SlideShare a Scribd company logo
1 of 1
Download to read offline
Joint Models for the Duration and Size of BC Forest Fires
Dexen Xi, Western University, London, ON
Charmaine Dean, Western University, London, ON
Steve Taylor, Pacific Forestry Centre, Saanich, BC
β€’ Recently there has been rapid development on the development of methods for the
joint analysis of linked outcomes, which has attracted the interest of researchers in
forest fire management because of the applicability of these tools for decision-making
processes related to wildland fire management.
β€’ Our goal is to jointly model time spent (Duration) and area burned (Size) from ground
attack to final control of a fire as a bivariate survival outcome using random effects to
link these outcomes.
β€’ The research question in a fire management context is to investigate the effects of
environmental variables on both the Duration and the Size of the β€œbig and long”
(Duration >2 days and Size > 4 hectares) lighting-caused fires.
Background and Objectives
Fig.1. Plot of fire locations, and a scatter plot of the outcomes with a log base 10 scale for
both axes. There is a moderate positive relationship between the two outcomes with a
Pearson correlation of 0.46 for the 912 β€œbig and long” (BAL) fires. Blue triangles are fires
considered to be ended by rain (accumulated > 12mm precipitation in 5 days).
Fig.2. Trajectories for the Drought Code (DC) and the Duff Moisture Code (DMC) of 100
randomly selected BAL fires through their burn days. The two indices represent the state of
the fuel available for combustion. Their values rise as the fire danger increases. The
trajectories demonstrate a positive linear trend, with sharp jumps to a lower value at certain
days, for some trajectories.
Acknowledgments
β€’ Historical government records from 1953 to 2000 are
obtained from about 200 provincial BC Wildfire Service
and 50 Environment Canada stations in BC
Fire Data
Preliminary Analysis
β€’ Survival (time-to-event, lifetime, life history, reliability) data: The time from an origin to
the occurrence of some event
β€’ Survivor function: 𝑆 𝑑 = 𝑃 𝑇 β‰₯ 𝑑
o Non-parametric: 𝑆 𝑑 =
# of observation β‰₯𝑑
𝑛
(where there is no censoring)
o Parametric: 𝑇 ~ Parametric familiy
β€’ Let π‘‡π‘–π‘˜, π‘˜ = 𝐷, 𝑆 be respectively the Duration and the Size outcome. We model them
separately through a log-location-scale model, or sometimes called, accelerated
failure time (AFT) model:
log π‘‡π‘–π‘˜ = πœ‡ π‘˜ + 𝛽 π‘˜
𝑇
π‘₯𝑖 + 𝜎 π‘˜ πœ€π‘–π‘˜
β€’ where
β€’ the baseline density functions, 𝑓0 πœ€π‘–π‘˜ , are chosen to be Normal(0, 1)
β€’ π‘₯𝑖 = 1 if fire 𝑖 is ended by rain, π‘₯𝑖 = 0 otherwise. 𝛽 π‘˜ is the corresponding coefficient
β€’ πœ‡ π‘˜ and 𝜎 π‘˜ are the overall location and scale parameters
β€’ Let 𝜽 π’Œ = πœ‡ π‘˜, 𝛽 π‘˜, 𝜎 π‘˜
𝑇. It follows that π‘‡π‘–π‘˜|π‘₯𝑖, 𝜽 π’Œ~ i.i.d.LogNormal πœ‡ π‘˜ + 𝛽 π‘˜
𝑇
π‘₯𝑖, 𝜎 π‘˜
2
Reference links for picture used:
NASA/GSFC/METI/Japan Space Systems, and U.S./Japan ASTER Science Team – http://asterweb.jpl.nasa.gov/gallery-detail.asp?name=okanagan
https://en.wikipedia.org/wiki/Fossil_record_of_fire#/media/File:Deerfire_high_res_edit.jpg
https://pixabay.com/en/british-columbia-canada-barkervillie-1155230/
Fig.3. The non-parametric (NP) and the parametric estimates of the survival functions of
the outcomes with a log base 10 scaling for the x-axis. The rain effect translates the
quantiles of the survival functions by a constant amount, which supports the using of the
AFT model. The Log-Normal distribution seems to fit well. LRT: Likelihood ratio test
statistic corresponding to test for significance of the covariate effect.
Joint Together, Joint With The Giants
β€’ If we want to estimate the two survival models simultaneously, we need to account
for the dependence between the two outcomes by conditioning on a latent variable.
β€’ We can treat the responses from an individual fire as a cluster of two outcomes, and
link the two survival models using a random effect, similar to the use of a frailty for
linking multiple individuals in the same cluster.
β€’ Furthermore, this framework is similar to that used for joint modeling of longitudinal
and survival data.
β€’ The use of the lognormal distribution for the outcome also enables a direct comparison
with a marginal approach where copulas are used to study the joint distribution of the
outcomes.
Joint Models
β€’ For instance, expand the previous model:
log π‘‡π‘–π‘˜ = πœ‡ π‘˜ + 𝛽 π‘˜
𝑇
π‘₯𝑖 + 𝑏𝑖 + 𝜎 π‘˜ πœ€π‘–π‘˜
β€’ where
β€’ π‘π’Š ~ i.i.d.Normal 0, πœŽπ‘
2
be the shared frailty
β€’ Let 𝜽 π’Œ = πœ‡ π‘˜, 𝛽 π‘˜, 𝜎 π‘˜, π‘π’Š
𝑇. It follows that π‘‡π‘–π‘˜|π‘₯𝑖, 𝜽 π’Œ~ i.i.d.LogNormal πœ‡ π‘˜ + 𝛽 π‘˜
𝑇
π‘₯𝑖 + π‘π’Š, 𝜎 π‘˜
2
β€’ To perform the estimation under a Bayesian paradigm using Gibbs Sampling in JAGS,
we initially assume naive priors. Parameter estimates and model diagnostics for πœŽπ‘
2
are displayed below.
β€’ For future work, one can model the trajectories of the time-varying covariates and
include its information as random effects in the survival models. For instance:
log π‘‡π‘–π‘˜ = πœ‡ π‘˜ + 𝜷 π‘˜
𝑇
𝒙𝑖 + 𝜢 π‘˜
𝑇
𝒄1𝑖 + 𝑏𝑖 + 𝜎 π‘˜ πœ€π‘–π‘˜
𝒛𝑖 𝑑 = 𝒄0𝑖 + 𝒄1𝑖 𝑑 + 𝝃𝑖
β€’ where
β€’ π’™π’Š = (π‘₯𝑖1, … , π‘₯𝑖𝑃) 𝑇 is a vector of 𝑃 time-constant covariates.
β€’ π’›π’Š 𝑑 = (π‘§π’ŠπŸ 𝑑 , … , π‘§π’Šπ‘Έ 𝑑 ) 𝑇, 𝑑 = 1, … , π‘šπ‘– is a vector of 𝑄 longitudinal variables
β€’ 𝒄0𝑖 = (𝑐0𝑖1, … , 𝑐0𝑖𝑄) 𝑇 and 𝒄1𝑖 = (𝑐1𝑖1, … , 𝑐1𝑖𝑄) 𝑇 are vectors of 𝑄 y-intercepts and 𝑄
slopes of the longitudinal model for π’›π’Š 𝑑 .
β€’ 𝜷 π‘˜
𝑇
= (𝛽 π‘˜1, … , 𝛽 π‘˜π‘ƒ) are the corresponding coefficients of π’™π’Š
β€’ 𝜢 π‘˜
𝑇
= (𝛼 π‘˜1, … , 𝛼 π‘˜π‘„) are the corresponding coefficients of 𝒄𝑖
β€’ 𝝃𝑖 = (πœ‰π’ŠπŸ, … , πœ‰π’Šπ‘Έ) 𝑇
where πœ‰π’Šπ’’ ~i.i.d.Normal 0, πœŽπœ‰
2
is the error term for the longitudinal
model
Fig.4. Selected JAGS output, with 3 chains, number of adapts = 10000, burn-in = 5000
and sample = 100000/3. The 95% credible interval for πœŽπ‘ does not contain 0, which
suggests that the frailty is significant and its variance does explain variability in each
outcome.
Parameter 95% Credible Interval
πœ‡ 𝐷 1.97 2.09
πœ‡ 𝑆 4.38 4.66
𝛽 𝐷 0.02 0.34
𝛽𝑆 -0.47 0.25
𝜎 𝐷 0.02 0.21
πœŽπ‘† 1.69 1.85
πœŽπ‘ 0.83 0.91
The authors acknowledge the assistance and support of the Pacific
Forestry Centre in conducting this research. Support also provided by the
Natural Sciences and Engineering Research of Council and the Ontario
Provincial Government. Thanks to Steve Taylor of the Pacific Forestry
Centre for very helpful discussions.

More Related Content

What's hot

Chap 1 intro_to_engineering_calculations_1_student
Chap 1 intro_to_engineering_calculations_1_studentChap 1 intro_to_engineering_calculations_1_student
Chap 1 intro_to_engineering_calculations_1_studentHelena Francis
Β 
Quark Model Three Body Calculations for the Hypertriton Bound State
Quark Model Three Body Calculations for the Hypertriton Bound StateQuark Model Three Body Calculations for the Hypertriton Bound State
Quark Model Three Body Calculations for the Hypertriton Bound StateIOSR Journals
Β 
Group Cohomology of the Poincare Group and Invariant States
Group Cohomology of the Poincare Group and Invariant States Group Cohomology of the Poincare Group and Invariant States
Group Cohomology of the Poincare Group and Invariant States James Moffat
Β 
Alam afrizal tambahan
Alam afrizal tambahanAlam afrizal tambahan
Alam afrizal tambahanAlam Afrizal
Β 
Erickson_FYP_Poster
Erickson_FYP_PosterErickson_FYP_Poster
Erickson_FYP_PosterSam Erickson
Β 
Chapter 1(3)DIMENSIONAL ANALYSIS
Chapter 1(3)DIMENSIONAL ANALYSISChapter 1(3)DIMENSIONAL ANALYSIS
Chapter 1(3)DIMENSIONAL ANALYSISFIKRI RABIATUL ADAWIAH
Β 
Energy Grid Theorem (Smart Grid)
Energy Grid Theorem (Smart Grid)Energy Grid Theorem (Smart Grid)
Energy Grid Theorem (Smart Grid)Anatoly Kazakov
Β 
Simulated annealing.ppt
Simulated annealing.pptSimulated annealing.ppt
Simulated annealing.pptKaal Nath
Β 
Unit-1 : Units and Dimensions
Unit-1 : Units and DimensionsUnit-1 : Units and Dimensions
Unit-1 : Units and DimensionsArosek Padhi
Β 
units and dimensions
units and dimensionsunits and dimensions
units and dimensionsKrishna Gali
Β 
Parameter Estimation for the Exponential distribution model Using Least-Squar...
Parameter Estimation for the Exponential distribution model Using Least-Squar...Parameter Estimation for the Exponential distribution model Using Least-Squar...
Parameter Estimation for the Exponential distribution model Using Least-Squar...IJMERJOURNAL
Β 
Density Functional Theory
Density Functional TheoryDensity Functional Theory
Density Functional Theorykrishslide
Β 
1 dimensions and units
1 dimensions and units1 dimensions and units
1 dimensions and unitsYusri Yusup
Β 
Simulated annealing
Simulated annealingSimulated annealing
Simulated annealingKirill Netreba
Β 
Quantum calculations and calculational chemistry
Quantum calculations and calculational chemistryQuantum calculations and calculational chemistry
Quantum calculations and calculational chemistrynazanin25
Β 

What's hot (20)

Chap 1 intro_to_engineering_calculations_1_student
Chap 1 intro_to_engineering_calculations_1_studentChap 1 intro_to_engineering_calculations_1_student
Chap 1 intro_to_engineering_calculations_1_student
Β 
Quark Model Three Body Calculations for the Hypertriton Bound State
Quark Model Three Body Calculations for the Hypertriton Bound StateQuark Model Three Body Calculations for the Hypertriton Bound State
Quark Model Three Body Calculations for the Hypertriton Bound State
Β 
Timeline of atomic models
Timeline of atomic modelsTimeline of atomic models
Timeline of atomic models
Β 
Group Cohomology of the Poincare Group and Invariant States
Group Cohomology of the Poincare Group and Invariant States Group Cohomology of the Poincare Group and Invariant States
Group Cohomology of the Poincare Group and Invariant States
Β 
Alam afrizal tambahan
Alam afrizal tambahanAlam afrizal tambahan
Alam afrizal tambahan
Β 
Erickson_FYP_Poster
Erickson_FYP_PosterErickson_FYP_Poster
Erickson_FYP_Poster
Β 
C1h2
C1h2C1h2
C1h2
Β 
Chapter 1(3)DIMENSIONAL ANALYSIS
Chapter 1(3)DIMENSIONAL ANALYSISChapter 1(3)DIMENSIONAL ANALYSIS
Chapter 1(3)DIMENSIONAL ANALYSIS
Β 
Energy Grid Theorem (Smart Grid)
Energy Grid Theorem (Smart Grid)Energy Grid Theorem (Smart Grid)
Energy Grid Theorem (Smart Grid)
Β 
Simulated annealing.ppt
Simulated annealing.pptSimulated annealing.ppt
Simulated annealing.ppt
Β 
Unit-1 : Units and Dimensions
Unit-1 : Units and DimensionsUnit-1 : Units and Dimensions
Unit-1 : Units and Dimensions
Β 
units and dimensions
units and dimensionsunits and dimensions
units and dimensions
Β 
Parameter Estimation for the Exponential distribution model Using Least-Squar...
Parameter Estimation for the Exponential distribution model Using Least-Squar...Parameter Estimation for the Exponential distribution model Using Least-Squar...
Parameter Estimation for the Exponential distribution model Using Least-Squar...
Β 
Concentrations
ConcentrationsConcentrations
Concentrations
Β 
Density Functional Theory
Density Functional TheoryDensity Functional Theory
Density Functional Theory
Β 
1 dimensions and units
1 dimensions and units1 dimensions and units
1 dimensions and units
Β 
Simulated annealing
Simulated annealingSimulated annealing
Simulated annealing
Β 
A0320105
A0320105A0320105
A0320105
Β 
Quantum calculations and calculational chemistry
Quantum calculations and calculational chemistryQuantum calculations and calculational chemistry
Quantum calculations and calculational chemistry
Β 
Lo #1
Lo #1Lo #1
Lo #1
Β 

Viewers also liked

The DeakinDigital Value Proposition
The DeakinDigital Value PropositionThe DeakinDigital Value Proposition
The DeakinDigital Value PropositionDeakinDigital
Β 
Certificate - RICS - Advanced Course in Quantity Survey
Certificate - RICS - Advanced Course in Quantity SurveyCertificate - RICS - Advanced Course in Quantity Survey
Certificate - RICS - Advanced Course in Quantity SurveySiddhartha Chakraborty
Β 
LincolnAnniversaryOpEd2015
LincolnAnniversaryOpEd2015LincolnAnniversaryOpEd2015
LincolnAnniversaryOpEd2015Dave Boe
Β 
CrashCourse_0622
CrashCourse_0622CrashCourse_0622
CrashCourse_0622Dexen Xi
Β 
27022015 - Resume Giovanni King
27022015 - Resume Giovanni King27022015 - Resume Giovanni King
27022015 - Resume Giovanni KingGiovanni King
Β 
GuΓ­a no.3 herramientas Ofimaticas
GuΓ­a no.3 herramientas OfimaticasGuΓ­a no.3 herramientas Ofimaticas
GuΓ­a no.3 herramientas Ofimaticasalejandrarc96
Β 
Ebcs broucher
Ebcs broucherEbcs broucher
Ebcs broucherRamesh Babu
Β 
Tipos de sistemas operativos jc
Tipos de sistemas operativos jcTipos de sistemas operativos jc
Tipos de sistemas operativos jcJosemanuel Cortes
Β 
Reset sesion 4 Elige Bien parte 1
Reset sesion 4 Elige Bien parte 1Reset sesion 4 Elige Bien parte 1
Reset sesion 4 Elige Bien parte 1Jonathan Ecv Vida
Β 
Ashram Ramesar Cv
Ashram Ramesar CvAshram Ramesar Cv
Ashram Ramesar Cvashram ramesar
Β 
Richard Horowitz: Why Charities Should Be On Instagram
Richard Horowitz: Why Charities Should Be On InstagramRichard Horowitz: Why Charities Should Be On Instagram
Richard Horowitz: Why Charities Should Be On InstagramRichard Horowitz
Β 
summer project_poster 0723dean
summer project_poster 0723deansummer project_poster 0723dean
summer project_poster 0723deanDexen Xi
Β 

Viewers also liked (13)

The DeakinDigital Value Proposition
The DeakinDigital Value PropositionThe DeakinDigital Value Proposition
The DeakinDigital Value Proposition
Β 
Certificate - RICS - Advanced Course in Quantity Survey
Certificate - RICS - Advanced Course in Quantity SurveyCertificate - RICS - Advanced Course in Quantity Survey
Certificate - RICS - Advanced Course in Quantity Survey
Β 
LincolnAnniversaryOpEd2015
LincolnAnniversaryOpEd2015LincolnAnniversaryOpEd2015
LincolnAnniversaryOpEd2015
Β 
CrashCourse_0622
CrashCourse_0622CrashCourse_0622
CrashCourse_0622
Β 
27022015 - Resume Giovanni King
27022015 - Resume Giovanni King27022015 - Resume Giovanni King
27022015 - Resume Giovanni King
Β 
GuΓ­a no.3 herramientas Ofimaticas
GuΓ­a no.3 herramientas OfimaticasGuΓ­a no.3 herramientas Ofimaticas
GuΓ­a no.3 herramientas Ofimaticas
Β 
Excel 01
Excel 01Excel 01
Excel 01
Β 
Ebcs broucher
Ebcs broucherEbcs broucher
Ebcs broucher
Β 
Tipos de sistemas operativos jc
Tipos de sistemas operativos jcTipos de sistemas operativos jc
Tipos de sistemas operativos jc
Β 
Reset sesion 4 Elige Bien parte 1
Reset sesion 4 Elige Bien parte 1Reset sesion 4 Elige Bien parte 1
Reset sesion 4 Elige Bien parte 1
Β 
Ashram Ramesar Cv
Ashram Ramesar CvAshram Ramesar Cv
Ashram Ramesar Cv
Β 
Richard Horowitz: Why Charities Should Be On Instagram
Richard Horowitz: Why Charities Should Be On InstagramRichard Horowitz: Why Charities Should Be On Instagram
Richard Horowitz: Why Charities Should Be On Instagram
Β 
summer project_poster 0723dean
summer project_poster 0723deansummer project_poster 0723dean
summer project_poster 0723dean
Β 

Similar to SSC20160524

R Packages for Time-Varying Networks and Extremal Dependence
R Packages for Time-Varying Networks and Extremal DependenceR Packages for Time-Varying Networks and Extremal Dependence
R Packages for Time-Varying Networks and Extremal DependenceWork-Bench
Β 
Logarithms in mathematics
Logarithms in mathematics Logarithms in mathematics
Logarithms in mathematics Hiethem Aliraqi
Β 
Elementary statistical inference1
Elementary statistical inference1Elementary statistical inference1
Elementary statistical inference1SEMINARGROOT
Β 
008a (PPT) Dim Analysis & Similitude.pdf
008a (PPT) Dim Analysis & Similitude.pdf008a (PPT) Dim Analysis & Similitude.pdf
008a (PPT) Dim Analysis & Similitude.pdfhappycocoman
Β 
"Application of Gaussian process regression for structural analysis" presente...
"Application of Gaussian process regression for structural analysis" presente..."Application of Gaussian process regression for structural analysis" presente...
"Application of Gaussian process regression for structural analysis" presente...TRUSS ITN
Β 
NIPS KANSAI Reading Group #5: State Aware Imitation Learning
NIPS KANSAI Reading Group #5: State Aware Imitation LearningNIPS KANSAI Reading Group #5: State Aware Imitation Learning
NIPS KANSAI Reading Group #5: State Aware Imitation LearningEiji Uchibe
Β 
Hydraulic similitude and model analysis
Hydraulic similitude and model analysisHydraulic similitude and model analysis
Hydraulic similitude and model analysisMohsin Siddique
Β 
NS 6141 - Physical quantities.pptx
NS 6141 - Physical quantities.pptxNS 6141 - Physical quantities.pptx
NS 6141 - Physical quantities.pptxcharleskadala21
Β 
ENCH 201 -ch 1.pdf
ENCH 201 -ch 1.pdfENCH 201 -ch 1.pdf
ENCH 201 -ch 1.pdfezaldeen2013
Β 
Units , Measurement and Dimensional Analysis
Units , Measurement and Dimensional AnalysisUnits , Measurement and Dimensional Analysis
Units , Measurement and Dimensional AnalysisOleepari
Β 
Nucleation and avalanches in film with labyrintine magnetic domains
Nucleation and avalanches in film with labyrintine magnetic domainsNucleation and avalanches in film with labyrintine magnetic domains
Nucleation and avalanches in film with labyrintine magnetic domainsAndrea Benassi
Β 
03 Data Mining Techniques
03 Data Mining Techniques03 Data Mining Techniques
03 Data Mining TechniquesValerii Klymchuk
Β 
Statistical Description of Turbulent Flow
Statistical Description of Turbulent FlowStatistical Description of Turbulent Flow
Statistical Description of Turbulent FlowKhusro Kamaluddin
Β 
IJCER (www.ijceronline.com) International Journal of computational Engineerin...
IJCER (www.ijceronline.com) International Journal of computational Engineerin...IJCER (www.ijceronline.com) International Journal of computational Engineerin...
IJCER (www.ijceronline.com) International Journal of computational Engineerin...ijceronline
Β 
Thesis Defense
Thesis DefenseThesis Defense
Thesis DefenseAaron Lu
Β 
Role of Tensors in Machine Learning
Role of Tensors in Machine LearningRole of Tensors in Machine Learning
Role of Tensors in Machine LearningAnima Anandkumar
Β 

Similar to SSC20160524 (20)

R Packages for Time-Varying Networks and Extremal Dependence
R Packages for Time-Varying Networks and Extremal DependenceR Packages for Time-Varying Networks and Extremal Dependence
R Packages for Time-Varying Networks and Extremal Dependence
Β 
Logarithms in mathematics
Logarithms in mathematics Logarithms in mathematics
Logarithms in mathematics
Β 
Av 738- Adaptive Filtering - Background Material
Av 738- Adaptive Filtering - Background MaterialAv 738- Adaptive Filtering - Background Material
Av 738- Adaptive Filtering - Background Material
Β 
Pakdd
PakddPakdd
Pakdd
Β 
Elementary statistical inference1
Elementary statistical inference1Elementary statistical inference1
Elementary statistical inference1
Β 
008a (PPT) Dim Analysis & Similitude.pdf
008a (PPT) Dim Analysis & Similitude.pdf008a (PPT) Dim Analysis & Similitude.pdf
008a (PPT) Dim Analysis & Similitude.pdf
Β 
"Application of Gaussian process regression for structural analysis" presente...
"Application of Gaussian process regression for structural analysis" presente..."Application of Gaussian process regression for structural analysis" presente...
"Application of Gaussian process regression for structural analysis" presente...
Β 
NIPS KANSAI Reading Group #5: State Aware Imitation Learning
NIPS KANSAI Reading Group #5: State Aware Imitation LearningNIPS KANSAI Reading Group #5: State Aware Imitation Learning
NIPS KANSAI Reading Group #5: State Aware Imitation Learning
Β 
Hydraulic similitude and model analysis
Hydraulic similitude and model analysisHydraulic similitude and model analysis
Hydraulic similitude and model analysis
Β 
Causality detection
Causality detectionCausality detection
Causality detection
Β 
NS 6141 - Physical quantities.pptx
NS 6141 - Physical quantities.pptxNS 6141 - Physical quantities.pptx
NS 6141 - Physical quantities.pptx
Β 
Presentation1
Presentation1Presentation1
Presentation1
Β 
ENCH 201 -ch 1.pdf
ENCH 201 -ch 1.pdfENCH 201 -ch 1.pdf
ENCH 201 -ch 1.pdf
Β 
Units , Measurement and Dimensional Analysis
Units , Measurement and Dimensional AnalysisUnits , Measurement and Dimensional Analysis
Units , Measurement and Dimensional Analysis
Β 
Nucleation and avalanches in film with labyrintine magnetic domains
Nucleation and avalanches in film with labyrintine magnetic domainsNucleation and avalanches in film with labyrintine magnetic domains
Nucleation and avalanches in film with labyrintine magnetic domains
Β 
03 Data Mining Techniques
03 Data Mining Techniques03 Data Mining Techniques
03 Data Mining Techniques
Β 
Statistical Description of Turbulent Flow
Statistical Description of Turbulent FlowStatistical Description of Turbulent Flow
Statistical Description of Turbulent Flow
Β 
IJCER (www.ijceronline.com) International Journal of computational Engineerin...
IJCER (www.ijceronline.com) International Journal of computational Engineerin...IJCER (www.ijceronline.com) International Journal of computational Engineerin...
IJCER (www.ijceronline.com) International Journal of computational Engineerin...
Β 
Thesis Defense
Thesis DefenseThesis Defense
Thesis Defense
Β 
Role of Tensors in Machine Learning
Role of Tensors in Machine LearningRole of Tensors in Machine Learning
Role of Tensors in Machine Learning
Β 

SSC20160524

  • 1. Joint Models for the Duration and Size of BC Forest Fires Dexen Xi, Western University, London, ON Charmaine Dean, Western University, London, ON Steve Taylor, Pacific Forestry Centre, Saanich, BC β€’ Recently there has been rapid development on the development of methods for the joint analysis of linked outcomes, which has attracted the interest of researchers in forest fire management because of the applicability of these tools for decision-making processes related to wildland fire management. β€’ Our goal is to jointly model time spent (Duration) and area burned (Size) from ground attack to final control of a fire as a bivariate survival outcome using random effects to link these outcomes. β€’ The research question in a fire management context is to investigate the effects of environmental variables on both the Duration and the Size of the β€œbig and long” (Duration >2 days and Size > 4 hectares) lighting-caused fires. Background and Objectives Fig.1. Plot of fire locations, and a scatter plot of the outcomes with a log base 10 scale for both axes. There is a moderate positive relationship between the two outcomes with a Pearson correlation of 0.46 for the 912 β€œbig and long” (BAL) fires. Blue triangles are fires considered to be ended by rain (accumulated > 12mm precipitation in 5 days). Fig.2. Trajectories for the Drought Code (DC) and the Duff Moisture Code (DMC) of 100 randomly selected BAL fires through their burn days. The two indices represent the state of the fuel available for combustion. Their values rise as the fire danger increases. The trajectories demonstrate a positive linear trend, with sharp jumps to a lower value at certain days, for some trajectories. Acknowledgments β€’ Historical government records from 1953 to 2000 are obtained from about 200 provincial BC Wildfire Service and 50 Environment Canada stations in BC Fire Data Preliminary Analysis β€’ Survival (time-to-event, lifetime, life history, reliability) data: The time from an origin to the occurrence of some event β€’ Survivor function: 𝑆 𝑑 = 𝑃 𝑇 β‰₯ 𝑑 o Non-parametric: 𝑆 𝑑 = # of observation β‰₯𝑑 𝑛 (where there is no censoring) o Parametric: 𝑇 ~ Parametric familiy β€’ Let π‘‡π‘–π‘˜, π‘˜ = 𝐷, 𝑆 be respectively the Duration and the Size outcome. We model them separately through a log-location-scale model, or sometimes called, accelerated failure time (AFT) model: log π‘‡π‘–π‘˜ = πœ‡ π‘˜ + 𝛽 π‘˜ 𝑇 π‘₯𝑖 + 𝜎 π‘˜ πœ€π‘–π‘˜ β€’ where β€’ the baseline density functions, 𝑓0 πœ€π‘–π‘˜ , are chosen to be Normal(0, 1) β€’ π‘₯𝑖 = 1 if fire 𝑖 is ended by rain, π‘₯𝑖 = 0 otherwise. 𝛽 π‘˜ is the corresponding coefficient β€’ πœ‡ π‘˜ and 𝜎 π‘˜ are the overall location and scale parameters β€’ Let 𝜽 π’Œ = πœ‡ π‘˜, 𝛽 π‘˜, 𝜎 π‘˜ 𝑇. It follows that π‘‡π‘–π‘˜|π‘₯𝑖, 𝜽 π’Œ~ i.i.d.LogNormal πœ‡ π‘˜ + 𝛽 π‘˜ 𝑇 π‘₯𝑖, 𝜎 π‘˜ 2 Reference links for picture used: NASA/GSFC/METI/Japan Space Systems, and U.S./Japan ASTER Science Team – http://asterweb.jpl.nasa.gov/gallery-detail.asp?name=okanagan https://en.wikipedia.org/wiki/Fossil_record_of_fire#/media/File:Deerfire_high_res_edit.jpg https://pixabay.com/en/british-columbia-canada-barkervillie-1155230/ Fig.3. The non-parametric (NP) and the parametric estimates of the survival functions of the outcomes with a log base 10 scaling for the x-axis. The rain effect translates the quantiles of the survival functions by a constant amount, which supports the using of the AFT model. The Log-Normal distribution seems to fit well. LRT: Likelihood ratio test statistic corresponding to test for significance of the covariate effect. Joint Together, Joint With The Giants β€’ If we want to estimate the two survival models simultaneously, we need to account for the dependence between the two outcomes by conditioning on a latent variable. β€’ We can treat the responses from an individual fire as a cluster of two outcomes, and link the two survival models using a random effect, similar to the use of a frailty for linking multiple individuals in the same cluster. β€’ Furthermore, this framework is similar to that used for joint modeling of longitudinal and survival data. β€’ The use of the lognormal distribution for the outcome also enables a direct comparison with a marginal approach where copulas are used to study the joint distribution of the outcomes. Joint Models β€’ For instance, expand the previous model: log π‘‡π‘–π‘˜ = πœ‡ π‘˜ + 𝛽 π‘˜ 𝑇 π‘₯𝑖 + 𝑏𝑖 + 𝜎 π‘˜ πœ€π‘–π‘˜ β€’ where β€’ π‘π’Š ~ i.i.d.Normal 0, πœŽπ‘ 2 be the shared frailty β€’ Let 𝜽 π’Œ = πœ‡ π‘˜, 𝛽 π‘˜, 𝜎 π‘˜, π‘π’Š 𝑇. It follows that π‘‡π‘–π‘˜|π‘₯𝑖, 𝜽 π’Œ~ i.i.d.LogNormal πœ‡ π‘˜ + 𝛽 π‘˜ 𝑇 π‘₯𝑖 + π‘π’Š, 𝜎 π‘˜ 2 β€’ To perform the estimation under a Bayesian paradigm using Gibbs Sampling in JAGS, we initially assume naive priors. Parameter estimates and model diagnostics for πœŽπ‘ 2 are displayed below. β€’ For future work, one can model the trajectories of the time-varying covariates and include its information as random effects in the survival models. For instance: log π‘‡π‘–π‘˜ = πœ‡ π‘˜ + 𝜷 π‘˜ 𝑇 𝒙𝑖 + 𝜢 π‘˜ 𝑇 𝒄1𝑖 + 𝑏𝑖 + 𝜎 π‘˜ πœ€π‘–π‘˜ 𝒛𝑖 𝑑 = 𝒄0𝑖 + 𝒄1𝑖 𝑑 + 𝝃𝑖 β€’ where β€’ π’™π’Š = (π‘₯𝑖1, … , π‘₯𝑖𝑃) 𝑇 is a vector of 𝑃 time-constant covariates. β€’ π’›π’Š 𝑑 = (π‘§π’ŠπŸ 𝑑 , … , π‘§π’Šπ‘Έ 𝑑 ) 𝑇, 𝑑 = 1, … , π‘šπ‘– is a vector of 𝑄 longitudinal variables β€’ 𝒄0𝑖 = (𝑐0𝑖1, … , 𝑐0𝑖𝑄) 𝑇 and 𝒄1𝑖 = (𝑐1𝑖1, … , 𝑐1𝑖𝑄) 𝑇 are vectors of 𝑄 y-intercepts and 𝑄 slopes of the longitudinal model for π’›π’Š 𝑑 . β€’ 𝜷 π‘˜ 𝑇 = (𝛽 π‘˜1, … , 𝛽 π‘˜π‘ƒ) are the corresponding coefficients of π’™π’Š β€’ 𝜢 π‘˜ 𝑇 = (𝛼 π‘˜1, … , 𝛼 π‘˜π‘„) are the corresponding coefficients of 𝒄𝑖 β€’ 𝝃𝑖 = (πœ‰π’ŠπŸ, … , πœ‰π’Šπ‘Έ) 𝑇 where πœ‰π’Šπ’’ ~i.i.d.Normal 0, πœŽπœ‰ 2 is the error term for the longitudinal model Fig.4. Selected JAGS output, with 3 chains, number of adapts = 10000, burn-in = 5000 and sample = 100000/3. The 95% credible interval for πœŽπ‘ does not contain 0, which suggests that the frailty is significant and its variance does explain variability in each outcome. Parameter 95% Credible Interval πœ‡ 𝐷 1.97 2.09 πœ‡ 𝑆 4.38 4.66 𝛽 𝐷 0.02 0.34 𝛽𝑆 -0.47 0.25 𝜎 𝐷 0.02 0.21 πœŽπ‘† 1.69 1.85 πœŽπ‘ 0.83 0.91 The authors acknowledge the assistance and support of the Pacific Forestry Centre in conducting this research. Support also provided by the Natural Sciences and Engineering Research of Council and the Ontario Provincial Government. Thanks to Steve Taylor of the Pacific Forestry Centre for very helpful discussions.