statistical inference and hypothesis testing for markov chains with interval censoring application...
TRANSCRIPT
Statistical inference and
hypothesis testing for Markov chains with interval censoring
Application to bridge condition data in the Netherlands.
Monika Skuriat-Olechnowska
Delft University of Technology
TU Delft 2July 28, 2005
Presentation outline
Importance of bridge management Data and Work plane Markov chains Markov deterioration model Conclusions Recommendations Questions
TU Delft 3July 28, 2005
Importance of bridge management
Bridge management is essential for today’s transportation infrastructure system
Bridge management systems (BMSs) help engineers and inspectors organize and analyze data collected about specific bridges
BMSs are used to predict future bridge deterioration patterns and corresponding maintenance needs
TU Delft 4July 28, 2005
Data
Statistical analysis of deterioration data (DISK) Condition rating scheme Data analysis
Include a total of 5986 registered inspection events for 2473 individual superstructures Ignoring the time between inspections there are 3513 registered transitions between
condition states
TU Delft 5July 28, 2005
Work plane
Check Markov property Create a deterioration model using Markov chains Determine and compare the expected deterioration over
time and annual probability of failure Separate bridges into sensible groups and determine the
effect of this grouping on the parameters of the Markov model
Take into account the inspector’s subjectivity into condition rating
TU Delft 6July 28, 2005
Markov chains
The Markov Chain is a discrete-time stochastic process
that takes on a finite or countable number of possible values.
, 0,1,2,...tX t
TU Delft 7July 28, 2005
Markov chains (cont’d)
The main assumption in modelling deterioration using a Markov chain is that the probability of a bridge moving to a future state only depends on the present state and not on its past states.
This is called Markov property and is given by:
1 1 1 1 1 0 0
1
Pr | , ,..., ,
Pr | ( ) 0,1,2,....
ij t t t t
t t ij
P X j X i X i X i X i
X j X i P t t
TU Delft 8July 28, 2005
Markov chains (cont’d)
Verification of the Markov property In order to test the Markov property we need to verify if
the transition probabilities from the present to future state don’t depend on past states.
Test based on the contingency tables Test to verify if a chain is of given order Test to verify if the transition probabilities are constant
in time
1 1Pr | ,t t tX m X j X i
TU Delft 9July 28, 2005
Markov chains (cont’d)
Test based on the contingency tables Generated frequency are from the same distributions; Result: can not reject
Test to verify if a chain is of given order The chain is of first-order; Result: can not reject
Test to verify if the transition probabilities are constant in time Transition probabilities do not depend on time; stationarity Result: can not reject
0 :H
0 :H
0 :H
TU Delft 10July 28, 2005
Markov deterioration model
Transition probability matrix
00 01 0
10 11 1
0 1
P
k
k
k k kk
p p p
p p p
p p p
0.4581 0.1181 0.2881 0.0978 0.0317 0.0062
0.3553 0.1684 0.2921 0.1276 0.0474 0.0092
0.2782 0.0989 0.3604 0.1890 0.0597 0.0137
0.2222 0.078
0 0.2813 0.3097 0.0993 0.0095
0.1096 0.0959 0.2877 0.3425 0.1164 0.0479
0.2857 0.1429 0.1429 0.2500 0.1429 0.0357
TU Delft 11July 28, 2005
Markov deterioration models
Modeling transition probability matrix
00 01 02 03 04 05
11 12 13 14 15
22 23 24 25
33 34 35
44 45
55
0
0 0
0 0 0
0 0 0 0
0 0 0 0 0
p p p p p p
p p p p p
p p p p
p p p
p p
p
01 01
12 12
23 23
34 34
45 45
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
0 0 0 0 1
0 0 0 0 0 1
p p
p p
p p
p p
p p
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
0 0 0 0 1
0 0 0 0 0 1
p p
p p
p p
p p
p p
TU Delft 12July 28, 2005
Markov deterioration models
Modeling transition probability matrixcase State-independent State-dependent
std lower upper state std lower upper
Alldata
0.154 0.00171 0.151 0.158
0 0.410 0.00823 0.394 0.427
1 0.310 0.00532 0.230 0.321
2 0.096 0.00198 0.092 0.100
3 0.033 0.00187 0.029 0.036
4 0.114 0.01435 0.086 0.142
TU Delft 13July 28, 2005
Markov deterioration models
5% lower Expected Lifetime 5% upper
State-independent 15 33,26 57
State-dependent 19 53,83 119
TU Delft 14July 28, 2005
Markov deterioration model Data analysis
Year of construction of the structure (we consider two groups of age; all bridges constructed before 1976 and all bridges constructed starting 1976)
Location of structure: bridges “in the road”- heavy traffic, against bridges “ over the road”- light traffic;
Type of bridge: separated into “concrete viaducts” and “concrete bridges” ;
Use, which means the type of traffic which uses the bridge: traffic only with trucks and cars, and mixture of traffic (also bikes, pedestrians, etc.);
Province in which bridge is located:
Groningen
Friesland
Drenthe
Overijssel
Gelderland
Utrecht
Noord-Holland
Zuid-Holland
Zeeland
Noord-Brabant
Limburg
Flevoland
West-Duitsland
Population density
Higher
Utrecht
Noord-Holland
Zuid-Holland
Noord-Brabant
Limburg
Proximity to the sea
Close to the sea
Groningen
Friesland
Noord-Holland
Zuid-Holland
Zeeland
Flevoland
Lower
Groningen
Friesland
Drenthe
Overijssel
Gelderland
Zeeland
Flevoland
West-Duitsland
Inland
Drenthe
Overijssel
Gelderland
Utrecht
Noord-Brabant
Limburg
West-Duitsland
TU Delft 15July 28, 2005
Markov deterioration model
Model parametersType of data Estimated parameters
All data [0.4104, 0.3102, 0.0963, 0.0328, 0.1143]
Province inwhich bridge
is located
Groningen [0.2912, 0.9476, 0.2608, 0.0050, 0.0000]
Friesland [0.2285, 0.3324, 0.0952, 0.0546, 0.0000]
Drenthe [0.4425, 0.3712, 0.1739, 0.0805, 0.0000]
Overijssel [0.2843, 0.2350, 0.2201, 0.0282, 0.0637]
Gelderland [0.2321, 0.3743, 0.1964, 0.0599, 0.0957]
Utrecht [0.8176, 0.4492, 0.0624, 0.0000, 0.0001]
Noord-Holland [0.5625, 0.2183, 0.0932, 0.0075, 0.1509]
Zuid-Holland [0.5324, 0.2211, 0.0629, 0.0351, 0.1231]
Zeeland [0.5575, 0.6394, 0.0512, 0.0000, 0.8807]
Noord-Brabant [0.5622, 0.3578, 0.0926, 0.0356, 0.3216]
Limburg [0.5059, 0.4049, 0.1290, 0.0263, 0.0000]
Flevoland [1.0000, 0.3461, 0.2899, 0.0688, 0.0000]
West-Duitsland [0.9785, 0.8198, 0.0000, 0.0183, 1.0000]
Type of data Estimated parameters
All data [0.4104, 0.3102, 0.0963, 0.0328, 0.1143]
Construction year
Built before 1976 [0.8026, 0.4135, 0.1027, 0.0335, 0.0816]
Built after 1976 [0.3988, 0.2542, 0.0842, 0.0305, 0.2073]
Location “in the road”- heavy traffic [0.4490, 0.3115, 0.0966, 0.0328, 0.1140]
“ over the road”- light traffic [0.3180, 0.3017, 0.0949, 0.0329, 0.1150]
Type “concrete viaducts” [0.3933, 0.2976, 0.0946, 0.0327, 0.1332]
“concrete bridges” [0.5883, 0.3867, 0.1025, 0.0332, 0.0493]
Use Only with trucks and cars [0.4193, 0.3029, 0.0951, 0.0326, 0.1152]
Mixture of traffic [0.3491, 0.3675, 0.1020, 0.0340, 0.1096]
Type of data Estimated parameters
All data [0.4104, 0.3102, 0.0963, 0.0328, 0.1143]
Population density
Higher [0.5398, 0.2790, 0.0789, 0.0255, 0.1422]
Lower [0.2677, 0.3856, 0.1817, 0.0477, 0.0847]
Proximity to the sea
Close to the sea [0.4739, 0.2501, 0.0787, 0.0233, 0.1176]
Inland [0.3465, 0.3725, 0.1223, 0.0416, 0.1111]
TU Delft 16July 28, 2005
Markov deterioration model
TU Delft 17July 28, 2005
Markov deterioration model
Logistic regression The goal of logistic regression is to correctly predict the
influence of the independent (predictor) variables (covariates) on the dependent variable (transition probability) for individual cases.
1 1 2 2
1 1 2 2
exp( ... )
1 exp( ... )
n n
n n
x x xp
x x x 1 1 2 2log ...1
n n
px x x
p
TU Delft 18July 28, 2005
Markov deterioration model
Logistic regression
Location
(Over the road) No influence No influence No influence No influence No influence
Type of bridge
(Concrete bridges)Influence Influence No influence No influence No influence
Type of traffic
(Mixture of traffic)No influence Influence No influence No influence No influence
Population density (High)
Influence Influence Influence Influence Influence
Construction year (built before 1976)
Influence Influence Influence No influence Influence
01p 12p 23p 34p 45p
TU Delft 19July 28, 2005
Markov deterioration model
Inspector’s subjectivity State assessed by inspector is uncertain
given an actual state.
This part of analysis was an extra subject and due to lack of time is not finish yet. We derived formula from which we can estimate transition probabilities, but it is very long and difficult so it will not be presented.
TU Delft 20July 28, 2005
Conclusions We may assume that Markov property holds for DISK deterioration
database
State-dependent model fits better to the data
Grouping bridges with respect to construction year, type of traffic, location, etc. has influence on the model parameters:
The most statistically significant are covariates “Population density (high)” and “Construction year (built before 1976)”
No influence for covariate “Location (over the road)”
Problem of taking into account subjectivity of inspectors into model parameters is not finish due to time constraints.
TU Delft 21July 28, 2005
Recommendations
Choose another combination of covariates in logistic regression and look on the changes (how this influence on transition probabilities).
Repeat tests for Markov property with a new deterioration data. We didn’t do this due to time constraints.
Model used in this analysis does not include maintenance. It would be interesting to incorporate this into Markov model.
TU Delft 22July 28, 2005
Questions ?
TU Delft 23July 28, 2005
Resulting models (cont’d)
Inspectors subjectivity
1 2 1 0 1Pr | , ,..., , Pr |t t t t t
t t t
X X X X X X X
X Y
0
11 1 1
1 1
( ,..., | ) ( ... Pr | Pr )n
m nm i i
j j j j j j ji j
L Y Y p X k X k
1 1
0
1 1
0
1, 1
1 1, 1
1
1( ,..., | ) ( ... ( ( ) Pr ))
... ( ) Prj j j j
n
j j j j
n
nmm
k k j j jni ju u
k k j j jj
Y Y p P t tp p
P t t