Simulated Car Crashes and Crash Predictors in Drivers with Parkinson’s Disease Qian Shi, Xia Mao, Zugui Zhang, Minggen Lu
Background - have many kinds of cognitive and visual impairments - can alter the abilities on which safe driving depends. Research on relationship of car accidents and neurological diseases - interested in estimating the risk of car crashes for drivers with and without neurological diseases
- mainly rely on results from driving simulations
Iowa Driving Simulator (IDS)
Objectives To estimate the risk (probability) of simulated car crashes for drivers with PD, as well as that for drivers without neurological diseases.
To test the hypothesis that older drivers with mild to moderate PD are at greater risk for simulated car crashes than control participants of similar ages.
Objectives To determine how such crashes are predicted by visual/cognitive measurements.
To compare the estimates of significant predictors obtained by Bayesian model and those obtained by frequentist method.
Data Subjects : 24 participants with PD (age: 66.58 10.31) and 70 participants without dementia (age: 68.59 6.24) Experiments : all participants drove in the same simulated environments with high-fidelity collision avoidance scenarios and were tested on the same batteries of cognitive and visual tasks. Outcome : counts of simulated car crashes within two groups.
Main covariates : age; education level; visual and cognitive measurements.
Methods - First phase Step 1: Determine a “best” transformation, and obtain estimates of simulated car cash risks and crude OR for the two groups.
Transformation(p[i])= alpha + beta×group[i]
Methods - First phase Step 2: Assess the association of simulated car crash risk and Parkinson’s disease status, after adjusting for age, gender and the education level. Association between covariates and response and predictor variables. logit(p[i])= alpha + beta.group×group[i]+beta.age×(age[i]-mean(age)
Alpha ~ dflat() Beta.group ~ dflat() Beta.age ~dflat()
Methods - Second phase Step 1: Determine significant predictors by stepwise selection with logistic regression in SAS . Step 2: Fit a frequentist multivariate logistic regression model including the significant predictors.
Step 3: Fit a Bayesian model including the significant predictors
Results - First phase Comparison of three transformations: Convergence is satisfied well for all three transformations. DICs are very similar (logit:124.297, Probit:124.273 , Cloglog:124.260 ). For ease of interpretation, we chose logit transformation to do subsequent analysis.
* Estimates are based on MCMC 1001-5000 iterations. Point estimates for OR’s are the medians.
Results - First phase Comparison of simulated car crash risk for the two groups Estimates of OR of car crash for the two groups
* Estimates are based on MCMC 3001-10000 iterations.
Results - Second phase Selection of significant predictors. Recall — 30 minutes delay score for Rey Auditory Verbal Learning Test, which is a rigorous measure of anterograde verbal memory.
CS – Contrast sensitivity (CS) is assessed using the Pelli-Robson chart. This test provides a measure of low to medium spatial frequency sensitivity.
Results - Second phase Comparison of frequentist method and Bayesian method Frequentist’s estimates of OR based on multivariate logistic regression model.
Bayesian Estimates of OR based on MCMC 2001-10000 iterations .
Example plots of convergence diagnoses
Conclusions The risk of simulated car crash for Parkinson’s patient is 79.16%, with a 95% credible set of (61.4%, 92.53%).The risk for the control group is 57.02%, with a 95% credible set of (45.16%, 68.61%).
Old drivers with mild to moderate PD are at greater risk for simulated car crashes than control participants of similar ages. (OR=2.989, 95% credible set=(1.059, 10.57))
Conclusions Anterograde verbal memory (recall) and contrast sensitivity are significant predictors of car crashes for people of these ages.
Frequentist method and Bayesian method based on non-informative priors yield similar point estimates of OR for Recall and CS. The Bayesian 95% credible set for CS is slightly shorter than frequentist 95% confidence interval for CS.