Comparisons of Predictive Power for Traffic Accident Involvement; Celeration Behaviour versus Age, Sex, Ethnic Origin, and Experience
Round 1
Reviewer 1 Report
I've read the manuscript with the interest. I thank author for that. Still, I suggest several things for improving as reader might not be deeply involved in the topic or specific approach.
Introduction:
1. Please, define celeration in depth. Are there any different methodological approaches in calculating it?
2. What is the logic of comparing the predictive power of behavioural variable versus demographic? Demographic variables lead to accident only through the behaviour. It is no surprise that behaviour predicts accidents better than demographic variables indirectly. O is there any other theory behind?
3. I couldn’t get the point for the sentence in lines 88-90.
Methods:
1. Please, define clearly what is the reliability of celeration and how is it calculated.
2. What is the sense of correlating celeration behaviour in 2001-2004 with culpable accidents in longer period of 1999-2005? Could it be concluded from the results that celeration patterns are rather unstable?
3. Line 142, please, explain the driver removing procedure. What are drivers with “few measurements”? Why should this be done?
4. Line 157 – “association was close to perfect”. Among which variables? Could we see this in the graph? Description is insufficient.
Discussion:
1. The discussion lacks explanation of the obtained results so much. What is meaning of the whole article? What is the main message?
2. I think author has to get back to the idea, why to consider these variables as antecedents of accidents after results have been introduced already.
3. Again, the issue of data variability has to be approached more detailed (line 170).
Author Response
Some comments to all reviewers: Thank you for your reviews of my paper about celeration and comparative predictive power. It was very nice to have as high a number as four different reviews, which were not only in good agreement with eachother, but also suggested some very good improvements to the paper. This is highly unusual.
The paper has now been substantially revised and extended. The reason for its originally brief and not particularly well-written form was that it was intended for a conference, with its limits on length. Also, I was not fully convinced about whether it would be important enough for a journal. However, re-visiting it after a few years, I realized that it had become more important, as my theory has not really caught on. Apparently, one of the reasons for this is that researchers do not seem to understand that it is testable (or the importance of this feature).
A very important change has been made in this version: I discovered an error of mine in one of the calculations. This was not computational as such, but faulty logic in applying the intended test. In Figure 1, the values should have been absolutes of the computed ones, because it is the size of the correlation which is important, not the sign (unless there is a change in sign between variables, in which case things get more complicated). This has now been pointed out in the introduction.
Reviewer 1
I've read the manuscript with the interest. I thank author for that. Still, I suggest several things for improving as reader might not be deeply involved in the topic or specific approach.
Introduction:
1. Please, define celeration in depth. Are there any different methodological approaches in calculating it?
A (rather long) section has been added about the basics of the theory and the differences between it and its closest relative, the safety event technique.
2. What is the logic of comparing the predictive power of behavioural variable versus demographic? Demographic variables lead to accident only through the behaviour. It is no surprise that behaviour predicts accidents better than demographic variables indirectly. Or is there any other theory behind?
I have now emphasised that this paper is mostly a model for how tests of the predictions should be made.
3. I couldn’t get the point for the sentence in lines 88-90.
There were two sentences on these lines, so I assume that the confusion arose because they are actually on different topics. I have therefore separated them.
Methods:
1. Please, define clearly what is the reliability of celeration and how is it calculated.
Three sections about reliability have been added, including values for this dataset.
2. What is the sense of correlating celeration behaviour in 2001-2004 with culpable accidents in longer period of 1999-2005? Could it be concluded from the results that celeration patterns are rather unstable?
I would not call it unstable, because it does correlate over several years, but there is apparently limits to its stability. I am interested in determining those limits, so that future research can take that into account. Some senctences have been added about this under Data.
3. Line 142, please, explain the driver removing procedure. What are drivers with “few measurements”? Why should this be done?
This has now been explained.
4. Line 157 – “association was close to perfect”. Among which variables? Could we see this in the graph? Description is insufficient.
This has been expanded upon.
Discussion:
1. The discussion lacks explanation of the obtained results so much. What is meaning of the whole article? What is the main message?
Thank you for asking. I have taken the opportunity to considerably expand upon this.
2. I think author has to get back to the idea, why to consider these variables as antecedents of accidents after results have been introduced already.
As in the previous point.
3. Again, the issue of data variability has to be approached more detailed (line 170).
Sorry, I do not understand this comment. It would seem like it refer to Figure 1, but there is not really any variability to take into account there. The lines connecting the dots, and the order of the variables in the Figure, do not indicate any sort of predicted absolute size differences between variables. The order of the variables could be sorted in any manner. The point is that if a certain variable tend to have a low correlation with accidents, it will also have a low correlation with celeration, and so on for high values.
Reviewer 2 Report
The paper is very interesting. One can see that the author is a long-term expert in the field. However, this is also a drawback, since some parts are less explained, and thus less clear for a reader with less experience.
In the first sentence of abstract, the author cites driver celeration theory. However, abstracts shoudl be self-explanatory, and I am not sure that general readers know the mentioned theory. It would be beneficial to re-phrase the abstract.
Regarding the correlation coefficients (p. 3, Types of analysis) - type of coefficient should be selected based on statistical distribution of studied variables.
Last comment - what is the practical outcome of knowing the safety implications of age, sex or ethnic origin? It should be mentioned in the text.
Author Response
Some comments to all reviewers: Thank you for your reviews of my paper about celeration and comparative predictive power. It was very nice to have as high a number as four different reviews, which were not only in good agreement with eachother, but also suggested some very good improvements to the paper. This is highly unusual.
The paper has now been substantially revised and extended. The reason for its originally brief and not particularly well-written form was that it was intended for a conference, with its limits on length. Also, I was not fully convinced about whether it would be important enough for a journal. However, re-visiting it after a few years, I realized that it had become more important, as my theory has not really caught on. Apparently, one of the reasons for this is that researchers do not seem to understand that it is testable (or the importance of this feature).
A very important change has been made in this version: I discovered an error of mine in one of the calculations. This was not computational as such, but faulty logic in applying the intended test. In Figure 1, the values should have been absolutes of the computed ones, because it is the size of the correlation which is important, not the sign (unless there is a change in sign between variables, in which case things get more complicated). This has now been pointed out in the introduction.
Reviewer 2
The paper is very interesting. One can see that the author is a long-term expert in the field. However, this is also a drawback, since some parts are less explained, and thus less clear for a reader with less experience.
In the first sentence of abstract, the author cites driver celeration theory. However, abstracts shoudl be self-explanatory, and I am not sure that general readers know the mentioned theory. It would be beneficial to re-phrase the abstract.
Several sentences have been added to the abstract about the theory.
Regarding the correlation coefficients (p. 3, Types of analysis) - type of coefficient should be selected based on statistical distribution of studied variables.
I have added some sentences about this. The problem lies in that the tests of the predictions mae it necessary to use a single coefficient, while the variables have very differing statistical features.
Last comment - what is the practical outcome of knowing the safety implications of age, sex or ethnic origin? It should be mentioned in the text.
I have expanded a bit upon the reasons for the study in various places in the manuscript. Basically, these variables are not very interesting in their own right, they are just vehicles for testing the predictions.
Reviewer 3 Report
The topic is very interesting. The structure of the paper is well organized, but it should be better specify some concepts. In particular:
1. the abstract should provide more clearly the justification for the choice of the comparative parameters;
2. some words present grammatical errors;
3. a description of the formulated analysis should be made from line 115 to line 131;
4. the meaning of the asterisks in tables 2 and 3 isn’t clear;
5. references should be extended and, at the same time, the auto citations should be reduced.
The paper is well done and with minor review can be published.
Author Response
Some comments to all reviewers: Thank you for your reviews of my paper about celeration and comparative predictive power. It was very nice to have as high a number as four different reviews, which were not only in good agreement with eachother, but also suggested some very good improvements to the paper. This is highly unusual.
The paper has now been substantially revised and extended. The reason for its originally brief and not particularly well-written form was that it was intended for a conference, with its limits on length. Also, I was not fully convinced about whether it would be important enough for a journal. However, re-visiting it after a few years, I realized that it had become more important, as my theory has not really caught on. Apparently, one of the reasons for this is that researchers do not seem to understand that it is testable (or the importance of this feature).
A very important change has been made in this version: I discovered an error of mine in one of the calculations. This was not computational as such, but faulty logic in applying the intended test. In Figure 1, the values should have been absolutes of the computed ones, because it is the size of the correlation which is important, not the sign (unless there is a change in sign between variables, in which case things get more complicated). This has now been pointed out in the introduction.
Reviewer 3
The topic is very interesting. The structure of the paper is well organized, but it should be better specify some concepts. In particular:
1. the abstract should provide more clearly the justification for the choice of the comparative parameters;
This has been done.
2. some words present grammatical errors;
Spell check has been run, as well as a language check by an English-speaking editor.
3. a description of the formulated analysis should be made from line 115 to line 131;
I have understood this as referring to the abstract, i.e. that the analysis as described in lines 115-131 should be summarized in the abstract. However, this getting to be difficult, because there is a 200 word limit. There is very little you can actually do to explain things within such a limit. I have had to delete the results to be able to describe the tests.
4. the meaning of the asterisks in tables 2 and 3 isn’t clear;
I forgot to add these. I always use the same standard.
5. references should be extended and, at the same time, the auto citations should be reduced.
A dozen references have been added. I have not seen the term 'auto citations' before, but I guess this refers to self-citation. This is difficult to do anything about, as I am (still) the only one who has published on the topic of celeration (as well as on a number of methodological issues). Removing some of these would mean that certain statements would be left without any supporting citations.
The paper is well done and with minor review can be published.
Reviewer 4 Report
The author tests the predictions of the celeration theory with demographic, celeration and accident data gathered from a sample of bus drivers of a bus company GUB in Uppsala, Sweden in between 1999-2005. I could not find the number of bus drivers included in the data from the manuscript.
The paper is well-written, I could find only two minor typos in the text (p.2, l.63: "[13]." and p.2,l.80: "Swedih"). However, clarity of expression could be improved at some parts, such as p.2, lines 53-65.
Rather uncommon feature in the current paper is that 15/24 of the references are to work by the author. All the publications are from reliable sources but there could be more discussions about the current theory's relationship to others' research (both in Introduction as well as in Discussion).
Personally I feel there is a need to try to explain in more detail the celeration theory and how the causal mechanism between celerations and accident risk works. It sounds a bit odd that celeration behavior would be the only significant predictor of accident risk. For instance, many of the recent naturalistic driving studies have found inattention as a significant predictor of crash risk (e.g., 100-car study, SHRP2 etc.). Is the link between inattention and crash risk moderated by changes in celeration behaviors, and which one is actually here the cause of increased crash risk? Are there any studies on this (etc.)?
Please clarify also, if the peak decelerations typical to crashes are included in the celeration data? This could explain at least partly the correlation between accidents and measured celerations, and also the observed time period effect.
Minor issues:
- It is odd to code one's ethnic origin based on one's name. Aren't there any individuals born in Sweden (that is, Swedish) that have names with ethnic origins?
- It would be nice to see more details about the regression models referred to in the text.
- Please illustrate the p-values ("from >.10 to <.001") in Figure 1.
Author Response
Some comments to all reviewers: Thank you for your reviews of my paper about celeration and comparative predictive power. It was very nice to have as high a number as four different reviews, which were not only in good agreement with eachother, but also suggested some very good improvements to the paper. This is highly unusual.
The paper has now been substantially revised and extended. The reason for its originally brief and not particularly well-written form was that it was intended for a conference, with its limits on length. Also, I was not fully convinced about whether it would be important enough for a journal. However, re-visiting it after a few years, I realized that it had become more important, as my theory has not really caught on. Apparently, one of the reasons for this is that researchers do not seem to understand that it is testable (or the importance of this feature).
A very important change has been made in this version: I discovered an error of mine in one of the calculations. This was not computational as such, but faulty logic in applying the intended test. In Figure 1, the values should have been absolutes of the computed ones, because it is the size of the correlation which is important, not the sign (unless there is a change in sign between variables, in which case things get more complicated). This has now been pointed out in the introduction.
Reviewer 4
The author tests the predictions of the celeration theory with demographic, celeration and accident data gathered from a sample of bus drivers of a bus company GUB in Uppsala, Sweden in between 1999-2005. I could not find the number of bus drivers included in the data from the manuscript.
It was stated in the second sentence under 'Data' that there are usually about 350 drivers employed. The number of drivers in each analysis is stated in each table and figure.
The paper is well-written, I could find only two minor typos in the text (p.2, l.63: "[13]." and p.2,l.80: "Swedih"). However, clarity of expression could be improved at some parts, such as p.2, lines 53-65.
Sorry, but I do not know exactly what to improve here.
Rather uncommon feature in the current paper is that 15/24 of the references are to work by the author. All the publications are from reliable sources but there could be more discussions about the current theory's relationship to others' research (both in Introduction as well as in Discussion).
Thanks for the suggestion - this is a topic which I have written about at length in another paper which is under review. It is indeed very important, and so I have added quite a bit about this.
Personally I feel there is a need to try to explain in more detail the celeration theory and how the causal mechanism between celerations and accident risk works. It sounds a bit odd that celeration behavior would be the only significant predictor of accident risk. For instance, many of the recent naturalistic driving studies have found inattention as a significant predictor of crash risk (e.g., 100-car study, SHRP2 etc.). Is the link between inattention and crash risk moderated by changes in celeration behaviors, and which one is actually here the cause of increased crash risk? Are there any studies on this (etc.)?
Hmm...I don't think I said that it would be the only significant predictor - that would obviously be wrong. What the theory predicts is that there can be no stronger predictor, and that all other predictors cause changes in celeration. My expectation would therefore be for distraction to increase celeration values, yes. Whether celeration or distraction is considered to be the cause of risk is more of a philosophical consideration. Maybe this is better handled by calling distraction a distal cause and celeration a direct cause.
Please clarify also, if the peak decelerations typical to crashes are included in the celeration data? This could explain at least partly the correlation between accidents and measured celerations, and also the observed time period effect.
There were no crashes during measurements so this is not an issue. However, even if there had been a crash, this would have had an extremely slight effect on the celeration value, as these are calculated as the average over time of driving. A crash might unfold over a time period of about two seconds. In the present data, that would mean that about eight values out of 10 000 would be strongly elevated.
So, the associations found here are not artefacts of the method used. However, this logic is indeed applicable to some of the naturalistic driving studies which I have referenced. They sometimes use methods which could be susceptible to this effect.
Round 2
Reviewer 1 Report
Thank you, all my comments were considered by the author. The manuscript has been expanded, therefore it is easier to follow the idea. I believe manuscript could be published in this form as it is now.
Author Response
Thank you.
Reviewer 2 Report
Thank you for the revisions, I do believe they helped improve the paper quality.
Author Response
Thank you.
Reviewer 3 Report
The current version does not require any further changes unless a minor revision of the English language.
The paper can be approved for publication.
Author Response
Thank you.
Reviewer 4 Report
The author has been able to address most of my earlier concerns and the quality of the presentation has improved much from the previous version. I still have doubts about the celeration theory in "that all behaviors of a driver which are related to safety can be measured as longitudinal and lateral speed changes, where the risk increases linearly with the size of the speed change" but let data prove me wrong.
Yet, and importantly, as the paper has been reformatted to focus on: "to describe in detail how two predictions from DCBT should be tested, within a methodological framework which takes into account the reliability of the variables included", the presentation of the statistical method and results (as correlations) should be further clarified. Sex and Ethnicity are categorical variables, how can one calculate Pearson correlations between these and celeration? If the correlations in Tables 2 and 3 would be betas based on regression analysis, I would better understand the results.
Author Response
Actually, regression cannot be used for this test, because the beta values influence eachother. Zero-order effects are needed.
Although Pearson correlations are usually not used for categorical variables, this is actually not a problem at all, they work perfectly well. I have added a short section about this under 'Types of analyses'.
