A latent class approach for driver injury severity analysis in highway single vehicle crash considering unobserved heterogeneity and temporal influence

Overview

• Why

  • Temporal variation - (major source) -> [unobserved heterogeneity](# unobserved heterogeneity) (need to be paid attention to)

• What

  • develop a [latent class](# latent class model，LCM) mixed logit model with temporal indicators - ( investigate ) - > single-vehicle crashes + effects of significant contributing factors - (to) -> driver injury severity

• How

  • Crash data: ( 2010 ~ 2016; Washington D.C ; 31,115 single-vehicle crashes )
  • A latent class random-parameter model with temporal indicators was motivated to incorporate the possible temporal variation as well as unobserved heterogeneity.
    • latent class structure captures the across-class unobserved heterogeneity
    • the incorporated random parameters with heterogeneity relaxed the ability of traditional latent class models in capturing within class unobserved heterogeneity
    • temporal indicators in class probability function demonstrate the temporal variation in the effects of significant factors. The current study adds to the growing studies that contributing to temporal instability.

• Achieve

  • model (two-class) is able to interpret within- & across- class unobserved heterogeneity + temporal variation

• Find

  • two temporal indicators (male & driver’s age indicators ) show significant influence on latent class probability functions
  • urban indicator and principle type are found to be random parameters

• backgroud

  • Single-vehicle crashes are more fatality-concentrated
  • Characteristically diffrenent from multi-vehicle crash

Previous Studies review

Methods	Applications
multinomial logit models	investigate effects of significant factors in single-vehicle crashes
nested logit models	address (✂️) the endogenous correlations among severity outcomes
ordered logit and probit models	considering the intuitive ordering (from 0 to death) of injury outcomes

Problem:

• temporal instability (params changing over time)
• in driver injury severity analysis domain, few efforts-> investigate heterogeneities of various impact factors introduced by temporal variation,
  • i.e., single-vehicle crash injury severity analysis allowing time-varying interactions among variables

Data

A total of 31,115 single-vehicle crash records were extracted, involving 131 fatality crashes, 534 serious injury crashes, 2474 evident injury crashes, 3485 possible injury crashes and 24,491 no injury crashes.

• general crash information

  • crash severity in terms of five accident-severity categories (i.e., no injury, possible injury, evident injury, serious injury and fatality)
  • collision type
  • temporal information
  • county name

• environmental information

  • weather
  • surface condition
  • lighting condition
  • speed limits
  • roadway characteristics
  • indicators for work zone

• vehicle information

  • vehicle type
  • vehicle age
  • airbag condition
  • ejection status

• driver and passenger information

  • age
  • gender
  • seat belt usage
  • license status
  • insurance
  • and passengers
  • restrain
  • sobriety conditions

Methodology

Results of data analysis

Terminology

• unobserved heterogeneity

  • heterogeneity: variableility, e.g Teams different from others
  • is unobserved when we don't observed its cost
  • Impact: *

• latent class model，LCM

  • 在统计学中，潜在类别模型（latent class model，LCM ），简称潜类模型，将一组观察到的（通常是离散的）多变量变量与一组潜变量联系起来（不能观察，只能推测出的变量） 。LCM是一种潜变量模型 。因为潜在变量是离散的，所以它被称为潜类模型。类的特征在于条件概率模式，其指示变量对特定值的可能性。