ●● ● ● ● ● Frequency ● ● ●●● ● ●●● ●● ● ● ● ● ●● ●● ●● ●●●● ● ● ● ● ● ●●●● ●● ● ●● ●● ● ●●● ● ●● ● ●● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●●● ● ● ● ● ● ●●● ● ● ●● ● ● ●● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0.06 0.08 0.10 0.12
● ● ●
2.5 2.6 2.7 2.8 0.80 0.85 0.90
x
bootstrap samples are simulated from a parametric model which mimics the data generating process as much as possible. Probably, parametric bootstrap should be the recommended method for hierarchical data.
The upper panel of Figure 10 displays a data set that has been analyzed by Gelman and Hill (2007). The data correspond to repeated measurements of HIV positive children during a period of two years. The outcome variable is a measurement of the immune system (CD4 percentage of cells). There are two treatment groups: The control group corresponds to children without zinc supplement dietary and the treatment group corresponds to children with zinc supplement. It is expected that a diet with zinc supplement will improve the response of the immune system. Clearly, the data are very noisy and it is difficult to observe differences between groups.
The lower panel of Figure 10 shows a marginal smoothing of the data. A surprising feature is spotted out in this plot, the group with zinc supplement seems to reach a minimum value after one year of treatment and then they start to recover, while the control group continually decays. We are going to analyze if this feature is just incidental or represents an important consequence of this treatment. We make a bootstrap analysis on the zinc supplement group and we set up the analysis by 30
Figure 10. Upper Panel: The Data Correspond to Repeated Measurements of HIV Positive Children During a Period of Two Years. The Outcome Variable is a Measurement of the Immune System (CD4 Percentage of Cells). There are Two Treatment Groups: One Control Group Cor- responds to Children Without Zinc Supplement Dietary and One Treatment Group Corresponds to Children With Zinc Supplement. Lower Panel: Smoothed Curves Represent the Average Response by Time Point for Each Treatment Group. The Group With Zinc Supplementary Diet Seems to Recover After One Year of Treatment
HIV positive children (1 to 5 years old)
0.0 0.5 1.0 1.5 2.0 control zinc
8
6
4 sqrt(cd4%) 2
0 0.0 0.5 1.0 1.5 2.0 time(years)
HIV positive children Control dieat Zinc supplement
5.4 5.2 5.0
y 4.8 4.6 4.4
0.0 0.5 1.0 1.5 2.0 time
fitting the following mode: 2 Yi,j = (β0 + ai) + (β1 + bi)ti,j + β2ti,j + i,j, 2 ai ∼ NOR(0, σa), 2 bi ∼ NOR(0, σb ), 2 i,j ∼ NOR(0, σ ). In this model each child has its own linear trajectory with common population in- 2 2 tercept β0, slope β1 and quadratic term β2. The components of variance σa and σb 2 are used to model between children variation and σ intra-children measurement error. The parameter of interest is β θ = − 1 , 2β2 31
Figure 11. Bootstrap Distribution of the Recovery Time Point of the Treatment Group With Zinc Supplementary Diet. Dashed Vertical Lines Correspond to the 95% CI Calculated With the Percentile Method
Bootstrap distribution: recovery point Density 0.0 0.5 1.0 1.5 2.0 2.5
0.0 0.5 1.0 1.5 2.0
min.time.c which corresponds to the theoretical time point where the minimum proportion of CD4 cells is achieved. If this value is stable, we expect that the children start to recover by following a zinc supplementary diet. In the parametric bootstrap approach model’s parameters are fitted by maximum likelihood (or restricted maximum likelihood). Bootstrap data is generated by simulating from the postulated probability model. A major problem is to estimate model parameters in each bootstrap sample in an efficient way. The function simulate of the package lme4, generates parametric bootstrap samples. The function refit can be used to calculate θ in each bootstrap repli- cate. We implement this approach in the function boot.min: boot.min <- function(model.fitted) { boot.data <- simulate(model.fitted) # Generate bootstrap data fn <- refit(model.fitted, boot.data) # Fit the model beta.0 <- fn@fixef[1] # Extract coefficients beta.1 <- fn@fixef[2] beta.2 <- fn@fixef[3] -1*beta.1/(2*beta.2) # Calculate theta } 32
The only argument of boot.min is the model object generated by lmer: M1 <- lmer(y ˜ time + I(timeˆ2) + (1 + time | person), data = hiv.dat, subset=tr=="zinc")
Once we implement the resampling version of θbwe can use the R function replicate to generate the bootstrap distribution as follows: # Take 1500 bootstrap samples and calculate theta min.time <- replicate(1500, boot.min(M1))
The vector min.time contains the 1500 replications of θ. The bootstrap distri- bution is presented in Figure 11. The bootstrap analysis shows that there is a strong evidence of a recovery point about one year after treatment.
7. Summary
In this tutorial review we have omitted several topics, such as: missing data, cross-validation, model choice, variable selection in regression, censor data, time series, spatial data, efficient computations,etc. But we hope that we have covered the following main points: • How automatic is the bootstrap? Probably it is fully automatic for estima- tion of standard error and bias. For more sophisticated statistical problems the human intervention that we called Analytics in Section 3 becomes very important. • A deep understanding of the bootstrap in a particular application is worth more than published simulation experiments and theoretical asymptotic re- sults. • Each application is a new problem and diagnostic techniques like those pre- sented in this paper become relevant when one apply bootstrap techniques. • Simulation techniques do not replace traditional ideas in statistics. Con- cepts like populations, parameters, pivot quantities, likelihood, and so on play a central role in bootstrap analysis.
Acknowledgments
The author is very grateful to the Editorial Board of Estad´ısticafor the invitation to write this tutorial paper, in particular to Ver´onicaBeritich for her help and patience during the editorial process and to Joelle Murray for her proof reading. This research supported in part by the German Research Foundation DFG Oh 39/11–1. 33
References
BABU, G. and SINGH, K. (1983). “Inference on means using the bootstrap.” Annals of Statistics. 11: 999–1003. BOSS, D. D. and MONAHAN, J. F. (1986). “Bootstrap methods using prior information.” Biometrika. 73: 77–83. CANTY, A. J. (2002). “Resampling methods in R: The boot package.” R News. 2(3): 2–7. DAVISON, A. and HINKLEY, A. (1997). Bootstrap Methods and Their Applica- tion. Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press.
DAVISON, A. and KUONEN, D. (2002). “An introduction to the bootstrap with applications in R.” Statistical Computing & Statistical Graphics Newsletter. 13: 6–11.
DICICCIO, T., MONTI, D., and YOUNG, G. (2006). “Variance stabilization for a scalar parameter.” Journal of the Royal Statistical Society. 68: 281–303. DICICCIO, T. and ROMANO, J. (1995). “On bootstrap procedures for second- order accurate confidence limits in parametric models.” Statistica Sinica. 5: 141– 160. EFRON, B. (1979). “Bootstrap methods: another look at the jackknife.” Annals of Statistics. 7: 1–16. EFRON, B. (1987). “Better bootstrap confidence intervals (with discussion).” Journal of the American Statistical Association. 82: 171–200. EFRON, B. (1992). “Jackknife-after-bootstrap standard errors and influence func- tions (with discussion).” Journal of the Royal Statistical Society. Series B. 54: 83– 127.
EFRON, B. (1993). “Bayes and likelihood calculations from confidence intervals.” Biometrika. 80: 3–26.
EFRON, B. and TIBSHIRANI, R. R. (1993). An Introduction to the Bootstrap. Chapman and Hall, New York.
GELMAN, A. and HILL, J. (2007). Data Analysis Using Regression and Multi- level/Hierarchical Models. Cambridge University Press.
HALL, P. (1987). “On the bootstrap and likelihood-based confidence regions.” Annals of Statistics. 18: 121–140.
HALL, P. (1988). “Theoretical comparison of bootstrap confidence intervals (with discussion).” Annals of Statistics. 16: 927–985. PAWITAN, Y. (2000). “Computing empirical likelihood from the bootstrap.” Statistics & Probability Letters. 47: 337–345.
34
R DEVELOPMENT CORE TEAM (2012). R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing. Vienna, Austria. SPIEGELHALTER, D. J., THOMAS, A., and BEST, N. (2004). Winbugs, version 1.4, upgraded to 1.4.1, user manual. MRC Biostatistics Unit, Cambridge. TIBSHIRANI, R. (1988). “Variance stabilization and the bootstrap.” Biome- trika. 75: 433–444. VENABLES, W. and RIPLEY, B. (2002). Modern Applied Statistics with S. Fourth Edition. Springer-Verlag, New York.
Invited paper Received July 2012 Revised May 2013
ESTADÍSTICA (2012), 64, 182 y 183, pp. 57-84 © Instituto Interamericano de Estadística
COMPARACIÓN DE MÉTODOS DE ESTIMACIÓN DEL MODELO DE LEE-CARTER (ARGENTINA)
BLACONÁ, M.T. Consejo de Investigaciones, Facultad de Ciencias Económicas y Estadística Universidad Nacional de Rosario, Argentina [email protected]
ANDREOZZI, L. Consejo de Investigaciones, Facultad de Ciencias Económicas y Estadística Universidad Nacional de Rosario, Argentina [email protected]
RESUMEN
Se estiman las tasas de mortalidad en la República Argentina para el período 1979- 2009 utilizando el modelo propuesto por Lee y Carter. Las estimaciones de los parámetros del modelo permiten describir la tendencia y el patrón de cambio de la mortalidad. Se obtienen estimaciones de los parámetros del modelo para total, varones y mujeres mediante el método clásico, mínimos cuadrados ponderados (MCP) y máxima verosimilitud-modelo log-bilineal Poisson (MV-LBP), a través de dos algoritmos iterativos BFGS y NM. El comportamiento de los residuos es similar para ambos métodos de estimación, y las medidas de error resultan levemente más pequeñas para el caso de la estimación por MV-LBP.La ventaja que presenta la estimación alternativa se refleja en el cálculo de las variancias. Las mismas resultan en todos casos menores a las variancias calculadas para las estimaciones por el método clásico, esto se puede deber a que este método contempla la heterocedasticidad presente en los datos.
Palabras Clave
Índice de mortalidad; máxima verosimilitud log-bilineal Poisson; mínimos cuadrados ponderados; algoritmos de optimización.
58 ESTADÍSTICA (2012), 64, 182 y 183, pp. 57-84
ABSTRACT
Mortality Rates in Argentina are estimated for the period 1979-2009 using the model proposed by Lee and Carter. Estimates of these parameters can describe the trend and pattern of change in mortality. Estimates of the parameters of the model for both gender, men and women are obtained through the traditional method of least squares (WLS) and maximum likelihood-log-bilinear Poisson model (MV- LBP) using two iterative algorithms BFGS and NM. The residuals behavior is similar for both methods of estimation and error measures are slightly smaller in the case of the MV estimation LBP. The advantage with the alternative estimate is reflected in the calculation of the variances. They are in all cases smaller than the variances for estimates calculated by the classical method, this may be because this method take into account the heteroscedasticity in the data.
Keywords
Mortality rate; maximum likelihood Poisson log-bilinear; weighted least squares; optimization algorithms.
I. Introducción
Predecir con exactitud el proceso de envejecimiento de la población es ahora más que nunca una preocupación de los gobiernos nacionales, por sus repercusiones económicas y sociales. El modelo de Lee-Carter (1992) es un método estadístico sólido, formal, relativamente reciente y ampliamente usado en diversas partes del mundo, que permite describir el comportamiento de la mortalidad a través del tiempo por género y edad. Además resulta de gran utilidad en epidemiología, por ejemplo, para estudiar el comportamiento de las tasas de mortalidad por causa de muerte, lo que es aplicado en la promoción y prevención de la salud.
El modelo ha sido perfeccionado a través de los años, utilizando nuevos y más eficaces métodos de estimación. El método de pronóstico de Lee y Carter tiene un desarrollo estadístico riguroso, debido a que está basado en un modelo explícito que permite no solamente el cálculo de pronósticos puntuales, sino que produce también medidas de incertidumbre y sirve como base para realizar inferencias en general.
En los diversos países donde se ha aplicado,E.E.U.U. (Lee y Carter, 1992), Canadá (Lee y Nault, 1993), Chile (Lee y Rofman, 1994), Japón (Wilmoth 1996), Bélgica (Brouhns y otros, 2001) y México (González y Guerrero, 2007), este método ha proporcionado mejores resultados que los métodos tradicionales, en términos de BLACONA et al.: Comparación de métodos de estimación del modelo... 59
precisión estadística de las proyecciones. Los modelos de predicción utilizados oficialmente presentan por lo general sobreestimación de la mortalidad y principalmente falta de medidas de sensibilidad e incertidumbre de las características estimadas. En este trabajo se avanza sobre la aplicación del modelo de Lee-Carter en Argentina (Andreozzi y otros, 2011) implementando métodos alternativos de estimación, que requieren de la utilización de algoritmos iterativos, para los cuáles se presentan varias opciones.
En la sección II se presenta el modelo de Lee-Carter, su método de estimación clásico y los métodos alternativos, los algoritmos iterativos utilizados, el software empleado para su implementación y las medidas de bondad de ajuste que se tienen en cuenta en la evaluación de las diferentes estimaciones obtenidas. En el punto III se desarrolla el análisis empírico, en el mismo se describen los datos, las tasas de mortalidad específicas estimadas por edad y género, se compara algoritmos y métodos de estimación y se culmina con el análisis de los residuos. En la sección IV se presentan las conclusiones.
II. Estimación del modelo
II.1 Primera estimación del modelo
Lee y Carter (1992) propusieron un modelo simple para describir el cambio secular en la mortalidad total, como función de un único parámetro kt que varía en el tiempo . A kt se lo denomina índice de mortalidad general. Dicho modelo describe el logaritmo de la serie de las tasas de mortalidad específicas por edad como:
=( % ) =++ ε = = fxt,ln m xt , abk x xt xt , x1,..., X yt1,..., T , (1)
% donde mx, t es la tasa de mortalidad específicaen el intervalo de edad x durante el tiempo t; ax describe el patrón general de la mortalidad promediado a través del tiempo; bx representa cuán rápido varía la mortalidad para cada intervalo de edad ε frente a cambios en el índice de mortalidad general; x, t es el término de error. Con X y T se indican el número de categorías de edad y la cantidad de años evaluados respectivamente.
60 ESTADÍSTICA (2012), 64, 182 y 183, pp. 57-84
La tasa de mortalidad específica se define como: d % = x, t mx, t , (2) Ex, t
donde dx, t es el número de muertos con edad x en el período t y el Ex, t número de individuos en la población con edad x en la mitad del período t.
En la ecuación (1) los parámetros bx y kt admiten infinitos valores posibles. Para que el modelo quede determinado se deben incluir restricciones para dichos b = 1 k = 0 parámetros. Para ello Lee-Carter (1992) proponen ∑ x x y ∑t t , también utilizadas por Butt y Haberman (2009).
Sujeto a estas restricciones, el modelo se puede ajustar minimizando la siguiente suma de cuadrados:
X T 2 − − ∑∑ fxt, a x bk kt . (3) x=1 t = 1
Lee y Carter propusieron este método para realizar el pronóstico del índice de mortalidad general, pero también se puede utilizar para pronosticar las tendencias según causas específicas de muerte.
Para obtener las estimaciones de los parámetros es necesario minimizar la suma de cuadrados (3). Como no hay variables observables del lado derecho de la ecuación que define el modelo (1), no se pueden utilizar los modelos de regresión ordinarios. Lee y Carter (1992), proponen una solución simple que se puede hallar utilizando el primer elemento de una Descomposición en Valores Singulares (Lawson y Hanson, 1974) de una matriz construida a partir de las tasas y las estimaciones de los parámetros ax .
En primer lugar se estiman los parámetros ax como:
T = 1 aˆx∑ f x, t . (4) T t=1
BLACONA et al.: Comparación de métodos de estimación del modelo... 61
A partir de estas estimaciones se define la matriz:
fa−ˆ... fa − ˆ 1,1 1 1,T 1 G = M O M . (5) − − faX,1ˆ X... fa XTX , ˆ
Las estimaciones de bx y kt se obtienen a partir de la descomposición en valores singulares de la matriz G(Koissi et al, 2006):
r = ρ DVSG xt, ∑ ixiti UV , , , (6) i=1 r= rango [G] i= {1,..., r } donde y ρi con son los valores singulares de la matriz U V G en orden decreciente. x, i y t, i son respectivamente los vectores singulares izquierdo y derecho correspondientes a ρi . La aproximación de la matriz se puede obtener mediante arreglos con dimensión máxima igual al rango de la matriz G , teniendo en cuenta la magnitud de los valores singulares. A partir de ρ ˆ = ˆ = ρ aproximar DVS G x, t ~ 1Ux ,1 V t ,1 se estima bx U x ,1 y kt1 V t ,1 . Para cumplir con la restricción de que la suma de los bx del modelo sea 1, se dividen las ˆ ˆ componentes de bx por su suma y se multiplica a kt por la suma de las ˆ componentes de bk para mantener la relación de igualdad (6).
Con esta metodología de estimación es necesaria una segunda etapa en la que se impone una restricción que permita obtener un índice de manera tal que el número de muertes observadas sea igual a las esperadas (Lee y Miller, 2001):
X =() + dt∑ E xt, exp abk xxt (7) x=1 donde dt es el total de muertes observadas en el año t y Ex, t son los expuestos al riesgo en el período t para el intervalo de edad x.
62 ESTADÍSTICA (2012), 64, 182 y 183, pp. 57-84
II.2. Métodos alternativos de estimación
II.2.1. Mínimos Cuadrados Ponderados
La implementación de Mínimos Cuadrados Ponderados (MCP) resuelve el problema que genera el uso de DVS, minimizando la siguiente suma de cuadrados de errores (Wilmoth, 1993):
X T 2 − − ∑∑ dxt, f xt , a x bk xt , (8) x=1 t = 1
Sujeta a las mismas restricciones que se impusieron a los parámetros en la estimación propuesta por Lee-Carter (1992).
Puede ocurrir que no se presenten muertes para un determinado año y una determinada categoría de edad, sin embargo, un número nulo de defunciones puede constituir una tasa de mortalidad. La elección de dx, t está estadísticamente justificada utilizando el siguiente resultado demostrado por Wilmoth (1993):
≈ 1 var fx, t . (9) dx, t
En este caso se aplica el método convencional de MCP y se evita la segunda etapa de estimación del método presentado en la sección II.1 para el cálculo del índice general de mortalidad.
II.2.2. Máxima Verosimilitud Log-Bilineal-Poisson
Para la estimación del modelo de Lee-Carter (1992), por el método clásico LC y por MCP, se supone que los errores tienen un comportamiento homocedástico, es decir poseen la misma variancia a través de todas las edades, supuesto que no siempre se cumple. Alho (2000) sugiere utilizar Máxima Verosimilitud Log- Bilineal-Poisson (MV-LBP). Este método se basa en suponer que la variable aleatoria Dx, t , número de defunciones en el intervalo de edad x en el período t, λ tiene una distribución de Poisson con media x, t . P ermite incorporar la λ = heterocedasticidad al modelo ya que xt,m xt , E xt , BLACONA et al.: Comparación de métodos de estimación del modelo... 63
m=exp( abk + ) donde xt, x xt , denominada tasa de mortalidad subyacente (Wilmoth, 1993).
La función de verosimilitud para una única combinación de edad-tiempo se puede escribir como:
λ d e−λ L() d ,λ = . (10) d !
De forma similar la función de log-verosimilitud es
ldd( ,λ) = ln( λ) − λ − ln( d ! ) . (11)
Asumiendo independencia entra las observaciones, se suma a través de las distintas edades y tiempos y se obtiene la log-verosimilitud total de la forma:
l= dln(λ) − λ − ln( d ! ) ∑ xt xt, xt ,, xt xt , . (12)
λ Entonces maximizar la log-verosimilitud (12) con respecto a x, t equivale a maximizar:
d ln (λ) − λ ∑ xt xt, xt , xt , . (13)
λ Si no hay restricciones sobre x, t , se verifica que la ecuación alcanza su máximo λ = valor cuando xt,d xt , . Por otro lado para el modelo de Lee Carter, se requiere
+ λ = = ax b x k t xt,,,mE xtxt e E xt , , (14) en consecuencia las estimaciones máximo verosímiles de los parámetros del + λ ax b x k t modelo de Lee-Carter se encuentran sustituyendo x, t por e E x, t en la ecuación (13) y maximizándola con respecto a ax , bx y kt . Este enfoque se conoce también como modelo log-bilineal de Poisson y se describe en Brouhns y otros (2002).
64 ESTADÍSTICA (2012), 64, 182 y 183, pp. 57-84
II.3. Algoritmos iterativos
Por ser el modelo no lineal se deben utilizar algoritmos de optimización, entre los algoritmos más difundidos se encuentran los métodos conocidos como “Quasi- Newton” y “Simplex”, ambas rutinas de minimización alcanzan resultados similares.
II.4. Software: paquetes y funciones
En el presente trabajo las estimaciones se obtienen a partir del uso de funciones de optimización disponibles en R (DevelopmentCoreTeam, 2008). Para la implementación de los métodos alternativos de estimación propuestos por Wilmoth (1993) se utilizan el algoritmo NM (Nelder-Mead, 1965) y el algoritmo BFGS (Broyden, 1970; Fletcher, 1970; Goldfarb, 1970 y Shanno, 1970). Existen distintos paquetes de R que aplican estos algoritmos. Para la elección de las funciones y los paquetes a aplicar se tomaron en cuenta aspectos tales como, los requisitos de los valores iniciales y la posibilidad de incluir o no restricciones. Los paquetes que finalmente fueron seleccionados son alabama (augmentedlagrangianadaptivebarrierminimizationalgorithm) con su función “auglag” para la aplicación del algoritmo BFGS y dfoptim (derivate free optimization) función “nmk”, que permite la aplicación del algoritmo NM.
La función “auglag” permite incluir restricciones a los parámetros de la función a optimizar de manera directa, ya sean igualdades o desigualdades, y es el mismo algoritmo el que las adapta con cada iteración. Mientras que en el caso de la función “nmk” se incluyen las restricciones de los parámetros mediante un sumando ponderado agregado a la misma función a optimizar. Para mantener la estabilidad del algoritmo la ponderación se debe ir incrementando en forma progresiva.
II.5 Bondad de ajuste
Para medir y comparar la bondad de ajuste de los modelos estadísticos se debe probar que los residuos son independientes e idénticamente distribuidos (i.i.d.) Estas condiciones se pueden evaluar utilizando gráficos de contorno. Los mismos emplean tonalidades de grises para indicar la magnitud de los residuos. Los tonos más claros indican residuos cercanos a cero y a medida que se hacen más oscuros indican valores de los residuos que se vuelven más grandes en valor absoluto. El sentido positivo o negativo se debe indicar con colores opuestos (en esta publicación se incluyen dichos gráficos en escala de grises). Si en los mismos se detecta un patrón en los residuos o bien franjas del mismo tono de gris, pueden BLACONA et al.: Comparación de métodos de estimación del modelo... 65
estar significando falta de independencia en los mismos y/o atribuirse a la existencia de interacción entre edad y tiempo.
Es posible calcular además de los residuos tradicionales (valores observados menos estimados bajo el modelo) residuos específicos para cada método de estimación, como por ejemplo en la estimación por MCP se calculan los residuos ponderados: