Perception-Reaction Time (PRT) Programs
The search for a simple answer to a complex problem has unending appeal. The 2008 financial crisis is a good example. It had many causes, but reliance on computer programs and their models was a prime factor. The "quants" (quantitative analysts) created models that estimated the riskiness of mortgage-backed securities based on limited historical data. Executives at investment banks as well as the quants failed to understand that a model may be somewhat like reality, but it is not reality. Models always contain both explicit and implicit assumptions about what is important and what can be ignored. The financial world made the mistake of relying completely on the models to guide and to justify decisions. As one former quant said, "they got lulled into thinking it was magic."
In the end, the models failed for two reasons. One was that the assumptions were inadequate and the data covered too narrow a range of scenarios. The second was that human behavior is too complex to be captured in a set of equations (as I well know from years of working on artificial intelligence modeling).
Ironically, the originators of financial modeling warned that it was just a tool and that models should not accepted without an accompanying reasoned analysis of the securities. This warning was completely ignored because it was just simpler and easier to take the models at face value than to do the work necessary to read and to analyze the underlying mortgage data upon which the securities were based. The result was the worst economic disaster since 1929.
Blind use of a computer program to determine perception-reaction time (PRT) repeats this mistake. In using a model, it might appear that the user is freed from the need to examine the underlying data and to understand anything about human behavior in general. However, this no more true of PRT than it was of securities. Perhaps the worst aspect of such a program is that it gives the user a sense of certainty that simply cannot exist in a behavior as complex and sensitive to many variables as PRT.
Computer models have limited value, and their results cannot be taken as gospel. The model is ultimately based on a number of underlying research studies which vary in methodology and ecological validity and which must be individually read and interpreted. For example, a program cannot capture most of the issues discussed in Green (2009) when interpreting the Olson & Sivak (1986) and AASHTO, and it cannot capture important factors such as action boundaries, affordances, response-conflicts. This is why programs cannot be used without understanding 1) their inputs, which in this case are the data from many studies of different methodology, and 2) human behavior in general. Does the computer program know that Broen and Chiang (1986) claimed to be studying PRT to unexpected obstacles, but that the authors state that they had told their subjects in advance, "an unexpected obstacle may appear in the vehicle's path, and in that event they should step on the brake and stop as quickly as possible." I have reviewed most of the articles used in computer programs and have concluded that none of the research can be taken at face value any more than Olson & Sivak or Broen & Chang can be taken at face value.
There are many other problems in applying research data to the real-world. Research subjects are abnormally alert and waiting for something to happen. Variability is artificially low. Research studies are designed to minimize variability so that it is easier to obtain a statistically significant result, which makes the research publishable. They routinely screen subjects, throw out aberrant data1, etc. The real world is far more variable than any research data suggest. They have demand characteristics which often force the subjects to act in an unnatural way. Moreover many studies are just as deficient as the subprime mortgages that brought down the financial system - the mere publication of study does not ensure its quality or validity. No one who has ever used a computer need be reminded of the old adage, "garbage in - garbage out," which is better stated as "ignorance in - ignorance out."
Further, a program cannot deal with every situation that is likely to arise on the roadway. It cannot distinguish, for example, between differences in detection and hence PRT of different objects, levels of adaptation, background clutter, experience with a situation, etc., etc. In fact, research studies tend to be conducted under very similar and relatively simple circumstances, so there are many scenarios where there simply are no reliable data. Moreover, there are very few data for PRT in nighttime conditions. One reason for this paucity of data is that canned PRT numbers have little value in situations where vision is constrained by low light and/or by low contrast (Green, 2009). The PRT clock cannot begin until the object is detected, but that time is difficult to determine because there are too many variables at work in contrast detection (and in conspicuity) to have standardized values.
In such cases, the behavior is limited by visibility and not PRT, so the real issue is visibility and not PRT. What, for example, would the PRT be for the 30% of drivers who failed to see the dark pedestrian in the Wood, Tyrrell & Carberry (2005)? The answer, of course, is infinity. How do you average infinity into a population mean for PRT? The answer is: you can't. This is one reason that such studies focus only on response distance, which can be objectively measured, and do not attempt to measure PRT. The researchers know that it is a hopeless task. Similarly, Olson (2006) has commented on the difficulty of determining a PRT in low visibility:
While Olson specifically mentions 'rain, snow, fog, night", there are many other visibility factors such as peripheral location, background clutter, glare, adaptation effects, age, crowding, switching costs, inhibition of return, psychological refractory period, overshadowing, etc. I have also emphasized two especially significant parts of this quote. Note Olson's comment "But when does it start?" In order to have a PRT, there must be a clock starting and clock stopping event. In low visibility, when does the clock starting event occur? That is often impossible to determine with much precision (Green, 2009). One thing is for certain: the twilight 3.2 lux value cannot be used with much confidence.
Further, the comment "gradually more visible" is also important. Anyone familiar with change blindness research will know that gradual changes are much less likely to be noticed. (See the many demonstrations on Youtube, for example). Moreover, Green et al. (2008) has discussed the general psychophysical issues of determining threshold and explained that there is no such thing as a fixed line between visible and not visible. Signal Detection Theory is based on the demonstrated fact that cognitive factors such expectations and predicted outcome modulate visibility. Most road detection situations are ascending method-of-limits tasks, which are known to delay visibility responses. (If you don't know what Signal Detection Theory and method-of-limits are, you should).
Such considerations have led Olson (1989) to further comment:
|"There are a number of conditions (e.g. rain, snow, fog, night) that can severely limit forward visibility and make it difficult to determine when some object or condition entered the driver's field of view. The most common of these is probably nighttime. During the daytime a potential hazard such as a pedestrian standing in the road would be visible for a considerable distance, but at night he/she becomes gradually more visible as the vehicle approaches. When the driver becomes aware of the pedestrian's presence this marks the end of the detection interval. But, when does it start?" |
Moreover, it is not possible to simply snip off PRT and examine it independently of basic perception, cognition and response. A computer program is no substitute for real knowledge of human behavior or for having read and critiqued the original studies. The output of a computer model can never be taken at face value without interpretation. On the surface, determining PRT from a program appears similar to the method described in Green (2008) for determining visibility level using a computer model that contains a set of curves which are fit from underlying contrast sensitivity data. The user inputs key variable values into the program and receives an increment threshold for the specified conditions. However, the problems are not analogous.
Contrast sensitivity is one of the most fundamental and hard-wired human abilities and is determined by low-level retinal processes. Most researchers use similar methodologies and when measured by good methodology, such as forced-choice, contrast sensitivity has very low variance across normal viewers. Variability is so low that most basic contrast sensitivity research employs only two or three viewers. Contrast sensitivity is so regular and predictable that it follows relatively simple basic rules, such as Weber's Law, DeVries-Rose Law, Rico's Law, Bloch's Law, etc.
Even in the relatively simple case of contrast detection, moreover, it is still not possible to blindly apply the model output (Green, 2008). The difficulties of applying a computer model of the Adrian/ANSI visibility equations require assessment of a very large number of factors that the program does not capture. In PRT, the number and complexity of possible factors is far, far greater: recognition, interpretation, expectation, response election, urgency, braking dynamics, mental models, complexity, memory, and of course, all the factors that affect contrast sensitivity. It is wishful thinking to believe that results from a computer program can model behavior as complex as PRT and can be directly applied without a significant amount of interpretation based on long experience and knowledge.
In sum, a computer program cannot be directly used to determine PRT because:
|"However, any attempt to produce hard estimates of perception-response time for a situation that has some complicating features is fraught with peril." |
I do not mean to entirely dismiss a computer program. The output of such a program cannot be blindly used, but it can be useful in the hands of someone who has 1) a background in basic human skills such as perception, attention and reaction, 2) actually read the original research studies and 3) the background to critique research methodology and to properly interpret the study and to assess its limitations. However, it is not a band-aid to cover lack of knowledge about PRT source data or about basic human abilities. Like most tools, it is only as good it its user. In sum, the reality has not changed since reviewed the PRT literature 15 years ago:
- It is based on research data that require interpretation and that cannot be taken at face value. It is necessary to read the underlying studies. It is also necessary to have the background to interpret them. The only way to achieve this background is to have experience running research studies;
- The source research employed different methodologies, clock starting and stopping times, definitions of expectation, etc. and cannot simply be averaged;
- PRT is often difficult to operationally define, e. g., clock starting times are often difficult to identify, especially in low visibility situations and where the hazard gradually emerges;
- There simply are no data for many common collision scenarios (e. g., approaching a stopped vehicle on a dark roadway), so no modeling is possible or computer generated answer is realistically possible;
- Many variables, like expectations, switching costs, response conflicts, etc. cannot be easily quantified and used as computer input;
- The PRT's are averages and do not necessarily apply to any specific circumstance;
- The program omits many important variables such as visibility level, response alternatives and avoidance-avoidance traps, affordances, action boundaries, TTC, retinal slip velocity, presence of distracters, background complexity, driver understanding of vehicle dynamics, urgency, size and nature of the obstacle, inhibition of return, psychological refractory period, Hick's Law, etc.;
- The program can't say "I don't know". It doesn't know when to conclude that there aren't enough data to extrapolate to the crash conditions. In real life, not all conclusions are held with the same strength of belief. But the program doesn't give a confidence level for its output, so it doesn't tell you whether its output is little more than a guess or a high certainty. And it can't say whether or not it's output is certain beyond a reasonable degree of scientific certainty, as an expert would be required to do; and
- The model has not been validated for specific situations to which it is frequently applied. In other words, it is at most an educated guess.
1My favorite example is a 1938 paper where the authors stated straight out and with no shame that they threw out some of the results because they "threatened to complicate the analysis." Those were the days when scientists were men!
Broen, N. L., & Chiang, D. P. (1996). Braking response times for 100 drivers in the avoidance of an unexpected obstacle as measured in a driving simulator. In Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 40, 900-904.
|" Unfortunately, no single study can reproduce the full complexity of human behavior and its sensitivity to environmental variables. Moreover, studies cannot be quantitatively combined because no mathematical formalism can capture the subtle effects of methodology and variable interaction or incorporate general knowledge from the basic science literature on RT, perception, and cognition. For the time being, RT estimation remains part science and part intuition, that is, part application of a general knowledge about human factors." (Green, 2000). |
Green, M. (2000). “How Long Does It Take to Stop?” Methodological analysis of driver perception-brake times. Tranmsportation Human Factors, 2, 195-216.
Green, M. (2009). Perception-reaction time: Is Olson (& Sivak) all you need to know? Collision, 4, 88-95.
Green, M. et al. (2008) Forensic Vision: With Applications To Highway Safety. Lawyers & Judges Publishing: Tucson.
Olson, P. L. (1989). Driver perception response time. SAE 890731.
Olson, P. (2006). Perception-reaction time. In International Encyclopedia of Ergonomics and Human Factors. 2. New York: CRC Press.
Wood, J. M., Tyrrell, R. A., & Carberry, T. P. (2005). Limitations in drivers' ability to recognize pedestrians at night. Human Factors, 47(3), 644-653.