Understanding the cell phone effect on vehicle fatalities: a Bayesian view

Abstract

This article examines the potential effect of various factors on motor vehicle fatality rates using a rich set of panel data and classical regression analysis combined with Bayesian Extreme Bounds Analysis (EBA), Bayesian Model Averaging (BMA) and Stochastic Search Variable Selection (SSVS) procedures. The variables examined in the models include traditional motor vehicle and socioeconomic factors. In addition, the models address the effects of cell phone usage on such accidents. The use of both classical and Bayesian techniques diminish the model and parameter uncertainties which afflict more conventional modelling methods which rely on only one of the two methods.

Keywords:

• motor vehicle fatality rates
• cell phones
• applied Bayesian methods

JEL Classification::

• L92
• C11
• I18

Acknowledgements

Loeb gratefully acknowledges the research support of a Rutgers University Research Council Grant. Tianda Xing and Stephanie Jensen provided research support funded by the Departments of Economics at the University of Utah and Rutgers University. An earlier version of this article was presented at the 2009 Annual Meeting of the Transportation and Public Utilities Group/Allied Social Sciences Meeting in San Francisco, CA. The authors appreciate the insightful comments from an anonymous referee.

Notes

¹ National Highway Traffic Safety Administration (NHTSA, 2008).

² See, for example, Keeler (1994), Loeb (1985, 1990), Loeb and Gilad (1984) and Garbacz and Kelly (1987).

³ See, for example, Lave (1985), Levy and Asch (1989), Fowles and Loeb (1989), among others.

⁴ See, for example, Cohen and Einav (2003), Evans (1996), Dee (1998) and Loeb (1993, 1995, 2001).

⁵ See, for example, Fowles and Loeb (1989) and Chaloupka et al. (1993).

⁶ See Cellular Telecommunications and Internet Association (CTIA, 2007, Available at http://www.ctia.org).

⁷ In addition, as of 2011, 30 states ban text messaging by drivers.

⁸ See, for example, Consiglio et al. (2003).

⁹ Glassbrenner (2005) has estimated that driver use of just hand-held phones increased from 5% in 2004 to 6% in 2005.

¹⁰ See Violanti (1998, p. 522).

¹¹ See Chapman and Schoefield (1998, p. 817).

¹² See Chapman and Schoefield (1998, p. 818).

¹³ See Poysti (2005, p. 50).

¹⁴ The models presented by Loeb et al. (2009) were evaluated for their conformity to the full ideal conditions associated with the error term, i.e. μ ∼ N(0, σ² I). To examine this, a set of specification error tests were applied to the models, i.e. the Regression Specification Error Test (RESET), the Jarque–Bera test and the Durbin–Watson test. Rejection of the null hypothesis of no specification errors by one or more of these tests resulted in the elimination of the models from consideration. These results were supported as well by Fowles et al. (2008) using Bayesian EBA.

¹⁵ The other fatality rate measures are fatalities per capita, fatalities per vehicle registrations and fatalities per licensed drivers. All measures exhibit, at the national level, a downward trend.

¹⁶ Newly available data on actual phone subscribers for the last 5 years have a correlation with the imputed data of 0.9943.

¹⁷ The per se law refers to a legislation that makes it illegal to drive a vehicle with a blood alcohol level at or above the specified BAC level. BAC is measured in grams per deciliter.

¹⁸ The regions are defined in Appendix B.

¹⁹ The fixed effects model is selected over a random effects model based on a Hausman test with a chi-square value of 639.44.

²⁰ Calculations of EBA were computed in Gauss using MICRO-EBA (Fowles, 1988). The Gauss code is available free on request. Details are provided in Appendix A.

²¹ Dropping a variable forces a very strong prior belief that the coefficient is exactly equal to zero with perfect precision.

²² In MICRO-EBA, the prior precision matrix was set equal to the identity matrix, so the priors are spherically symmetric, centred at zero.

²³ c _i and τ _i are choice variables. In this article the reported results are for c _i  = 10 and τ _i  = (2 log(c)(c ²/(l − c ²))⁻⁵σ _{β i} where the parameter σ _{β i} is the OLS coefficient SD. This choice is consistent with George and McCulloch (1993) and follows their notation.

²⁴ In this article, SSVS was implemented via MCMC methods using R. This code is available on request.

²⁵ The first 500 iterations were deleted as a break-in period so there were a total of 9500 iterations employed in the results reported.

²⁶ The average value of p for this set was 0.12.

²⁷ Figure 2 in Appendix C highlights the findings when all variables are doubtful and presents data favoured EBA bounds.

²⁸ See, for example, Loeb (1995, 2001).

²⁹ Lave (1985), Fowles and Loeb (1989) and Levy and Asch (1989), among others, have examined the effect of speed versus speed variance as well. Data on average speed and the 85% speed are no longer collected by US Department of Transportation (USDOT) and as such speed and speed variance could not be investigated in the current study.

³⁰ Suicide may reflect measures of self worth and thus be associated with risk taking behaviours.

³¹ See Loeb et al. (1994).

³² See Chaloupka et al. (1993) on the effect of alcohol control policies.

³³ In this article, τ is a vector with one 1 and k − 1 zeros that corresponds with the ith parameter of interest.