Results 1 to 1 of 1

Thread: UBC/IVOAC Analysis of ICBC's Study That Claims RHD/JDM Cars Are Higher Risk than LHD

  1. #1

    User Info Menu

    Default UBC/IVOAC Analysis of ICBC's Study That Claims RHD/JDM Cars Are Higher Risk than LHD

    OK guys it turns out I did have the original ICBC report and UBC's analysis saved in an old archive.
    For those who don't know or remember ICBC created a report "THE SAFETY OF RIGHT-HAND-DRIVE VEHICLES IN BRITISH COLUMBIA" which claims RHD/JDM cars are more likely to be involved in accidents and says driving on the right side is more dangerous essentially.

    Basically UBC finds that ICBC's methodology and data are questionable and concludes ICBC's report simply calls for more proper study, research and investigation but doesn't agree with ICBC's conclusion.

    Here is the full text of UBC's Analaysis of ICBC's Report:
    (original file attached)



    Assessment of the ICBC Report Regarding the Safety of Right
    Hand Drive Vehicles in BC
    Mohua Podder, PhD and Rick White, M.Sc.
    Statistical Consulting and Research Laboratory
    Department of Statistics
    University of British Columbia
    6356 Agricultural Road, Vancouver, BC
    March 26, 2010
    1 Introduction
    This report evaluates the analyses done by ICBC in their repo
    rt titled “The Safety of Right-Hand-
    Drive Vehicles in British Columbia”. The request for the eva
    luation came from the Imported
    Vehicle Owners Association of Canada (IVOAC). The number of
    right-hand-drive (RHD) vehicles
    being imported into Canada over the past few years has been in
    creasing. Imported vehicles over
    15 years old are exempt from the requirement of having a manuf
    acturers plate or letter stating
    that the vehicle is compliant with the Canadian motor vehicl
    e safety standards. This allows RHD
    vehicles to retain their RHD control configuration. The beli
    ef is that potential issues related to
    these vehicles might lead to a greater accident risk for oper
    ators of RHD vehicles and people in
    RHD vehicles might be more severely injured than people in LH
    D vehicles.
    ICBC conducted a study in order to assess the safety issues of
    RHD vehicles in BC and published
    a report in early 2008. The main purpose of the study was to com
    pare RHD to LHD vehicles in
    terms of their risk of crash involvement and their occupant p
    rotection potential. The study applied
    three different methodologies to try to answer these questio
    ns: a Relative Risk Analysis comparing
    RHD and LHD culpable crash rates for the same group of drivers
    ; a Survival Analysis to compare
    the time to first culpable crash between drivers of RHD and LHD
    vehicles; and a Poisson Regression
    Analysis to compare the risk of RHD vehicles to a similar grou
    p of LHD vehicles and their drivers.
    1
    2 Data Issues
    The data for all three analyses were obtained from the ICBC cr
    ash-claim data. Does this database
    accurately reflect all crashes in BC or just those that are rep
    orted? Drivers may choose not to
    report their crash to avoid an increase in insurance rates be
    cause the value of the vehicle is low
    or the repair is inexpensive relative to the increased insur
    ance cost, perhaps the driver doesn’t
    have the appropriate insurance coverage. If the tendency to
    not report a crash is related to the
    drive configuration of the vehicle, it may introduce a bias in
    the data that could be reflected in the
    analyses.
    All three analyses rely on the definition of culpability. A ve
    hicle is culpable in a crash if it is
    assigned at least 50% of the responsibility during the claim
    adjustment process. Is blame assigned
    by objective criteria or is there a subjective component? If
    the assignment of responsibility is
    affected by the drive configuration of the cars involved in the
    crash then all analyses based on this
    definition will be inherently biased. The assignment must be
    based on the circumstances of the
    crash only. That means if a LHD car involved in a crash is deeme
    d non-culpable then that status
    would not change if that car had been a RHD vehicle.
    3 Relative Risk Analysis
    The first method used looks at the relative risk of culpable cr
    ashes to non-culpable crashes between
    RHD and LHD vehicles operated by the same driver. This method
    compares crashes within drivers
    to control for driver differences. All drivers involved in cr
    ashes while operating a RHD vehicle
    since 2001 were included in the analysis. In addition the pri
    ncipal operators (POs) of these RHD
    vehicles, if not already included, were selected to provide
    a set of drivers not involved in a crash
    while driving a RHD vehicle. The complete crash history of ea
    ch driver since 2001 was used in the
    analysis. Each crash was classified as culpable or non-culpa
    ble from the target driver’s perspective.
    After applying several constraints to the data, 359 crashes
    for RHD and 1204 crashes for LHD
    vehicles were identified. Of the 1204 LHD vehicle crashes, 32
    4 were from drivers with RHD ve-
    hicle crashes and 880 were from the RHD PO group, the drivers w
    ithout a crash while driving a
    RHD vehicle. The total number of drivers in each group is not m
    entioned. A cross tabulation of
    culpable/non-culpable crashes versus RHD/LHD vehicle was
    created. A Chi-squared test of inde-
    pendence was used to determine if the type of crash was associ
    ated with the drive configuration
    2
    of the vehicle. This analysis assumes that all events in the t
    able are independent events not con-
    nected by any other factor whereas here we actually have seve
    ral 2x2 tables, one for each driver.
    The data within each table depends on a specific driver but the
    tables themselves are independent.
    In essence, we have a stratified sample. In order to perform th
    e Pearson’s Chi-squared test for
    independence, the stratification is ignored and the individ
    ual 2x2 tables are summed across drivers
    to make a single 2x2 table. A more appropriate analysis for th
    is type of data would account for the
    stratification within the data. A possible analysis method w
    ould be the Cochran-Mantel-Haenszel
    test.
    Another possible concern for this data is if two vehicles inv
    olved in the same accident are
    included in the data. This is probably a rare event and theref
    ore a minor concern but should
    probably be checked. The 2x2 tables for any drivers involved
    in the same accident would no longer
    be independent and the analysis would need to be adjusted acc
    ordingly.
    4 Survival Analysis
    The second method of analysis uses a Cox proportional hazard
    regression model to compare the
    instantaneous crash rate of RHD and LHD vehicles after they a
    re first insured. The response data
    is the time to a culpable crash following the initial insuran
    ce policy purchase by a PO for each
    vehicle. The model has the advantage of including all vehicl
    es even those that never had a culpable
    crash over the course of the study. In addition, the model can
    also include adjustments for other
    factors that might modulate the effect. In this analysis data
    were collected from “all RHD POs
    aged 20 years and older at the time of first policy and all vehic
    les (RHD and LHD) for which
    they were listed as POs”. Only culpable crashes were include
    d as an event. Age and gender were
    included as adjustments in the final model.
    The sampling method described in section 2.2 sounds like a sm
    all set of POs was selected for
    the analysis, each having been a PO of several vehicles of whi
    ch at least one was a RHD vehicle.
    The description of the analysis in section 3.2 implies a diffe
    rent set of drivers was included. Section
    3.2 claims “A total of 23717 drivers were included in the anal
    ysis of which 2882 were associated
    with RHD vehicles”. This section also describes the results
    in terms of vehicles. It is not clear who
    has been sampled or what is being modeled here. Is it time to fir
    st culpable crash of a driver or
    vehicle? Are multiple vehicles of the same driver or PO inclu
    ded in the model? Without a clear
    description of the sampling scheme or data being collected,
    it is hard to comment on the methods.
    3
    If the data is time to first culpable crash of a vehicle and seve
    ral vehicles are associated with
    the same PO, a correlation structure is induced into the data
    which needs to be adjusted for in
    the model. Another concern with POs having multiple vehicle
    s is they may split their time at risk
    between their vehicles which further complicates the analy
    sis. If we think of POs as a stratification
    variable, a stratified Cox model can be applied to deal with so
    me of these issues.
    Ignoring the stratification issue, the analysis assumes tha
    t each vehicle is at risk at all times
    when in fact a vehicle is only at risk when it is being driven. A
    vehicle driven 5 days a week is
    more likely to have a crash than one driven once a week. The cra
    sh risk of a vehicle is also affected
    by where and when it is driven. A vehicle driven in rush hour on
    Monday morning probably has a
    different risk than one driven on Sunday morning, as would a ve
    hicle driven in downtown Vancouver
    compared to one driven in White Rock. Is there a way to guarant
    ee that RHD vehicles are driven
    the same amount of time, under the same conditions and locati
    ons as LHD vehicles? If not, what
    is the affect of these factors on the results?
    Age and Gender were included in the model as covariates. It is
    unclear if the effects presented in
    Table 3 are from a model that contains only main effects or one t
    hat contains interactions with RHD
    vehicles. As main effects, they modulate both the LHD and RHD c
    rash rates. So all statements
    pertaining to the effect of age and gender on RHD vehicles appl
    y equally to LHD vehicles. A model
    with interaction allows separate age and gender effects to be
    estimated and compared for the RHD
    and LHD vehicles. The report states that the RHD vehicle grou
    p contained more males and was
    younger. This causes partial confounding between age, gend
    er and RHD vehicles. Confounding
    affects the model estimates especially in a model with only ma
    in effects. The model estimates
    the crash rate to be higher for males and younger drivers. Con
    founding may cause the model to
    attribute some of these effects to RHD vehicles.
    5 Poisson Regression Analysis
    In the final analysis, a Poisson Regression model is used to co
    mpare the crash rates between RHD
    vehicles and a group of similar LHD vehicles. A RHD group cons
    isting of 748 vehicles, was identified
    and a LHD group consisting of 8933 vehicles, was selected by m
    odel, make, year and body style so
    the proportion of vehicles in that group matched the proport
    ions in the RHD group. The response
    data is the number of crash-claims in a two year period follow
    ing the date of a PO’s first policy
    with that vehicle. Each crash was classified as injury or mate
    rial damage only and culpable or
    4
    non-culpable. Covariates considered in the model were gend
    er, age, region (lower mainland or not),
    speeding contraventions and non-speeding contraventions
    .
    The main concern of the model is to compare the crash rate betw
    een RHD and LHD vehicles.
    However matching model-year-style in vehicles does not gua
    rantee that both RHD and LHD vehicles
    are exposed to equal amount of driving time or are driven in th
    e same locations or in the same
    traffic conditions. The confounding affect of gender and age wi
    th RHD vehicle operation also affects
    a Poisson regression model in a similar fashion as a Cox model
    . The distribution and effects for
    region and traffic contraventions is not presented in the repo
    rt.
    Another concern with Poisson regression models is overdisp
    ersion in the response data. If the
    counts are overdispersed than an overdispersed Poisson or N
    egative Binomial regression model
    needs to be fitted otherwise the standard errors of the effects
    will be underestimated and effect
    significance will be overstated. The report does not indicat
    e if overdispersion was checked in the
    model.
    The analysis is done by vehicle not by driver. The Poisson reg
    ression model assumes the data
    for each vehicle is independent from each other. However a ve
    hicle can have many drivers and a
    driver can operate many vehicles. This may introduce a corre
    lation between the vehicles if the same
    driver operated more than one vehicle in the study. If there a
    re many such drivers, the induced
    correlation could become an issue for this model as well.
    The report says principal operators were also examined to in
    vestigate the differences between
    RHD and LHD vehicles at the driver level but it doesn’t explai
    n how this is done. Does a RHD
    driver always operate a RHD vehicle? If not, how is a crash in a
    LHD vehicle by a RHD driver
    dealt with in the analysis?
    6 Summary
    There are several issues that could affect the results of the s
    tudy conducted by ICBC in the report
    “The Safety of Right-Hand-Drive Vehicles in British Columb
    ia”. It is possible the conclusion would
    remain unchanged even after these issues were resolved but w
    e do not know without actually doing
    those analyses.
    Reporting a crash to ICBC is not mandatory if an insurance cla
    im is not made. Are the factors
    5
    that affect the tendency to make a claim related to the drive co
    nfiguration of the vehicle? How is
    blame assigned in a crash claim? If blame assignment has a sub
    jective component, is it related to
    drive configuration of the vehicle? If either is related to dr
    ive configuration then the data contains
    a bias that could affect the results of any analysis based on th
    e data and may not reflect the true
    risk of a culpable crash of LHD or RHD vehicles in the province
    as a whole.
    In addition to data issues, there are some issues with the ana
    lyses themselves. The relative
    risk analysis completely ignores the repeated measures wit
    hin each subject. This analysis is easily
    corrected by using a Cochran-Mantel-Haenszel test instead
    of a Chi-squared test. There is also a
    minor issue of some data coming from the same crash but this is
    probably a rare event and will
    have little impact on the results.
    The problems with the other two analyses are more subtle and t
    echnical. The drive configuration
    of the vehicle was not evenly distributed between gender and
    age. Although the model adjusted for
    these factors, imbalance can still affect the estimates in th
    e model. If the data had been matched
    on age and gender, the results would be more trustworthy. A st
    ratified or paired analysis might
    then be more appropriate in these cases. A More subtle issue i
    s risk exposure. A vehicle is only at
    risk of a culpable crash when it is driven. The risk exposure a
    lso depends on driving locations and
    conditions. A vehicle driven to work every day during rush ho
    ur is at greater risk than one driven
    once a week on Sunday morning. It is unclear if adequate data c
    an be obtained to adjust for these
    factors but they could have a big impact on the results. Drive
    rs using multiple vehicles or vehicles
    driven by multiple drivers is another issue. This introduce
    a correlation into the data that needs
    to be taken into account. If there are several such vehicles o
    r drivers in the data and the repeated
    measures are properly adjusted for, the results of the analy
    ses could be quite different. Again it is
    unclear if such data is available.
    Overall, the ICBC report suggests that RHD vehicles and thei
    r operators are at a greater risk
    than their LHD counterparts but issues with the data and the a
    nalyses suggest that further study
    is needed. Causation is difficult to establish with observati
    onal data. I would caution the use of
    the ICBC report as anything more than an indication that furt
    her study is needed.
    6


    Attached Files Attached Files
      To view attachments your post count must be 1 or greater. Your post count is 0 momentarily.
    Last edited by jdmvip; 11-13-2017 at 01:48 PM.

Tags for this Thread

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •