My Research

Statistical Methods

A recurring theme in my research is the development and/or implementation of statistical approaches to solving safety problems, as well as the integration of multiple datasets to answer research questions. For example, I have worked on methods applied to the Crash Injury Research Engineering Network (CIREN) data for NHTSA. This work included collaboration with Dr. Michael Elliott (UM Biostatistics and Survey Research Center), Dr. Jonathan Rupp, and a doctoral student, Alexa Resler, on development of a pseudo-weighting method for combining CIREN and NASS-CDS cases and applying weights to the CIREN cases in the combined database (Elliott, Resler, Flannagan & Rupp, 2009). In a new project, funded by the Michigan Institute for Data Science (MIDAS), Dr. Elliott and I are working with a student, Ali Rafei, to develop pseudo-weights for our large naturalistic driving datasets based on the National Household Travel Survey.

In other areas, I have developed a method of using injury data to estimate the crash severity (delta-V) distribution for datasets, such as NASS-GES, that do not include delta-V. This method was used in work done for NHTSA on safety benefits of a Collision Mitigation Braking system for heavy trucks, as well as implementation of the UTMOST model. In the same project, I also developed a method of creating a simulated baseline crash dataset on which to exercise crash-avoidance algorithms.

I was PI on a two-year, $450K project funded by the National Cooperative Highway Research Program (NCHRP), in which I developed a research program on linkage of data systems at the state level. This research was done in collaboration with the University of Utah. The specific target of this data linkage was to tie crash data to medical outcome data to better predict serious injuries in crashes. However, the broad goal of data linkage at the state level serves many additional purposes. The project has involved an interesting combination of statistical applications and understanding logistical challenges in implementation. This work has been presented at the Transportation Research Board (TRB), the Lifesavers Conference, the Traffic Records Forum, and a meeting of the Standing Committee on Highway Traffic Safety (SCOHTS). It was also referenced in the Federal Highway Administration Notice of Proposed Rulemaking (FHWA NPRM) for National Performance Management Measures; Highway Safety Improvement Program [Docket No. FHWA–2013–0020].

The largest opportunity for development of statistical methods is in analysis of driving data. UMTRI has conducted several large driving studies over the last decade, and these studies provide rich data sources for understanding driver behavior. However, the data collected are complex to analyze, and analytical methods are undergoing a period of rapid growth in the literature. I have a number of past and current projects exploring different analytical approaches to understanding driver behavior and response to warning systems. My colleagues and I are employing extreme-value modeling, functional data analysis, hidden Markov models, and a variety of classification algorithm and decision-theoretic approaches (e.g., logistic regression, support vector machines, Bayesian Additive Regression Trees, and random forests). I also collaborate with faculty and students in Statistics and Biostatistics to explore additional methods for analyzing driving data.

Modeling Injury Risk

A continuing core element of my research is modeling of injury outcomes in vehicle crashes. The goal of these modeling efforts is to investigate the best predictors of a particular injury in specific types of crashes, such as fetal death, abdominal injury, or hip injury in full frontal and offset-frontal crashes. These analyses are generally targeted at understanding the benefits of new occupant-protection systems, such as seatbelts and airbags, or mapping the effects of demographic, vehicle, and seating characteristics on injury risk.

One of the key areas of injury risk prediction I have been involved in has been development and assessment of triage algorithms. In 2011, I collaborated with a researcher at General Motors to develop a triage algorithm that has since been implemented by OnStar (Kononen, Flannagan & Wang, 2011). This algorithm uses information available from the Event Data Recorder (EDR) or collected by the OnStar adviser to predict the probability of individuals of a specified age, gender, and stature sustaining a serious injury in a given crash. In conjunction with this work, I served on a Centers for Disease Control (CDC) Expert Panel to design a national study to evaluate Automatic Collision Notification systems.

More recently, I was Principal Investigator on a $713k project to compare injury risk in US- and EU-regulated vehicles. This work was done in collaboration with research groups in Sweden, France, and the UK, with the goal of providing data-driven input for negotiators and decision-makers who are working on the Transatlantic Trade and Investment Partnership (TTIP) agreement (a trade agreement currently being negotiated between the US and EU). In conjunction with that work, I briefed EU and US trade negotiators three times, and the program and results have been documented in articles in the US and European new media. In addition to being high profile and high impact, this work required that I solve several technical problems, including harmonizing international crash databases and developing statistical models from datasets in which raw data could not be shared. For the latter, I led a team of four analysts in Europe to provide summary statistics from each of their datasets that together could be used to produce models as though the original data were combined into one dataset.

Safety Benefits

In 2009, I developed a model called the Unified Theory Mapping Opportunities for Safety Technology (UTMOST). This model was designed to look at the combined or integrated benefits of multiple approaches to automotive safety, including active and passive vehicle-based systems, laws and enforcement, and roadway improvements. This conceptual framework has been used in a number of subsequent projects. For example, in 2013, I studied the safety trade-offs associated with low-mass vehicles and determined that crash avoidance systems in low-mass vehicles can offset the increase in injury risk associated with being in a smaller car. In this study, I looked at how crash-avoidance systems with varying effectiveness will change the nature of crashes experienced by low-mass vehicles and therefore the nature of occupant protection needs for such a vehicle. UTMOST now serves as the unifying theme of work being conducted under multiple projects, each of which expands the capability of the model in some way.

Also in the area of crash-avoidance systems, I have worked on development of crash-avoidance algorithms (e.g., forward collision warning and intersection warnings), as well as assessment of safety benefits of for a wide variety of technologies. Safety-benefits assessment for crash avoidance may involve crash-data analysis, analysis of behavior in response to systems, and analysis of long-term behavioral changes in response to using an active vehicle safety system. Recent projects on which I serve as PI include a $1.2M study funded by the National Highway Traffic Safety Administration (NHTSA), which involves collaborating with GM and OnStar on Lane-Departure Warning (LDW) and Forward-Collision Warning (FCW), and a $900K follow-on project on Crash Imminent Braking. In addition, I was co-I on a study for NHTSA related to Collision Mitigation Braking (CMB) in heavy trucks and I worked with Toyota on a pre-collision sensing system.

One of the unique and crucial elements of a NHTSA-funded study on forward collision warning (FCW) and lane-departure warning (LDW) is that it is the first-ever large-scale Field Operational Test (FOT) using targeted, event-driving data collected telematically. In addition to answering questions about driver response and use of warning systems, this study demonstrated a unique means of collecting data on these systems across a very large sample throughout the US. This work represents another growth opportunity as vehicle technologies proliferate and need to be tested inexpensively.

Public Health

As a result of my relationship with the UM Injury Center and the Michigan Office of Highway Safety Planning (OHSP), I have engaged in a number of analyses related to public-health policy issues. Recent publications report on the potential reduction in fatalities realized from a fleet-wide alcohol interlock program and the predicted effect of reducing the current Michigan primary seatbelt law to a secondary belt law. I am working with several collaborators on an extensive analysis of the consequences of Michigan’s 2012 repeal of the motorcycle helmet law. My recent appointment to the Michigan Impaired Driving Commission allows me to make use of this expertise in practice.

Data Access and Open Data

I am passionate about open data, and a key role of CMISST is to provide access to transportation data for researchers and data users in industry, government, academia and the general public. A critical project funded by the Michigan State Police Office of Highway Safety Planning (OHSP), provides open access to Michigan crash data via a website developed and hosted by CMISST staff. Under my leadership, the project has grown from $250K per year to $600K in FY2016. In particular, we are funded to make significant upgrades to the web tool over the next three years. In addition, OHSP has added $100K/year for technical support for CMISST staff statisticians and $50K/year for special analysis projects that I lead. These projects typically result from legislative issues that arise during the year, and have included an analysis of how crashes might be affected if the 0.08 BAC (blood alcohol content) law sunsets to a 0.01 BAC law, an analysis of lives saved by the primary seat belt law, and an analysis of how crashes might be affected by a shift from a primary to a secondary seat belt law.

I have attended a number of meetings promoting and discussing open data. Recent events include a conference of domain repository holders run by UM’s Inter-university Consortium for Social and Political Research (ICPSR), an Open Data Roundtable hosted by USDOT, and an AV Data Sharing Roundtable hosted by USDOT.

Occupant Posture and Position in Vehicles

My early research at UMTRI primarily dealt with driver posture and position modeling. In that research, I created a model of driver seat position as a function of driver stature and vehicle interior package geometry. This model has been codified as part of the Society of Automotive Engineers (SAE) Recommended Practice J1517. I also modeled driver eye location as a function of stature and vehicle package geometry and helped to update the SAE J941 driver eyellipse, which is used to construct sight lines in vehicle design. Although I no longer do much work in this area, the lessons from these efforts (especially around statistically accounting for variability across humans in design) inform my current work.