The April 2010 edition of the ITE Journal has an interesting article in it about LED traffic signal indications (titled Analyzing the Effects of LED Traffic Signals on Urban Intersection Safety). The authors (Eustace, Griffin, & Hovey) found a 71% increase in crashes due to replacing incandescent light bulbs with LED indications. I normally skim the ITE (Institute of Transportation Engineers) Journal, but this article caught my eye and I read the whole thing.
The authors studied the crashes at ten intersections in Middletown, Ohio where the incandescent bulbs were replaced with LED indications in 2003 and 2005. There are two pages of statistical analysis related to the empirical Bayes method (assuming a negative binomial distribution) used. This is a little over my head, so I'll trust the analysis is correct.
There is a relatively small sample size and there could be other factors contributing to the increased crash levels, but this is still an alarming study. I hope FHWA and other agencies will study this further. Traffic engineers assumed the brighter LED indications would be safer than incandescent bulbs(because of better visibility – the issue of snow sticking to them aside). We need to go back to the drawing board if the current style of LED indications are causing a 71% increase in crashes.
Most likely the cause of the increased crashes had nothing to do with the installation of LEDs. Perhaps there were increasing crash trends throughout the study area. Unless all other variables were isolated, anything could have contributed to the increased crashes.
It seems to me that the methodology used in this study was inappropriate. The use of empirical bayes appropriate for studies where regression-to-the-mean has an effect on the data. This is common for locations that are selected for study based on high collision occurance. However, in this study (at least based on the information in the ITE Journal article – which may be incomplete), the intersections selected for study where based on having had the signal heads changed from incandescent to LED, not on collision history. There is no evidence (at least based on the information from the ITE article) that regression-to-the-mean is relevant to these locations. If it is not, then the estimated expected collisions after the conversion is much too low. But regardless of whether empirical bayes was the appropriate methodology or not, the sentence “71% increase in crashes due to replacing incandescent light bulbs with LED indications” is incorrect as the number of collisions actually went down (marginally). It would have been better stated that there were a 71% greater crashes compared to the estimated expected number of crashes. And thus, focus on the way that the expected number of collisions were calculated would be more evident.